Image Simplification Using Kohonen Maps: Application to Satellite Data for Cloud Detection and Land Cover Mapping

Abstract

In this chapter, we present an application of Kohonen maps to two well-known problems in Earth Observation: cloud detection and land cover mapping. We based our methodology on a self-organised map (SOM) algorithm able to learn from data containing quantitative and qualitative variables, denoted MTM (Mixed Topological Map algorithm). After completion of basis radiometric information by additional variables and creation of appropriate databases for learning and labelling phases of the Kohonen map, we applied this algorithm to low spatial resolution imagery of the MODIS sensor (Moderate Resolution Imaging Spectroradiometer) on board the US Aqua satellite and to the high spatial resolution imagery of the HRG sensor (Haute Resolution Geometrique) on board the French SPOT 5 satellite (Satellite Pour l’Observation de la Terre). In the case of MODIS, we could compare the performance of the MTM algorithm used for cloud detection to the official MODIS cloud mask product, showing that the MTM algorithm significantly decreases indetermination and over-detection on bright surfaces (snow and deserts). In the case of HRG, we used the MTM algorithm for land cover mapping, but as no real-time reference mask of land cover is available, we validated our results by projection on Google Earth cartographic background, verifying the qualitative performance of our classification on relatively timely stable structures, like rivers, forests, and down-town areas. In both cases, i.e. cloud and land cover detection, we showed that using a Kohonen map significantly improves the operational implementation of corresponding processing chains, as it gives equivalent to better results than classical algorithms, with significantly easier development and application processes. Moreover, because these processes are shortened, and because the output information of Kohonen maps is radiometrically simplified, their use allows us the development of further solutions for change detection, data fusion, and image processing.