summertime heat across the united states by
Post on 30-Oct-2021
1 Views
Preview:
TRANSCRIPT
SUMMERTIME HEAT ACROSS THE UNITED STATES
by
Tiffany T. Smith
A dissertation submitted to Johns Hopkins University in conformity with the requirements
for the degree of Doctor of Philosophy
Baltimore, Maryland
August, 2016
© Tiffany T. Smith
All Rights Reserved
ii
ABSTRACT
Extreme summertime heat has been the most deadly natural hazard in the United States
over the past 30 years and is projected to become more intense, more frequent and longer
lasting in the second half of the century. We take this as motivation to improve our
understanding of the drivers in summertime heat across the Continental United States
(CONUS), and provide a framework to discussing results from studies with diverse
motivations. This dissertation attempts to (1) create a baseline in understanding in the way
heat waves are defined and how this impacts conclusions of the patterns and trends of heat
waves, (2) investigate large scale drivers of summertime temperature on seasonal
timescales across variable-informed regions of CONUS, and (3) identify the impact of North
Atlantic Oscillation (NAO) definition on local heat waves as defined by two relevant
definitions. We find that (1) positive trends in heat waves are seen across most of the
United States where spatial patterns differ between definitions, (2) temperature variability
is sensitive to climate processes across CONUS regions, notably though that nonlinear
models produce improvement in explaining these relationships, and (3) that by defining
NAO by its centers of action, rather than phase, we increase our ability to model heat waves
in Baltimore, MD. This work will provide an outline for discussing results from heat wave
studies with diverse motivations, as well as deepen our understanding of the large-scale
drivers of summertime heat with the intention of informing and improving seasonal
forecasting and ultimately mitigate negative impacts heat has on the human population.
iii
Advisor:
Dr. Benjamin F. Zaitchik, Department of Earth and Planetary Sciences
Thesis Reader:
Dr. Darryn W. Waugh, Department of Earth and Planetary Sciences
iv
ACKNOWLEDGMENTS
As the saying goes, it takes a village. So I will attempt to thank my village, without whom I
would not be here today.
I would first like to thank my advisor, Ben Zaitchik, for being the academic role model I
needed for this journey. Ben took a chance on me when he decided that I would be the first
student he brought into Hopkins to begin building his lab group. Since then, Ben’s lab group
has grown and flourished and I am glad to have been part of that. Without the support,
patience and encouragement from Ben throughout this process, I would not be here today,
submitting this dissertation.
I would also like to thank the other members of my advisory committee, Darryn Waugh and
Carlos del Castillo. I would like to thank Darryn for his integral role in advocating for EPS to
be more involved in interdisciplinary research, without which I would have had no place
here. I would like to thank Carlos for having me as one of his “grasshoppers” and for always
reminding me that I do, in fact, know the answer.
In addition, I would like to thank the many colleagues I had the pleasure to collaborate with
throughout my dissertation work. In particular, Seth Guikema and Julia Gohlke, though
there are many more. Seth and Julia were integral to the science done for this dissertation,
and are the reason I was able to investigate heat waves in such an interdisciplinary matter.
Julia opened her world of public health research to me, which went on to be the driving
motivation throughout my dissertation. Seth gave me the most important, tangible skill set
v
I acquired in grad school, which is first and foremost his love of statistics, but also his
ability to think outside of the box and always ask questions in new and interesting ways.
I would like to thank the EPS Front Office, Kim Trent, Jean Light, Teresa Healy, and
especially Kristen Heisey for making my grad student tenure as headache-free as possible.
In addition, I would like to thank Anand Gnanadesikan for always being willing to discuss
ideas no matter how periphery to his research they may have been.
This work would not have been possible without the financial support of the Department of
Earth and Planetary Sciences, The National Institute of Environmental Health Sciences
Grant R21 ES020205 and the National Science Foundation Hazard SEES Type 2 Grant #
1331399.
To Sara Rivero, Alex Fuller, Scott Pitz, Greg Henkes, Erin Urquhart, and all of the other EPS
students who I have built friendships and navigated grad school with, I am thankful. I am
particularly grateful to Saleh Satti, Asha Jordan, and Sophie Lehmann without whom I
would not have survived the many “off-hours” time spent in Olin over the past 13 months.
To my coworkers at Constellation, most importantly to the Fundamentals team, thank you
for your support and understanding. To my boss, Ed Fortunato, thank you for taking a leap
of faith in hiring me and for advocating and supporting the completion of my Ph.D. despite
the limited impact this milestone will have on our work.
vi
I would like to thank my family and friends for their support and understanding over the
last six years. I would first like to thank my family, my parents Bill & Carolyn Townsend, my
brother Jason Townsend and his partner Kevin Dang – thank you for your encouragement
and support as I navigated my doctoral education. To my parents, who have encouraged me
in whatever I chose to do over the past 29 years, while never once questioning if I could do
what I’d set my mind to (looking back on the stubbornness they endured through my
childhood, I’m sure they’re glad to see me put that quality to good use). And to my dad
especially, had I not been handed down your tenacity, inquisitiveness and penchant for
precision, I would not be the scientist I am today – I am proud that I’ll always be Little Bill
Fred. To my in-laws, Dixon, Kiki, Zach and Cragan Smith, thank you for always being a
breath of fresh air, especially when that air was from San Diego. To my friends, Sarah Isbell,
Jess & Eddie Nie, Anna Holland, Lauren & Rudi Greenberg, Katie Weber, Marcelle Empey,
Amy Bond, Molly Finch, Charlotte Smith, Alec Cronin, Teddy Davidson, Sean McCullagh,
Sara Atwater, Jason Vodzak, Kendie Bauer, Andrew & Kay Vassallo, Jimmy & Keisha Lewis,
Aaron Larrimore, Jen Gilbert, Lucy Rose Davidoff and Tori & Chris Hidalgo, thank you for
my sanity. I am so thankful my friends have been there, ready to drink a beer, hit the
weights, run the miles, go to the dog park, listen to me vent my academic stress and
everything in between. Over the past 13 months, my family and friends have received an
endless string of, “no’s” from me, but in reply I have heard only, “we’ll miss you, but we
understand”, “you got this” or “I’m so proud of you” – and for this, I am forever grateful and
indebted to you all.
vii
Last but not least, I would like to thank my husband and my best friend, Tucker Smith.
There are no words to accurately express the love and support Tucker has given me over
the past 12+ years, but especially over these last six. I arrived at Hopkins for new student
orientation five days after our wedding – our honeymoon was the first in a string of
sacrifices Tucker has so graciously made for me as I pursued my doctoral education. None
of these sacrifices has been so big as the decision to live here, in Baltimore, some 40 miles
from his place of work (~120,000 commuting miles over my grad school tenure…but,
who’s counting). He has propped me up when I needed it – the impossible problem sets, the
“conditional” passes that felt like failures, the moments of self-doubt and the tears of
frustration – and he has cheered for me when I deserved it – the acceptance letter, the
conference presentations, the first first-author publication, the Masters degree, the
interview offers and finally this dissertation. It is no secret how challenging it is to work full
time while completing your dissertation, but I simply would not have survived these past
13 months had Tucker not been there keeping me afloat. Tucker – thank you for riding this
roller coaster with me, and I cannot wait to see what post-Ph.D. life has in store for us.
viii
TABLE OF CONTENTS
ABSTRACT …………………………………………………………………………………………………………………… ii
ACKNOWLEDGMENTS ……………………...……………………………………………………………………….…. iv
LIST OF TABLES …………………………………………………………………………………………………………... xi
LIST OF FIGURES ………………………………………………………………………………………………………... xii
1. CHAPTER 1: INTRODUCTION ………………………………………………………………………………...… 1
1.1. Heat wave definitions ……………………………………………………………………………………..… 1
1.2. Drivers of summertime temperature across CONUS regions ……………………………….. 2
1.3. The impact of NAO on heat waves in Baltimore ………………………………………..……...… 3
1.4. Dissertation outline ……………………………………………………………………………………...…… 4
2. CHAPTER 2: HEAT WAVES IN THE UNITED STATES: DEFINITIONS, PATTERNS AND
TRENDS ………………………………………………………………………………………………………………...... 7
2.1. Introduction …………………………………………………………………………………………….............. 8
2.2. Materials and Methods ……………………………………………………………………………............ 11
2.2.1. Data ……………………………………………………………………………………………………… 11
2.2.2. Heat wave indices ………………………………………………………………………………….. 14
2.2.2.1. Relative thresholds ……………………………………………………………………... 14
2.2.2.2. Absolute thresholds ……………………………………………………………………. 16
2.2.3. Statistical analysis …………………………………………………………………………………. 17
2.3. Results ……………………………………………………………………………….…...……………………… 20
2.3.1. Average annual heat wave days ……………………………………………………………… 20
2.3.2. Temporal trends ……………………………………………………………………………………. 22
2.4. Discussion ………………………………………………………………………………………………………. 24
ix
2.5. Conclusions …………………………………………………………………………………………………….. 28
3. CHAPTER 3: LARGE-SCALE DRIVERS OF INTERANNUAL SUMMERTIME TEMPERATURE
VARIABILITY ACROSS THE CONTINENTAL UNITED STATES ……………………………….….. 33
3.1. Introduction …………………………………………………………………………………………………… 34
3.2. Methods …………………………………………………………………………………………………………. 37
3.2.1. Datasets ………………………………………………………………………………………………... 37
3.2.2. Regionalization ……………………………………………………………………………………... 41
3.2.3. Prediction ……………………………………………………………………………………………... 42
3.2.4. Leading Indicators ………………………………………………………………………………… 44
3.3. Results …………………………………………………………………………………………………………… 45
3.3.1. Regionalization ……………………………………………………………………………………... 45
3.3.2. Prediction ……………………………………………………………………………………………... 46
3.3.2.1. Linear Trend Present (LTP) analysis ……………………………………………. 46
3.3.2.2. Linear Trend Removed (LTR) analysis ………………………………………… 51
3.4. Discussion ………………………………………………………………………………………………………. 57
3.4.1. Regionalization ……………………………………………………………………………………... 57
3.4.2. Model Structure …………………………………………………………………………………….. 58
3.4.3. Predictors & Mechanism ………………………………………………………………………... 60
3.5. Conclusions …………………………………………………………………………………………………….. 64
4. CHAPTER 4: THE IMPACT OF THE NORTH ATLANTIC OSCILLATION ON HEAT WAVES
IN BALTIMORE ……………………………………………………………………………………………………… 71
4.1. Introduction …………………………………………………………………………………………………… 72
4.2. Methods …………………………………………………………………………………………………………. 75
x
4.2.1. Data ……………………………………………………………………………………………………… 75
4.2.2. Heat wave indices ………………………………………………………………………………….. 77
4.2.3. Data evaluation and model fit ………………………………………………………………… 78
4.2.4. Prediction ……………………………………………………………………………………………... 79
4.3. Results …………………………………………………………………………………………………………… 80
4.3.1. Model fit ……………………………………………………………………………………………….. 80
4.3.2. Prediction ……………………………………………………………………………………………... 84
4.4. Discussions and conclusions ………………………………………………………………………..….. 87
5. CHAPTER 5: CONCLUSIONS ………………………………………………………………………………….... 98
5.1. Future work ………………………………………………………………………………………………….. 100
REFERENCES ……………………………………………………………………………………………………………. 101
AUTHOR’S CURRICULUM VITAE ………………………………………………………………………………... 110
xi
LIST OF TABLES
Table 2.1 Definitions of heat wave indices. *HI16 did not have enough data to include in this analysis.
Table 2.2 Average number of annual heat wave days, divided by region. Bold indicates region with highest
frequency of heat waves days for each HI. Regions are: Northwest (NW), Southwest (SW), Great Plains (GP),
Midwest (MW), Southeast (SE) and Northeast (NE).
Table 3.1 Bold indicates intraregional correlations Italics indicates interregional correlations
Table 3.2 Average (standard error) MSE, for holdout with LTP data; bold indicates top-performing model.
Table 3.3 Swings from BEST model using the LTP data, with relative rank in parenthesis, where (1) is most
important and (8) is least important, top three in bold.
Table 3.4 Average (standard error) MSE, for holdout with LTR data; bold indicates top-performing model.
Table 3.5 Swings from BEST model using the LTR data, with relative rank in parenthesis, where (1) is most
important and (8) is least important, top three in bold.
Table 4.1 Root mean square error (RMSE) for four GLM. Table 4.2 Mean square error (MSE) values from holdout analysis for HI02; * indicates top performing model.
Table 4.3 Mean square error (MSE) values from holdout analysis for HI11; * indicates top performing model.
Table 4.4 % impact on MSE each variable has for RF models for HI02. Higher values indicate variables of
higher importance.
Table 4.5 % impact on MSE each variable has for RF models for HI11. Higher values indicate variables of
higher importance.
xii
LIST OF FIGURES
Figure 2.1 Regional division of the Continental United States (CONUS).
Figure 2.2 (a) Average daily Tmax (ºC) over the time period 1979-2011 (b) Standard deviation of Tmax over
the time period 1979-2011 (c) 95th percentile of Tmax(ºC).
Figure 2.3 1979-2011, annual average number of heat wave days. Note the varying scales.
Figure 2.4 Trends in the number of annual heat wave days, over the period 1979-2011. White areas indicate
results below 95% significance. Units are days/year.
Figure 2.5 Average 95% significant trends in the number of annual heat wave days, over the period 1979-
2011, divided by region. The value printed in each cell is the trend value (in days/year). The color of the cell
represents positive (red) and negative (blue) trends. The shades of red and blue represent the landmass
percentage covered by this significant trend given by the scale bar. Regions are: Northwest (NW), Southwest
(SW), Great Plains (GP), Midwest (MW), Southeast (SE) and Northeast (NE).
Figure 3.1 (a) Map of regions and (b) the corresponding dendrogram. Y-axis in (b) is the sum of squared
distances within all regions, and is a measure of intra-regional variance.
Figure 3.2 Timeseries of Tmin, LTP (dashed red line) and Tmin, LTR (solid black line) for regions (a) NW, (b)
SW, (c) NGP, (d) South, and (e) NE.
Figure 3.3 Timeseries of LTP (dashed, red line) and LTR (black line) for all covariates: (a) ENSO, (b) NAO, (c)
PDO, (d) PNA, (e) GMSST, (f) AMO, and (g) AO.
Figure 3.4 Partial dependence plots for the three leading indicators in each region for models built with LTP
data: NW – (a) GMSST, (b) PNA, (c) NAO; SW – (d) GMSST, (e) PDO, (f) AO; NGP – (g) GMSST, (h) AO, (i) SM;
South – (j) GMSST, (k) AO, (l) NAO; NE – (m) GMSST, (n) ENSO, (o) SM.
Figure 3.5 Scatterplots of actual versus predicted Tmin for the best LTR models in (a) SW, (b) NGP, (c) South,
and (d) NE regions.
Figure 3.6 Temperature anomaly plots for (a) ENSO+, (b) ENSO-, (c) PDO+, (d) PDO-, (e) AO+, (f) AO-.
xiii
Figure 3.7 Partial dependence plots for the three leading indicators in each region for models built with LTR
data: SW – (a) GMSST, (b) PDO, (c) AMO; NGP – (d) GMSST, (e) AO, (f) PNA; South – (g) GMSST, (h) AMO, (i)
NAO; NE – (j) ENSO, (k) GMSST, (l) SM.
Figure 3.8 Composite plots showing 300 hPa anomalies for the top five warmest years for (a) SW, (b) NGP,
(c) South, and (d) NE.
Figure 4.1. Timeseries of annual (July-August) heat wave (HW) day counts for (A) HI02 and (B) HI11.
Figure 4.2. Correlations between all variables included in models. Top right corner shows correlations,
where boxes are shaded according to the scale bar. Bottom left corner prints correlation values, also shaded
according to the scale bar, where values not printed were insignificant.
Figure 4.3 Actual versus fitted data for (A) HI02 and (B) HI11 where blue dots indicate Ind7 model results
and red dots indicate COA model results. A 1:1 line is provided for reference of a “perfect” fit.
Figure 4.4 P-values of covariates included (A) GLM-Ind7 and (B) GLM-COA. Stars indicate results for HI02,
and triangles indicate results for HI11. The horizontal dashed line represents 95% significance level; any
symbol below the dashed line is significant.
Figure 4.5. Partial dependence plots for the top two most important variables for HI02 for (A-B) RF-Ind7,
and (C-D) RF-COA.
Figure 4.6. Partial dependence plots for the top two most important variables for HI11 for (A-B) RF-Ind7,
and (C-D) RF-COA.
Figure 4.7 Composite anomaly plots for the top five hottest years according to (A) HI02 and (B) HI11 where
colors represent surface pressure and arrows represent vector winds. A box is located over the Azores for
reference.
Figure 4.8 Composite anomaly plots of 300 hPa geopotential heights for the top five hottest years according
to (A) HI02 and (B) HI11.
Figure 4.9 Composite anomaly plots of surface pressure (colors) and vector winds (arrows) for study time
period, 1950-2014.
1
1. CHARPTER 1: INTRODUCTION
Extreme summertime heat has been the most deadly natural hazard in the United States
over the past 30 years (NOAA, 2015). As mean global temperatures rise (IPCC, 2007), there
has been increased attention to the frequency of extreme heat and the resulting impacts on
society. Between 1951-2003, statistically significant increases in minimum and maximum
temperature were seen over 40-75% of global land area (Alexander et al., 2006; Trenberth
et al., 2007). Over the Continental United States (CONUS), it has been projected that by mid-
Centry, 50% of all summers will be as hot at the top 5% of summers in the historic baseline
(Duffy and Tebaldi, 2012). In addition to increased frequency, it has been shown that
super-extreme (>3σ over mean) events have shifted from covering 1% of global landmass
to >10% of landmass by 2011 (Hansen et al., 2012).
1.1. Heat wave definitions
When discussing changes and trends of extreme summertime temperatures, it is
important to note that these discussions are related to studies of heat waves, but not
synonymous. This is partly due to the inherent differences between temperature,
typically evaluated as a continuous variable, and a heat wave event, understood as the
exceedance of a defined threshold. More importantly, there is no consensus definition of
heat wave throughout the literature due to the correspondingly diverse set of reasons
researchers study heat waves. Climate scientists, who are primarily interested in the
evolving statistics of weather and climate change, tend towards definitions that include
a probability of exceedance in some relatively straightforward metric, usually defined
relative to a long-term mean (e.g. Shar et al., 2004; Hansen et al., 2012). In contrast,
2
public health researchers are interested in the aspects of a heat wave that are most
relevant to human health and well being. These definitions are more likely to be
regionally specific than climate-focused definitions because people living in different
climates are known to experience heat waves differently, where mortality in the mild
climate of the Northeast increases 6.76% on heat wave days versus and increase of
1.84% in the warm, humid climate of the South (Anderson & Bell, 2011). These health
definitions are therefore more often informed by recognized links between heat and
morbidity/mortality, where it has been shown that relative threshold definitions
represent increased effects of temperature better than absolute threshold definitions
(Kent et al., 2014). As such, quantitative definitions of heat waves differ in (1) the
metric of heat used, (2) the manner in which thresholds of exceedance are defined,
and/or (3) the role of duration in a heat wave event. This diversity of definitions can
lead to some confusion in the broader discussion of climate trends and impacts of
climate change. So while there is value in monitoring and projecting each of these kinds
of heat extremes, there is also a need to be clear about definitions such that reported
patterns and trends can be understood in the context of other studies.
