Diapositive 1

1
Monitoring land use and land cover changes in
oceanic and fragmented lanscapes with
reconstructed MODIS time series R. Lecerf, T.
Corpetti, L. Hubert-Moy and V. Dubreuil COSTEL
UMR CNRS 6554 LETG Multitemp-2005, 16-18 May 2005
Biloxi, Mississippi USA
COSTEL
2
INTRODUCTION
Focus Winter bare soils increase losses of
nutrients, pesticides or soils from agricultural
fields and damage water quality.
Monitoring winter bare soils at meso-scale with
medium resolution image time series such as
MODIS in farming areas located in fragmented
landscapes. Time series extracted from
original images are often corrupted and hence
not exploitable (atmospheric and geometric
distortions and others artifacts) Objective
? Reconstruction of high accurate NASA
EOS/MODIS time series based on selected and
preprocessed daily images. ?Experiment of
twofold cost function using Robust estimators
3
INTRODUCTION

• Data NASA EOS/MODIS time series
• High resolution images ? low revisit frequencies
  and small coverage
• Coarse resolution (1km) images ? spatial
  resolution too large to study fragmented
  landscapes
• Medium resolution images ? spatial resolution
  more adapted to fragmented landscapes and high
  revisit frequency
• - 16 days NDVI composite (MOD13Q1)
• - Daily surface reflectance images pre-processed
  by NASA (MOD09GQK)
• - Daily Level 1B images corrected by us
  (MOD02QKM)
• The study area Britanny
• - an intensive farming region
• - highly fragmented landscape
• - 350 km x 150 km large
• - oceanic climate with frequent winter cloud
  cover

4
SUMMARY

1. DATA AND PRE-PROCESSING STEPS
2. ROBUST SMOOTHING TEMPORAL DATA
3. RESULTS OBTAINED FROM SINGLE DATE DATA SETS
4. RESULTS OBTAINED FROM TIME SERIES DATA SETS
5. CONCLUSION

5
DATA AND PREPROCESSING STEPS MODIS 16 days NDVI
synthesis (MOD13Q1)
MODIS 16 days NDVI synthesis at 250m (Constraint
View Maximum Value Composite) ? Are used to
calculate monthly NDVI by three steps 1.
Masking of residual clouds with QA bands 2.
Selection of NDVI maximum values 3. Selection of
minimum viewing zenithal angles ? The MOD13Q1
product which is the nearest of the considered
month is selected ? Images are projected into a
french coordinate system
Monthly composites
CV-MVC
6
DATA AND PREPROCESSING STEPS MODIS daily images
selection

• MODIS daily surface reflectance images at 250m
  MOD09GQK
• MODIS daily level 1B products MOD02 also called
  MODIS L1B
• - Daily images are often completely cloudy
• - High viewing zenithal angles are degrading
  edges, reflectance values.
• ? MODIS daily images are selected based upon two
  criteria
• ? less than 50 cloud cover
• ? within a radius of 200km centered on Brittany
• Through the Eos Data Gateway
• Cloud cover is given for each MODIS tile and not
  only for a subsetted area.
• Viewing zenithal angles are not available for a
  precise location.

7
DATA AND PREPROCESSING STEPS MODIS daily images
pre-processing

• MOD09GQK ? MODIS daily data atmospherically and
  geometrically preprocessed by NASA (Daily surface
  reflectance images)
• Atmospheric corrections are applied using PGE 11
  algorithm
• Algorithm changed 11 times since the first image
  of the time series
• Corrections depend on other MODIS products like
  cloud mask
• Red and NIR bands at 250m are used to compute
  NDVI
• NDVI (NIR-Red)/(NIRRed)

