REDUCTION OF THE DIMENSION OF THE FEATURE SYSTEM AT CLASSIFICATION OF HYPERSPECTRAL DATA OF THE EARTH REMOTE SENSING USING NEURAL NETWORKS
V. I. Kozik, E. S. Nejevenko
Institute of Automation and Electrometry, Siberian Branch, Russian Academy of Sciences, Novosibirsk, Russia
Keywords: Earth remote sensing, hyperspectral images, classification, neural networks, training, genetic algorithm, reduction of the features number
Abstract
The hyperspectral method for analyzing the Earth's surface is very effective in solving problems of classification of both objects located on it and the state of these objects (for example, agricultural crops). However, a full-scale hyperspectral analysis is a very expensive job, and the search for ways to reduce the cost of this procedure is quite understandable. The most logical way is to reduce the number of spectral components - classification features - by choosing (or forming from them) the most informative ones. In this paper, to implement it by using neural network technologies is proposed. By an example of processing a 200-channel hyperspectral image, it is shown that reducing the dimension of the feature space using these technologies makes it possible to achieve high-accuracy classification with the accuracy exceeding that obtained by other known methods.
|
