Characterization Of Spatial Heterogeneity and Structure at Landscape Scale

The monitoring of land surface dynamic processes at global scale, such as primary production, carbon and water fluxes, requires high temporal frequency remote sensing observations. Because of technological constraints, the sensors are characterized by coarse spatial resolution, i.e. a resolution from few hundred meters (MERIS/ENVISAT, MODIS/TERRA) up to one or few kilometres (VEGETATION/SPOT, SEVIRI/MSG). However, the scenes observed at this range of scales, present spatial heterogeneity which may have a great influence on land surface characteristic estimation from remotely sensed data. Therefore the characterisation of spatial heterogeneity is an important concern to scale non linear land surface processes. The aim of this study is to discuss a geostatistical approach based on two complementary tools to characterize spatial structure of remote sensing data at the landscape scale. The high spatial resolution NDVI (vegetation index) of SPOT/HRV images (20m resolution) is used to characterize the ground spatial structure of different landscapes. These NDVI images are then aggregated in order to describe the evolution of their structure with the spatial resolution. A classical method consists in describing the image spatial heterogeneity by a geostatistic tool: the variogram. The interest of the variogram is that it jointly allows to model the spatial distribution of a scene as well as to quantify the spatial heterogeneity as a function of the spatial resolution. A typology of spatial heterogeneity is derived from the variogram model parameters computed over several types of landscapes. To account for the availability of multiple wavebands, a multivariate description of the spatial heterogeneity could also be proposed. A first limit of the variogram approach is the assumption of spatial stationarity, necessary for modelling the variogram. Spatial stationarity can be checked by:

Dividing the image into local windows and adjusting the corresponding variogram model parameters for a range of window size.

Computing the variogram on increasing size of the same scene.

A second limit of the variogram approach is that different models of spatial random fields can share the same variogram function. This is for example the case of the Gaussian random field with an exponential variogram and the mosaic model with Poisson random polyhedra. We show that these two models and their linear mixture are undistinguishable if the histogram and the variogram are the only tools used for characterizing the heterogeneity. In this work we propose to use the first order variogram to discriminate between these two models. Moreover, we show that it is possible to model a wide range of landscapes as a mixture of these models and to estimate their parameters and the proportion of the mixture. This new way of characterization of landscape spatial structure and heterogeneity is discussed with possible application to land surface characteristic estimation from coarse resolution observations. Key Words: remote sensing, spatial heterogeneity, landscape, scaling, spatial resolution, variogram, spatial random field simulation, non linear process