PolSOM and TexSOM in polarimetric SAR classification

The polarimetric observables in a SAR image possess an intrinsic physical information, what makes polarimetric data fit to unsupervised classification, without need of a-priori information on the scene. Indeed, in natural targets, like vegetation, surface, volume and sometimes double-bounce scattering mechanisms are mixed, while backscattering from man-made targets can be usually attributed to dihedrons, trihedrons and bare surfaces. In many cases a radar resolution cell hosts more than one mechanism, although an average or dominant scattering mechanism can be identified for the purposes of classification. Following Chandrasekhar's pioneering target decomposition and the generalized and systematic theory by Huynen, a number of approaches to the interpretation of the scattering processes and to the identification of scatterers have appeared in the open literature. Target decomposition theory laid down the basis for the classification of radar images. In particular, the formalism worked out by Cloude, led to the introduction of an unsupervised classification scheme, further augmented and improved by subsequent contributions, also connecting the fuzzy logic theory with Wishart's statistical approach and electromagnetic modeling. Neural Network Algorithms (NNA) have been used in multispectral images classification and for change maps, but their application to polarimetric SAR image classification is more limited. In supervised schemes, the NNA were trained by Huynens parameters, or by the polarimetric coherence matrix [T], H and alpha. Unsupervised Neural Net classifiers, based on Self-Organizing Maps (SOM), have exploited MÏ ller matrix directly, polarization signatures, or parameters derived from decomposition, like Huynen's, or Freeman's. In this Thesis two novel unsupervised classification algorithms, named PolSOM and TexSOM, for polarimetric data are proposed. Both algorithms are SOM-based and have been tested on complex Italian landscapes, where classification can become quite challenging and a limited use of polarimetric data has been reported for undulating, heterogeneous and fragmented scenarios. AIRSAR fully polarimetric data from MAC-Europe Campaign and RADARSAT-2 data acquired for a SOAR project (SOAR-1488) have been classified and confusion matrices have been computed from ground truth maps. PolSOM and TexSOM performances have been compared with each other and with consolidated and commonly used classification method, to assess their potential. The Neural Network algorithms have been carefully designed based on an in-depth analysis of their operation and, for the first time at the author's knowledge, both object-based and pixel-based information are jointly used in Radar polarimetric image analysis. The proposed classification algorithms are proving to be fairly versatile and not strictly confined to polarimetric images, like the other considered algorithms.