8
DATA AND PREPROCESSING STEPS MODIS daily images
pre-processing
MOD02QKM ? Daily images without atmospheric and
geometric corrections - Atmospheric corrections
using MODIS land rapid response algorithm
(crefl) - Computing of NDVI with MODIS land rapid
response algorithm (ndvi_evi) - Geometric
corrections using MODIS Swath Reprojection Tool
(NASA EDCS) and geolocation file (MOD03). Note
on atmospheric corrections ? crefl algorithm is
a 5S like model.
9
DATA AND PREPROCESSING STEPS High resolution
images (SPOT, LANDSAT, ASTER)
High resolution images used for validation of
MODIS time series ? SPOT/HRVIR ?
LANDSAT/ETM ? ASTER All images are
atmospherically and geometrically corrected ?
Atmospheric corrections using 5S for SPOT and
LANDSAT Red and Near infrared bands are
aggregated using mean values to get NDVI values
for cells of 250m.
10
Conclusion on daily MODIS time series

• Daily image time series are not homogeneous
• ? Revisit frequency is not regular
• Data are still corrupted by clouds, cloud
  shadows, aerosol, angles variations

Yellow cases Images selected for year 2003
11
Robust smoothing temporal data General
reconstruction formulation
Reconstruction of time series using a twofold
cost function
Reconstructed time series y(t) Observed time
series x(t)
Favors smooth solutions Solution with the lowest
temporal derivative a orders the influence of the
smoothing function
12
Robust smoothing temporal data General
reconstruction formulation

• Function not adapted to high outliers data
• Clouds
• Cloud shadows
• Detector errors
• High variations are kept into the time series.

Next step integration of a robust function
13
Robust smoothing temporal data Robust
reconstruction formulation
Robust estimator Imposes different penalisations
for coherent and incoherent data For high
outliers attenuates the contribution of the
observed values
14
Robust smoothing temporal data Robust
reconstruction formulation
With a quadratic function, resulting time series
depend on high outliers With a robust estimator
the weight of high outliers is minimal and
particularly with a Geman/Mc Clure M-estimator
15
Robust smoothing temporal data Robust
reconstruction formulation
This function is not convex and leads to
minimization difficulties We can cope with this
problem by using semi-quadratic functions
16
Results FROM SINGLE DATE DATA SETS
Comparison of MODIS images with a Landsat ETM
image, March 2003
Monthly composite (MOD13Q1)
r 0.64
Landsat ETM
Daily image (MOD09GQK)
r 0.70
NDVI
Reference images
Daily image (MOD02QKM)
r 0.86
17
Results FROM SINGLE DATE DATA SETS

• ? Daily images pre-processed with MODIS land
  rapid response algorithm show better accuracy
  than MODIS surface reflectance products.
• ? Similar results are obtained with SPOT/HRVIR
  and ASTER images
• ? Then, NDVI time series have been reconstructed
  using these products

18
Results FROM SINGLE DATE DATA SETS
Before reconstruction
After reconstruction
r 0.77
r 0.86
March 2003
December2003
NDVI
Reconstructed images are more reliable than
surface reflectance images (MOD09GQK) distributed
through the EOS Data Gateway
19
Results FROM TIME SERIES DATA SETS
1- Meadows
2- Corn
3- Wheat
4- Forest
20
Experimental results FROM TIME SERIES DATA SETS

• Based on ground truth and meteorological data,
  NDVI profiles depend on
• land cover types
• mixture of land cover types
• precipitations and temperatures

21
Experimental results FROM TIME SERIES DATA SETS
Corn
Winter Wheat
NDVI profiles are coherent with canopy
development and crop phenology - Corn begins to
grow up in may and is senescent in September -
Winter wheat begins to grow up in February/March
and is harvested in July
22
Conclusion
? MODIS daily images are more reliable than
monthly composites to determine bare soils in
winter in fragmented landscapes. ? Daily images
pre-processed with MODIS land rapid response
algorithm show better accuracy than MODIS surface
reflectance products. ? The use of a robust
estimator to reconstruct NDVI time series appears
to be useful to minimize contribution of high
outliers. ? The smoothed NDVI time series
obtained are suitable to monitor vegetation
phenology in fragmented landscapes.
23
Acknowledgments
Thanks to Jacques Descloitres (MODIS Land Rapid
Response) to provide us the atmospheric
correction algorithm