Satellite image time series classification and analysis using an adapted graph labeling

Temporal sequences of images called Satellite Image Time Series (SITS) afford a large amount of information compared to individual images in the context of temporal behavior of land cover components. Besides, graph represents a powerful tool for modeling such structured data. It offers the possibility to model the spatio-temporal relationship in a simple way for further analysis. In this paper, an adapted graph labeling is used to encode SITS versatile information. This labeling extends a multitemporal classification approach for Very High Resolution (VHR) SITS to use several feature types instead of a single value labeling. The new proposed multitemporal classification method takes advantage from an original graph-based SVM classification. Therefore, graph kernel designed for graphs with simple labels is generalized for complex graph structures. The experimental results have been conducted on synthesized and real data proving the accuracy of the proposed approach.