Fuzzy nearest-mean reclustering of SAR pixel features: Assessments on land use classification

This paper describes a non-parametric algorithm based on fuzzy-reasoning concepts and suitable for land use classification, either supervised or unsupervised, starting from pixel features derived from SAR observations. Pixel vectors composed by simple features calculated from the backscattering coefficient(s) in one or more bands and/or polarizations are iteratively clustered. At each iteration step, pixels in the scene are classified based on the minimum attained by a weighted Euclidean distance from the centroid representative of each cluster. Upgrade of centroids is iteratively obtained both from the previously obtained classification map and by thresholding a membership function of pixel vectors to each cluster. Such a function is inversely related to the weighted Euclidean distances from the centroid representative of each cluster. To yield the weighted distances from a pixel vector, its features are weighted by means of progressively refined coefficients, whose calculation still relies on the membership function through a least squares algorithm. Possible "a priori" knowledge coming from ground truth data may be used to initialize the procedure, but is not required. Experiments on MAC-91 NASA/JPL AIRS AR data on the Montespertoli test site show that a total of nine features derived from C-HV, L-HV, and P-HV observations are capable to discriminate seven agricultural cover classes with an overall pixel accuracy of 70%. with one tenth of the ground truth data used for training and the remaining nine tenths for testing the classification accuracy.