1.2. Drivers of summertime temperature across CONUS regions
Increased summertime temperatures have been associated with a wide range of
negative consequences including increased morbidity and mortality (Curriero et al.,
2002; Peng et al., 2011), increased wildfire activity (Westerling et al., 2006) and
decreased agricultural yields (Lobell et al., 2013). Because of the diversity of impacts, it
is important to clarify our understanding of the drivers of temperature trends and
3
variability. These drivers can include remote large-scale modes (LSM) of climate
variability (Barnston 1996; Drosdowsky and Chambers, 2001; Kenyon and Hegerl,
2008) and local feedback mechanisms (Fischer et al., 2007; Portmann et al., 2009).
Historically, the most well studied LSM is the El Nino Southern Oscillation (ENSO).
However, more recently studies of CONUS temperatures have begun to include a wider
variety of large-scale modes of variability, including the Pacific-North American mode
(PNA), the Northern Annular mode (NAM), the Pacific Decadal Oscillation (PDO), the
North Atlantic Oscillation (NAO), and the Atlantic Multidecadal Oscillation (AMO)
(Loikith and Broccoli, 2014; Kenyon and Hegerl, 2008; Zhang et al., 2007; Sutton and
Hodson, 2005). Many studies highlight relationships between these LSM and
wintertime temperatures due to the inherent nature of the LSM being more active
during winter months (Visbeck et al., 2001) and the relationship between the LSM and
temperature being more robust (Becker et al., 2013). Studies of the relationship
between LSM and summertime climate have become increasingly common (Wang et al.,
2007; Krishnamurthy et al., 2015) in response to the plethora of evidence showing the
negative impacts of summertime climate (USGCRP, 2016).
1.3. Impact of North Atlantic Oscillation on heat waves in Baltimore
As noted above, people living in the mild climate of the Northeast experience a stronger
negative impact during a heat wave than those living in the warm, humid climate of the
South (Anderson & Bell, 2011). On the East coast, the intersection of the cool, dry
climate of the Northeast and the warm, humid climate of the South are found in the Mid-
4
Atlantic region, centered near Baltimore, MD. Because of the multitude of negative
impacts summertime heat has in Baltimore, it is important we hone our understanding
of the drivers of the trends and variability in these heat waves.
As discussed in Section 1.2, a variety of work exists surrounding the impacts of LSM on
summertime heat. As explained in Section 1.1, summertime heat and heat wave events
are related but not always synonymous. As such, the current literature investigating the
impacts of LSM on heat wave events is thin.
Event-based research surrounding the definition of the LSM has become increasingly
prevalent, where event-to-event differences in ENSO spatial patterns and evolution has
lead to ENSO definitions on a continuum versus two distinct modes of variability
(Ashok et al., 2007; Singh et al., 2011; Capotondi et al., 2015). Similarly, temperatures
have been shown to be sensitive to not only the phase of the NAO, but also to the
location of the NAO’s centers of action (COA) (Castro-Diez et al., 2002). Asymmetry has
also been found between the NAO COA over the Azores High and the Icelandic Low,
where movements centered on the Icelandic Low are highly correlated to NAO phase
while the relationship with the Azores High was insignificant (Hameed and Pinotkovski,
2004). Recently, summertime temperatures were found to be more sensitive than
wintertime temperatures to the definitions of NAO (Pokorna and Huth, 2015).
1.4. Dissertation outline
5
This dissertation is comprised of five chapters. Chapter 1 is an introduction to extreme
summertime temperatures and the motivation for this work.
Chapter 2 represents published work (Smith et al., 2013) on the definition of heat
waves. In this work, we analyze geographic patterns and trends over CONUS for fifteen
different, previously published heat wave indices. The objective of this study is to
describe and explain how the choice of definition influences conclusions regarding the
observed frequency of extreme heat events in different regions of CONUS in order to
provide a baseline for interpreting studies that project future trends in extreme heat
events.
Chapter 3 represents submitted work (Smith et al., 2016 under review) where a novel
regionalization of CONUS is presented followed by an investigation of LSM drivers to
summertime temperatures over these regions. Here we focus on a metric of seasonal
average temperatures in order to achieve stronger statistical relationships (Pepler et
al., 2015) and explain large-scale drivers of variability on seasonal timescales. Because
these LSM are known to persist, this work will inform future studies of seasonal
forecasting and climate change impacts specifically relevant across heat-vulnerable
regions of CONUS.
Chapter 4 presents a deep dive into the large-scale drivers of heat waves in Baltimore,
MD. Specifically, this work compares the effect of two NAO definitions, classically
defined NAO and COA-defined NAO, to heat waves in Baltimore, where a heat wave is
6
defined by both a relative and an absolute definition from Chapter 2. We also build a
framework for a predictive model of heat waves in Baltimore, and with this in mind,
include a variety of LSM as this will likely increase the predictive skill of these models.
Finally, Chapter 5 concludes the dissertation work.
7
2. CHAPTER 2: HEAT WAVES IN THE UNITED STATES: DEFINITIONS, PATTERNS AND
TRENDS1
ABSTRACT
High temperatures and heat waves are related but not synonymous concepts. Heat waves,
generally understood to be acute periods of extreme warmth, are relevant to a wide range
of stakeholders because of the impacts that these events have on human health and
activities and on natural environments. Perhaps because of the diversity of communities
engaged in heat wave monitoring and research, there is no single, standard definition of a
heat wave. Experts differ in which threshold values (absolute versus relative), duration and
ancillary variables to incorporate into heat wave definitions. While there is value in this
diversity of perspectives, the lack of a unified index can cause confusion when discussing
patterns, trends, and impacts. Here, we use data from the North American Land Data
Assimilation System to examine patterns and trends in 15 previously published heat wave
indices for the period 1979-2011 across the Continental United States. Over this period the
Southeast region saw the highest number of heat wave days for the majority of indices
considered. Positive trends (increases in number of heat wave days per year) were greatest
in the Southeast and Great Plains regions, where more than 12% of the land area
experienced significant increases in the number of heat wave days per year for the majority
of heat wave indices. Significant negative trends were relatively rare, but were found in
portions of the Southwest, Northwest, and Great Plains.
1 Smith, T. T., B. F. Zaitchik and J. M. Gohlke. (2013) Heat waves in the United States: definitions, patterns and trends. Climatic Change, 118:811-825.
8
2.1. Introduction
As mean global temperatures rise (IPCC, 2007), there has been increased attention to
the frequency of heat extremes and their social and environmental impacts. In 2012
alone, the Intergovernmental Panel on Climate Change (IPCC) issued the full text of
their Special Report on Managing the Risks of Extreme Events and Disasters to Advance
Climate Change Adaptation (SREX), the Natural Resources Defense Council (NRDC)
released the report “Killer Summer Heat: Projected Death Toll from Rising
Temperatures in America Due to Climate Change”, and Climate Communications
released the report, “Heat Waves and Climate Change.” These studies provide
assessments of recent patterns and trends in heat extremes. They also address the
complexities involved in evaluating and projecting the frequency and intensity of heat
extremes in a changing climate. The occurrence of multiple high profile extreme heat
waves in recent years (e.g., Chicago, 1995; Europe, 2003; Russia, 2010) has highlighted
the importance of understanding and projecting patterns in extreme heat anomalies.
With respect to recent trends, the IPCC SREX cites the work of Trenberth et al. (2007)
and Alexander et al. (2006), who found that 70-75% of global land area with data
coverage saw a statistically significant increase in minimum temperatures (Tmin),
while 40-50% of global land area experienced an increase in maximum temperatures
(Tmax) over the period 1951-2003. However, Alexander et al. (2006) shows that some
regions departed from these trends, two of which were the eastern United States and
central North America. In these regions, negative trends in Tmin and Tmax were
observed, supporting Pan et al. (2004) who termed the phrase “warming hole” over this
9
region. In parallel, a number of studies have looked specifically at trends in heat wave
events rather than in temperature means (e.g., Meehl & Tibaldi 2004; Hansen et al.,
2012). The recent work in Hansen et al. (2012) describes that future (warmer) climates
will experience the emergence of a new category of extreme outliers (>3σ over the
mean). During the baseline period of 1951-1980, heat wave events classified as >3σ
over the mean, covered only 1% of landmass. However, in recent years (2006-2011) the
average landmass coverage of these extreme outliers has been 10%. Model results from
Meehl and Tebaldi (2004) show that heat waves will become more intense, more
frequent and longer lasting in the second half of the 21st century.
Trends in Tmin and Tmax are related, but not necessarily synonymous with trends in
heat wave events. In part this is due to the inherent difference between temperature,
typically evaluated as a continuous variable, and a heat wave event, understood as an
exceedance of some defined threshold. However, it is also a product of the fact that
there is no consensus on the definition of a heat wave. While a heat wave is generally
understood to be a period of extreme and unusual warmth, quantitative definitions of
heat waves differ in (1) the metric of heat used—e.g., daily mean temperature (Tmean)
(Anderson & Bell, 2011), daily Tmax (Peng et al, 2011; Meehl & Tebaldi, 2004),
temperature and humidity (Grundstein et al, 2012; Rothfusz, 1990) and apparent
temperature (Steadman, 1984)—(2) the manner in which thresholds of exceedence are
defined—as an absolute threshold (e.g., Tmax greater than 40.6°C (Robinson, 2001)), or
as a relative threshold, (e.g., the 95th percentile (Anderson & Bell, 2011))—and/or (3)
10
the role of duration in a heat wave definition (e.g., one day (Tan et al., 2007), or multiple
length criteria (Meehl & Tibaldi, 2004)).
The diversity of definitions reflects the diversity of reasons that heat waves are studied.
Climate scientists, who are primarily interested in the evolving statistics of weather in
climate change, tend towards definitions that include a probability of exceedance in
some relatively straightforward metric, usually defined relative to a long term mean
(e.g., Hansen et al., 2012; Schar et al., 2004). Health researchers, in contrast, are
interested in the aspects of a heat wave that are most relevant to human well-being.
These heat wave definitions are also more likely to be regionally specific than climate-
focused definitions, and may also target vulnerable subsets of the population. These
health definitions are often informed by recognized links between heat and morbidity
or mortality (Semenza et al., 1999; Hajat et al., 2006; Medina-Ramon and Schwartz,
2007, Anderson and Bell, 2009), which makes it possible to select or customize heat
wave indices for the purpose of predicting health impacts in specific geographic settings
or in particular populations (Barnett et al., 2010; Metzger et al., 2010; Williams et al.,
2012). Such studies can inform the use of heat wave indices in operational warning
systems.
Because of the diversity of stakeholders involved in the study of heatwaves, researchers
and health experts are able to collaborate to evaluate local health risks posed by climate
variability and change. However, the diversity of definitions can also lead to some
confusion in the broader discussion of climate trends and the impacts of climate change.
11
Indices that give a high weight to daily Tmax may be higher in hot, dry regions while
those focused on daily Tmin or apparent temperature would tend to be higher in humid
zones. There is value in monitoring and projecting each of these kinds of heat extremes,
but there is also a need to be clear about definitions such that reported patterns and
trends can be understood in the context of other studies.
In this paper we analyze geographic patterns and trends for the Continental United
States (CONUS) in fifteen different previously published heat wave indices (HI). The
objective is to describe and explain how the choice of definition influences conclusions
regarding the observed frequency of extreme heat events in different regions of the
CONUS over the past thirty-three years, in order to provide a baseline for interpreting
studies that project future trends in extreme heat events.
2.2. Materials and Methods
2.2.1. Data
The meteorological data used in this study are from Phase 2 of the North American
Land Data Assimilation System (NLDAS-2). The NLDAS-2 data have a spatial
resolution of 1/8 degree (~12.5km) and an hourly temporal resolution (Xia et al,
2012; Mitchell et al, 2004). As heat waves are the focus of this study, only data from
the warm season, May 1 – Sept 30, were considered. The analysis period is 1979-
2011.
12
NLDAS-2 was used for this study because it provides the most reliable, spatially
complete dataset of its kind. The six NLDAS-2 forcing fields used were temperature,
surface pressure, specific humidity, downward shortwave radiation, U-wind and V-
wind. From these fields, wind speed, net extra radiation per unit area of body
surface, relative humidity and vapor pressure were derived.
NLDAS-2 forcing fields are derived from the National Centers for Environmental
Prediction (NCEP) North American Regional Reanalysis (NARR) data. NARR fields
have a 32km spatial resolution and 3-hourly temporal resolution. The NARR fields
are derived from the 3-hourly NCEP Eta Data Assimilation System (EDAS) when
available (~92% of the time), otherwise the 3-hourly and 6-hourly Eta mesoscale
model forecast fields are used (Cosgrove et al., 2003). The NARR assimilation
system is fully cycled including prognostic land states with a 3-hourly forecast from
the previous cycle serving as a first guess for the next cycle (Mesinger et al., 2006).
Additional information about the NARR field processing can be found in Mesinger et
al. (2006).
The NLDAS processing system performs a spatial interpolation to downscale NARR
fields to the NLDAS 0.125° grid. For all fields used in this study, a bilinear
interpolation procedure is used. This interpolation method conserves the area-
averaged values, while simultaneously allowing for interpretation to a higher
resolution grid (Cosgrove et al., 2003). Following this, a temporal interpolation is
performed to adjust to the NLDAS hourly resolution. Once the other interpolations
13
are completed, Geostationary Operational Environmental Satellite (GOES) based
shortwave radiation data are used to produce that NLDAS forcing fields (Cosgrove et
al., 2003).
The topography of the NLDAS 12.5km grids is different from the NARR 32km grids,
and therefore temperature, pressure, humidity and longwave radiation fields must
be adjusted. First, a lapse rate of -6.5 K km-1 is applied to adjust temperature over
the change in elevation. Using the adjusted temperature, the 2m pressure fields are
adjusted using the Hydrostatic Approximation and Ideal Gas Law. Specific humidity
fields are adjusted by assuming constant relative humidity across the change in
elevation, as well as the equations of state for water vapor and dry air, the definition
of specific humidity and Wexler’s saturated water vapor pressure equation.
Downward longwave radiation is adjusted using the Stefan Boltzman law.
Further details and equations for the interpolations and elevation adjustments can
be found in Cosgrove et al., 2003. The implication of the interpolation method for
this study is that heat indices defined by absolute thresholds will be responsive to
lapse rate corrections, while indices defined using local statistical distributions will
be insensitive to the static in time local corrections. We note that for all indices the
resolution of NLDAS-2 is appropriate for regional studies but is inadequate to
address questions related to urban heat islands, within-municipality variability in
extreme heat exposure, or localized effects of contrasts in topography and land
cover.
14
Quality control of the NLDAS-2 forcing fields is based on the Assistance for Land
surface modeling Activities (ALMA) forcing data conventions. In addition, validation
studies have shown NLDAS forcing fields to be extremely realistic (Cosgrove et al.,
2003). One such study focuses on the NLDAS forcing fields validated against in-situ
measurements from the Oklahoma Mesonet monitoring network (Luo et al., 2003).
The fields of air temperature, surface pressure, specific humidity, downward
shortwave and longwave radiation all had an R ≥ 0.92, while wind speed has an
R=0.75. Similar validation studies are being completed on NLDAS-2 forcing fields
(http://ldas.gsfc.nasa.gov/nldas/NLDAS2valid.php).
2.2.2. Heat wave indices
Table 2.1 shows the sixteen, previously published HI used in this study. The indices
are divided into two main groups according to the type of threshold used in defining
heat wave days: relative thresholds and absolute thresholds. Data from the warm
season (1 May – 30 September) was used for the years 1979-2011.
2.2.2.1. Relative thresholds
All relative threshold heat indices use the climatological mean over the 1979-
2011 time period to calculate the percentiles.
Indices HI01 through HI06 are drawn from Anderson and Bell (2011). For each,
a threshold defined on the basis of the long-term local temperature record must
15
be met for at least two consecutive days. HI01 uses daily Tmean and the 95th
percentile as the threshold. HI02, HI03 and HI04 apply the same rule to the 90th,
98th, and 99th percentiles of daily Tmean, respectively. HI05 uses the 95th
percentile of Tmin and HI06 uses the 95th percentile of Tmax.
HI07 (Peng et al, 2011; Meehl and Tebaldi, 2004) applies a three-step process to
define heat wave days. First, for every day in a heat wave the Tmax must be over
the 81st percentile. Second, the Tmax must exceed the 97.5th percentile for at
least three consecutive days within the heat wave, but also allowing for days
with Tmax < 97.5th percentile. Lastly, the whole time period classified as a heat
wave must have an average Tmax greater than the 81st percentile.
HI08, HI09 and HI10 are based on thresholds of Apparent Temperature (AT).
Following Steadman (1984), AT is calculated as:
𝐴𝑇 = −1.8 + 1.07𝑇 + 2.4𝑉𝑃 − 0.92𝑣 + 0.044𝑄𝑔 (2.1)
where AT is measured in ºC, T is temperature in ºC, VP is vapor pressure in kPa,
v is wind speed in m s-1, and Qg is net extra radiation per unit area of body
surface in W m-2. AT is calculated every hour throughout the day, and then daily
ATmax is classified into one of three categories: hot (> 85th percentile), very hot
(> 90th percentile) and extremely hot (>95th percentile).
16
2.2.2.2. Absolute thresholds
HI11 uses a straightforward approach: everyday that Tmax is over 35ºC is
classified as a heat wave day (Tan et al, 2007).
HI12 (Robinson, 2001) uses both Tmin and Tmax to define heat wave days: the
threshold value for Tmax is 40.6°C and for Tmin is 26.7°C. At least one of these
thresholds must be met on at least two consecutive days to classify each of those
days as a heat wave day.
HI13 through HI16 use the National Weather Service’s (NWS) heat index
(INWS). INWS is calculated using temperature and relative humidity, however
other factors such as vapor pressure, wind speed, characteristics of human
activity levels, sweating rate etc. were used to parameterize the equation
(Rothfusz, 1990). The heat index is calculated using Equation 2.2:
𝐼𝑁𝑊𝑆 = −42.379 + 2.04901523𝑇 + 10.14333127𝑅 − 0.22475541𝑇𝑅 − 0.00683783𝑇2 −
0.05481717𝑅2 + 0.00122874𝑇2𝑅 + 0.00085282𝑇2𝑅2 − 0.00000199𝑇2𝑅2 (2.2)
where INWS is the heat index in ºF, T is the temperature in ºF and R is relative
humidity measured as a percentage (Rothfusz, 1990; Steadman, 1979). INWS is
calculated at each hour throughout the day, and then the daily maximum INWS is
classified into one of four categories: 1) Caution: >80ºF [HI13], 2) Extreme
17
Caution: >90ºF [HI14], 3) Danger: >105ºF [HI15] and 4) Extreme Danger:
>130ºF [HI16].
