Feature Extraction From Multitemporal SAR Images Using Selforganizing Map Clustering and Object-Based Image Analysis

We introduce a new architecture for feature extraction from multitemporal synthetic aperture radar (SAR) data. Its the purpose is to combine classic SAR processing and geographical object-based image analysis to provide a robust unsupervised tool for information extraction from time series images. The architecture takes advantage from the characteristics of the recently introduced RGB products of the Level-1<inline-formula><tex-math notation="LaTeX"> $\alpha$</tex-math></inline-formula> and Level-1<inline-formula><tex-math notation="LaTeX">$\beta$</tex-math> </inline-formula> families, and employs self-organizing map clustering and object-based image analysis. In particular, the input products are clustered using color homogeneity and automatically enriched with a semantic attribute referring to clusters’ color, providing a preclassification mask. Then, in the frame of an application-oriented object-based image analysis, opportune layers measuring scattering and geometric properties of candidate objects are evaluated, and an appropriate rule-set is implemented in a fuzzy system to extract the feature of interest. The obtained results have been compared with those given by existing techniques and turned out to provide high degree of accuracy and negligible false alarms. The discussion is supported by an example concerning small reservoir mapping in semiarid environment.