MODIS LAI and FPAR: Product Analysis

1
MODIS LAI and FPAR Product Analysis
N. Shabanov, W. Yang, B.Tan, H. Dong, Y.
Knyazikhin, R. Myneni Boston University,
Geography Department rmyneni_at_crsa.bu.edu
Abstract
Product Validation
Leaf Area Index (LAI) and Fraction of
Photosynthetically Active Radiation 400-700 nm
absorbed by vegetation (FPAR) are key
environmental variables used in climate research
to characterize exchange of fluxes of energy,
mass and momentum in Earth system. MODerate
resolution Imaging Spectroradiometer (MODIS)
provide set of global remote sensing measurements
to generate 16 land geophysical products
including LAI and FPAR. Current MODIS LAI and
FPAR research integrates three major components
LAI and FPAR algorithm development, product
analysis and product validation. Product analysis
is the central activity, which is closely
interconnected with algorithm development and
validation. The objective of this presentation is
to illustrate this interconnection, targeted for
continuous product refinement from Collection 3
through current Collection 4 and into future
Collection 5 of the LAI and FPAR products.
Product analysis is presented at global scale,
regional scale (MODIS tile) and local scale
(validation). At global scale we routinely
analyzed spatial coverage of the main algorithm,
studied seasonality of LAI and FPAR products, and
analyzed impact of snow and cloud conditions on
product retrievals. MODIS tile scale is
convenient for detailed study of product
retrievals as function of the algorithm and its
input uncertainties and provide direct feedback
to algorithm refinement. Finally we summarized
reported in peer-reviewed literature results of
the validation activities, performed by LAI/FPAR
team as well as external validators.
Validation, in general, refers to assessment of
the uncertainty of higher-level
satellite-sensor-derived products by comparison
to field measurements scaled to MODIS
resolutions. Validation of the MODIS LAI and FPAR
products is performed by MODIS LAI/FPAR team as
well as external validators. Below results from
three validation exercises are detailed. Summary
of completed and reported in per-reviewed
literature (to the best of our knowledge)
validation exercises is given at the end of this
section.

• Site Ruokolahti, Finland
• Measurements Dates June 14-21, 2000
• Land Cover Type Coniferous forests
• Results Comparison of aggregated high-resolution
  LAI map and corresponding MODIS LAI retrievals
  suggests satisfactory behavior of the MODIS LAI
  algorithm although variation in MODIS LAI product
  is higher than expected. An improoved correlation
  between field measurements and reduced simple
  ratio (RSR) suggests that shortwave infrared
  (SWIR) may improve accuracy of LAI retrievals for
  needleaf forests.
• Publication Wang et al. (2004). Evaluation of
  the MODIS LAI algorithm at a coniferous forest
  site in Finland. Remote Sensing of Environment,
  91, 114-127.