There are two instances where an adjustment is made to the calculation of INWS
(http://www.hpc.ncep.noaa.gov/heat_index/hi_equation.html). First, if R is less
than 13% and T is between 80°F and 112°F, then the following adjustment value
is subtracted:
13−𝑅
4∗ √
17−|𝑇−95|
17 (2.3)
Second, if R is greater than 85% and T is greater than 80°F but less than 87°F,
then the following adjustment value is added:
𝑅−85
10∗87−𝑇
5 (2.4)
The highest threshold value, HI16, was too rare to inform statistical analyses of
frequency or trends. Because of this, HI16 results will not be shown.
2.2.3. Statistical analysis
For each HI, the number of heat wave days was summed annually (warm season) at
the NLDAS grid cell scale. To gain insight into regionalized patterns of the HI, CONUS
was divided into six regions: Northwest (Washington, Oregon, Idaho), Southwest
(California, Nevada, Utah, Arizona, New Mexico and Colorado), Great Plains (North
Dakota, South Dakota, Montana, Wyoming, Nebraska, Kansas, Oklahoma, Texas),
Midwest (Minnesota, Iowa, Missouri, Wisconsin, Illinois, Indiana, Ohio, Michigan),
18
Southeast (Arkansas, Louisiana, Mississippi, Alabama, Tennessee, Kentucky,
Georgia, Florida, South Carolina, North Carolina, Virginia) and Northeast
(Pennsylvania, New Jersey, New York, Rhode Island, Connecticut, Massachusetts,
Vermont, New Hampshire, Maine, Maryland, West Virginia, Delaware). These
regions are depicted in Figure 2.1 and approximately match the six geographical
regions used for regional climate change analysis in the United States Global Change
Research Program (USGCRP) report Global Climate Change Impacts in the United
States (2009).
Figure 2.1 Regional division of the Continental United States (CONUS).
For each of the fifteen HI, the total number of annual heat wave days was averaged
over the 33-year timespan at the NLDAS grid cell scale. These results were then
19
averaged over the six CONUS regions to arrive at the average number of heat wave
days unique for each heat wave index and region.
These heat wave day averages were then assessed for their trends over the 1979-
2011 time period using ordinary least squares (OLS) regression. As the OLS
residuals exhibited non-normality for several indices (as shown by Shapiro-Wilks
normality testing), significance tests were performed using the Mann-Kendall tau
test. The Mann-Kendall tau test is a nonparametric test that does not assume an
underlying probability distribution of the data, and is also robust to outliers
(Moberg et al., 2006). Because of this, it is valuable when assessing trends in climate
data and therefore has been used in previous studies of trends in extreme
temperature indices (El Kenawy et al, 2011; Efthymiadis et al, 2011; Kuglitsch et al,
2010). A trend was considered statistically significant if the p-value was smaller
than the significance level α of 0.05. Based on the results of the Mann-Kendall
testing, all insignificant trends were masked out. The average of the resulting
significant trends was computed for each of the six CONUS regions for all fifteen HI.
Results are reported for trends calculated over the entire 1979-2011 period of
record available at the time the study was performed. The sensitivity of trends to
choice of time period was assessed by repeating the trend analysis with the
beginning date shifted between 1979 and 1981 and the end date shifted between
2007 and 2012 (provisional data used for 2012). It was found that results were very
similar for all analyses that included data through at least 2009. For periods that
20
excluded recent years the geographic pattern and direction of trends was similar,
but statistical significance of trends tended to be reduced.
To compliment these trend values, landmass coverage was calculated. These
landmass coverage percentages represent the number of cells covered by significant
trends (which were averaged) divided by the total number of cells in that region.
2.3. Results
To provide an example of the prevailing climate characteristics used to define the
relative HI, Figure 2.2 shows the average daily Tmax, standard deviation of the daily
Tmax, and the 95th percentile threshold for Tmax for May-September 1979-2011. That
is to say, for HI06 the temperature field mapped in Figure 2.2c must be met for at least
two consecutive days for those days to count as a heat wave day.
Figure 2.2 (a) Average daily Tmax (ºC) over the time period 1979-2011 (b) Standard deviation of Tmax
over the time period 1979-2011 (c) 95th percentile of Tmax(ºC).
2.3.1. Average annual heat wave days
21
Figure 2.3 shows the results of the analysis of the average number of annual heat
wave days; note that scale bars differ depending on the definition so as to
accentuate the spatial patterns. To get a quantitative look at these differences, Table
2.2 shows the results of this analysis for each of the six CONUS regions.
Figure 2.3 1979-2011, annual average number of heat wave days. Note the varying scales.
22
The Northwest region experienced the highest frequency of heat wave occurrence
for HI01-HI05, HI08 and HI09, the Southeast experienced the highest frequency for
HI06, HI07, HI10, HI11, and HI13-HI15 and the Southwest experienced the highest
frequency for HI12. The Great Plains, Midwest and Northeast regions did not see the
highest frequency of heat wave occurrence for any HI. The range of averages
between the six regions was much smaller for relative threshold definitions than the
range seen between absolute threshold definitions.
2.3.2. Temporal trends
Figure 2.4 shows the results of temporal trend analysis. Only the trends with
significance over 95%, based on the Mann-Kendall test, are shown in color. The
majority of the trends are positive, indicating that from 1979 to 2011 the average
number of annual heat wave days has increased. Figure 2.5 summarizes this trend
information for defined CONUS regions in terms of magnitude and extent of
significant trends. The value printed in each cell is the trend value (in days/year).
The color of the cell represents the direction of the trend, where red represents
positive trends and blue represents negative trends. The shades of red and blue
represent the landmass percentage covered by this significant trend, increasing in
percentage from the lightest shades to the darkest shades.
23
Figure 2.4 Trends in the number of annual heat wave days, over the period 1979-2011. White areas
indicate results below 95% significance. Units are days/year.
The majority of the trends observed across CONUS regions are positive, with the
largest positive trends found in the Southeast and Great Plains regions. For the
positive trends, the percentage of landmass covered ranges from 0-38%. Positive
24
trends were found across more than 12% of the landmass area for the majority of HI
for the Southeast and Great Plains regions; coverage over 12% was found for twelve
HI in the Southeast and eight indices in the Great Plains. The largest magnitude
negative trend was found in the Southwest region, with other negative trends seen
in the Northwest and Great Plains regions. Landmass coverage of these negative
trends ranged from 0 -12%.
Figure 2.5 Average 95% significant trends in the number of annual heat wave days, over the period
1979-2011, divided by region. The value printed in each cell is the trend value (in days/year). The color
of the cell represents positive (red) and negative (blue) trends. The shades of red and blue represent the
landmass percentage covered by this significant trend given by the scale bar. Regions are: Northwest
(NW), Southwest (SW), Great Plains (GP), Midwest (MW), Southeast (SE) and Northeast (NE).
2.4. Discussion
25
The results of the index intercomparison presented in this study are generally
consistent with warming trends observed in previous studies (Meehl and Tebaldi, 2004;
Schar et al., 2004; Hansen et al., 2012). Our study shows that in the past 33 years the
frequency of heat waves days has increased across most CONUS regions, according to
the majority of HI. Some of our results show regional exceptions to these overall trends:
negative trends were found for HI01 and HI02 in the Northwest, HI13 in the Southwest
and Great Plains and HI14 and HI15 in the Southwest.
In looking at the frequency of heat wave days over the past 33 years, the geographical
patterns are varied between HI. Most of the patterns shown in Figure 2.3 have a
sensible explanation based on what information was included in the definition of a heat
wave day. For example, the geographical patterns of high heat wave day frequency in
HI13 and HI14 correspond to the areas of CONUS with high humidity; this is consistent
with the fact that these definitions include both temperature and relative humidity data.
Our results do differ from some previous studies of heat wave trends in the United
States. These apparent inconsistencies can be attributed to differences in time period of
analysis, data source, spatial resolution of analysis, and heat wave definitions. For
example, Alexander et al. (2006) found that the Southeast region has experienced a
decrease in heat waves, but their study used an earlier time period, 1951-2003, and
defined a heat wave on the basis of single day Tmax. Robinson (2001) also found
decreases in heat waves throughout the US South. That study uses the HI12 definition of
a heat wave day, but is based on an earlier time period, 1951-1990, and uses station
26
data that resulted in a coarser temporal resolution. Such discrepancies make it difficult
to compare results across studies and can lead to some confusion in assessments of the
magnitude and geography of heat wave trends.
The diversity of HI found in the literature is understandable, considering the range of
reasons that heat waves are studied. While climate scientists are often concerned with
trends in the statistics of high temperature, health experts focus on indices that capture
impacts on human well-being, which are frequently influenced by social factors such as
acclimation and exposure. Collaboration between climate and health communities is
particularly valuable in this context. For example, a number of health-oriented studies
have demonstrated that the largest mortality effects due to increased heat occur in
northern cities in the United States, or areas with milder summers (Anderson and Bell,
2011; Medina-Ramon and Schwartz, 2007; Curriero et al., 2002). Here, we find that in
the NLDAS record, ten of the fifteen indices show that the percent area experiencing
significant increases in heat waves is larger in the Southeast than in the Northeast, and
that for nine of these ten indices the magnitude of trend has been greater in the
Southeast than the Northeast. Therefore, according to the majority of indices
considered in this study the Southeast has seen larger and more spatially extensive
trends than the Northeast. Future studies can pair this information with the relative
health effects of heat waves in differing regions to grasp a complete understanding for
planning health interventions and climate change adaptation strategies. The Northwest
region is another example of the potential value in communication across fields: while
several of the objective indices in this study had their highest frequency in the
27
Northwest, health experts recognize that heat waves have not been a major health
concern in that region relative to other parts of the country.
There are also discrepancies between the research literature and operational health
warnings. Davis et al. (2006) found that relative threshold HI are better predictors of
health impacts than those that use absolute thresholds. However, HI13-HI16, which use
absolute thresholds, are of particular importance to the general public because they
align with the heat alerts that are issued by the National Weather Service and broadcast
by organizations such as The Weather Channel. In a recent study of US football player
deaths, Grundstein et al. (2012) found that 45% of these deaths occurred in conditions
that did not trigger an NWS alert (meaning conditions less than the HI13 threshold).
The spatial patterns and trends observed in HI13-HI15 do not align particularly closely
with any other HI, which raises the question of whether the public is receiving the most
meaningful information regarding heat wave events.
The specific results of our study are limited by the dataset used in the analysis: NLDAS-
2 is the preeminent gridded land surface reanalysis for the United States, but it relies on
relatively coarse NARR fields as the foundation for meteorological estimates and its
downscaling routines do not account for localized differences in lapse rate or surface
properties or for nonstationarities such as land-use change. For this reason the patterns
and trends identified in this study must be understood as mesoscale results that do not
account for phenomena such as urban heat islands that are relevant to local health
impacts. In addition, only monotonic trends in heat wave days were considered; higher
28
order trend analysis could provide further insight on recent climate changes.
Nevertheless, it is shown that choice of index is critically important to the resulting
analysis of patterns and trends. In addition, these index comparisons could be
translated easily to different datasets for more focused local analyses.
2.5. Conclusions
This study demonstrates that estimates of the frequency, trends, and geographical
patterns of heat waves in the CONUS strongly depend on how heat waves are defined.
This fact, combined with discrepancies between studies in the time period considered,
meteorological datasets used, and spatial resolution of analysis, has led to a wide range
of conclusions regarding frequency and trends of extreme heat events in the United
States.
This study shows that across CONUS regions the range in average number of heat wave
days is greater for absolute HI than for relative HI. For both the relative and absolute HI,
the Southeast saw the highest values of average heat wave days, indicating that the
Southeast has experienced more heat wave days from 1979-2011 than any other
CONUS region. From the trend analysis, this study has shown that the Southeast and
Great Plains regions have experienced both the largest magnitude and most widespread
increases in heat wave days per year according to most indices.
Characterization and regionalization of heat waves across the United States is essential
to expanding our knowledge of climate change processes and impacts. Understanding
29
the role that definitions play in such studies is important for interpreting seemingly
contradictory results and for enhancing the quality of communication between climate
scientists, health researchers, and the general public. In this study, we have applied one
consistent dataset to compare patterns and trends across fifteen previously published
HI. Similar comparisons can be performed for other meteorological datasets and for
future climate projections in order to explore the full range of heat wave impacts
associated with climate variability and change.
30
Table 2.1 Definitions of heat wave indices. *HI16 did not have enough data to include in this analysis.
Heat Wave
Indices (HI) Temperature Metric Threshold Duration HI Type Reference(s)
HI01 Mean daily temperature > 95th
percentile 2+ consecutive days Relative Anderson and Bell (2011)
HI02 Mean daily temperature > 90th
percentile 2+ consecutive days Relative Anderson and Bell (2011)
HI03 Mean daily temperature > 98th
percentile 2+ consecutive days Relative Anderson and Bell (2011)
HI04 Mean daily temperature > 99th
percentile 2+ consecutive days Relative Anderson and Bell (2011)
HI05 Minimum daily temperature > 95th
percentile 2+ consecutive days Relative Anderson and Bell (2011)
HI06 Maximum daily temperature > 95th
percentile 2+ consecutive days Relative Anderson and Bell (2011)
HI07 Maximum daily temperature
T1: > 81st
percentile Everyday, >T1; 3+
consecutive days, >T2;
Avg Tmax >T1 for whole
time period
Relative Peng et al (2011); Meehl and Tebaldi
(2004) T2: > 97.5th
percentile
HI08 Maximum daily apparent
temperature >85
th percentile 1 day Relative Steadman (1984)
HI09 Maximum daily apparent
temperature >90
th percentile 1 day Relative Steadman (1984)
HI10 Maximum daily apparent > 95th
percentile 1 day Relative Steadman (1984)
31
temperature HI11 Maximum daily temperature > 35°C 1 day Absolute Tan et al (2007)
HI12 Minimum & maximum daily
temperature
Tmin > 26.7°C ≥ 1 threshold for 2+
consecutive days Absolute Robinson (2001)
Tmax > 40.6°C
HI13 Maximum daily heat index >80°F 1 day Absolute National Weather Service, Rothfusz
(1990); Steadman (1979)
HI14 Maximum daily heat index >90°F 1 day Absolute National Weather Service, Rothfusz
(1990); Steadman (1979)
HI15 Maximum daily heat index >105°F 1 day Absolute National Weather Service, Rothfusz
(1990); Steadman (1979)
HI16* Maximum daily heat index > 130°F 1 day Absolute National Weather Service, Rothfusz
(1990); Steadman (1979)
32
Table 2.2 Average number of annual heat wave days, divided by region. Bold indicates region with highest
frequency of heat waves days for each HI. Regions are: Northwest (NW), Southwest (SW), Great Plains (GP),
Midwest (MW), Southeast (SE) and Northeast (NE).
NW SW GP MW SE NE
HI01 3.92 1.54 2.23 2.73 1.76 2.56
HI02 11.95 7.77 8.82 9.39 7.00 9.79
HI03 0.77 0.16 0.30 0.54 0.26 0.30
HI04 0.24 0.05 0.07 0.15 0.06 0.05
HI05 3.92 1.45 1.44 1.97 0.21 2.02
HI06 3.42 1.96 3.12 3.73 4.59 2.71
HI07 0.50 0.27 0.78 1.42 2.26 0.49
HI08 22.53 20.01 20.98 21.54 21.08 21.72
HI09 12.47 9.90 11.01 12.09 12.13 11.96
HI10 3.60 2.23 2.89 3.76 3.94 3.53
HI11 2.25 21.76 23.06 3.80 6.76 0.16
HI12 0.10 5.94 4.48 0.46 4.49 0.01
HI13 4.00 15.92 52.93 58.75 113.63 28.54
HI14 0.23 2.04 23.58 26.48 72.10 7.61
HI15 0.00 0.16 1.29 2.35 4.66 0.17
33
3. CHAPTER 3: LARGE-SCALE DRIVERS OF INTERANNUAL SUMMERTIME
TEMPERATURE VARIABILITY ACROSS THE CONTINENTAL UNITED STATES2
ABSTRACT
High summertime temperatures are associated with myriad impacts on humans and the
natural environment across the Continental United States (CONUS). As such, it is important
to improve understanding of the drivers of temperature variability in the summertime
season. Here, an objective regionalization of CONUS based on summertime temperature
variability is performed, yielding five regions that exhibit distinctly different behavior.
These regions are used to develop statistical models of July-August average daily minimum
temperatures (Tmin) in each region, from 1950-2012, as a function of large-scale climate
modes. Recognizing that summertime climate variability can be difficult to predict using
standard linear statistical techniques, multiple parametric and non-parametric statistical
modeling approaches are applied to capture non-linear responses and interactions
between climate modes. Results indicate that detrended summertime Tmin variability is
sensitive to different climate processes in different regions, including the Pacific Decadal
Oscillation variability in the Southwest, Arctic Oscillation in the Northern Great Planes,
Atlantic Multidecadal Oscillation in the South, and late spring soil moisture anomalies in
the Northeast. Notably, nonlinear models yielded the best results in all regions: a random
forest model is found to be most robust in the SW and South, while a generalized additive
model is most robust for the NGP and Northeast. These results are valuable for informing
2 Smith T.T., B. F. Zaitchik and S. D. Guikema, (2016) Large-scale drivers of interannual summertime temperature variability across the Continental United States. Journal of Applied Meteorology and Climatology. (Under review)
34
future improvements in seasonal forecasting of summertime temperatures and are
relevant for projections of summertime climate change across climatically distinct regions
of CONUS.
3.1. Introduction
It is well documented that mean global temperatures are on the rise (IPCC, 2007). Over the
time period 1951-2003, statistically significant increases in minimum and maximum
temperatures were seen over 40-75% of global land area (Alexander et al., 2006; Trenberth
et al., 2007). Over the Continental United Sates (CONUS), it has been projected that by the
mid-21st Century 50% of summers will be as hot as the top 5% of summers in the historic
baseline (Duffy and Tebaldi, 2012). Increased summertime temperatures have been
associated with consequences ranging from human morbidity and mortality (Peng et al.,
2011; Curriero et al., 2002) to increased wildfire activity (Westerling et al., 2006) and
decreased agricultural yields (Lobell et al., 2013).
Because of the diverse impacts that high summertime temperatures have in CONUS, it is
important to clarify our understanding of the drivers of temperature trends and variability.
These drivers can include remote large-scale modes (LSM) of climate variability (e.g.,
Barnston 1996; Drosdowsky and Chambers, 2001; Kenyon and Hegerl 2008) and local
feedback mechanisms (e.g., Fischer et al., 2007; Portmann et al., 2009).
Historically, the most well studied LSM is the El Nino Southern Oscillation (ENSO). For
several decades now, ENSO has been associated with changes in extreme temperature
35
frequencies across CONUS (Gershunov and Barnett, 1998). In addition, ENSO has been
shown to increase the predictability of climate variables, with the general consensus that
higher predictability is found in years when ENSO is strongest (Barnett et al., 1997;
Brankovic and Palmer, 2000; Visbeck et al., 2001; Becker et al., 2013). This result has been
broken down further to highlight the asymmetry between El Nino and La Nina events. For
instance, the low-level jet over the Great Plains was found to have a stronger relationship
with ENSO during El Nino events (Krishnamurthy et al., 2015), while hot and dry summers
in the central US have been attributed to La Nina events (Wang et al., 2007).
