SVM spatio-temporal classification of HR satellite image time series using graph based kernel

Satellite Image Time Series (SITS) are a very useful source of information for geoscientists especially for land cover monitoring. In this paper a new multi-temporal classification approach for High Resolution (HR) SITS is proposed. It is mainly two stages original approach using two different kernels based SVM algorithms. The first step of this approach consists in applying multiband RBF kernel based SVM classification on individual images. Then, for each cartographic region of the first classified image, a graph characterizing its temporal evolution is built using texture features and radiometry for graph labeling. In the second stage, a graph kernel based SVM algorithm is used to analyze and classify the temporal behaviors of these regions that are modeled by different graphs aspects. The resulted temporal map discern between cartographic regions behaviors (stable, periodic, growing, etc.), which is very beneficial in many applications fields. The experimental results have been conducted on synthesized and real data proving the accuracy of the proposed approach.