Analysis of the Performance of the MODIS LAI and FPAR Algorithm
MODIS Science Team Meeting,
BWI Airport Marriott, Baltimore, MD, July 15-16, 2003
N. Shabanov, W. Yang, B. Tan, H. Dong,
R. B. Myneni, Y. Knyazikhin /Boston University
S. W. Running, J.Glassy, P. Votava, R. Nemani /University of Montana,
NASA Ames Research Center
Roadmap of the Presentation
1. Status of the MODIS TERRA and AQUA LAI/FPAR
2. Analysis of the MODIS Terra LAI/FPAR Collection 3 Data Time Series from November 2000 to December 2002
3. Assessment of the Performance of the MODIS LAI Algorithm as a Function of the Input Uncertainties: Case Study with Grasses
4. Analysis of the Performance of the MODIS LAI/FPAR Algorithm over Broadleaf Forests
1. Status of the MODIS Terra and Aqua LAI/FPAR
Status
TERRA LAI/FPAR (MOD15A2):
• Collection 3-- Generation completed. Coverage: November 2000 – December 2002, Validation status = “Validated Stage 1”, QA status = “Inferred Passed”
• Collection 4-- Generation In progress. Released to public as of March 7, 2003. Coverage: March 2000 – December 2001 and January 2003 – present. Year 2002 will be reprocessed before end of this year. Validation status = “Provisional”, QA status = “Inferred Passed”
AQUA LAI/FPAR (MYD15A2):
• Product generated since October 24, 2002. Global coverage available since January 2003. Data are available for internal evaluation only
• Currently new AQUA LUTs are under testing by LDOPE QA team. After end of testing AQUA LAI/FPAR product is planned for public release
Documentation:
• PI website ( http://cybele.bu.edu ), FLUXNET site with ASCII subsets of LAI, EDC DAAC site were updated with material for collection 4 (including user guide)
Biome: validated by July, 2002
Transect: validated by July, 2004
Grasses/ Cereal Crops
Konza, USA SAFARI 2000 wet season
Gourma, Mali
Shrubs Puechabon, France
Broadleaf Crops Bondville, USA
Savannas SAFARI 2000 wet season Australia (planned)
Broadleaf Forests Harvard Forest, USAJaervselja, Estonia
Siberia, Russia
Needle ForestsRuokolahti, FinlandFlakaliden, Sweeden
Siberia, Russia
Validation
• All 6 biomes have been sampled at field • Collaborators from Europe (VALERI), Jeff Privette team, BigFoot team • Continue to analyze field data and compare with collection 3 and 4
TERRA MODIS LAI/FPAR Collection 4 Improvements
• Non-physical peaks at high LAI values for herbaceous vegetation (biome 1 - 4) were removed
• Validation feedback (BigFoot): improved agreement with field measurements (KONZA, grasses, ARGO, crops, etc.)
• Retrievals with main algorithm increased by 20% compared to collection 3 data
Collection 3 green Collection 4 red
LUT Tuning for the Main and Back-up Algorithms
TERRA MODIS LAI/FPAR Collection 4 Improvements (Cont.)
Improvements to QA Scheme
• Reduced redundancy between MODLAND and SCF_QC quality flags• SCF_QC is more clearly structured: Main (M), Saturation (S), and Back-up (B)
Update for Input Land Cover
• At-launch AVHRR based IGBP land cover was replaced with 6-bime land cover generated from one year of MODIS data
• Cross-walking from IGBP to 6-biome was eliminated
• New LC has less uncertainties
New 8-day compositing scheme
• Compositing over best quality retrievals, instead of all retrievals• Lowers LAI values, decreases saturation and number of pixels generated by back-up
algorithm
MOD15A2, Collection 4 Data, July 20-27, 2001
LAI
QC
Achievements
• Spatial coverage of main algorithm increased by ~20% due to LUTs tuning and new compositing scheme:
M=72%, S=5%, B=23% (collection 4) M+S=60%, B=40% (collection 3)
• Improved consistency with field observations over herbaceous vegetation
Future Improvements (collection 5)…
• Decrease dominance of back-up algorithm retrievals over woody vegetation (broadleaf and needle leaf forests)
• Further improve agreement with with field data
• Research on retrievals under snow condition (resolve needle leaf forests seasonality)
TERRA MODIS LAI/FPAR Collection 4 Improvements (Cont.)
2. Analysis of the MODIS Terra LAI/FPAR Collection 3 Data Time Series from November 2000 to December 2002
Statement of the Problem
Objective
• Collection 3 MODIS TERRA LAI/FPAR product provides about two years of data time series, valuable of the assessment of the product quality. We performed analysis of the product spatial coverage, seasonality of LAI and FPAR for different vegetation types. Special attention was given to retrievals under snow and cloudy conditions.
Data Used
• MOD15A2, 8-day LAI composite, collection 3, November 2000- December 2002
Retrieval Index Seasonality
• Retrieval Index, RI = Pixels (Main algorithm) / Pixels (Main + Back-up algorithm)
• Main algorithm fails significantly less on herbaceous vegetation (grasses & cereal, shrubs, broadleaf crops and savannas), compared to woody vegetation (broadleaf and needle leaf forests)
• Strong seasonality in retrievals for latitudes > 50 degrees North is due to snow and other factors
RI by biomes RI by latitudinal band
LAI Seasonality
LAI by biomes LAI by latitudinal band
• The LAI/FPAR profiles for each biome type and latitude band have the expected shape.
• Needle leaf forests show high seasonality, which is also pronounced for the highest
latitudinal band.