Product Analysis at Global Scale
Product Analysis at MODIS Tile Scale
Collection 3 LAI Anomalies and its
Causes Collection 3 product analysis and field
measurements indicate significant LAI
overestimation over some areas with herbaceous
vegetation and dominance of back-up algorithm
retrievals over woody vegetation. Uncertainties
in LAI and FPAR product are due to the following
resons a) uncertainties in the LAI/FPAR
algorithm inputs (surface reflectance data and
land cover map) and b) Uncertainties of the model
underlying LAI/FPAR algorithm. We illustrate the
impact of the above uncertainties for the three
sites (a) KONZ site in Illinois (Grasses, tile
h10v05) b) ARGO site in Kansas (Croplands, tile
h11v04) and c) HARV site in Massachusetts
(Broadleaf forests, tile h12v04).
ETM Surface Reflectances, 1x1 km
Collection 3 and Collection 4 MODIS LAI and FPAR
products were routinely analyzed as part of
product quality control activities evaluation of
spatial coverage of the main algorithm,
seasonality of LAI and FPAR products, and
analysis of the impact of snow and cloud
conditions on product retrievals. Global
distribution of LAI, biome-by-biome, is shown for
Collection 3 and Collection 4 for July 20-27,
2001. Collection 4 reprocessing improvements
resulted in elimination of retrieval anomalies
(LAI overestimation) over herbaceous vegetation
(for details refer to Product analysis at tile
scale section ). Key indicator of the quality
of LAI and FPAR retrievals is Retrieval Index
(RI), defined as ratio of main algorithm
retrievals to total retrievals. RI for Collection
4 product was improved by about 20 compared to
Collection 3. Biome-by-biome analysis of RI
indicate that main algorithm retrievals over
herbaceous vegetation are more frequent than over
woody vegetation (algorithm is highly sensitive
to precision of surface reflectances especially
in the case of woody vegetation). Global LAI
time series demonstrates expected seasonal
variation. The biome profiles have generally the
expected shape. Needle leaf forests, however,
show anomalous seasonality. Low LAI and Retrieval
index during winter time is due to low
availability (low solar zenith angle) and poor
quality (snow and cloud contamination) of input
reflectance data.
Patches, 1x1 km
Collection 3 Collection 4
1) MODIS LAI/FPAR algorithm references land cover
to select vegetation parameters required for LAI
retrievals for particular biome type. Land cover
misclassification is the first reason,
responsible for errors in LAI estimation.
Collection 3 LAI/FPAR algorithm referenced
at-launch AVHRR based LC, while
ETM LAI, 1x1 km
ETM LAI, 10x10 km
LAI

• Site Senegal, Western Sudano-Sahelian zone
• Measurements Dates June-November in years 2001
  and 2002
• Land Cover Type grass savanna
• Results Seasonal dynamics of both in situ LAI
  and FPAR were captured well by MODIS LAI and
  FPAR. MODIS LAI is overestimated by approximately
  2-15 and the overall level of FPAR is
  overestimated by 8-20
• Publication Fensholt, R., Sandlholt, I.,
  Rasmussen, M.S. (2004). Evaluation of MODIS LAI,
  fAPAR and the relation between fAPAR and NDVI in
  a semi-arid environment using in situ
  measurements. Remote Sens. Environ. vol. 91, pp.
  490-507

At-launch LC Tile h10v05, 20x20 km
LC for Collection 4 LAI Tile h10v05, 1200x1200 km
LC for Collection 4 LAI Tile h10v05, 20x20 km
At-launch LC Tile h10v05, 1200x1200 km
Collection 4 algorithm references Collection 3 LC
which is based on 1 year of MODIS observations.
At-launch LC has significant misclassification
between grasses and broadleaf crops, as can be
seen in tile h10v05, which contains KONZ site.

2) Atmospheric correction of surface reflectances
only partially reduces impact of clouds and
aerosols on data quality. Variations of retrieved
LAI were found linearly proportional to
variations in surface reflectanes due to
uncertainties of atmospheric correction for
broadleaf crops in tile h11v04. Quality of LAI
retrievals are consistent with input surface
reflectance quality good quality input data
translate to low uncertainty of retrieved LAI and
vice versa. Therefore, second factor which set
limit to the quality of the LAI and FPAR
retrievals is precision of surface reflectances.
Broadleaf Crops Tile h11v04, 1200x1200 km

• Site Alpilles, France
• Measurements Dates February 26 March 15, 2001
• Land Cover Type Croplands
• Results Collection 4 MODIS LAI is accurate to
  within 0.3LAI precision is 20, uncertainty is
  25. Biome misidentification deteriorates the
  accuracy by factor of 2. Results also indicate
  that small variations in input due to model and
  observation inaccuracies result in low precision
  of retrieved LAI.
• Publication Tan et al. (2004). Validation of
  MODIS LAI product in croplands of Alpilles,
  France. Journal of Geophysical Research, (
  Submitted).