More recently, studies of CONUS temperatures have begun to include a wider variety of
large-scale modes of variability, including the Pacific-North American mode (PNA), the
Northern Annular mode (NAM), the Pacific Decadal Oscillation (PDO), the North Atlantic
Oscillation (NAO), and the Atlantic Multidecadal Oscillation (AMO) (Loikith and Broccoli,
2014; Kenyon and Hegerl, 2008; Zhang et al., 2007; Sutton and Hodson, 2005). Because of
the wealth of research surrounding ENSO, some studies focus on how a particular index
interacts with ENSO. Again, asymmetry has been found in these relationships, such as
differing correlation between precipitation and ENSO during the warm and cool phases of
the AMO (Hu and Feng, 2012).
Importantly, wintertime climate characteristics are known to have a more robust
relationship with LSM (Becker et al., 2013) due to the higher activity of most LSM during
the winter months (Visbeck et al., 2001). However, studies of LSM relationships to
summertime climate have become increasingly common (Wang et al., 2007; Krishnamurthy
36
et al, 2015) in response to the plethora of evidence showing negative impacts of
summertime climate on human health (USGCRP, 2016), and the increasingly extreme
patterns of summertime climate (Meehl & Tebaldi, 2004).
Here, we characterize and attempt to explain temperature variability in the two hottest
months of the year (July-August). We focus on variability in minimum daily temperature
(Tmin) for two reasons. First, though there is no consensus on what characteristics of heat
exposure are most harmful to human health (e.g., Anderson and Bell, 2011; Kent et al.,
2014), a number of studies suggest that elevated Tmin during a heat wave has specific
physiological impacts because the human body is given no relief at night (Schwartz, 2005;
Medina-Ramon et al., 2007). Second, it has been shown that Tmin is strongly influenced by
large-scale patterns whereas local effects have a major influence on maximum temperature
(Tmax) (Alfero et al., 2006), suggesting that there is greater potential to predict Tmin. In
addition, we chose to focus on the July-August average, as opposed to each month
individually, as multi-month means have been found to have higher predictive ability than
one-month means, in part on account of higher signal-to-noise ratio (Barnston, 1994;
Kumar et al., 2000; Becker et al., 2013).
In order to contend with the difficulty of predicting local climate variability as a function of
LSM in summer months we employ two statistical approaches to improve predictive power
of our models. First, we apply objective climate regionalization (Manning et al., 2008;
Dezfuli 2011; Badr et al., 2015) to divide CONUS into regions that are coherent with respect
to summertime temperature variability. This allows us to train and apply statistical models
37
using response regions that are optimally selected for homogenous temperature
variability. The use of standard climate regions (e.g., Melillo et al., 2014) or political
boundaries to define regions is not necessarily optimal for this problem. Second, we
employ a suite of parametric and non-parametric regression techniques to capture
nonlinear responses and interactions between predictors. Nonlinear statistical techniques
and machine learning algorithms have become increasingly common in climate analysis
and forecast application (Lobell et al., 2010; Rasouli et al., 2012; Nicholson, 2014; Badr et
al., 2014). To our knowledge, however, this is the first study that compares multiple
parametric and non-parametric techniques to study summertime temperature variability
in CONUS.
We note that this work is complementary to and distinct from extreme value analysis
applied to heat wave analysis and prediction (Tarleton and Katz, 1995; Cheng et al., 2014;
Hansen, 2012). Here we focus on a metric of seasonal average temperatures in order to
achieve stronger statistical relationships (Pepler et al., 2015) and explain large-scale
drivers of variability on seasonal timescales. Because these LSM are known to persist, this
work will inform future studies of seasonal forecasting and climate change impacts
specifically relevant across heat-vulnerable regions of CONUS.
3.2. Methods
3.2.1. Datasets
Temperature data for the study is from the Climate Research Unit (CRU) monthly
dataset (CRU TS3.21). This global dataset has a 0.5-degree spatial resolution and spans
38
the period 1901-2012. We choose CRU because of its long record, widespread use, and
global coverage, which makes it applicable for future studies outside of CONUS. This
study focuses on minimum temperature (Tmin) during the height of boreal summer,
July-August, for the years 1950-2012.
Large-scale mode (LSM) data for the study were gathered from the National Oceanic
and Atmospheric Administration (NOAA) Earth System Research Laboratory (ESRL);
additional information can be found at
http://www.esrl.noaa.gov/psd/data/climateindices/list. This study uses seven LSM
that have plausible associations with CONUS temperature variability: the El Nino
Southern Oscillation (ENSO), North Atlantic Osciallation (NAO), Pacific Decadal
Oscillation (PDO), Pacific-North American Mode (PNA), Atlantic Multidecadal
Oscillation (AMO), Arctic Oscillation (AO), and Global Mean Land/Ocean Temperature
Index (GMSST).
Pressure-based indices used in this study are NAO, PNA and AO. The NAO is an index of
sea level pressure (SLP) between centers of action over Iceland and the subtropical
Atlantic, near the Azores. The index if in its positive phase when SLP is below normal
near Iceland and above normal near the Azores; the negative phase is defined by the
opposite pattern (Barnston and Livezey, 1987). NAO varies on sub-seasonal timescales.
Here we use the July-August average as an indicator of the preferred orientation of NAO
in each year. The PNA is a pattern of anomalies in the 500 hPa geopotential height field
with centers of action over the Aleutian Islands and Southeast CONUS (Barnston and
39
Livezey, 1987). Positive phase PNA brings above normal geopotential heights over
western CONUS and below normal geopotential heights over eastern CONUS. This
pattern in the geopotential height fields leads to cold, Canadian air spilling south into
eastern CONUS while western CONUS experiences above normal temperature; the
opposite is true during negative phase PNA. The AO, also known as the Northern
Annular Mode (NAM), is defined as the 1st empirical orthogonal function of the 100 hPa
height field poleward of 20N, and is normalized by the standard deviation of the
monthly 1979-2000 baseline monthly values (Higgins et al., 2000).
Temperature-based indices used in this study are PDO, AMO and GMSST. The PDO
consists of the first principal component (PC) of monthly sea surface temperature (SST)
anomalies in the Pacific Ocean poleward of 20N (Zhang et al., 1997). Positive phase
PDO, also referred to as the warm phase, is associated with a cold tongue of SST in the
interior North Pacific, with warm SSTs along the eastern edge of the Pacific basin. The
AMO is an index of North Atlantic (0-70N) SSTs, computed using Kaplan SST data
(Enfield et al., 2001); for this study, the unsmoothed version of the dataset was used.
The GMSST is an anomaly index produced by the Goddard Institute for Space Studies
(GISS) where positive (negative) values refer to a positive (negative) anomaly relative
to a 1951-1980 baseline (Hansen et al., 2010). Because values of these indices are
sometimes revised, it is important to note the dataset used herein was obtained on 11
April 2013.
40
In addition, this study uses the Multivariate ENSO Index (MEI) for the ENSO variable.
MEI uses the first, unrotated principal component of six observed fields over the central
Pacific Ocean: sea-level pressure, zonal surface winds, meridional surface winds, sea
surface temperature, and total cloudiness fraction of the sky (Wolter 1987). Positive
MEI values represent the well-known warm phase, El Nino, while negative MEI values
represent the cool phase, La Nina.
Lastly, this study includes local soil moisture as a proxy for local-scale feedback systems
that may be present in addition to large-scale forcings. The soil moisture (SM) data are
drawn from the NOAA Climate Prediction Center (CPC) Soil Moisture V2 dataset. CPC
Soil Moisture V2 is a global dataset with 0.5 degree spatial resolution that includes
monthly averaged soil moisture water height equivalents from 1948 to present (Fan
and van den Dool, 2004;
http://www.esrl.noaa.gov/psd/data/gridded/data.cpcsoil.html).
As the purpose of this study is explanatory modeling rather than forecasts, we apply all
predictors synchronously: July-August LSM are used to predict July-August Tmin. In
some cases lead-time predictors may actually have greater statistical power than
synchronous predictors, but exploratory analysis to support this study found that the
large majority of associations were strongest for synchronous analysis, so we do not
consider lead-time predictors in our models. The one exception to this is soil moisture.
Because local soil moisture anomalies are mechanistically linked to temperature
anomalies in many regions, July-August soil moisture is not an independent predictor of
41
July-August Tmin. For this reason we use April-May soil moisture anomaly as a
predictor, with the understanding that late spring soil moisture has the potential to
impact late summer temperature through vegetation-mediated feedbacks (Fischer et
al., 2007).
3.2.2. Regionalization
For this study, regions were defined by the variable of interest, interannual variability
in July-August Tmin, using a hierarchical clustering analysis. Previous studies have
proposed objective regionalizations of CONUS based on mean climate conditions or
patterns of variability (Fovell and Fovell, 1993; Bukovsky et al., 2013), but for this study
it was important to utilize the variable of interest, specifically, to draw regional
boundaries so as to optimize the scope of this work. Regions for this study were defined
by Ward’s minimum variance clustering algorithm (Ward, 1963; Murtagh, 1983) as
implemented in the HiClimR package for R (Badr et al., 2015). Ward’s method aims to
find compact, n-spherical clusters by utilizing a minimum-variance method, and has
been used for similar climate regionalization applications in the past (Ward, 1963).
Regionalization was performed on detrended (linear trend removed) and standardized
Tmin data. Ward’s method was then applied, with the number of regions prescribed by
the user. Selection of the number of regions is a subjective process that depends on
application. Here we recognized that we needed a relatively small number of regions
(e.g., 4-7) in order to align our study with the scale of analysis used in the National
Climate Assessment (NCA) and other widely recognized CONUS climate analyses
42
(Melillo et al., 2014). This number of regions is also appropriate for associations with
LSM, which tend to have impacts across large areas within CONUS. Within these
subjective constraints, the final regions are selected based on objective metrics of
intraregional homogeneity versus inter-regional correlation that are produced by the
clustering algorithm.
Following initial regionalization, additional pre-processing was applied to the dataset to
remove cell values that were not well correlated with the region’s time series. The
motivation for this pre-processing was to remove noise and end with the most cohesive
regions to analyze. As such, the mean time series for each region was correlated to the
time series at each grid cell within the region using Pearson’s product moment to
calculate the correlation coefficient. All grid cells correlated less than 0.6 were then
removed from the dataset. Final regional time series were then re-calculated using only
the remaining grid cells.
3.2.3. Prediction
This study used a suite of statistical models to investigate the predictability of the
regionalized Tmin time series. The use of varied statistical techniques offers advantages
for capturing nonlinear effects, diagnosing variable importance, and accounting for
interacting relationships between predictors. The ability to capture these effects and
interactions is particularly important when studying diverse drivers of climate
variables, which often exhibit nonlinear and interacting effects.
43
This study includes three types of models--linear, additive, and tree-based--for a total of
five models. The first is a generalized linear regression model (GLM), which is an
ordinary least squares model with an added link function that allows for the model to
respond based on the distribution of the response variable. The response variable,
Tmin, displays a Gaussian (normal) distribution, and as such the identity link function is
used. The additive model used is a generalized additive model (GAM) that utilizes both
linear terms and terms with a cubic natural regression spline applied. The GAM is a
semi-parametric regression model that is based on a GLM, but adds the functionality of
a smoothing function (spline) to summarize the trends of the response variable against
the covariate(s) (Hastie and Tibshirani, 1986). When used, the smoothing functions are
fit using penalized likelihood maximization to prevent overfitting the model.
This study also used two tree-based models, a classification and regression tree (CART)
and a random forest (RF). Tree-based models use recursive binary partitioning of the
dataset to build either a single tree or an ensemble of trees. The CART model builds a
single tree by partitioning the dataset to reduce the sum of squared errors (SSE) and
then pruning this tree by removing nodes in an effort to reduce out of sample
prediction error. For this study, the RF model was built to an ensemble of 500 trees
using the mean square error (MSE) to determine the optimal partitioning of the dataset.
As a baseline comparison tool, a mean model (CLIM) was also calculated by taking the
average value of the response variable, Tmin.
44
To determine the model with the best predictive accuracy, a 100-fold holdout cross-
validation analysis was performed where a randomly selected 10% of the data (seven
years) were held out for each iteration. The models then predicted Tmin using the
remaining 90% of the data (56 years). Predictive accuracy was assessed using the
average and standard error of the mean squared error (MSE) of the 100-fold holdout
analysis.
To test for statistical significance, a student’s t-test was used to quantify p-values for
each model. These p-values where evaluated at the 95% confidence level for
independent tests, but also at two incrementally more stringent measures that adjust
for multiple comparisons. The first measure is false discovery rate (FDR), sometimes
referred to as Banjamini & Hochberg (BH), which controls for the expected proportion
of false discoveries amongst the rejected hypothesis (Benjamini and Hochberg, 1995).
The most stringent p-value adjustment method used was the Bonferroni correction
(BF), which submits strong control of the family-wise error rate by multiplying the p-
values by the number of comparisons. For the remainder of this study, significance is
defined by FDR unless otherwise noted.
3.2.4. Leading Indicators
Using the best model, as indicated by the holdout analysis, variable selection was
applied in each region to determine the leading indicators. To do so a method was
adapted from Shortridge et al. (2015), in which partial dependence plots were
developed by fitting the model using all covariates and then quantifying the marginal
45
influence that changing the covariate of interest while keeping all other covariates
equal has on model predictions. The influence of each covariate was then quantified as
the range of partial dependence value associated with that covariate, divided by the
total swing over all covariates in that model (Equation (3.1)).
influence𝑚𝑛 =max(𝑃𝐷𝑚𝑛)−min(𝑃𝐷𝑚𝑛)
∑ Swing𝑚𝑛𝑛 (3.1)
These swings were then used to rank variable importance level, where large (small)
swings indicate more (less) influence on model predictions.
3.3. Results
3.3.1. Regionalization
Figure 3.1a shows the final regions analyzed for this study, where associations were
determined using Ward’s method and cells correlated < 0.6 with the region’s time series
are masked out (no color). Figure 3.1b shows the corresponding dendrogram, where
the horizontal line indicates the cutoff used to differentiate between the number of
regions to include, where each stem the horizontal line touches represents one of the
five final regions. Table 3.1 shows the intraregional correlations and interregional
correlations for the final five regions. Intraregional correlations ranged from 0.60 to
0.81, while interregional correlations ranged from (absolute value) 0.12 to 0.56. NGP
had the highest intraregional correlation, while its highest interregional correlation was
with NE. Despite producing the lowest interregional correlation, the NW also had the
lowest intraregional correlation.
46
Figure 3.1 (a) Map of regions and (b) the corresponding dendrogram. Y-axis in (b) is the sum of squared
distances within all regions, and is a measure of intra-regional variance.
3.3.2. Prediction
3.3.2.1. Linear Trend Present (LTP) analysis
−120 −100 −80 −70
20
30
40
50
60
Latitude
Longitude
a.
NW
SW
NGP
South
NE0
100
200
300
40
0
b.
47
First, the holdout analysis was completed using the time series with the linear trend
present, referred to hereafter as LTP. Figure 3.2 shows the time series of Tmin LTP
and Linear Trend Removed (LTR). Table 3.2 shows the average and standard error
of the MSE values resulting from the holdout analysis. Because the MSE is based on
the response variable (Tmin), which varies from region to region, it is only
appropriate to compare MSE values within regions and not across regions. Low
average and standard error MSE values indicate the model has high predictive
accuracy. GLM is the best performing model for the NW, SW and South regions,
while GAM has the lowest MSE for the NGP and NE regions; all of these models
significantly outperform the null model, CLIM (p < 0.05 with FDR adjustment).
48
Figure 3.2 Timeseries of Tmin, LTP (dashed red line) and Tmin, LTR (solid black line) for regions (a)
NW, (b) SW, (c) NGP, (d) South, and (e) NE.
1950 1960 1970 1980 1990 2000 2010
−2
−1
01
2
Tm
in
a.
1950 1960 1970 1980 1990 2000 2010
−2
−1
01
2
b.
1950 1960 1970 1980 1990 2000 2010
−3
−2
−1
01
2
Tm
in
c.
1950 1960 1970 1980 1990 2000 2010−
2−
10
1
d.
1950 1960 1970 1980 1990 2000 2010
−2
−1
01
23
Tm
in
e.
LTP LTR
49
Figure 3.3 Timeseries of LTP (dashed, red line) and LTR (black line) for all covariates: (a) ENSO, (b)
NAO, (c) PDO, (d) PNA, (e) GMSST, (f) AMO, and (g) AO.
Table 3.3 shows the swing value of each covariate for the BEST LTP model for each
region. High value swings indicate that covariate had a large influence on model
predictions, while low value swings indicate the covariate had a small influence on
1950 1960 1970 1980 1990 2000 2010
−2
01
23
EN
SO
a.LTP LTR
1950 1960 1970 1980 1990 2000 2010
−2
01
23
NA
O
b.
1950 1960 1970 1980 1990 2000 2010
−2
01
23
PD
O
c.
1950 1960 1970 1980 1990 2000 2010
−2
01
23
PN
A
d.
1950 1960 1970 1980 1990 2000 2010
−2
01
23
GM
SS
T
e.
1950 1960 1970 1980 1990 2000 2010
−2
01
23
AM
O
f.
1950 1960 1970 1980 1990 2000 2010
−2
01
23
AO
g.
50
model predictions. Partial dependence plots for the three most influential covariates
(largest swings) for each region are shown in Figure 3.4.
Figure 3.4 Partial dependence plots for the three leading indicators in each region for models built
with LTP data: NW – (a) GMSST, (b) PNA, (c) NAO; SW – (d) GMSST, (e) PDO, (f) AO; NGP – (g)
GMSST, (h) AO, (i) SM; South – (j) GMSST, (k) AO, (l) NAO; NE – (m) GMSST, (n) ENSO, (o) SM.
a. c. b.
d. f. e.
g. i. h.
j. l. k.
m. o. n.
51
The leading indicator for all regions, regardless of model, was overwhelmingly
GMSST. For the NW region, PNA and NAO were the second and third most influential
variable. For the SW, PDO and AO were the second and third most influential
variable. For the NGP, AO and SM were the second and third most influential
variable. For the South, AO and NAO were the second and third most influential
variable. For the NE, ENSO and SM were the second and third most influential
variable.
3.3.2.2. Linear Trend Removed (LTR) analysis
Table 3.4 shows the average and standard error of MSE values resulting from the
holdout analysis for models trained to predict LTR temperature as a function of LTR
LSM. For all regions except the NW the best performing model—RF in NW and
South, GAM in NGP and NE—significantly outperforms CLIM. For simplicity, we will
sometimes refer to the top-performing model as the “BEST” model in the remainder
of the paper. Since we were not able to identify a model that significantly
outperforms CLIM in the NW we do not perform any additional analysis for LTR in
that region. Figure 3.5 shows actual Tmin versus modeled Tmin for the best LTR
models. In the SW and South (Fig 5(a,c)) models tend to under-predict extremely
warm and over-predict extremely cold summers. For the NGP and Northeast (Fig
5(b,d)) the dynamic range of models is a close match to observations.
