Self−organizing maps of polarimetric SAR imagery

New−generation SAR missions will provide polarimetric data, that, besides other features, have the potential of improving the accuracy of land cover mapping. Processing of polarimetric data for classification purposes has been carried out by a variety of algorithms which span from Bayesian Maximum Likelihood to Fuzzy Logic to Support Vector Machines to Neural Networks. Target decomposition provides an alternative way to single out scattering mechanisms, hence to discriminate the observed surface types in an unsupervised fashion [Cloude and Pottier, An Entropy−based classification scheme for land applications of polarimetric SAR, IEEE TGARS, 1997]. Land cover segmentation by polarimetric SAR data has been extensively conducted in different sites, mainly located in relatively flat and piecewise homogeneous areas, while more limited results have been reported for undulating, heterogeneous and fragmented landscapes, where classification can become quite challenging. This paper reports on the pixel−by−pixel unsupervised classification of a hilly area in Central Italy by multi−angle, polarimetric SAR data at three frequencies. Our aim was to test self−organizing neural networks [Kohonen, Self−Organizing Maps , 2001] in a rather difficult situation, to discuss the results and assess the accuracy of this method when applied to complex landscapes. The performance of the H/A/−Wishart unsupervised procedure on the same data set is used as suitable benchmark. The test site in Central Italy, set for the MAC−Europe Campaign, was overflown in the summer of 1991 by the JPL/NASA AirSAR [Baronti et al, SAR polarimetric features of agricultural areas, I.J.R.S, 1995}. The area is dominated by hills surrounding the river Pesa valley. The main land uses are agriculture, forestry and urban. Agricultural plots include vineyards, olive groves and crops like wheat, sunflower, rapeseed, alfalfa, maize, sorghum and pasture, with a typical dimension of parcels around 4 hectares. The images were acquired on three different summer dates at P−, L− and C−band at about 20, 35 and 50 degrees incidence angles at the center of the swath. Spatial resolutions are 12 and 6.6 meters in azimuth and range respectively, the number of looks is 16 and the calibration accuracy is around 1 dB. The polarimetric data sets have been extracted from the database created by the ERA−ORA European project [http://eraora.disp.uniroma2.it/]. The polarimetric data relative to each frequency, angle and acquisition date have been processed by two−dimensional self organizing neural networks (SOM) in a hierarchical scheme and the classification results have been cast in terms of confusion matrices. The effects of the net topology, of the processing parameters (like learning rate, neighborhood function, number of training epochs) and clustering procedure on the attained accuracy are examined. The performance of the algorithm is also discussed with reference to the probability distribution function of the input measurement vector components. The results obtained from the present case study hint at the optimal radar configuration in terms of frequency and observation angle, also with respect to the crop development stage and to the features of the rugged terrain. The accuracy attained by the self−organizing map method compares favourably against that of the H/A/−Wishart procedure and allows a finer discrimination among crops. The observed discrepancies between the results yielded by the two methods are discussed also by considering the local prevailing scattering mechanisms.