Optimum Learning Rate in Back-Propagation Neural Network for Classi cation of Satellite Images (IRS-1D)

Abstract

Remote sensing data are essentially used for land cover and vegetation classi cation. However, classes of interest are often imperfectly separable in the feature space provided by the spectral data. Application of Neural Networks (NN) to the classi cation of satellite images is increasingly emerging. Without any assumption about the probabilistic model to be made, the networks are capable of forming highly non-linear decision boundaries in the feature space. Training has an important role in the NN. There are several algorithms for training and the Variable Learning Rate (VLR) is one of the fastest. In this paper, a network that focuses on the determination of an optimum learning rate is proposed for the classi cation of satellite images. Di erent networks with the same conditions are used for this and the results showed that a network with one hidden layer with 20 neurons is suitable for the classi cation of IRS-1D satellite images. An optimum learning rate between the ranges of 0.001-0.006 was determined for training the VLR algorithm. This range can be used for training algorithms in which the learning rate is constant.