Back Propagation Neural Network for Classification of IRS-1D Satellite Images

The suitability of Back Propagation Neural Network (BPNN) for classification of remote sensing images is explored in this paper. An approach that consists of three steps to classify IRS-1D images is proposed. In the first step, features are extracted from the firstorder histogram measures. The next step is feature classification based on BPNN, and in the finally step the results are compared with the maximum likelihood classification (MLC) method. The statistical features in this paper are based on the first-order distribution measure: mean, standard-deviation, skew-ness, kurtosis, energy, and entropy. The network contains 3 layers. The extracted features are fed to input layer that consists of 18 neurons. The back propagation neural network was trained on six classes of the IRS-1D image base on known features and the trained network was used to classify the entire image. The method of this paper is tested for regions of Iran. The IRS-1D 8-bits bands 2, 3 and 4 of LISS-Ш sensor were fused with pan data to construct an image with 5.8 m spatial resolution. Experimental results show that BPNN method is more accurate than MLC and is more sensitive to training sites. 1.INTRODUCTION Based on biological theory of human brain, artificial neural networks (NN) are models that attempt to parallel and simulate the functionality and decision-making processes of the human brain. In general, a neural network is referred to as mathematical models of theorized mind and brain activity. Neural network features corresponding to the synapses, neuron, and axons of the brain are input weights. Processing Elements (PE) is the analogs to the human brain's biological neuron. A processing element has many input paths, analogous to brain dendrites. The information transferred along these paths is combined by one of a variety of mathematical functions. The result of these combined inputs is some level internal activity (I) for the receiving PE. The combined input contained within the PE is modified by the transfer function (f) before being passed to other connected PEs whose input paths are usually weighted(W) by the perceived synaptic strength of neural connections. Neural networks have been applied in many applications such as: automotive, aerospace, banking, medical, robotics, electronic, and transportation. An other application of NN is in remote sensing for classification of images. Many methods of classification have been already proposed. Bendiktsson et al.(1990) compared neural network and statistical approaches to classify multi-spectral data. They noted that conventional multivariate classification methods cannot be use in processing multi-source spatial data because of their often different distribution properties and measurement scales. Heermann and Khazenie(1992) compared NN with classical statistical techniques. They concluded that the back propagation network could be easily modified to accommodate more features or to include spatial and temporal information. Bischof et al (1992) included texture information in the NN process and concluded that neural network were able to integrate other sources of knowledge and use then in classification. Hepner et al (1990) compared the use of NN back propagation with maximum likelihood method for classification. The result showed that a single training per-class neural network classification was comparable to a four-training site per-class in conventional classification. Ritter and Hepner (1990) used feed-forward neural network model for classification. The results showed that the neural network had the ability to distinguish small linear pattern, which were apparent on the TM image. In this paper, artificial neural network for classification of IRS-1D data has been implemented. The back propagation algorithm is applied for classification of the images. A good method for training is an important problem in the classification of IRS data with neural network. TrainLM method has been applied on using back propagation neural networks algorithm on IRS images. Specifications of the training method are discussed in section 2, the study area and experimental results are explored in section 3 and the conclusions are presented in section 4. 2-THE MODEL OF NEURAL NETWORK First, this section presents the architecture of the back propagation algorithm. Back propagation was created by generalizing the Widrow-Hoff learning rule to multiple layer network and non linear differentiable transfer function. Input vectors and corresponding target vectors are used to train a network until it can approximate a function, associate input vectors with specific output vectors, or classify input vectors in an appropriate way as defined in this study. Networks with biases, a sigmoid layer and a linear output layer are capable of approximating any function with a finite number of discontinuities. The back propagation algorithm consists of two paths; forward path and backward path. Forward path contain creating a feed forward network, initializing weight, simulation and training the network. The network weights and biases are updated in backward path. (Rumelhart,1986) A single layer network with 4 inputs is shown in figure 1. Figure 1. Single layer network Feed forward networks often have one or more hidden layers of sigmoid neurons followed by output layer of linear neurons. Multiple layers of neurons with non linear transfer functions allow the network to learn non linear and linear relationships between input and output vectors. The linear output layer lets the network produce values outside the range -1 to +1 (Figure 2).