52
Figure 3.5 Scatterplots of actual versus predicted Tmin for the best LTR models in (a) SW, (b) NGP,
(c) South, and (d) NE regions.
The notably stronger performance of nonlinear models for LTR is highlighted in
Figure 3.6, which maps July-Aug Tmin anomalies for the top (bottom) five years for
the positive (negative) phase of ENSO, PDO and AO, all of which are important
●
●●
●
●
●
●
●
●●●●
● ●●● ●
●●
●
●●●
●●●
●●
●
●●
●
●
● ●●
●
●●
●●●
●●
●
●
●
●
●
●
●
●● ●
●
●
●
●● ●
●
●
●
−2 −1 0 1 2
−1
01
2
Tmin (actual)
Tm
in (
mo
de
l)
a. ●
● ● ●●
●
●
●●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●● ●
●
●
●
●
●
●
●
●
●
●● ●
●
●●
●
●
●
●
●
●
−2 −1 0 1 2
−2
−1
01
2
Tmin (actual)
Tm
in (
mo
de
l)
b.
●
●●
●
●●
●
●
●
●
●●
●
●
●● ●●
● ●
●●●
●
●●● ●
●
●
●●
●
●●
● ●
●●
● ●
●
● ●●
●●
●
●
●
●
●● ●
●
●
●
●
●● ●
●
●
−2 −1 0 1 2
−1
01
2
Tmin (actual)
Tm
in (
mo
de
l)
c. ●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●●
●
●
●●
●
−2 −1 0 1 2
−1
.5−
0.5
0.5
1.5
Tmin (actual)
Tm
in (
mo
de
l)
d.
53
predictors in some regions. For ENSO, a clear East-West pattern is seen for the
negative phase whereas the positive phase experiences a widespread negative
anomaly (Figure 3.6a,b). For PDO, a widespread positive anomaly is seen in the
negative phase while the positive phase experiences a cool anomaly throughout the
Rockies (Figure 3.6c,d). The AO shows a symmetric anomaly throughout the NGP,
but the area extending from Texas up the Eastern seaboard to the Mid-Atlantic
region experiences a warm anomaly in the negative phase, while the opposite is not
true of the positive phase (Figure 3.6e,f).
54
Figure 3.6 Temperature anomaly plots for (a) ENSO+, (b) ENSO-, (c) PDO+, (d) PDO-, (e) AO+, (f) AO-
.
Table 3.5 shows the swing of each covariate for the BEST LTR model for each region.
Partial dependence plots for the three most influential covariates (largest swings)
for each region are shown in Figure 3.7. For the SW region, the leading indicator of
Longitude
Latitu
de
30°N
35°N
40°N
45°N
a. b.
c. d.
30°N
35°N
40°N
45°N
120°W 110°W 100°W 90°W 80°W 70°W
e. f.−2.0
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
2.0
55
Tmin was GMSST, followed by PDO and AMO. For the NGP region, the leading
indicator of Tmin was GMSST, followed by AO and PNA. For the South region, the
leading indicator of Tmin was GMSST, followed by AMO and NAO. For the NE region,
the leading indicator was ENSO, followed by GMSST and SM.
a. c. b.
d. f. e.
g. i. h.
j. l. k.
56
Figure 3.7 Partial dependence plots for the three leading indicators in each region for models built
with LTR data: SW – (a) GMSST, (b) PDO, (c) AMO; NGP – (d) GMSST, (e) AO, (f) PNA; South – (g)
GMSST, (h) AMO, (i) NAO; NE – (j) ENSO, (k) GMSST, (l) SM.
Composite plots of detrended 300 hPa geopotential height anomaly for hot years in
each of the predictable LTR regions provide some indication of how LSMs influence
circulation patterns relevant to high temperatures (Figure 3.8). For each region
there is a clear high geopotential anomaly within or upstream of the region that
would be associated with a tendency to deflect the jet stream to the north.
57
Figure 3.8 Composite plots showing 300 hPa anomalies for the top five warmest years for (a) SW, (b)
NGP, (c) South, and (d) NE.
3.4. Discussion
3.4.1. Regionalization
−150 −100 −50
−2
02
06
01
00
−60
−40
−20
0
20
40
60
a.
−150 −100 −50
−2
02
06
01
00
−60
−40
−20
0
20
40
60
b.
−150 −100 −50
−2
02
06
01
00
−60
−40
−20
0
20
40
60
c.
−150 −100 −50
−2
02
06
01
00
−60
−40
−20
0
20
40
60
d.
58
Numerous regionalizations of CONUS can be found in the literature, with the choice of
regions driven by the study’s motivation. For example, the National Climate Assessment
defines six regions that are drawn to be consistent with state borders, since the
Assessment was designed to be relevant to decision makers that include State officials
(Melillo et al., 2014). Bukovsky and Karoly (2011) divided North America into a smaller
set of regions in order to capture dominant eco-regions that might have differing
sensitivities to climate. Since the focus of this study was summertime temperature
variability, we performed our own objective regionalization based on interannual
variability in July-August Tmin. The exercise was found to be useful for developing
skillful and informative statistical models.
Some features of the regionalization are unsurprising: the mountain West is highly
heterogeneous, and this resulted in large parts of the NW and SW being masked out by
our minimum intraregional correlation requirement (Figure 3.1). It also might explain
why the NW proved to be a difficult region to predict. Other aspects of the
regionalization were unexpected. For example, instead of seeing distinct regions for the
Southern Great Plains, coastal Southeast, and Texas-Mexico (TexMex) areas, we found a
unified greater Gulf of Mexico region that extends from Texas to the East Coast (our
“South” region). It is also interesting that the NW and SW separate from each other
relatively cleanly, indicating that these two portions of the West have quite different
sensitivities for summertime temperature.
3.4.2. Model Structure
59
The clearest result drawn from the multiple model comparisons performed in this
study is that a linear regression model (GLM) is generally adequate for predicting
interannual variability in July-August Tmin when data are not detrended (the LTP
analyses). It is the top performing model in three regions and is close to the top
performing model in the other two (Table 3.2). This reflects the fact that GLM captures
the robust and, to first order, linear relationship between global SST and summertime
temperature in every region (see Figure 3.4), and is sufficient to provide good model fit
and strong out-of-sample predictive skill. This is a useful point of reference for certain
applications of statistical models. For example, climate change projections include
global climate model representations of both the global temperature trend and
changing patterns of large-scale climate modes. Statistical models are frequently used
to interpret these projections, and our results here suggest that linear statistical
formulations should be adequate in those applications.
For the second iteration of this study, using LTR data, GLM was no longer found to have
the most robust predictive accuracy for any region (Table 3.4). Instead, nonlinear or
nonparametric approaches—GAM in NGP and NE and RF in SW and South—showed
significant advantages in out-of-sample predictive skill. Notably, the GAM models in
NGP and NE captured the range of observed historical temperature variability, where
RF in the SW and South provided predictive skill but tended to underestimate extreme
years (Figure 3.5). The fact that nonlinear approaches are more valuable for analysis of
detrended temperature records is not surprising, since the presence of a strong linear
trend in the record favors linear models and its removal exposes nonlinear interactions
60
between LSMs in their influence on CONUS temperature variability; some of these
nonlinearities are evident in the composite plots shown in Figure 3.6. The finding is
useful, however, in that it demonstrates that linear approaches are suboptimal for
understanding drivers of interannual temperature variability or for predicting
variability on a year-to-year basis, when the long term trend is less relevant.
3.4.3. Predictors & Mechanism
As described in Section 3.3, the leading predictor of Tmin variability varied between
regions (Table 3.3). CONUS is known to be heterogeneous in climate variability, and the
time series plots in Figure 3.2 show that summertime Tmin variability is not consistent
across regions. The statistical models also indicate that the temperature response to
LSM variability is often nonlinear, such that GAM or RF were more skillful than GLM.
When predicting LTP data, these nonlinearities proved not to be a dominant issue. The
significant warming trend in all regions results in significant and reasonably linear
correlations with rising global GMSST over the period of analysis; this result is
consistent with previous studies (van Oldenborgh and van Ulden, 2003). There are
influential predictors beyond GMSST in these models, and they differ in interesting
ways between regions (Figure 3.4; Table 3.3). In the NW, both PNA and NAO have a
positive correlation to Tmin. In the SW, PDO has a negative correlation with Tmin while
AO has a positive correlation. In the NGP, both AO and SM have a positive correlation to
Tmin. In the South, AO is positively correlated to Tmin while NAO is negatively
61
correlated. In the NE, ENSO is negatively correlated with Tmin while SM is positively
correlated.
These results need to be interpreted in their proper context. If the objective is to
develop a model that predicts evolving Tmin over time, that can explain the recent
increase in CONUS Tmin, or that can be applied to project Tmin trends in coming years,
then it is useful to build a model that includes information on the GMSST trend.
However, the skill of these models overstates their ability to predict interannual
variability on shorter time scales, since much of the skill derives from the long term
trend. This is the reason that we pursued LTR models: the presence of the trend in LTP
models means that this approach is not optimized to explain shorter term variability or
to inform development of operational seasonal forecasts.
Shifting to the LTR results, we see that GLM is no longer the top performing result in
any region (Table 3.4). GMSST is still a highly influential variable—it has the largest
swing for three of the four regions in which skillful LTR prediction was possible—but
other predictors in the LTR iteration had a swing value magnitude closer to GMSST,
indicating that GMSST no longer dominates the models.
Using LTR data for the SW region, the top three leading indicators were GMSST, PDO,
and AMO. As shown by the partial dependence plots in Figure 3.7(a-c), Tmin is
positively associated with GMSST and AMO but negatively associated with PDO. The
advantage of the RF model is notable here as seen by the flattening of the relationship
62
between Tmin and the leading indicators for extreme values. This finding is supported
by Chylek et al. (2014), which suggests positive AMO leads to increased summertime
temperatures in the SW, and also Kurtz (2015), which found both AMO and PDO to be
influential over this same region. In addition, Figure 3.8(a) shows the composite plot of
the 300 hPa anomalies during the top five warmest years in the SW, where high
pressure is seen along the entire west coast, suggesting a tendency towards blocking
events that create clear-sky conditions and hence warmth over the West.
For the NGP region, the top three leading indicators were GMSST, AO, and PNA. As
shown by the partial dependence plots in Figure 3.7(d-f), Tmin is positively associated
with GMSST and AO, while the relationship with PNA is not monotonic. The positive
relationship exhibited between warmer Tmin and positive AO is consistent with the
accentuation of the jet stream during positive AO events, restricting colder arctic air
from flowing into the NGP. The positive relationship observed with both extremely
negative and positive PNA events is consistent with the understanding that the exact
location of the PNA’s ridge-trough delineation may vary such that neutral events see
little response in Tmin.
In Figure 3.8b 300 hPa anomalies of the top five warmest years are shown. The broad
high pressure feature seen over the northern area, covering NGP, is consistent with
positive AO and a deflection of the jet stream to the north of the NGP, allowing for
intrusion of warmer air from the south and for relatively stagnant conditions across the
region.
63
For the South region, the top three leading indicators were GMSST, AMO, and NAO. As
shown by the partial dependence plots in Figure 3.7(g-i), Tmin is positively associated
with GMSST and AMO, and negatively associated with NAO. These results suggest
positive phase AMO events lead to warmer Tmin, which is consistent with the well-
known relationship of positive AMO leading to warmth in the Gulf of Mexico region, and
extending to the land area north of the Gulf (Kurtz, 2015). Tmin association with
negative NAO is consistent with NAO-associated blocking and increased heat wave
activity (Wright et al., 2014). This blocking tendency is visible in the high pressure 300
hPa anomaly upstream of the South (Figure 7(c)).
For the NE region, the top three leading indicators were ENSO, GMSST, and April-May
SM. As shown by the partial dependence plots in Figure 3.7(j-l), Tmin is negatively
associated with ENSO, and positively, yet asymmetrically, associated with GMSST and
SM. This asymmetry again shows the importance of; non-linear techniques when
quantifying these relationships. In particular, the relationship between Tmin and SM is
flat when SM is negative and positive when SM is positive. Because this study uses
antecedent SM (April-May) and summertime Tmin (July-August), this relationship is
consistent with summertime warming due to early greening given a wet spring (Loikith
and Broccoli, 2014). The 300 hPa anomalies for warm years in the NE (Figure 3.8d)
show a high pressure blocking system over this region that is consistent with the jet
stream being deflected leading to stagnant, clear sky conditions that enhance warming.
64
3.5. Conclusions
This study began with an objective regionalization of CONUS on the basis of interannual
variability in summertime temperature. This allowed us to extract timeseries of the
variable of interest (Tmin) from regions that are coherent in Tmin variability, and these
timeseries were then used as the response variable for developing statistical models of the
large-scale drivers of summertime Tmin. Next, we compared multiple regression
approaches to determine the form of statistical model that best describes the relationship
between Tmin and the studied LSM. Using the best model, which was allowed to vary
between regions, the leading indicators of changes to Tmin were identified by calculating
the contribution of each LSM to the model’s predictive capacity.
Results show that detrended summertime Tmin is associated with different large-scale
climate modes in different regions of CONUS. In the SW, cooler summers are associated
with positive phase PDO. In the NGP, cooler summer Tmin is found when the jet stream is
accentuated during negative AO events. In the South, positive phase AMO is most closely
associated with warm summers. In the NE, wet spring conditions are associated with warm
summer Tmin. When the long-term trend is retained, Tmin in all regions is most closely
associated with global mean SST.
These results contribute to understanding and, potentially, prediction and projection of
summertime temperatures in three ways. First, the application of objective regionalization
65
using the specific climate variable of interest makes it possible to define spatial variability
in a way that is physically meaningful and relevant to predictive modeling.
Second, distinguishing between predictions with and without the long-term warming trend
included is important when developing and interpreting statistical models. Both types of
models have useful applications: when the trend is present, we see strong linear
relationships between GMSST and summertime temperature in all regions, which informs
climate projections. It also indicates that within the range of historically observed
variability a relatively simple linear regression is adequate to capture this relationship.
Removing the linear trend makes draws attention to the importance of large-scale climate
models that drive year-to-year variability and can add skill to seasonal forecasts.
Third, our comparison of multiple parametric and non-parametric regression techniques
shows that nonlinear techniques can be particularly useful for predicting variability in
detrended summertime temperatures. The prediction of summertime temperature is
known to be challenging, as LSM tend to be weak during this season and local climate
conditions are sensitive to geographic setting and land-atmosphere interactions. In this
context linear models fail to identify relationships between Tmin and LSM, but nonlinear
approaches can provide mechanistic insight and significant predictive skill.
66
Table 3.1 Bold indicates intraregional correlations Italics indicates interregional correlations
NW SW NGP South NE
NW 0.60 - - - -
SW 0.30 0.68 - - -
NGP -0.26 0.38 0.81 - -
South -0.41 0.44 0.53 0.71 -
NE -0.18 0.12 0.56 0.43 0.79
67
Table 3.2 Average (standard error) MSE, for holdout with LTP data; bold indicates top-performing model.
NW SW NGP South NE
GLM 0.461
(0.026)
0.489
(0.027)
0.572
(0.029)
0.456
(0.025)
0.530
(0.029)
GAM 0.520
(0.025)
0.550
(0.028)
0.505
(0.034)
0.486
(0.032)
0.477
(0.031)
CART 0.593
(0.032)
0.596
(0.032)
0.877
(0.045)
0.669
(0.043)
0.797
(0.038)
RF 0.515
(0.029)
0.505
(0.028)
0.615
(0.037)
0.581
(0.055)
0.651
(0.043)
CLIM 0.996
(0.047)
1.085
(0.056)
0.983
(0.050)
1.023
(0.085)
1.070
(0.055)
68
Table 3.3 Swings from BEST model using the LTP data, with relative rank in parenthesis, where (1) is most
important and (8) is least important, top three in bold.
NW SW NGP South NE
Model GLM GLM GAM GLM GAM
SM 0.204 (7) 0.013 (8) 1.335 (3) 0.060 (8) 1.133 (3)
ENSO 0.089 (8) 0.200 (5) 0.023 (8) 0.276 (6) 2.285 (2)
NAO 0.829 (3) 0.167 (6) 0.035 (7) 0.762 (3) 0.813 (4)
PDO 0.748 (4) 1.021 (2) 0.866 (6) 0.252 (7) 0.092 (6)
PNA 0.899 (2) 0.580 (4) 0.957 (5) 0.327 (5) 0.124 (5)
GMSST 2.686 (1) 2.552 (1) 2.746 (1) 2.411 (1) 2.997 (1)
AMO 0.259 (6) 0.102 (7) 1.088 (4) 0.476 (4) 0.000 (7)
AO 0.370 (5) 0.584 (3) 1.542 (2) 0.840 (2) 0.000(7)
69
Table 3.4 Average (standard error) MSE, for holdout with LTR data; bold indicates top-performing model.
NW SW NGP South NE
GLM 0.973
(0.064)
0.824
(0.045)
0.912
(0.058)
0.760
(0.036)
0.859
(0.043)
GAM 0.916
(0.051)
0.827
(0.046)
0.897
(0.060)
0.794
(0.042)
0.799
(0.048)
CART 1.463
(0.083)
1.287
(0.070)
1.326
(0.072)
1.243
(0.062)
1.175
(0.075)
RF 0.880
(0.052)
0.780
(0.045)
0.898
(0.054)
0.715
(0.036)
0.803
(0.045)
CLIM 0.990
(0.058)
1.002
(0.055)
1.131
(0.060)
0.994
(0.051)
1.105
(0.062)
70
Table 3.5 Swings from BEST model using the LTR data, with relative rank in parenthesis, where (1) is most
important and (8) is least important, top three in bold.
SW NGP South NE
Model RF GAM RF GAM
PNA 0.231 (7) 2.013 (3) 0.204 (8) 0.041 (8)
AO 0.356 (5) 2.209 (2) 0.347 (5) 0.985 (5)
GMSST 0.725 (1) 2.402 (1) 0.984 (1) 1.733 (2)
ENSO 0.375 (4) 0.432 (7) 0.214 (7) 2.197 (1)
NAO 0.108 (8) 0.083 (8) 0.570 (3) 0.984 (6)
PDO 0.588 (2) 1.096 (6) 0.428 (4) 0.272 (7)
AMO 0.575 (3) 1.164 (5) 0.726 (2) 1.001 (4)
SM 0.245 (6) 1.326 (4) 0.282 (6) 1.729 (3)
71
4. CHAPTER 4: THE IMPACT OF THE NORTH ATLANIC OSCIALLATION ON HEAT
WAVES IN BALTIMORE
ABSTRACT
Extreme summertime temperatures are the most deadly natural hazard in the United
States, and are known to negatively impact human health. Heat-related mortality increases
asymmetrically between warm and cool climates, and as such previous work shows that
heat waves defined with a relative (not absolute) threshold, represent increased effects of
heat. Here, we use summertime temperature data from the BWI airport to first, define heat
waves by two definitions, one absolute and one relative, over the period 1950-2014. Using
these two heat wave indices, we investigate the relationship they have with several large-
scale climate modes. We focus on the relationship heat waves in Baltimore have with the
North Atlantic Oscillation by replacing the classically defined NAO index with six derived
indices that shows variations in pressure, latitude and longitude of the Azores High and
Icelandic Low. This work shows that different heat wave indices are attributed to different
LSM, where the relative index used is attributed to the Azores High pressure and latitude,
while the absolute heat wave index used is attributed to the Pacific Decadal Oscillation. We
find that the centers of action approach to the NAO index produces better model fit for both
heat wave indices. Lastly, we built the groundwork for further investigation into creating a
model with the ability to use these large-scale climate modes to predict heat waves in
Baltimore. These results are valuable for informing future improvements to seasonal
forecasting of heat waves, which bring us closer to decreasing the negative effects heat
waves currently have on human health.
