競合型ニューラルネットワークによる土地被覆分類法

Abstract

In this paper, we propose a method to apply a competitive neural network to land cover mapping of remote sensing data. This neural network is trained by the Learning Vector Quantization (LVQ) method. For the neural network, several weight vectors of neurons in the competitive layer represent a category. We employ LVQ1 and OLVQ1 in learning algorithms of LVQ. OLVQ1 is introduced in order to obtain high speed convergence compared with LVQ1. The three types of pattern distance functions and number of neurons are considered to find an optimal LVQ method. To evaluate the classification accuracy, the LVQ1 and OLVQ1, Maximum Likelihood Method, Back Propagation Method (3 layered neural network) and Nearest Neighbor method are considered. We classify LANDSAT TM data using TR1 or TR2 type training data where TR1 includes the same number of training data for each category, and TR2 is produced by picking up pixels on a grid in a certain processing area. As a result of this experiment, the OLVQ1 using the Mahalanobis' generalized distance and TR2 outperform the other methods with respect to overall accuracy despite a very small number of neurons, for example 24. The LVQ1 and OLVQ1 using TR1 produce better classification results regarding with average accuracy compared to the other methods. This method produces excellent classification images which are more realistic and noiseless compared with the conventional methods.