Introduction Methodology Results Conclusion Other information
An Assessment of Deforestation Models forReducing Emissions from Deforestation & forest
Degradation (REDD)A Case Study of Chiquitanıa, Bolivia
Oh Seok Kim
Department of Geography, University of Southern California
O Kim (Geography@USC) Deforestation Models for REDD 1
Introduction Methodology Results Conclusion Other information
Background
What’s REDD?
I Carbon credit is a new monetary system [UN FCCC 1998].
I A carbon offset project must prove additional carbon benefitto generate carbon credits [IPCC 2000].
- Afforestation- Reforestation- REDD: Reducing Emissions from Deforestation & Degradation
I UN-REDD
- The United Nations Collaborative Programme on ReducingEmissions from Deforestation and Forest Degradation inDeveloping Countries [UN-REDD 2008].
- Website: http://www.un-redd.org
O Kim (Geography@USC) Deforestation Models for REDD 2
Introduction Methodology Results Conclusion Other information
Background
What’s REDD?
I Carbon credit is a new monetary system [UN FCCC 1998].I A carbon offset project must prove additional carbon benefit
to generate carbon credits [IPCC 2000].
- Afforestation- Reforestation
- REDD: Reducing Emissions from Deforestation & Degradation
I UN-REDD
- The United Nations Collaborative Programme on ReducingEmissions from Deforestation and Forest Degradation inDeveloping Countries [UN-REDD 2008].
- Website: http://www.un-redd.org
O Kim (Geography@USC) Deforestation Models for REDD 2
Introduction Methodology Results Conclusion Other information
Background
What’s REDD?
I Carbon credit is a new monetary system [UN FCCC 1998].I A carbon offset project must prove additional carbon benefit
to generate carbon credits [IPCC 2000].
- Afforestation- Reforestation- REDD: Reducing Emissions from Deforestation & Degradation
I UN-REDD
- The United Nations Collaborative Programme on ReducingEmissions from Deforestation and Forest Degradation inDeveloping Countries [UN-REDD 2008].
- Website: http://www.un-redd.org
O Kim (Geography@USC) Deforestation Models for REDD 2
Introduction Methodology Results Conclusion Other information
Background
What’s REDD?
I Carbon credit is a new monetary system [UN FCCC 1998].I A carbon offset project must prove additional carbon benefit
to generate carbon credits [IPCC 2000].
- Afforestation- Reforestation- REDD: Reducing Emissions from Deforestation & Degradation
I UN-REDD
- The United Nations Collaborative Programme on ReducingEmissions from Deforestation and Forest Degradation inDeveloping Countries [UN-REDD 2008].
- Website: http://www.un-redd.org
O Kim (Geography@USC) Deforestation Models for REDD 2
Introduction Methodology Results Conclusion Other information
Background
Why LUCC modeling?
I In the case of REDD, the additional carbon benefit iscalculated based on business-as-usual deforestation(i.e., baseline).
I Land-use and land-cover change (LUCC) modeling cansystematically simulate business-as-usual deforestation.
I Its result can be later combined with the correspondingcarbon content’s map.
O Kim (Geography@USC) Deforestation Models for REDD 3
Introduction Methodology Results Conclusion Other information
Background
Why LUCC modeling?
I In the case of REDD, the additional carbon benefit iscalculated based on business-as-usual deforestation(i.e., baseline).
I Land-use and land-cover change (LUCC) modeling cansystematically simulate business-as-usual deforestation.
I Its result can be later combined with the correspondingcarbon content’s map.
O Kim (Geography@USC) Deforestation Models for REDD 3
Introduction Methodology Results Conclusion Other information
Background
Why LUCC modeling?
I In the case of REDD, the additional carbon benefit iscalculated based on business-as-usual deforestation(i.e., baseline).
I Land-use and land-cover change (LUCC) modeling cansystematically simulate business-as-usual deforestation.
I Its result can be later combined with the correspondingcarbon content’s map.
