Plenary lecture 2: neural network models for pattern recognition in satellite and aerial imagery with environment and defense applications
Victor-Emil Neagoe · 2010
This lecture is an approach dedicated to the improvement and experimentation of several neural network pattern recognition models for satellite and aerial imagery. One considers the following neural network classifiers: Multilevel Perceptron (MLP), Radial Basis Function (RBF) neural net, supervised Self-Organizing Map (SOM), and the system of Concurrent Self-Organizing Maps (CSOM). CSOM was previously proposed by the author of this lecture; it is in fact a model of Concurrent Neural Classifiers (CNC), representing a collection of small neural networks, which use a global winner-takes-all strategy. Each neural module is trained to correctly classify the patterns of one class only and the number of modules equals the number M of classes. One considers the case of choosing the SOM (Self-Organized-Map) as a neural module. We built M training pattern sets and each neural module is trained with the pattern set characterized by the corresponding class label. We have implemented and evaluated the above mentioned neural classifier models for two kind of applications. First application investigates multispectral satellite image classification for environment monitoring. The implemented neural classifiers are evaluated using a LANDSAT 7 ETM+ image. One takes in consideration both the interband and also the intraband pixel correlation of the 7-band image. There is a subset containing labeled pixels, corresponding to several thematic categories: urban areas, agricultural fields, woods, water, bushes, meadows and barren fields. The best experimental result is obtained by CSOM model and it corresponds to the recognition rate of 99.11 %. Second application evaluates the considered neural network models for Automated Target Recognition (ATR) based on spot Synthetic Aperture Radar (SAR) imagery. One uses the MSTAR (Moving and Stationary Target Acquisition and Recognition) database for three categories of military vehicles: BMP2 (Tank), BTR 70 (Armored car) and T72 (tank). The best performance corresponds also to the CSOM model and it leads to the recognition rate of 95.81%.