1. Estimation of surface characteristics over heterogeneous landscapes from medium resolution sensors. F. Baret 1, S. Garrigues 1, D. Allard 2, R. Faivre 3, L. Duval 1 1 INRA-CSE, Agroparc, 84914 Avignon, France 2 INRA-Biométrie, Agroparc, 84914, Avignon, France 3 INRA-BIA, Centre de Toulouse, 31326, Castanet Tolosan, France

2. Introduction Regional scale estimation of canopy characteristics required for: –Process model validation/forcing/assimilation Dynamic Global Vegetation Models Crop production models Meteorological models Hydrological models … Availability of medium resolution observations 250m; 1km ; 7km VEGETATION, MERIS, MODIS, SEAWIFS, POLDER, AVHRR… Heterogeneity of land surfaces: impact on accuracy of canopy characteristcs retrieval?

3. The heterogeneity: evidence 1 km Land surfaces are generally very heterogeneous at a range of scales SPOT image Classification (6 classes) YW GW SR BS MA GW

4. Heterogeneity and scaling NDVI image at 4 spatial resolution 20m 200m 500m 1000m Heterogeneity within pixels Depend on resolution

5. Heterogeneity: characterization LAI Nb. pixels Statistical distribution Semi variogram Dispersion variance

6. Non linearity: evidence K lai is the extinction coefficient: 0.638449 NDVI  of NDVI value for infinite LAI : 0.804545 NDVI s : bare soil NDVI value:0.240425 00.10.20.30.40.50.60.70.8 0 1 2 3 4 5 6 NDVI Value LAI Value Biase: e LAI true LAI est NDVI Systematic under-estimation of LAI values

7. =200m =500m =1000m (Heterogenity)  (non linearity): effect Biase in LAI estimation increases with spatial resolution (heterogeneity)

8. Objectives Analyse the effect of heterogeneity on canopy characteristics estimation from medium resolution sensors Propose methods to correct for this heterogeneity effect: Based on prior knowledge derived at higher spatial resolution »Variogram: bias correction »Land cover: desagregation

9. Bias correction using variogram True LAI Estimated LAI Theoretical expression of the bias expectancy LAI biase Local Dispersion variance for resolutions

10. Results =200m =500m =1000m For higher spatial resolution, problems of local estimation of the dispersion variance

11. using desagregation N classes of vegetation Each class is characterized by mean and variance of its properties LAI Red reflectance NIR reflectance NDVI NDVI values for class i The problem is split in two sub-problems: 1.Estimation of intra-pixel NDVI values for each class i 2.Agregation at the resolution

12. Desagregating: background Reflectance of pixel i Abondance of class k in pixel i Reflectance of class k in pixel i Error for pixel i Production of individual estimates using the BLUP method Estimation of using Faivre’s method

13. = 200m, 6 classes = 200m, 4 classes Desagregating: results = 500m, 6 classes = 500m, 4 classes = 1000m, 6 classes = 1000m, 4 classes Singular values 500 m is optimal for these conditions 1000 m has not enough inter-pixel variability 200 m Problems due to spatial dependance not accounted for Possible regularization by using ancillary values for  i 2 (measurement error) Spectral consistency to be improved

14. Results on LAI per class per pixel ( LAI i k ) = 200m = 500m = 1000m

15. = 200m Results on LAI per pixel ( LAI i ) = 500m = 1000m

16. Desagregation: Conclusion Good general performances Problems when little inter-pixel variability –adaptation of the spatial resolution of observations –Medium resolution better adapted? Possible improvements –Accounting for spatial correlation –Accounting for spectral consistency: Ensemble retrieval –Regularization of error variance (  ²  ) Taking into account the measurement uncertainties –Distribution of the variables

17. Comparison of approaches Desagregation generally better performing when desagregation OK Problems with bias correction when stationarity of dispersion variance

18. Conclusion Exercise to be repeated over a range of situations –selected within the VALERI database Extension to more sophisticated retrieval methods –Radiative transfer model inversion Need higher spatial resolution for: –Desagregation: land cover maps derived from other sources of information well coregistred with satellite observations Possibly derived from the temporal time course of medium resolution sensors Variability within one land cover class? –Bias correction: associated simple imager (red-nir) At high spatial resolution Interest of desagregation techniques with regards to: The relative stability of the land cover map The potential applications (values per land cover class)