O Kim (Geography@USC) Deforestation Models for REDD 3
Introduction Methodology Results Conclusion Other information
Objectives
General scope of the research
1. This paper proposes a modular framework to assess accuracyof a spatially-explicit REDD baseline by demonstrating thecomparison of two LUCC modeling approaches.
2. It compares the performance of “GEOMOD Modeling (GM)”and “Land Change Modeler (LCM)” for simulating baselinedeforestation of multiple transitions based on model structureand predictive accuracy.
O Kim (Geography@USC) Deforestation Models for REDD 4
Introduction Methodology Results Conclusion Other information
Objectives
General scope of the research
1. This paper proposes a modular framework to assess accuracyof a spatially-explicit REDD baseline by demonstrating thecomparison of two LUCC modeling approaches.
2. It compares the performance of “GEOMOD Modeling (GM)”and “Land Change Modeler (LCM)” for simulating baselinedeforestation of multiple transitions based on model structureand predictive accuracy.
O Kim (Geography@USC) Deforestation Models for REDD 4
Introduction Methodology Results Conclusion Other information
Objectives
Research questions
1. How do GM’s linear extrapolation and LCM’s Markov Chaincompare in terms of quantity of LUCC (for multiple transitionmodeling)?
2. How do GM’s empirical frequency, LCM’s logistic regression,and LCM’s multilayer perceptron compare in terms of(spatial) allocation of LUCC, measured by relative operatingcharacteristics, figure of merit, and multiple resolutionanalysis?
O Kim (Geography@USC) Deforestation Models for REDD 5
Introduction Methodology Results Conclusion Other information
Objectives
Research questions
1. How do GM’s linear extrapolation and LCM’s Markov Chaincompare in terms of quantity of LUCC (for multiple transitionmodeling)?
2. How do GM’s empirical frequency, LCM’s logistic regression,and LCM’s multilayer perceptron compare in terms of(spatial) allocation of LUCC, measured by relative operatingcharacteristics, figure of merit, and multiple resolutionanalysis?
O Kim (Geography@USC) Deforestation Models for REDD 5
Introduction Methodology Results Conclusion Other information
Study area and data
Chiquitanıa, Bolivia: land-cover of 1986,1994 & 2000
0 10 205Km
Persistent ForestPersistent SavannaForest Disturbed between 1986 and 1994Forest Disturbed between 1994 and 2000Savanna Disturbed between 1986 and 1994Savanna Disturbed between 1994 and 2000Persistent Anthropogenic DisturbanceOther Transitions between 1994 and 2000
O Kim (Geography@USC) Deforestation Models for REDD 6
Introduction Methodology Results Conclusion Other information
Study area and data
Environmental variables
43,700 0
41,500 0
Distance from disturbance (meter) Distance from roads (meter)a b
13,600 0
124,500 0
Distance from streams (meter)c Distance from cities (meter)d
Elevation (meter)e Slope (degree)f
870 220
61 0
O Kim (Geography@USC) Deforestation Models for REDD 7
Introduction Methodology Results Conclusion Other information
Strategy
Multiple transition modeling
I Transition refers to a process in which something undergoes achange from one land-type (e.g., forest) to another (e.g.,anthropogenic).
I There is a need of modeling multiple transitions in REDDprojects since there can be numerous transitions due todifferent types of land management.
O Kim (Geography@USC) Deforestation Models for REDD 8
Introduction Methodology Results Conclusion Other information
Strategy
Calibration and validation
I The common essence of these validation processes isseparating data for calibration and validation.
O Kim (Geography@USC) Deforestation Models for REDD 9
Introduction Methodology Results Conclusion Other information
Model structure
Quantity of LUCC
I Linear extrapolation (GM)
I Markov Chain (LCM)
O Kim (Geography@USC) Deforestation Models for REDD 10
Introduction Methodology Results Conclusion Other information
Model structure
Allocation of LUCC (or transition potential map)
I Empirical frequency (EmpFreq) equation
- Is clear to interpret the output;- Has information of variables’ contributions;- Subjective input: number of bins/width of bins.
I Logistic Regression (LogReg) equation
- Is easier to interpret the output than MLP(i.e.,linear estimation);
- Has information of variables’ contributions;- Subjective input: statistical significance test.