72
4.1. Introduction
Over the past 30 years, extreme summertime heat has been the most deadly natural
hazard in the United States (NOAA, 2015). This extreme heat is known as a heat wave,
but has myriad definitions throughout the scientific community, ranging from relative
to absolute thresholds, single-day to multi-day lengths and varying temperature-related
metrics (Smith et al., 2013). These definitions depend on what the focus of the study is,
climatological, public health, agriculture or otherwise. When studying human morbidity
and mortality, relative thresholds definitions are known to represent increased effects
of temperature better than absolute threshold definitions (Kent et al., 2014). In
addition, people living in different climates are known to experience heat waves
differently, where mortality in the mild climate of the Northeast increases 6.76% on
heat wave days versus an increase of 1.84% in the warm, humid climate of the South
(Anderson & Bell, 2011).
On the East coast, the intersection of the cool, dry climate of the Northeast and the
warm, humid climate of the South is found in the Mid-Atlantic region, centered in the
Baltimore-Washington corridor where 9.6 million people reside. As such, this work will
focus on the area of Baltimore, MD. Out of 11 cities throughout the Eastern Seaboard,
Baltimore has been shown to have the highest increase in mortality due to temperature,
where a large difference was seen even when compared to nearby city, Washington,
D.C. (Curriero et al., 2002). In addition to direct-effects of heat on humans, indirect
effects have also been found. In Baltimore, maximum temperature has the highest
correlation to increases in violent crime out of any meteorological variable, leading to
73
implications of hospital and police staffing (Michel et al., 2016). Globally, it is well
known that extreme temperatures are expected to increase in frequency, where Duffy
and Tebaldi (2012) expect by mid-century 50% of summers will be as hot as the
current top 5% of summers. Likewise, maximum temperatures that are currently a
once-per-season occurrence in Baltimore are expected to increase by 28% in frequency
by mid-century (Horton et al., 2015).
Because of the multitude of negative impacts summertime heat has in Baltimore, it is
important to clarify our understanding of the drivers of the trends and variability of
these heat waves. Previous work shows that drivers of summertime heat across the
United States can include large-scale modes (LSM) of climate variability (e.g. Barnston,
1996; Drosdowsky and Chambers, 2001; Kenyon and Hegerl, 2008) as well as local
feedback mechanisms (e.g. Fischer et al., 2007; Portmann et al., 2009). Throughout the
literature, several LSM are studied as potential drivers to climate variables, such as El
Nino Southern Oscillation (ENSO), North Atlantic Oscillation (NAO), Atlantic
Multidecadal Oscillation (AMO), Arctic Oscillation (AO) and Pacific Decadal Oscillation
(PDO) (Hurrell and van Loon, 1997; Gershuov and Barnett, 1998; Sutton and Hodson,
2005; Kenyon and Hegerl, 2008; Loikith and Broccoli, 2014). Historically, ENSO has
been the best studied LSM in the context of its impact on climate variables. ENSO has
been associated with extreme temperature frequencies across the United States
(Gershunov and Barnett, 1998), and has been shown to increase predictability of
climate variables (Barnett et al., 1997). More recently, research has highlighted a need
to better understand the event-to-event differences in ENSO spatial patterns and
74
evolution, as ENSO is increasingly defined as a continuum versus two distinct modes of
variability (Ashok et al., 2007; Singh et al., 2011; Capotondi et al., 2015).
Parallel to our increasing understanding of ENSO-driven climate variability, researchers
have begun to amass a knowledge base of NAO-driven climate impacts. The NAO is an
index of sea level pressure (SLP) between centers of action over Iceland and the
subtropical Atlantic, near the Azores. The index is in its positive phase when SLP is
below normal near Iceland and above normal near the Azores, where the negative
phase is defined when this pattern is relaxed (Barnston and Livezey, 1987). In the
landmark NAO paper from 1997, Hurrell and van Loon found that wintertime
circulation, and therefore temperature and precipitation, were impacted profoundly
due to variations in the NAO. Since then additional studies have worked to understand
these same phenomena during summertime as well as over other regions. Specifically,
NAO-related, upstream circulations were found to influence summertime temperatures
in the US Southwest (Myoung et al., 2015). In addition, cold Atlantic Ocean surface
temperatures (indicative of NAO patterns) are linked to stationary positions of the Jet
Stream that favors the development of high temperatures over Central Europe (Duchez
et al., 2016).
As our understanding of the NAO-driven impacts to climate variables increases, so does
our awareness of the limitations of the classical definition of NAO. Likewise to what has
emerged in the ENSO community, researchers of the NAO are turning their attention to
the way NAO is defined. In 2002, Castro-Diez et al. found that temperatures in southern
75
Europe are sensitive not only to the phase of the NAO, but also to the location of the
NAO’s centers of action (COA). Additional research indicates a non-symmetric response
between the COA over the Azores High and the Icelandic Low, where movements
centered on the Icelandic Low are highly correlated to NAO while the relationship with
the Azores High is insignificant (Hameed and Pinotkovski, 2004). More recently, it was
found that sensitivity to NAO definition was higher for summertime temperatures than
winter temperatures (Pokorna and Huth, 2015).
In this paper, we compare the effect of two NAO definitions, classically-defined NAO and
COA-defined NAO, to heat waves in Baltimore, where a heat wave is defined by both a
relative and an absolute definition. Other previously mentioned LSM are included in
this study as it is important to put these NAO results in the context of other
teleconnections. We also build the framework for a predictive model of heat waves in
Baltimore, and with this in mind, the inclusion of other LSM is likely to increase the
predictive skill.
4.2. Methods
4.2.1. Data
Temperature data for this study was acquired from the National Oceanic and
Atmospheric Administration (NOAA) Global Historical Climate Network (GHCN), a
dataset that spans back to 1880 and includes data from over 90,000 stations. The
GHCN dataset was developed for climate analysis and monitoring studies that
require sub-monthly time resolution, and as such was appropriate for this study
76
(Menne et al., 2012). For this study, we used the daily temperature (mean and
maximum) from the station located at Baltimore-Washington International airport
(BWI).
Seven of the large-scale climate mode (LSM) datasets were gathered from the NOAA
Earth System Research Laboratory (ESRL), these include the El Nino Southern
Oscillation (ENSO), North Atlantic Osciallation (NAO), Pacific Decadal Oscillation
(PDO), Pacific-North American Mode (PNA), Atlantic Multidecadal Oscillation
(AMO), Arctic Oscillation (AO), and Global Mean Land/Ocean Temperature Index
(GMSST). Additional information on these data can be found in Section 3.2.1 and at
http://www.esrl.noaa.gov/psd/data/climateindices/list.
The remaining six indices used in this study are NAO-based indices. These are
objective indices of pressure, latitude and longitude for the Azores High and
Icelandic Low. Additional information about the derivation of these indices can be
found in Hameed and Piontkovski (2004).
All data were analyzed for July-August months over the years 1950-2014. As
explained in Section 3.5, all predictors were detrended for this analysis. Because the
purpose of this study is to attribute LSM to heat waves, we apply all predictors
synchronously, where the July-August average of the LSM are used to predict counts
of heat wave days over July-August.
77
4.2.2. Heat wave indices
Because there is no consensus definition of a heat wave event (Smith et al., 2013),
this study utilized two, exemplary and previously published indices for analysis. As
referenced in Chapter 3 and Smith et al. (2013), HI02 and HI11 are used herein.
HI02, initially from Anderson and Bell (2011), uses a relative threshold defined as
the 90th percentile of the long-term mean temperature, where the threshold must be
met for at least two consecutive days. For this study, the long-term average was
calculated on the 30-year baseline of 1980-2010. HI11, initially from Tan et al.,
2007, uses an absolute threshold defined as everyday that maximum temperatures
are greater than 35º C is classified as a heat wave day (Tan et al., 2007). Figure 4.1
shows the 1950-2014 timeseries of annual heat wave (HW) day counts for both
indices.
For this analysis the heat wave indices, HI02 and HI11, were defined as the number
of heat wave days, totaled over each summer (July-August) so that counts varied
annually between 0 and 62 days.
78
Figure 4.1. Timeseries of annual (July-August) heat wave (HW) day counts for (A) HI02 and
(B) HI11.
4.2.3. Data evaluation and model fit
First, we evaluated the correlations between HI02, HI11 and the thirteen covariates
through Pearson’s correlation method, where correlations were considered
# H
W d
ays
1950 1958 1966 1974 1982 1990 1998 2006 2014
010
20
30
A.# H
W d
ays
1950 1958 1966 1974 1982 1990 1998 2006 2014
010
20
30
B.
79
significant if p-values were less than 0.05 (95% level). Next, a generalized linear
model (GLM) was used to test the fit of the data. A GLM is an ordinary least squares
model with an added link function that allows for the model to respond based on the
distribution of the response variable. Because HI02 and HI11 exhibit a Poisson
distribution, a log link function was used for the GLM. Four GLM were built: HI02
with the original seven indices (Ind7), HI02 with the NAO centers of action indices
(COA), HI11-Ind7, and HI11-COA. These models are assessed for their fit of the data
through comparison of their root-mean square error (RMSE) values. From these
models, the p-value was used to indicate variable importance to each model, where
p-values less than 0.05 indicate statistical significance. We then use this variable
importance information to trim the number of variables included in the GLM, and
reevaluate the model fit.
4.2.4. Prediction
While the techniques discussed above are appropriate to understand and attribute
the LSM to heat waves, we also present a framework for predicting these heat
waves. To do so, the use of varied statistical techniques expands our ability to
capture nonlinear effects and account for interacting relationships between
predictors. As such, this work uses a similar suit of statistical models as explained in
Section 3.2.3 to investigate the predictability of HI02 and HI11. For this, we includ
the Poisson GLM explained in Section 4.2.3, but also a generalized additive model
(GAM), a classification and regression tree model (CART) and a random forest (RF).
To determine the model with the best predictive accuracy, a holdout cross-
80
validation analysis was performed. To test for statistical significance, a student’s t-
test was used to quantify p-values, which were evaluated at various confidence
levels. Additional details about these models and the holdout process can be found
in Section 3.2.3.
The model with the best predictive accuracy was determined by taking the average
mean square error (MSE) from the holdout analysis, where the model with the
lowest average MSE is best. Using this model, we determine the relative importance
of each variable by evaluating its impact on MSE by assigning the variable vales by
random permutation to evaluate the increase/decrease on MSE. Next, we
investigate partial dependence plots, as explained in Section 3.2.4, to visualize the
influence of important covariates over the response variables, HI02 and HI11.
4.3. Results
4.3.1. Model fit
Figure 4.2 shows the correlations between all variables and covariates in the dataset
used for this study. For HI02 we found that only AH.p and AH.lon were statistically
significant, with correlations of -0.34 and 0.25 respectively. For HI11 we found that
only PDO was statistically significant, with a correlation of 0.24.
81
Figure 4.2. Correlations between all variables included in models. Top right corner shows
correlations, where boxes are shaded according to the scale bar. Bottom left corner prints correlation
values, also shaded according to the scale bar, where values not printed were insignificant.
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
HI02
HI11
IL.p
IL.lon
IL.lat
AH.p
AH.lon
AH.lat
ENSO
AO
AMO
GMSST
NAO
PDO
PNA
0.9
−0.06
−0.24
0.03
−0.34
0.25
−0.02
0.05
0.04
0.18
0.24
−0.03
0.15
−0.01
−0.03
−0.19
−0.07
−0.21
0.22
−0.11
0.12
0.02
0.12
0.21
−0.07
0.24
−0.02
−0.02
−0.57
−0.15
−0.13
−0.32
0.15
−0.69
−0.02
−0.05
−0.63
−0.01
0.08
−0.17
0.12
−0.29
−0.12
−0.1
0.05
0.13
0.17
−0.3
0.01
0.22
0.37
0.39
0.63
−0.01
0.67
−0.21
−0.18
0.81
0.16
−0.31
0.05
0.19
0.26
0.27
−0.43
−0.18
0.43
0.26
0.01
0.65
0.11
0.22
−0.09
0
0.16
0.24
−0.24
−0.2
0.41
0.04
0.12
0.41
−0.13
−0.19
−0.23
−0.06
0.11
0.07
0.59
0.07
−0.04
−0.03
0.58
−0.03
−0.37
0.72
−0.26
−0.15
0.28
−0.21
−0.1
0.17
0.11
−0.220.21
82
To evaluate which models fit the data better, Table 4.1 shows the root mean square
error (RMSE) of the four variations of the GLM. For both HI02 and HI11, the RMSE
was improved (reduced) when the traditional NAO index was replaced with the COA
indices. To visualize this result, Figure 4.3 shows the relationship between actual
and fitted data from the four models.
Figure 4.3 Actual versus fitted data for (A) HI02 and (B) HI11 where blue dots indicate Ind7 model
results and red dots indicate COA model results. A 1:1 line is provided for reference of a “perfect” fit.
In evaluating the four iterations of the GLM, we gain insight into which variables are
driving these models. Figure 4.4 shows the p-value of each covariate included in the
GLMs. For the GLM-Ind7 (Figure 4.4a) we see that ENSO, PDO, PNA and GMSST are
all significantly at the 95% level for both HI02 and HI11. For GLM-COA (Figure
4.4b), we see that for HI02, AO, AMO, PNA, GMSST, IL.lon, AH.p, AH.lat and AH.lon
●
●●
●
●
●
●●
●
●
●●
●
●
●
● ●●
●
●
●
●
●●● ●●
●●●
●●
●
●
●
●
●
●
●●
●
●
● ●●
●
● ●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
0 5 10 15 20 25
05
10
15
20
25
Actual
Fitte
d
●
●●
●
●
●
●
●
●●●
●●
●
●●
●● ●●●● ●
●●●
●
●
●●
●
●
●
●
● ● ●
●
●
●
●●
●
●
●
●
●●
●
●● ●●
●
●
●
●●
●
●●●
●
●
●
A.
●
●
Ind7
COA
●
●●
●
● ●
●
●
●
●
●
●
●
●
● ●●
●
●
●
● ● ●●● ●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●
●●
●
●
●●
●
●
●●●
●●
●●
●●
●
0 5 10 20 30
05
10
20
30
Actual
Fitte
d
●
●●
●
●
●●
●
●
●●●
●
●
●●
●
●●●
●●●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●● ●
●
●
●
●
●●
●
●
● ●
●●
●
B.
83
are all significant at the 95% level. For the GLM-COA for HI11, we see that ENSO, AO,
AMO, PDO, GMSST, IL.lon, AH.lat and AH.lon are all significant at the 95% level.
Figure 4.4 P-values of covariates included (A) GLM-Ind7 and (B) GLM-COA. Stars indicate results for
HI02, and triangles indicate results for HI11. The horizontal dashed line represents 95% significance
level; any symbol below the dashed line is significant.
p−
valu
e
ENSO AO AMO NAO PDO PNA GMSST
00.2
0.6
1
HI02
HI11
A.
p−
valu
e
ENSO AO AMO PDO PNA GMSST IL.p IL.lat IL.lon AH.p AH.lat AH.lon
00.2
0.6
1
B.
84
Using the information from Figure 4.4, we trim the number of covariates in the
GLMs to include only those that are significant at the 95% level, but found the RMSE
were not improved in any of the GLMs for either heat wave index.
4.3.2. Prediction
Table 4.2 shows the average, mean square error (MSE) from the 100 holdouts for
each of the five models used for HI02. For both HI02-Ind7 and HI02-COA, the RF
performed the best as indicated by having the lowest MSE. Neither model run
produced a model that significantly outperformed CLIM at p <0.05. When comparing
the best performing model from each run for HI02, the COA-RF and the Ind7-RF, we
find them to be 99% similar.
Table 4.3 shows the average MSE from the 100 holdouts for each of the five models
used for HI11. For both HI11-Ind7 and HI11-COA, the RF performed the best as
indicated by having the lowest MSE. Neither model run produced a model that
significantly outperformed CLIM at p <0.05. When comparing the best performing
model from each run for HI11, the COA-RF and Ind7-RF, we find them to be 92%
similar.
Using the best performing model for each iteration (RF for all), we then determined
the most important variables in the model by evaluating the impact that variable has
on the model’s MSE. Table 4.4 shows the percent change in MSE for the models
evaluating HI02. Higher values indicate variables of higher importance. Figure 4.5
85
shows the partial dependence plots for the top two variables for the RF-Ind7 and
RF-COA models using HI02, where the leading indicators are GMSST and AMO for
RF-Ind7 and AH.p and AMO for RF-COA.
Figure 4.5. Partial dependence plots for the top two most important variables for HI02 for (A-B) RF-
Ind7, and (C-D) RF-COA.
−2 −1 0 1 2
34
56
78
9
GMSST
HI0
2
A. −2 −1 0 1 2
34
56
78
9AMO
HI0
2B.
−3 −2 −1 0 1 2
34
56
78
9
AH.p
HI0
2
C. −2 −1 0 1 2
34
56
78
9
AMO
HI0
2
D.
86
Table 4.6 shows the percent change in MSE for the models evaluating HI11. Figure
4.6 shows the partial dependence plots for the top two variables for the RF-Ind7 and
RF-COA for HI11, where the leading indicators are PDO and AMO for RF-Ind7, and
PDO and AH.p for RF-COA.
Figure 4.6. Partial dependence plots for the top two most important variables for HI11 for (A-B) RF-
Ind7, and (C-D) RF-COA.
−2 −1 0 1 2 3
68
10
12
14
PDO
HI1
1
A.−2 −1 0 1 2
68
10
12
14
AMO
HI1
1
B.
−2 −1 0 1 2 3
68
10
12
14
PDO
HI1
1
C.−3 −2 −1 0 1 2
68
10
12
14
AH.p
HI1
1
D.
87
4.4. Discussions and conclusions
In this study, we examined the effect of two NAO definitions, classically-defined and
COA-defined, to heat waves in Baltimore, defined by both a relative and an absolute
definition. Results show that the inclusion of COA-defined NAO indices increases model
fit of the heat wave data, however results are not consistent across heat wave
definitions, where NAO is only important for the relative heat wave definition.
