Validation of the use of multiple linear regression as a tool for unmixing coarse spatial resolution images

Abstract

When estimating agricultural production, we are faced with a problem of scale, be they scales of observation, modeling, or representation. In the context of the farming landscape of Europe, the relationships between scales and modeling take on special importance as the aim is to extract information on a smaller scale than the scale of observation, the latter necessarily being large relative to the average size of fields in order to obtain the required time resolution. This article presents a method for unmixing coarse resolution signals (of the NOAA-AVHRR type) through the use of multiple linear regression. This allows the signal for each mixed coarse resolution pixel to be broken down thanks to a knowledge of land use and a linear mixture model. It is then possible to calculate the individual radiometric contribution of each constituent of a mixed pixel. The work presented here is an assessment of the performance of this method using SPOT-HRV data, degraded to simulate the coarse resolution and as normally provided to produce the land use information. Simulation with SPOT-HRV data also enables various parameters to be calculated, from which it is possible to verify the interest and reliability of multiple linear regression as a method for spectral unmixing.