I Multilayer perceptron (MLP)
- Is complex and often limits interpretation(i.e.,nonlinear & black box-like estimation);
- Has no information of variables’ contributions;- Subjective input: network training parameters.
O Kim (Geography@USC) Deforestation Models for REDD 11
Introduction Methodology Results Conclusion Other information
Model structure
Allocation of LUCC (or transition potential map)
I Empirical frequency (EmpFreq) equation
- Is clear to interpret the output;- Has information of variables’ contributions;- Subjective input: number of bins/width of bins.
I Logistic Regression (LogReg) equation
- Is easier to interpret the output than MLP(i.e.,linear estimation);
- Has information of variables’ contributions;- Subjective input: statistical significance test.
I Multilayer perceptron (MLP)
- Is complex and often limits interpretation(i.e.,nonlinear & black box-like estimation);
- Has no information of variables’ contributions;- Subjective input: network training parameters.
O Kim (Geography@USC) Deforestation Models for REDD 11
Introduction Methodology Results Conclusion Other information
Model structure
Allocation of LUCC (or transition potential map)
I Empirical frequency (EmpFreq) equation
- Is clear to interpret the output;- Has information of variables’ contributions;- Subjective input: number of bins/width of bins.
I Logistic Regression (LogReg) equation
- Is easier to interpret the output than MLP(i.e.,linear estimation);
- Has information of variables’ contributions;- Subjective input: statistical significance test.
I Multilayer perceptron (MLP)
- Is complex and often limits interpretation(i.e.,nonlinear & black box-like estimation);
- Has no information of variables’ contributions;- Subjective input: network training parameters.
O Kim (Geography@USC) Deforestation Models for REDD 11
Introduction Methodology Results Conclusion Other information
Predictive accuracy assessment
Map comparison measurements
I Relative operating characteristic (ROC) ex. equation
- Assesses the predictive accuracy of a transition potential mapby comparing the transition potential map with a referencemap that shows the occurrence of binary events;
- Can be roughly summarized by an Area Under the ROC Curve(AUC).
I Figure of merit ex. equation
- Assesses the predictive accuracy of a simulated LUCC bysimply overlaying the land-cover map of 1994, the land-covermap of 2000 and the predicted map of 2000;
- Has both visual and numeric expressions (Figure of meritstatistics).
O Kim (Geography@USC) Deforestation Models for REDD 12
Introduction Methodology Results Conclusion Other information
Predictive accuracy assessment
Map comparison measurements(con’d.)
I Multiple resolution analysis ex.
- Assesses how closely in space the simulated LUCC is to thereference LUCC;
- Compares each simulation map to its reference map byaggregating pixels to coarser resolutions;
- Assumes “persistence” as the prediction of future LUCC thatexperiences no change;
- Employs a “null resolution,” and more precise LUCCsimulations have finer null resolutions.
O Kim (Geography@USC) Deforestation Models for REDD 13
Introduction Methodology Results Conclusion Other information
Quantity of LUCC
Linear extrapolation vs. Markov Chain
Linear extrapolation makes sense when there is only onetransition of land-cover change, e.g., undisturbed forest todisturbed forest.
Markov Chain is applicable for multiple transition modeling.