Our results show that only two indices were significantly correlated to HI02, AH.p and
AH.lon, where both are COA indices. In addition, when evaluating the GLM models,
RMSE values were improved (reduced) when the classically defined NAO index was
replaced with the COA indices. According to GLM-COA for HI02 (Figure 4.4b), we found
the following covariates were significant at the 95% level: AO, AMO, PNA, GMSST, IL.lon,
AH.p, AH.lat and AH.lon. This is supportive of the initial descriptive statistics that found
AH.p and AH.lon as significantly correlated to HI02. Through our attempt to gain
predictive skill, we found several non-linear relationships associated with the heat
wave indices. For HI02, AH.p was again confirmed as the most important variable in RF-
COA, and Figure 4.5c shows the non-linear dependence where low AH.p values are
associated with high counts of HI02. To visualize this, Figure 4.7 shows composite
anomaly plots of surface pressure (colors) and vector winds (arrows); for reference, a
box is located over the Azores.
88
Figure 4.7 Composite anomaly plots for the top five hottest years according to (A) HI02 and (B) HI11
where colors represent surface pressure and arrows represent vector winds. A box is located over the
Azores for reference.
The HI02 composite anomaly in Figure 4.7a shows where the area over the Azores
experiences lower pressure during hot years, as expected. This aligns with the
200 250 300 350
20
40
60
80
-40
-20
0
20
40
A.
200 250 300 350
20
40
60
80
-40
-20
0
20
40
B.
89
mechanism we present through composite anomaly plots of 300 hPa geopotential
height anomalies in Figure 4.8.
Figure 4.8 Composite anomaly plots of 300 hPa geopotential heights for the top five hottest years
according to (A) HI02 and (B) HI11.
In these composite anomaly plots, we see the patter of a Rossby wave train, initiated in
the tropical Pacific. This feature persists during warm years, not in a typical “blocking”
200 250 300 350
20
40
60
80
-60
-40
-20
0
20
40
60
A.
200 250 300 350
20
40
60
80
-60
-40
-20
0
20
40
60
B.
90
fashion, but in a way so as to organize the circulation regimes and allow heat to
stagnant over Baltimore. The wind vectors in Figure 4.7a support this mechanism
where anomaly winds are flowing against the typical southerly winds (shown in Figure
4.9) seen along the Eastern seaboard, indicating increased stagnation of this air
movement.
Figure 4.9 Composite anomaly plots of surface pressure (colors) and vector winds (arrows) for study
time period, 1950-2014.
Our results show that only PDO is statistically significantly correlated to HI11, but that
the replacement of NAO with the COA improved the fit of the GLM through a reduced
RMSE value. According to the GLM-COA for HI11 (Figure 4.4b), we found the following
covariates to be significant at the 95% level: ENSO, AO, PDO, GMSST, IL.lon, AH.lat,
AH.lon. This is supportive of the initial descriptive statistics that found PDO as
significantly correlated to HI11. We see in Figure 4.8b where 300 hPa anomalies for
HI11 hot years follow the same general pattern as for HI02, but shifted westward so
200 250 300 350
20
40
60
80
-100
0
100
200
300
400
91
that the anomaly ridge covers a greater area of the United States. On the surface
however, the HI11 anomalies look quite different than HI02, thus supporting our
findings of different important variables. Specifically, in the Pacific basin we see an
anomalous high pattern indicative of positive phase PDO, and in alignment with the
dependence plots results from our RF models (Figure 4.6 a,c).
As mentioned above, this study showed improvement in modeling the heat wave
indices by replacing NAO with the COA indices. As seen in Figure 4.3, the GLM using
COA was better able to fit summers with high counts of heat wave days. For HI02
(Figure 4.3a) we see where GLM-Ind7 never produces a value higher than 9 days, where
GLM-COA models counts up to 19 days. For HI11 (Figure 4.3b) we see where GLM-Ind7
produces values up to 16 days, where GLM-COA models counts up to 25 days. In
practice, the ability to model summers with higher
From the results discussed above, this study attempts to build on the ability to model
the summers by completing a comprehensive investigation into predicting HI02 and
HI11. The baseline metric for predictability is to outperform the climatologic average
(CLIM), which was not achieved for either HI02 or HI11. However, we found that again
COA improved models when replacing NAO as shown in Table 4.3 where the MSE value
is improved (lowered) for RF-COA over RF-Ind7. As such, it is recommended that
further research be done to improve predictability. The consistency of results across
techniques strengthens the results found in the first part of this study, but also support
the impetus behind building a predictive model. Recommendations for further work
92
include increasing the sample size through clustering station data from nearby cities
and evaluating the dataset in the same way, or to use the same BWI timeseries but the
employ more advanced statistical techniques, such as extreme value analysis, given the
inherent extreme nature of heat wave data.
In the end, this study was able to show that models of heat waves in Baltimore are more
robust when including COA indices over NAO in general, but also that the inclusion of
AH.p helps quantify blocking over the Atlantic, which leaves heat built up over the US.
We also showed that results differ between heat wave indices, indicating that results
across studies can only be compared when the heat wave index is defined consistently.
Lastly, we built the groundwork for further investigation into creating a model with the
ability to use these LSM to predict heat waves in Baltimore. These results are valuable
for informing future improvements to seasonal forecasting of heat waves, which bring
us closer to decreasing the negative effects heat waves currently have on human health.
93
TABLES Table 4.1 Root mean square error (RMSE) for four GLM.
HI02 HI11
Ind7 4.93 7.69
COA 3.95 7.03
94
Table 4.2 Mean square error (MSE) values from holdout analysis for HI02; * indicates top performing model. COA Ind7
GLM 35.42 31.61
CART 46.12 43.15
RF 26.39* 25.88*
CLIM 30.64 25.90
GAM 40.95 34.33
95
Table 4.3 Mean square error (MSE) values from holdout analysis for HI11; * indicates top performing model.
COA Ind7
GLM 119.71 81.63
CART 88.47 91.03
RF 61.63* 71.02*
CLIM 68.92 73.07
GAM 99.38 107.28
96
Table 4.4 % impact on MSE each variable has for RF models for HI02. Higher values indicate variables of
higher importance.
COA Ind7
AMO 1.69 2.82
AO 0.43 0.24
ENSO -0.06 0.39
GMSST 1.24 5.52
NAO - 0.42
PDO 0.80 1.18
PNA -0.34 0.41
AH.lat 0.35 -
AH.lon 0.76 -
AH.p 3.31 -
IL.lat 0.13 -
IL.lon 0.93 -
IL.p 0.76 -
97
Table 4.5 % impact on MSE each variable has for RF models for HI11. Higher values indicate variables of
higher importance.
COA Ind7
AMO 2.26 6.62
AO -0.24 -1.38
ENSO -1.01 -0.43
GMSST 2.00 1.96
NAO - -0.59
PDO 10.49 9.35
PNA 0.03 1.64
AH.lat 1.95 -
AH.lon 1.51 -
AH.p 4.52 -
IL.lat 0.38 -
IL.lon 1.09 -
IL.p 0.23 -
98
5. CHAPTER 5: CONCLUSIONS
Extreme summertime heat is the most deadly natural hazard in the United States (NOAA,
2015), where it has been projected that by mid-century, 50% of all summers will be as hot
as the top 5% of summers in the historic baseline (Duffy and Tebaldi, 2012). The preceding
chapters take this knowledge as motivation to improve our understanding of the drivers of
summertime heat across the United States. The intended outcome of this research is to
provide an outline for discussing results from studies with diverse motivations, as well as
creating a framework of understanding the large-scale drivers of summertime heat to
inform and improve seasonal forecasting.
Chapter 2 describes and explains how the choice of definition influences conclusions
regarding the observed frequency of extreme heat events in different CONUS regions in
order to provide a baseline for interpreting studies that project future trends in extreme
heat events. Over the 1979-2011 time period investigated, the Southeast region saw the
highest number of heat wave days for the majority of indices considered. Positive trends
(increases in number of heat wave days per year) were greatest in the Southeast and Great
Plains regions, where more than 12% of the land area experienced significant increases in
the number of heat wave days per year for the majority of heat wave indices. Significant
negative trends were relatively rare, but were found in portions of the Southwest,
Northwest, and Great Plains.
Chapter 3 first presents a novel regionalization of CONUS, where regions were selected by
temperature-informed hierarchical clustering analysis and yielded five regions that exhibit
99
distinctly different behavior. Seasonal average temperatures were then investigated to
explain large-scale driver of variability over these regions on seasonal timescales. Results
indicate that summertime temperature variability is sensitive to different climate
processes in different regions, including the Pacific Decadal Oscillation variability in the
Southwest, Arctic Oscillation in the Northern Great Planes, Atlantic Multidecadal Oscillation
in the South, and late spring soil moisture anomalies in the Northeast. Notably, nonlinear
models yielded the best results in all regions: a random forest model is found to be most
robust in the SW and South, while a generalized additive model is most robust for the NGP
and Northeast. These results are valuable for informing future improvements in seasonal
forecasting of summertime temperatures and are relevant for projections of summertime
climate change across climatically distinct regions of CONUS.
Chapter 4 presents a deep dive into the large-scale drivers of heat waves in Baltimore, MD.
Specifically, this work compares the effect of two NAO definitions, classically defined NAO
and COA-defined NAO, to heat waves in Baltimore, where a heat wave is defined by both a
relative and an absolute definition from Chapter 2. This work shows that replacement of
classically defined NAO definition with the COA-defined NAO indices increases model fit for
both heat wave definitions. This phenomenon is explained by the Azores High experiencing
lower pressure during hot years and the resulting Rossby wave train, which emulates an
atypical blocking pattern by organizing circulation regimes that allow heat to stagnate over
Baltimore. This work also built a framework for a predictive model of heat waves in
Baltimore, and despite not gaining predictive skill over climatology, we were able to
100
investigate non-linear variable responses to heat waves, which support our explanatory
mechanisms explained above.
5.1. Future work
The work presented in this thesis allows for multiple avenues of improvement by
extending these analyses. From Chapter 2, future studies can pair our findings with the
relative health effects of heat waves in differing regions to grasp a complete
understanding for planning health interventions and climate change adaptation
strategies. From the last two chapters, these works aimed to inform future studies of
seasonal forecasting and climate change impacts specifically relevant across heat-
vulnerable regions of CONUS. Thus, we suggest further work to incorporate these
findings into seasonal forecasting models to improve predictive skill in forecasting
summertime heat. One way to improve on the predictive skill of the models presented
in Chapter 4 is to automate a process to include a stepwise reduction in variables to find
the optimal conditions to predicting summertime heat waves. Extension of these
methods will significantly improve the prediction of summertime heat across the
United States and ultimately decrease the negative impacts heat has on the human
population.
101
REFERENCES
Alexander, L. V., and Coauthors, 2006a: Global observed changes in daily climate extremes of temperature and precipitation. Journal of Geophysical Research-Atmospheres, 111, 22.
Alfaro, E. J., A. Gershunov, and D. Cayan, 2006: Prediction of summer maximum and minimum temperature over the central and western United States: The roles of soil moisture and sea surface temperature. Journal of Climate, 19, 1407-1421.
Anderson, B. G., and M. L. Bell, 2009: Weather-Related Mortality How Heat, Cold, and Heat Waves Affect Mortality in the United States. Epidemiology, 20,205-213.
Anderson, G. B., and M. L. Bell, 2011: Heat Waves in the United States: Mortality Risk during Heat Waves and Effect Modification by Heat Wave Characteristics in 43 U.S. Communities. Environmental Health Perspectives, 119, 210-218.
Ashok, K., S. K. Behera, S. A. Rao, H. Y. Weng, and T. Yamagata, 2007: El Nino Modoki and its possible teleconnection. Journal of Geophysical Research-Oceans, 112.
Badr, H. S., B. F. Zaitchik, and S. D. Guikema, 2014: Application of Statistical Models to the Prediction of Seasonal Rainfall Anomalies over the Sahel.Journal of Applied Meteorology and Climatology, 53, 614-636.
Badr, H. S., B. F. Zaitchik, and A. K. Dezfuli, 2015: A tool for hierarchical climate regionalization. Earth Science Informatics, 8, 949-958.
Barnett, A. G., S. Tong, and A. C. A. Clements, 2010: What measure of temperature is the best predictor of mortality? Environmental Research, 110, 604-611.
Barnett, T. P., K. Arpe, L. Bengtsson, M. Ji, and A. Kumar, 1997: Potential predictability and AMIP implications of midlatitude climate variability in two general circulation models. Journal of Climate, 10, 2321-2329.
Barnston, A. G., 1994: LINEAR STATISTICAL SHORT-TERM CLIMATE PREDICTIVE SKILL IN THE NORTHERN-HEMISPHERE. Journal of Climate, 7, 1513-1564.
Barnston, A. G., and R. E. Livezey, 1987: CLASSIFICATION, SEASONALITY AND PERSISTENCE OF LOW-FREQUENCY ATMOSPHERIC CIRCULATION PATTERNS. Monthly Weather Review, 115, 1083-1126.
Barnston, A. G., and T. M. Smith, 1996: Specification and prediction of global surface temperature and precipitation from global SST using CCA. Journal of Climate, 9, 2660-2697.
102
Becker, E. J., H. van den Dool, and M. Pena, 2013: Short-Term Climate Extremes: Prediction Skill and Predictability. Journal of Climate, 26, 512-531.
Benjamini, Y., and Y. Hochberg, 1995: CONTROLLING THE FALSE DISCOVERY RATE - A PRACTICAL AND POWERFUL APPROACH TO MULTIPLE TESTING.Journal of the Royal Statistical Society Series B-Methodological, 57, 289-300.
Brankovic, C., and T. N. Palmer, 2000: Seasonal skill and predictability of ECMWF PROVOST ensembles. Quarterly Journal of the Royal Meteorological Society, 126, 2035-2067.
Bukovsky, M. S., and D. J. Karoly, 2011: A Regional Modeling Study of Climate Change Impacts on Warm-Season Precipitation in the Central United States.Journal of Climate, 24, 1985-2002.
Bukovsky, M. S., D. J. Gochis, and L. O. Mearns, 2013: Towards Assessing NARCCAP Regional Climate Model Credibility for the North American Monsoon: Current Climate Simulations. Journal of Climate, 26, 8802-8826.
Capotondi, A., and Coauthors, 2015: Understanding ENSO Diversity. Bulletin of the American Meteorological Society, 96, 921-938.
Castro-Diez, Y., D. Pozo-Vazquez, F. S. Rodrigo, and M. J. Esteban-Parra, 2002: NAO and winter temperature variability in southern Europe. Geophysical Research Letters, 29.
Chen, X. Y., and J. M. Wallace, 2015: ENSO-Like Variability: 1900-2013. Journal of Climate, 28, 9623-9641.
Cheng, L. Y., A. AghaKouchak, E. Gilleland, and R. W. Katz, 2014: Non-stationary extreme value analysis in a changing climate. Climatic Change, 127, 353-369.
Chylek, P., M. K. Dubey, G. Lesins, J. N. Li, and N. Hengartner, 2014: Imprint of the Atlantic multi-decadal oscillation and Pacific decadal oscillation on southwestern US climate: past, present, and future. Climate Dynamics, 43, 119-129.
Cosgrove, B. A., and Coauthors, 2003: Real-time and retrospective forcing in the North American Land Data Assimilation System (NLDAS) project. Journal of Geophysical Research-Atmospheres, 108, 12.
Curriero, F. C., K. S. Heiner, J. M. Samet, S. L. Zeger, L. Strug, and J. A. Patz, 2002: Temperature and mortality in 11 cities of the eastern United States.American Journal of Epidemiology, 155, 80-87.
Diaz, J., A. Jordan, R. Garcia, C. Lopez, J. C. Alberdi, E. Hernandez, and A. Otero, 2002: Heat waves in Madrid 1986-1997: effects on the health of the elderly. International Archives of Occupational and Environmental Health, 75, 163-170.
103
Drosdowsky, W., and L. E. Chambers, 2001: Near-global sea surface temperature anomalies as predictors of Australian seasonal rainfall. Journal of Climate,14, 1677-1687.
Duffy, P. B., and C. Tebaldi, 2012: Increasing prevalence of extreme summer temperatures in the U.S. Climatic Change, 111, 487-495.
Efthymiadis, D., C. M. Goodess, and P. D. Jones, 2011: Trends in Mediterranean gridded temperature extremes and large-scale circulation influences.Natural Hazards and Earth System Sciences, 11, 2199-2214.
El Kenawy, A., J. I. Lopez-Moreno, and S. M. Vicente-Serrano, 2011: Recent trends in daily temperature extremes over northeastern Spain (1960-2006).Natural Hazards and Earth System Sciences, 11, 2583-2603.
Enfield, D. B., A. M. Mestas-Nunez, and P. J. Trimble, 2001: The Atlantic multidecadal oscillation and its relation to rainfall and river flows in the continental US. Geophysical Research Letters, 28, 2077-2080.
Fan, Y., and H. van den Dool, 2004: Climate Prediction Center global monthly soil moisture data set at 0.5 degrees resolution for 1948 to present. Journal of Geophysical Research-Atmospheres, 109.
Fischer, E. M., S. I. Seneviratne, P. L. Vidale, D. Luethi, and C. Schaer, 2007: Soil moisture - Atmosphere interactions during the 2003 European summer heat wave. Journal of Climate, 20, 5081-5099.
Fovell, R. G., and M. Y. C. Fovell, 1993: CLIMATE ZONES OF THE CONTERMINOUS UNITED-STATES DEFINED USING CLUSTER-ANALYSIS. Journal of Climate, 6, 2103-2135.
Gershunov, A., and T. P. Barnett, 1998: ENSO influence on intraseasonal extreme rainfall and temperature frequencies in the contiguous United States: Observations and model results. Journal of Climate, 11, 1575-1586.
Grundstein, A. J., and Coauthors, 2012: A retrospective analysis of American football hyperthermia deaths in the United States. International Journal of Biometeorology, 56, 11-20.
Hajat, S., and Coauthors, 2006: Impact of high temperatures on mortality - Is there an added heat wave effect? Epidemiology, 17, 632-638.
Hameed, S., and S. Piontkovski, 2004: The dominant influence of the Icelandic Low on the position of the Gulf Stream northwall. Geophysical Research Letters, 31, 4.
Hansen, J., M. Sato, and R. Ruedy, 2012: Perception of climate change. Proceedings of the National Academy of Sciences of the United States of America,109, E2415-2423.
104
Hansen, J., R. Ruedy, M. Sato, and K. Lo, 2010: GLOBAL SURFACE TEMPERATURE CHANGE. Reviews of Geophysics, 48.
Hattis, D., Y. Ogneva-Himmelberger, and S. Ratick, 2012: The spatial variability of heat-related mortality in Massachusetts. Applied Geography, 33, 45-52.
Higgins, R. W., J. K. E. Schemm, W. Shi, and A. Leetmaa, 2000: Extreme precipitation events in the western United States related to tropical forcing.Journal of Climate, 13, 793-820.
Horton, R. M., E. D. Coffel, J. M. Winter, and D. A. Bader, 2015: Projected changes in extreme temperature events based on the NARCCAP model suite.Geophysical Research Letters, 42, 7722-7731.
Hu, Q., and S. Feng, 2012: AMO- and ENSO-Driven Summertime Circulation and Precipitation Variations in North America. Journal of Climate, 25, 6477-6495.