Collection 5
Collection 4
Collection 4
3) The third reason responsible for Collection 3
product anomalies is mismatch between model
simulated and MODIS surface reflectances. This
mismatch still cased low rate of Collection 4
main algorithm retrievals during summer over
broadleaf forests in tile h12v04, which contains
HARV site. Collection 3 LAI/FPAR algorithm
parameters were based on SeaWiFS 8-km surface
reflectances. Optimization of the algorithm over
woody vegetation is complicated due to
sensitivity of LAI to uncertainties in surface
reflectances. These modifications will be
implemented in Collection 5.
Collection 4 LUTs Tile h12v04,
1200x1200 km
ETM Surface Reflectances, 3x3 km
Collection 5 LUTs Tile h12v04,
1200x1200 km
LAI sampling 3x3 km area
Retrievals under snow conditions 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 processed
by the back-up algorithm. Main algorithm
recognize non vegetation signal in data.
Cumulative LAI (LAI x amount of pixels) retrieved
under snow condition is 100 times smaller than
one for snow free condition. Overall, LAI
retrievals under snow should be considered as
unreliable (due
QC
to back-up algorithm retrievals)- they introduce
spurious seasonality especially for needle leaf
forests. Further improvements of the LAI
algorithm are required to improve LAI retrievals
for needle leaf forests during boreal winter time.
In summary, the above analysis demonstrate the
impact of three factors (land cover, surface
reflectances and model uncertainties) on the
quality of LAI retrievals. We want to emphasis
that all three factors combined are responsible
for the observed anomalies for each site.
QC
LAI, 10x10 km
Main
Saturation
Back-Up
Fill
Summary of Product Improvements
Summary of Validation
Collection 4
Collection 3
LAI and FPAR algorithm and its inputs were
revised for Collection 4 as follows a) the
at-launch AVHRR-based land cover map was replaced
with MODIS observation-based LC b) atmospheric
correction of surface reflectanes incorporates
improved cloud screening and compositing
algorithms, time series of surface reflectances
were extended to cover earlier periods of the
Terra mission c) LAI and FPAR algorithm was
optimized to better simulate features of MODIS
surface reflectances, composting scheme was
improved to select best quality retrievals.
Collection 4 refinements were focused on
improvements of LAI and FPAR retrievals over
herbaceous vegetation. Optimization of the
algorithm for woody vegetation is more involved
(due to saturation of surface reflectances) and
will be implemented in Collection 5.Below we
review the impact of the above improvements on
LAI retrievals for each site. KONZ a)
Collection 3 anomaly of LAI overestimation over
grasses was resolved in Collection 4 b) The
proportion of best quality main algorithm
retrievals was increased. ARGO a) Collection 3
anomaly of LAI overestimation over croplands was
also resolved in Collection 4 b) The proportion
of best quality main algorithm retrievals was
increased. Some areas have fill values- they will
be eliminated in Collection 5. HARV a)
Optimization of the algorithm for woody
vegetation is scheduled for Collection 5 b)
Results of algorithm optimization over broadleaf
forests indicate substantial increase of main
algorithm coverage (saturation) and good
agreement of retrieved LAI with field
measurements.
Col. 3 Col. 4 BigFoot
Publication Site Vegetation Type Results
Privette et al., RSE, 83, 232-243, 2002 IGBP Kalahari Transect Semiarid woodlands savannas C1 MODIS LAI algorithm correctly accommodates structural and phenological variability in semiarid woodlands and savannas, and is accurate to within the uncertainty of the validation approach used
Huemmrich et al., IJRS, 2004 IGBP Kalahari transect, Mongu, Bolivia woodlands C3 MODIS and ground-measured LAI time series (2000-2002) corresponded well, while there was a significant bias between MODIS and ground-measured FPAR
Scholes et al, CCB, 10, 292-302, 2004 IGBP Kalahari transect Six savanna sites on deep, sandy soils There was good agreement between LAI observed in the field using a line ceptometer and the C3 LAI inferred by the MODIS sensor on the Terra satellite platform, 2 months later in the same season
Wang et al., RSE, 91, 114-127, 2004 Ruokolahti, Finland Needle leaf forest mixed with large and small lakes C4 MODIS LAI is accurate to within 0.5LAI, precision 48
Tan et al., JGR, 2004 (submitted) Alpilles, France Agricultural area C4 MODIS LAI is accurate to within 0.3LAI precision 20, uncertainty 25. Biome misclassification deteriorates the accuracy by factor of 2
Cohen et al., RSE, 88, 233-255, 2003 BOREAS NSA, Harvard Forests, MA Konza, KS, Bondville, IL Cropland prairie grassland boreal needle leaf temperate forests C3 MODIS-based LAI estimates were considerably higher than those based on ETM LAI. Samples of C4 LAI products were examined and found to consist of significantly improved LAI predictions for KONZ, and to some extend for ARGO
Fensholt et al., RSE, 91, 490-507, 2004 Senegal, Western Sudano-Sahelian zone grass savanna Seasonal dynamics of both in situ LAI and FPAR were captured well by MODIS LAI and FPAR. MODIS LAI is overestimated by approximately 2-15 and the overall level of FPAR is overestimated by 8-20.
Kang et al., RSE, 86(2) 232-242 Korea temperate mixed forests Minimal cross-validation errors between the predicted and MODIS-based timings of onset were found at a mean absolute error (MAE 3.0 days) and bias ( 1.6 days). This study demonstrates the utility of MODIS land products as tools for detecting spatial variability in phenology across climate gradients.
Pandya et al., Current Science, India, 85 (12) 1777-1782 Madhya Pradesh, India Agricultural areas The results indicated significant positive correlation between LAI derived from LISS-III data and MODIS data, with an overestimation in the MODIS product, root mean square error of 0.92 to 1.26 for the Bhopal site and 0.20 to 0.33 for the Indore site.
Fernandes et al., IGARSS 2002 BOREAS Needle leaf forests C3 MODIS LAI overestimates LANDSAT derived LAI by 33
Retrievals under cloudy conditions 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 MODIS cloud mask. LAI
values retrieved under cloudy condition are
spurious. Back-up
Grasses Collection 4, Main, Saturation, Back-Up,
Tile h10v05, 1200x1200 km
QC
QC
c
d
Collection 3
Collection 4
Col. 3 Col. 4 BigFoot
retrievals under cloudy conditions indicate bad
quality, but main algorithm retrievals may
indicate cloud cover overestimation by MODIS
cloud mask. The difference between LAI retrieved
under cloudy and clear condition is biome
dependant, but has biome independent maximum
during boreal summer time.