O Kim (Geography@USC) Deforestation Models for REDD 14
Introduction Methodology Results Conclusion Other information
Allocation of LUCC
Transition potential: Forest to anthropogenic
0.930
0.930
0.930
a Empirical frequency b Logistic regression c Multilayer perceptron
0 10 205Km
0 10 205Km
0 10 205Km
O Kim (Geography@USC) Deforestation Models for REDD 15
Introduction Methodology Results Conclusion Other information
Allocation of LUCC
Transition potential: Savanna to anthropogenic
0.770
0.770
0.770
a Empirical frequency b Logistic regression c Multilayer perceptron
0 10 205Km
0 10 205Km
0 10 205Km
O Kim (Geography@USC) Deforestation Models for REDD 16
Introduction Methodology Results Conclusion Other information
Map comparison
ROC: Forest to anthropogenic logic
O Kim (Geography@USC) Deforestation Models for REDD 17
Introduction Methodology Results Conclusion Other information
Map comparison
ROC: Forest to anthropogenic (Zoom-in)
O Kim (Geography@USC) Deforestation Models for REDD 18
Introduction Methodology Results Conclusion Other information
Map comparison
ROC: Savanna to anthropogenic
O Kim (Geography@USC) Deforestation Models for REDD 19
Introduction Methodology Results Conclusion Other information
Map comparison
ROC: Savanna to anthropogenic (Zoom-in)
O Kim (Geography@USC) Deforestation Models for REDD 20
Introduction Methodology Results Conclusion Other information
Map comparison
Figure of merit logic
0 10 205Km
0 10 205Km
0 10 205Km
Error due to observed change predicted as persistenceCorrect due to observed change predicted as changeError due to observed persistence predicted as changeCorrect due to observed persistence predicted as persistence
Error due to observed change predicted as persistenceCorrect due to observed change predicted as changeError due to observed persistence predicted as changeCorrect due to observed persistence predicted as persistence
Error due to observed change predicted as persistenceCorrect due to observed change predicted as changeError due to observed persistence predicted as changeCorrect due to observed persistence predicted as persistence
a Empirical frequency b Logistic regression c Multilayer perceptron
O Kim (Geography@USC) Deforestation Models for REDD 21
Introduction Methodology Results Conclusion Other information
Map comparison
Multiple resolution analysis: EmpFreq logic
O Kim (Geography@USC) Deforestation Models for REDD 22
Introduction Methodology Results Conclusion Other information
Map comparison
Multiple resolution analysis: LogReg
O Kim (Geography@USC) Deforestation Models for REDD 23
Introduction Methodology Results Conclusion Other information
Map comparison
Multiple resolution analysis: MLP
O Kim (Geography@USC) Deforestation Models for REDD 24
Introduction Methodology Results Conclusion Other information
Summary
AUCs: Forest to anthropogenic
AUC AUCp AUCt
EmpFreq .800 .000274 .00444LogReg .791 .000304 .00477MLP .813 .000300 .00479
O Kim (Geography@USC) Deforestation Models for REDD 25
Introduction Methodology Results Conclusion Other information
Summary
AUCs: Savanna to anthropogenic
AUC AUCp AUCt
EmpFreq .706 .001415 .01088LogReg .741 .001422 .01266MLP .703 .001034 .00772
O Kim (Geography@USC) Deforestation Models for REDD 26
Introduction Methodology Results Conclusion Other information
Summary
Figure of merit & null resolution
Figure of merit Null resolution
EmpFreq 6.62% 2.4∼4.8kmLogReg 8.0% 1.2∼2.4kmMLP 6.57% about 2.4km
O Kim (Geography@USC) Deforestation Models for REDD 27
Introduction Methodology Results Conclusion Other information
Revisit
Research questions
1. How do GM’s linear extrapolation and LCM’s Markov Chaincompare in terms of quantity of LUCC (for multiple transitionmodeling)?
2. How do GM’s empirical frequency, LCM’s logistic regression,and LCM’s multilayer perceptron compare in terms of(spatial) allocation of LUCC, measured by relative operatingcharacteristics, figure of merit, and multiple resolutionanalysis?
O Kim (Geography@USC) Deforestation Models for REDD 28
Introduction Methodology Results Conclusion Other information
Revisit
Research questions
1. How do GM’s linear extrapolation and LCM’s Markov Chaincompare in terms of quantity of LUCC (for multiple transitionmodeling)?
2. How do GM’s empirical frequency, LCM’s logistic regression,and LCM’s multilayer perceptron compare in terms of(spatial) allocation of LUCC, measured by relative operatingcharacteristics, figure of merit, and multiple resolutionanalysis?
O Kim (Geography@USC) Deforestation Models for REDD 28
Introduction Methodology Results Conclusion Other information
Conclusion
LCM is better than GM!
Based on the model structure and predictive accuracy comparisons,the LCM seems more suitable than the GM to construct a REDDbaseline when considering multiple transitions.
