A soft approach to ERA algorithm for hyperspectral image classification

This work presents a novel supervised algorithm to solve the problem of crop classification using Hymap hyperspectral images acquired under controlled conditions in the Barrax area (Spain). We considered a multilayer perceptron (MLP) trained with the expanded range approximation (ERA) algorithm where the training set was varied in order to avoid local minima on the function error, which is an important drawback of the classical backpropagation (BP) algorithm. Soft modifications to ERA are proposed in order to improve its performance and robustness by tailoring smooth homotopy functions of the training set expansion. Our proposal is benchmarked with MLP trained by the BP and ERA algorithms and with other classical techniques, such as the radial basis function (RBF) network and the maximum-likelihood Gaussian classifier (MLGC).