LAI Retrievals Under Snow Condition
• About 50-60% of vegetated pixels north of 40 degrees North are identified as having snow during peak winter period.
• The majority of snow pixels are retrieved by backup algorithm. Main algorithm recognize non vegetation signal in data.
• Cumulative LAI retrieved under snow condition is 100 times smaller than one for snow free condition
LAI Retrievals Under Cloudy Condition
• About 50 to 60% of the vegetated pixels are identified as cloud free, 15% as partially cloudy and 25-35% are cloud covered
• LAI/FPAR algorithm performs retrievals regardless of cloud conditions
• LAI values retrieved under cloudy condition are spurious. The difference between LAI retrieved under cloudy and cloud free conditions depends on biome type.
Crops Needle Leaf Forests
3. Assessment of the Performance of the MODIS LAI Algorithm as a Function of Input Uncertainties:
Case Study with Grasses
Statement of the Problem
The Problem• As reported by BigFoot team, MOD15A2 product, collection 3 substantially
overestimate field measured LAI at Konza site (grasses, 5x5 km area):
a) Field measurements: LAI ~3
b) MODIS product: LAI =5.7 +/- 0.7
Solution Approach• Analysis of uncertainties in LAI, collection 3, was performed as function of input
uncertainties: land cover misclassification and uncertainties in input surface reflectances
Data Used• MOD15A2, 8-day LAI composite, collection 3, tile h10v05, composite July 04-11,
2001
• MODAGAGG, daily surface reflectance, collection 3, tile h10v05, days July 04-11, 2001
• MOD12Q1, 6-biome classification map, at launch and new version, tile h10v05
Impact of Biome Misclassification on LAI Retrievals
• MODIS LAI/FPAR algorithm references LC to select vegetation parameters from LUTs. Misclassification leads to errors in LAI estimation
• Collection 3 LAI used MODIS at-launch IGBP LC (AVHRR-based), cross-walked to 6 biome LC
• Collection 4 LAI referencing MODIS 6-biome LC product (based on one year MODIS data)
• Significant misclassification occur at local scale (5x5 km) for at-launch LC: this map predicts 24% of the pixels are grasses, while field measurements indicates that 64% of the pixels are grasses.
1200x1200 km 1200x1200 km
20x20 km20x20 km
LC for collection 3 LAI LC for collection 4 LAI
Biome 1 Biome 2 Biome 3 Biome 4 Biome 5 Biome 6
Impact of Uncertainties in Surface Reflectances on LAI Retrievals
• Definition: “Good data” are surface reflectance data with MODAGAGG QA = “Product produced at ideal quality” or “Product produces, less than ideal quality”
• Good quality data have lower uncertainty than poor quality data
• Uncertainties in LAI retrievals are proportional to uncertainties in surface reflectance variations
How to Reduce Impact of Input Uncertainties on LAI Retrievals?
• However averages over sufficiently large regions (in this case study- over the tile) can smooth uncertainties in retrieved LAI Due to uncertainties in input land cover / surface reflectances at the scale of few MODIS pixels errors in LAI are possible. Selecting larger spatial patches generally helps accumulate sufficient amount of correctly classified pixels and reduces errors in LAI
• Additionally, collection 4 LAI product has more accurate input data, and improvements to the algorithm were made. We got good agreement with field data: MODIS LAI collection 4 over 5x5 km subset during July 04-11, 2001 is 2.97+/-1.5 (field data: LAI ~3).
4. Analysis of the Performance of the MODIS Terra LAI/FPAR Algorithm over Broadleaf Forests
Statement of the Problem
The Problem• Dominance of back-up retrievals for broadleaf forests during summer time
Solution Approach• Investigate the properties of surface reflecatnces
Data Used• MOD15A2, 8-day LAI composite, collection 4, tile h12v04, all the data during year
2001
• MODAGAGG, daily surface reflectance, collection 4, tile h12v04, May 01-08, 2001 and July 12-19, 2001
• MOD12Q1, 6-biome classification map, collection 3, tile h12v04
Broadleaf Forests of New England
LC LAI: May 01-08, 2001
QC: May 01-08, 2001 QC: July 12-19, 2001
LAI: July 12-19, 2001
Problem:
Dominance of back-up retrievals for broadleaf forests during summer time (MODIS tile h12v04 is shown).
0.0 0.1 0.3 0.5 0.8 1.2 1.6 2.1 2.8 3.4 4.4 5.4 6.0 7.0
LAI
Saturation Back-UpMain
QC
Biome 1
LC
Biome 2 Biome 3 Biome 4 Biome 5 Biome 6
Broadleaf Forests of New England
• Collection 4 data for broadleaf forest, tile h12v04 are shown
• LAI/FPAR algorithm correctly captures seasonality in LAI
• However during summer retrievlas are performrmed mostly with back-up algorithm (main fails, picture on the bottom)
• What changes in surface reflectances are responsible for decrease in Main algorithm retrievals during transition from early spring to summer when LAI reaches its maximum?
Saturation Back-UpMain
QC
Broadleaf Forests of New England
VariableMay 01-08
July 12-19
Change, %
Red 0.047 0.034 -27.5%
NIR 0.226 0.387 +71.2%
NDVI 0.646 0.835 +29.1%
LAI 2.97 5.59 +88.2%
May 01- 08, 2001 July 12 - 19, 2001
Broadleaf Forests of New England
• Analysis of surface reflectances indicates that predominant location of MODIS observations (in Red/NIR spectral space) during summer time mismatch the model predictions as stored in LUTs of the algorithm.
• LUTs will be updated to be in agreement with observed values of surface reflectances. The proposed changes will be implemented in collection 5