Hurrell, J. W., and H. VanLoon, 1997: Decadal variations in climate associated with the north Atlantic oscillation. Climatic Change, 36, 301-326.
Kent, S. T., L. A. McClure, B. F. Zaitchik, T. T. Smith, and J. M. Gohlke, 2014: Heat Waves and Health Outcomes in Alabama (USA): The Importance of Heat Wave Definition. Environmental Health Perspectives, 122, 151-158.
Kenyon, J., and G. C. Hegerl, 2008: Influence of modes of climate variability on global temperature extremes. Journal of Climate, 21, 3872-3889.
Kim, Y.-M., S. Kim, H.-K. Cheong, and E.-H. Kim, 2011: Comparison of temperature indexes for the impact assessment of heat stress on heat-related mortality. Environmental health and toxicology, 26, e2011009.
Koster, R. D., and M. J. Suarez, 1992: MODELING THE LAND SURFACE BOUNDARY IN CLIMATE MODELS AS A COMPOSITE OF INDEPENDENT VEGETATION STANDS. Journal of Geophysical Research-Atmospheres, 97, 2697-2715.
——, 1994: THE COMPONENTS OF A SVAT SCHEME AND THEIR EFFECTS ON A GCMS HYDROLOGICAL CYCLE. Advances in Water Resources, 17, 61-78.
Krishnamurthy, L., G. A. Vecchi, R. Msadek, A. Wittenberg, T. L. Delworth, and F. R. Zeng, 2015: The Seasonality of the Great Plains Low-Level Jet and ENSO Relationship. Journal of Climate, 28, 4525-4544.
Kuglitsch, F. G., A. Toreti, E. Xoplaki, P. M. Della-Marta, C. S. Zerefos, M. Turkes, and J. Luterbacher, 2010: Heat wave changes in the eastern Mediterranean since 1960. Geophysical Research Letters, 37, 5.
105
Kumar, A., A. G. Barnston, P. T. Peng, M. P. Hoerling, and L. Goddard, 2000: Changes in the spread of the variability of the seasonal mean atmospheric states associated with ENSO. Journal of Climate, 13, 3139-3151.
Kurtz, B. E., 2015: The Effect of Natural Multidecadal Ocean Temperature Oscillations on Contiguous US Regional Temperatures. Plos One, 10.
Lobell, D. B., and M. B. Burke, 2010: On the use of statistical models to predict crop yield responses to climate change. Agricultural and Forest Meteorology,150, 1443-1452.
Lobell, D. B., G. L. Hammer, G. McLean, C. Messina, M. J. Roberts, and W. Schlenker, 2013: The critical role of extreme heat for maize production in the United States. Nature Climate Change, 3, 497-501.
Loikith, P. C., and A. J. Broccoli, 2014: The Influence of Recurrent Modes of Climate Variability on the Occurrence of Winter and Summer Extreme Temperatures over North America. Journal of Climate, 27, 1600-1618.
Luo, L. F., and Coauthors, 2003: Validation of the North American Land Data Assimilation System (NLDAS) retrospective forcing over the southern Great Plains. Journal of Geophysical Research-Atmospheres, 108, 10.
Medina-Ramon, M., and J. Schwartz, 2007: Temperature, temperature extremes, and mortality: a study of acclimatisation and effect modification in 50 US cities. Occupational and Environmental Medicine, 64, 827-833.
Meehl, G. A., and C. Tebaldi, 2004: More intense, more frequent, and longer lasting heat waves in the 21st century. Science, 305, 994-997.
Menne, M.J., I. Durre, B. Korzeniewski, S. McNeal, K. Thomas, X. Yin, S. Anthony, R. Ray, R.S. Vose, B.E.Gleason, and T.G. Houston, 2012: Global Historical Climatology Network - Daily (GHCN-Daily), Version 3.22. NOAA National Climatic Data Center. http://doi.org/10.7289/V5D21VHZ [11 May 2016].
Mesinger, F., and Coauthors, 2006: North American regional reanalysis. Bulletin of the American Meteorological Society, 87, 343-+.
Metzger, K. B., K. Ito, and T. D. Matte, 2010: Summer Heat and Mortality in New York City: How Hot Is Too Hot? Environmental Health Perspectives, 118,80-86.
Michel, S. J., and Coauthors, 2016: Investigating the relationship between weather and violence in Baltimore, Maryland, USA. Injury-International Journal of the Care of the Injured, 47, 272-276.
Mitchell, K. E., and Coauthors, 2004: The multi-institution North American Land Data Assimilation System (NLDAS): Utilizing multiple GCIP products and partners in a
106
continental distributed hydrological modeling system. Journal of Geophysical Research-Atmospheres, 109, 32.
Moberg, A., and Coauthors, 2006: Indices for daily temperature and precipitation extremes in Europe analyzed for the period 1901-2000. Journal of Geophysical Research-Atmospheres, 111, 25.
Murtagh, F., 1983: A SURVEY OF RECENT ADVANCES IN HIERARCHICAL-CLUSTERING ALGORITHMS. Computer Journal, 26, 354-359.
Myoung, B., S. H. Kim, J. Kim, and M. C. Kafatos, 2015: On the Relationship between the North Atlantic Oscillation and Early Warm Season Temperatures in the Southwestern United States. Journal of Climate, 28, 5683-5698.
National Oceanic and Atmospheric Administration (2015), U.S. weather related fatalities. [Available athttp://www.nws.noaa.gov/om/hazstats.shtml, Accessed on July 23, 2016.]
Nicholson, S. E., 2014: The Predictability of Rainfall over the Greater Horn of Africa. Part I: Prediction of Seasonal Rainfall. Journal of Hydrometeorology,15, 1011-1027.
Pan, Z. T., R. W. Arritt, E. S. Takle, W. J. Gutowski, C. J. Anderson, and M. Segal, 2004: Altered hydrologic feedback in a warming climate introduces a ''warming hole''. Geophysical Research Letters, 31, 4.
Peng, R. D., J. F. Bobb, C. Tebaldi, L. McDaniel, M. L. Bell, and F. Dominici, 2011: Toward a Quantitative Estimate of Future Heat Wave Mortality under Global Climate Change. Environmental Health Perspectives, 119, 701-706.
Pokorna, L., and R. Huth, 2015: Climate impacts of the NAO are sensitive to how the NAO is defined. Theoretical and Applied Climatology, 119, 639-652.
Portmann, R. W., S. Solomon, and G. C. Hegerl, 2009: Spatial and seasonal patterns in climate change, temperatures, and precipitation across the United States. Proceedings of the National Academy of Sciences of the United States of America, 106, 7324-7329.
Rasouli, K., W. W. Hsieh, and A. J. Cannon, 2012: Daily streamflow forecasting by machine learning methods with weather and climate inputs. Journal of Hydrology, 414, 284-293.
Robinson, P. J., 2001: On the definition of a heat wave. Journal of Applied Meteorology, 40, 762-775.
Schar, C., and G. Jendritzky, 2004: Climate change: Hot news from summer 2003. Nature, 432, 559-560.
107
Schar, C., P. L. Vidale, D. Luthi, C. Frei, C. Haberli, M. A. Liniger, and C. Appenzeller, 2004: The role of increasing temperature variability in European summer heatwaves. Nature, 427, 332-336.
Schwartz, J., 2005: Who is sensitive to extremes of temperature? A case-only analysis. Epidemiology, 16, 67-72.
Semenza, J. C., J. E. McCullough, W. D. Flanders, M. A. McGeehin, and J. R. Lumpkin, 1999: Excess hospital admissions during the July 1995 heat wave in Chicago. American Journal of Preventive Medicine, 16, 269-277.
Shortridge, J. E., S. M. Falconi, B. F. Zaitchik, and S. D. Guikema, 2015: Climate, agriculture, and hunger: statistical prediction of undernourishment using nonlinear regression and data-mining techniques. Journal of Applied Statistics, 42, 2367-2390.
Singh, A., T. Delcroix, and S. Cravatte, 2011: Contrasting the flavors of El Nino-Southern Oscillation using sea surface salinity observations. Journal of Geophysical Research-Oceans, 116.
Smith, T. T., B. F. Zaitchik, and J. M. Gohlke, 2013: Heat waves in the United States: definitions, patterns and trends. Climatic Change, 118, 811-825.
Steadman, R. G., 1979: ASSESSMENT OF SULTRINESS .2. EFFECTS OF WIND, EXTRA RADIATION AND BAROMETRIC PRESSURE ON APPARENT TEMPERATURE. Journal of Applied Meteorology, 18, 874-885.
——, 1984: A UNIVERSAL SCALE OF APPARENT TEMPERATURE. Journal of Climate and Applied Meteorology, 23, 1674-1687.
Sutton, R. T., and D. L. R. Hodson, 2005: Atlantic Ocean forcing of North American and European summer climate. Science, 309, 115-118.
Tan, J. G., Y. F. Zheng, G. X. Song, L. S. Kalkstein, A. J. Kalkstein, and X. Tang, 2007: Heat wave impacts on mortality in Shanghai, 1998 and 2003.International Journal of Biometeorology, 51, 193-200.
Tarleton, L. F., and R. W. Katz, 1995: STATISTICAL EXPLANATION FOR TRENDS IN EXTREME SUMMER TEMPERATURES AT PHOENIX, ARIZONA. Journal of Climate, 8, 1704-1708.
Trenberth, K. E., and P. D. Jones, 2007: Observations: Surface and Atmospheric Climate Change. Climate Change 2007: the Physical Science Basis, S. Solomon, and Coauthors, Eds., Cambridge Univ Press, 235-336.
USGCRP, 2016: The Impacts of Climate Change on Human Health in the United States: A Scientific Assessment. Crimmins, A., J. Balbus, J.L. Gamble, C.B. Beard, J.E. Bell, D. Dodgen,
108
R.J. Eisen, N. Fann, M.D. Hawkins, S.C. Herring, L. Jantarasami, D.M. Mills, S. Saha, M.C. Sarofim, J. Trtanj, and L. Ziska, Eds. U.S. Global Change Research Program, Washington, DC, 312 pp. http://dx.doi.org/10.7930/J0R49NQX
Van Oldenborgh, G. J., and A. Van Ulden, 2003: On the relationship between global warming, local warming in the Netherlands and changes in circulation in the 20th century. International Journal of Climatology, 23, 1711-1724.
Vaneckova, P., G. Neville, V. Tippett, P. Aitken, G. Fitzgerald, and S. L. Tong, 2011: Do Biometeorological Indices Improve Modeling Outcomes of Heat-Related Mortality? Journal of Applied Meteorology and Climatology, 50, 1165-1176.
Visbeck, M. H., J. W. Hurrell, L. Polvani, and H. M. Cullen, 2001: The North Atlantic Oscillation: Past, present, and future. Proceedings of the National Academy of Sciences of the United States of America, 98, 12876-12877.
Wang, Z., C. P. Chang, and B. Wang, 2007: Impacts of El Nino and La Nina on the U.S. climate during northern summer. Journal of Climate, 20, 2165-2177.
Ward, J. H., 1963: HIERARCHICAL GROUPING TO OPTIMIZE AN OBJECTIVE FUNCTION. Journal of the American Statistical Association, 58, 236-&.
Westerling, A. L., H. G. Hidalgo, D. R. Cayan, and T. W. Swetnam, 2006: Warming and earlier spring increase western US forest wildfire activity. Science,313, 940-943.
Williams, S., and Coauthors, 2012: Heat and health in Adelaide, South Australia: Assessment of heat thresholds and temperature relationships. Science of the Total Environment, 414, 126-133.
Wolter, K., 1987: THE SOUTHERN OSCILLATION IN SURFACE CIRCULATION AND CLIMATE OVER THE TROPICAL ATLANTIC, EASTERN PACIFIC, AND INDIAN OCEANS AS CAPTURED BY CLUSTER-ANALYSIS. Journal of Climate and Applied Meteorology, 26, 540-558.
Wolter, K., and M. S. Timlin, 2011: El Nino/Southern Oscillation behaviour since 1871 as diagnosed in an extended multivariate ENSO index (MEI.ext).International Journal of Climatology, 31, 1074-1087.
Wright, C. K., K. M. de Beurs, and G. M. Henebry, 2014: Land surface anomalies preceding the 2010 Russian heat wave and a link to the North Atlantic oscillation. Environmental Research Letters, 9, 10.
Xia, Y. L., and Coauthors, 2012: Continental-scale water and energy flux analysis and validation for the North American Land Data Assimilation System project phase 2 (NLDAS-2): 1. Intercomparison and application of model products. Journal of Geophysical Research-Atmospheres, 117, 27.
109
Zhang, R., and T. L. Delworth, 2007: Impact of the Atlantic Multidecadal Oscillation on North Pacific climate variability. Geophysical Research Letters, 34, 6.
Zhang, Y., J. M. Wallace, and D. S. Battisti, 1997: ENSO-like interdecadal variability: 1900-93. Journal of Climate, 10, 1004-1020.
110
AUTHOR’S CURRICULUM VITAE
Tiffany T. Smith EDUCATION
2016 Ph.D. Earth Science
Department of Earth and Planetary Sciences, Johns Hopkins University
Advisor: Dr. Benjamin Zaitchik Dissertation: Summertime heat across the United States
2013 M.A. Earth Science
Department of Earth and Planetary Sciences, Johns Hopkins University
2009 B.S. Earth Science (Minor: Chemistry)
Department of Geosciences, Oregon State University
EXPERIENCE
2015 – present Market Analyst
Fundamentals and Analytics, Constellation Energy
2016 Lab Mentor
Department of Meteorology & Atmospheric Science, Penn State Univ Course: Weather Risk and Financial Markets (METEO 460)
2015 Instructor
Betamore Course: Introduction to Data Analysis Using R
2011, 2013 Teaching Assistant
Department of Earth and Planetary Science, Johns Hopkins University
Courses: Oceans and Atmospheres (AS.270.108); Introduction to Sustainability
(AS.270.107)
2010 – 2015 Graduate Research Assistant
Department of Earth and Planetary Sciences, Johns Hopkins University
Advisor: Dr. Benjamin Zaitchik Dissertation: Summertime heat across the United States
2009-2010 Research Intern
Hydrologic Sciences Branch, NASA Goddard Space Flight Center (GSFC)
Supervisor: Dr. Matthew Rodell Research: Variations in precipitation over India; compilation of groundwater
level data in Mississippi River Basin
Global Modeling and Assimilation Office, NASA GSFC
Supervisor: Dr. Randal Koster
111
Research: Global analysis of water-limited versus energy-limited drought
environments
2009 GIS Assistant
Transboundary Freshwater Dispute Database, Oregon State University Research: Geographic and historic data conversion to online GIS system
2008 Instructor
Annapolis Sailing School Course: Adult learn-to-sail program, 5-day and 2-day sessions
2006-2009 Researcher’s Assistant
Center for Genome Research & Biocomputing, Oregon State University
James C. Carrington Lab Research: Investigation of RNA mutations in Arabidopsis thaliana
2006 Data Collector
Parks and Recreation Department, City of Boulder, CO Research: Data collection in the field to better understand the effects of the
bubonic plague on prairie dog populations
PUBLICATIONS
Smith TT, BF Zaitchik, and SD Guikema (2016) Large-scale drivers of interannual summertime
temperature variability across the Continental United States. Journal of Applied Meteorology and
Climatology. Under Review
Kent ST, LA McClure, BF Zaitchik, TT Smith, and JM Gohlke (2014) Heat Waves and Health
Outcomes in Alabama (USA): The Importance of Heat Wave Definition. Environmental Health
Persepctives. DOI:10.1289/ehp.1307262
Smith TT, BF Zaitchik, and JM Gohlke (2013) Heat waves in the United States: definitions,
patterns and trends. Climatic Change 118 (3-4):811-825. DOI: 10.1007/s10584-012-0659-2
Cuperus JT, TA Montgomery, N Fahlgren, RT Burke, T Townsend, CM Sullivan and JC
Carrington (2010) Identification of MIR390a Precursor Processing-Defective Mutants in
Arabidopsis by direct genome sequencing. Proc Natl Acad Sci. DOI: 10.1073/pnas.0913203107.
PRESENTATIONS
Smith TT (2015) Characterization, forcings, and feedbacks on summertime temperatures across
the United States. Invited George Mason University Climate Dynamics Seminar. March 25.
Fairfax, VA.
Smith TT, BF Zaitchik, and JA Santanello (2014) The role of land-atmosphere interactions
during the CONUS 2012 summertime heat wave. AGU Fall Meeting. December 15-19. San
Francisco, CA.
112
Smith TT, BF Zaitchik, and SG Guikema (2014) Remote forcings on summertime heat waves
across the United States. AMS 26th
Conference on Climate Variability and Change. February 2-
6. Atlanta, GA
Gohlke JM, ST Kent, TT Smith, LA McClure, and BF Zaitchik (2013) Heat waves and health
outcomes: The importance of heat wave definition. ISPRS: Climate Variability and Health.
August 25-29. Arlington, VA.
Smith TT, BF Zaitchik, JM Gohlke, and SG Guikema (2013) Heat waves in the United States.
5th
Annual Atmosphere-Ocean Science Days. June 6-7. Baltimore, MD
Smith TT, BF Zaitchik, and JM Gohlke (2013) Heat waves in the United States: definitions,
patterns and trends. AMS Fourth Conference on Environment and Health. January 6-10. Austin,
TX.
Smith TT, BF Zaitchik, MC Anderson, MT Yilmaz, CA Alo, and M Rodell (2011) Remotely
Sensed Terrestrial Water Balance of the Nile Basin. AGU Fall Meeting. December 5-9. San
Francisco, CA.
PROFESSIONAL DEVELOPMENT
July 2013 WRF Users Tutorial
NCAR Foothills Laboratory
Boulder, CO
July 2013 Fifth Biannual Colloquium On Climate And Health
NCAR Foothills Laboratory/Center for Disease Control
Boulder, CO
January 2013 Extreme Weather, Climate and Health: Putting Science into Practice
National Institute of Health/Center for Disease Control
Washington, DC
November 2012 Climate Dynamics of Tropical Africa: Present Understanding and Future
Direction
Department of Earth and Planetary Sciences, Johns Hopkins University
Baltimore, MD
July 2011 Climate Resilience in the Blue Nile/Abay Highlands
Addis Ababa University/Johns Hopkins Univ. Global Water Program
Bahir Dar, Ethiopia
Apr 2011 Managing Climate Change Impacts on Water Resources
American Water Resources Association
Baltimore, MD
113
Mar 2009 Current State of Palestinian Water Resources
Department of Earth and Environmental Sciences, Al-Quds University
Abu Dis, West Bank
HONORS AND AWARDS
2013 Graduate Student Summer Field Grant
Department of Earth & Planetary Sciences, Johns Hopkins University
PROFESSIONAL ACTIVITIES AND AFFILIATIONS
Journal Reviewer
Climatic Change
International Journal of Climatology
Journal of Health Geographics
Member
American Meteorological Society
American Geophysical Union
SKILLS
Mac, Windows & Unix environments
R
Python
MATLAB
NCAR Command Language (NCL)
Land Information System (LIS)
Weather Research and Forecasting Model (WRF)
NASA Unified WRF (NU-WRF)
top related