1. The embedded Markov Chain can estimate the quantity ofLUCC for multiple transitions, while it is challenging for linearextrapolation to do so;
2. The LCM’s LogReg has the highest predictive accuracy inmost cases (it is important to note that MLP contains astochastic element).
O Kim (Geography@USC) Deforestation Models for REDD 29
Introduction Methodology Results Conclusion Other information
Discussion
Predictive accuracy
I From a land change scientist’s point of view, the differences,measured by ROC, figure of merit, or multiple resolutionanalysis, must be statistically tested to check out howsignificantly they differ.
I It is invalid to employ a traditional t test or Mann-Whitney Utest to measure such differences since those are notindependent samples of observations.
O Kim (Geography@USC) Deforestation Models for REDD 30
Introduction Methodology Results Conclusion Other information
Discussion
Carbon implication
I Most likely, a carbon mass map of rainforests will contain aconfidence interval to indicate its measurement precision.
I If the difference (in terms of carbon mass) between two LUCCmodeling approaches falls into the confidence interval of thecarbon mass measured then it seems reasonable to regard thedifference (in terms of predictive accuracy) between the twoLUCC modeling approaches as insignificant.
O Kim (Geography@USC) Deforestation Models for REDD 31
Introduction Methodology Results Conclusion Other information
Appendix
Empirical frequency (EmpFreq)
R(i) =
∑Aa=1WaPa(i)∑A
a=1Wa
(1)
where R(i) is a transition potential value in pixel (i), a indicates aparticular environmental variable, A indicates the total number ofenvironmental variables, Wa is the weight of environmental variablea, and Pa(i) indicates the percent of LUCC during the calibrationinterval in the bin to which pixel i belongs for variable a.
back
O Kim (Geography@USC) Deforestation Models for REDD 32
Introduction Methodology Results Conclusion Other information
Appendix
Logistic regression (LogReg)
P (y = 1|X) =exp
∑BX
1 + exp∑
BX(2)
where y is a binary event, P is the probability of the binary eventgiven column vector X, X is a column vector that give the valuesof the environmental variables, and B is a row vector that givesthe estimated coefficients.
back
O Kim (Geography@USC) Deforestation Models for REDD 33
Introduction Methodology Results Conclusion Other information
Appendix
Area Under the ROC Curve
AUC =
n∑i=1
(xi+1 − xi)×
{yi +
(yi+1 − yi)
2
}(3)
where xi is the false positives for the threshold i, yi is the truepositives for threshold i, and n+ 1 is the number of thresholds.
back
O Kim (Geography@USC) Deforestation Models for REDD 34
Introduction Methodology Results Conclusion Other information
Appendix
Figure of Merit
Figure of Merit statistics = B/(A+B + C) (4)
where A is a number of pixels for “error due to observed changepredicted as persistence” (or misses), B is a number of pixels for“correct due to observed change predicted as change” (or hits), Cis a number of pixels for “error due to observed persistencepredicted as change” (or false alarms). The Figure of Meritstatistics range from 0 to 100 percent, where 100 percent indicatesperfect prediction.
back
O Kim (Geography@USC) Deforestation Models for REDD 35
Introduction Methodology Results Conclusion Other information
Bibliography
Intergovernmental Panel on Climate Change (IPCC) 2000Summary for Policymakers–Land Use, Land-Use Change andForestry.
United Nations Framework Convention on Climate Change(UN FCCC) 1998 Kyoto Protocol to the United NationsFramework Convention on Climate Change.
United Nations Collaborative Programme on ReducingEmissions from Deforestation and Forest Degradation inDeveloping Countries (UN-REDD) 2008 FrameworkDocument.
O Kim (Geography@USC) Deforestation Models for REDD 36
Introduction Methodology Results Conclusion Other information
Thank you!!!!
Author’s information
I Research interest
- REDD
I Education
- B.A. in Geography education, B.S. in Statistics- M.A. in GISc for Development & Environment- Ph.D. Student in Geography
I Contact
- E-mail: [email protected]
I Questions?
O Kim (Geography@USC) Deforestation Models for REDD 37