A Detailed Comparison of Neuro-Fuzzy Estimation of Sub-pixel Land-Cover Composition from Remotely Sensed Data

The paper reports on the experimental comparison of two neuro-fuzzy classification schemes in the estimation of sub-pixel land cover composition in remotely sensed images. The two neuro-fuzzy classifiers, chosen from the literature and representing quite different combinations of fuzzy set-theoretic principles with neural network learning from data, are: the Fuzzy Multilayer Perceptron (FMLP) proposed by Pal and Mitra; and a Two-Stage Hybrid (TSH) learning scheme whose unsupervised first stage consists of the Fully Self-Organizing Simplified Adaptive Resonance Theory clustering network. FMLP and TSH classifiers are also compared with the traditional MLP algorithm to assess whether neuro-fuzzy techniques can be considered alternatives to conventional neural models in pixel unmixing. Classification performance are compared on a standard set of synthetic images, named CLASSITEST, consisting of pure and mixed pixels featuring known geometry and radiometry (i.e., known pixel location and proportions of land cover). Accuracy results are assessed by means of specific evaluation tools designed to extend conventional descriptive and analytical statistical estimators to the case of multi-membership in classes.