1 of 5 CORSISCA : Classification Of Remotely Sensed Images - a Soft Computing Approach Aditya Saurabh, Raghu B.V., and Anupam Agrawal, Member, IEEE E-mail of contact author: asaurabh_00@iiita.ac.in Indian Institute of Information Technology, Deoghat, Jhalwa, Allahabad - 211 011. Abstract The classification of multispectral satellite images is a challenging problem and has a number of applications such as feature identification, change detection, etc. We apply modified Neural Network Algorithms: GA-BP (Genetic Algorithm as precursor to the Back Propagation) and Modular Artificial Neural Network (MNN) to classify the LISS-3 image of Allahabad area. We also classify the resolution merged image (LISS-3 with PAN) using the same algorithms. By using Genetic algorithm as a precursor to ANN, we increase the probability of reaching to the global minimum, thus reducing the problem of a stuck neural network in the local minimum. MNN models the human brain more closely to apply task decomposition to the satellite images as well. The output of the above techniques are generated and analyzed. 1. Introduction The ability of learning in artificial neural network (ANN) provides an interesting alternative to the conventional classification methods [1,8]. In remote-sensing data classification, neural network models exploit the following features. Neural network classifiers do not require any prior knowledge of the class statistical distribution in data sources since they are nonparametric classifiers. The neural network approach avoids the problem in statistical multi-source analysis of specifying how much influence each source should have on classification. This implies that the neural network approach becomes more preferable for multi-source remote-sensing data classification. A model of reasoning based on the human brain, Neural Networks aims to teach a machine through the techniques used by biological neural networks in learning and processing. There are various models through which human brains are represented, and based on their representation strategy, they can be differentiated. Neural Networks using steepest descent (greedy step) tries to solve an NP-Complete problem of error minimization by weight adjustment. Thus, eventually, Neural Network with Back Propagation remains an approximation. Thus ANN efficiency and reachability depend upon the choice of initial weights, which if chosen in the vicinity of local optimum will lead to a stuck Network. For a point chosen in the vicinity of global minimum, reachability is fast and certain in steepest descent. In contrast, the dependence of evolutionary algorithms, such as GA, on initial conditions is much less than the feed-forward (BP) neural network. In the trade-off, an evolutionary algorithm takes much longer to converge. Thus, a hybrid GA-BP where GA is used for initialization and BP is used for fine tuning is better than both of them alone in multi local-optimum search space such as satellite image data. Reverse engineering of brain has proved its modularity on different spatial scales, ranging from dendrites as cluster of synapses to functionally distinct areas. Such fault tolerant, competing-cooperating, scalable and extendible Modular Neural Network such as human brain motivates us to apply task decomposition and modularized learning for satellite images. Multi-spectral image classification using Modular Neural Network has been carried out. LISS-3 (Linear Image Self Scanning) is the multispectral image with a resolution of 23.5 m captured by IRS-1C/1D satellites. It covers the wavelength bands of green (0.52-0.59 µm), red (0.62- 0.68 µm) and NIR (Near Infrared, 0.77-0.86 µm). A fourth band of resolution 70m is also available in the LISS-3 image resampled to 23.5 m resolution, resulting in haziness, thus the use of fourth band is discouraged in classifiers. PAN (Panchromatic) is the higher spatial resolution image with resolution 5.8m covering the range of wavelengths 0.50-0.75 µm. The images are cut based on area of interest, geo-referenced, resolution- merged and classified using modified neural networks algorithms. The classified maps are compared and accuracy parameters are computed. The following subtasks were performed in the work:  Procurement of Data – Satellite Images: LISS-3 and PAN  Preparation of Data – Image to Map Registration – Resolution merge ¬ PCA