• As a part of our collaboration with climate
  modelers (Robert Dickinson group) we performed
  comparison of seasonal and spatial variations of
  LAI and FPAR from MODIS and Common Land Model.
  Collection 4 MODIS LAI is generally consistent
  with the CLM during the snow-free periods but may
  be underestimated in the presence of snow,
  especially for evergreen forests. On average, the
  model underestimates FPAR in the Southern
  Hemisphere and overestimates FPAR over most areas
  in the Northern Hemisphere compared to MODIS
  observations during all seasons except northern
  middle latitude summer.

QC
QC
Collection 5 QC Tile h12v04, 1200x1200 km
Col. 5 BigFoot
e
July 20-27, 2001

• Publications
• Shabanov et al. (2004). Optimization of the MODIS
  LAI and FPAR Algorithm Performance over Broadleaf
  Forests. IEEE Trans. Geosci. Remote Sens., in
  review.
• Huang et al. (2004). Evaluation of Collection 3
  MODIS LAI Product with Respect to Input Data
  Uncertainties Case Study for Grasses. J.
  Geophys. Research, (submitted).
• Tan et al. (2004). Analysis of Collections 3 and
  4 MODIS Broadleaf Crops LAI Products A Case
  Study of the Bondville Site. Photogrammetric
  Engineering and Remote Sensing, (submitted).

MODIS LAI in July
LAI difference (CLM-MODIS) in July

• Publications
• Yang et al. (2004). Analysis of Global MODIS Leaf
  Area Index and Fraction Absorbed PAR Time Series
  Data from February 2000 to December 2003. J.
  Geophysical Research (in review).
• Tian et al., (2004). Comparison of seasonal and
  spatial variations of leaf area index and
  fraction of absorbed photosynthetically active
  radiation from Moderate Resolution Imaging
  Spectroradiometer (MODIS) and Common Land Model.
  J. Geophys. Research, 109 D01103.

Broadleaf forests QC Collection 5, Tile
h12v04, 10x10 km
July 20-27, 2001