METHOD OF FORMING CLASSES OF SPATIALLY DISTRIBUTED OBJECTS FOR THE PURPOSE OF THEIR CLASSIFICATION AND RECOGNITION ON HIGH-RESOLUTION MULTISPECTRAL SPACE IMAGES

Abstract

The paper discusses methods for automated classification of vegetation types and soil cover based on high-resolution aerospace data and geoinformation technologies. The study was conducted using IKONOS satellite imagery and includes the formation of training and test samples for 12 classes of vegetation and soils characteristic of the region under study. A statistical analysis of the spectral characteristics of clusters and an assessment of their representativeness and separability were performed. It was shown that the uneven distribution of training examples and the overlap of spectral features of individual classes reduce the stability of classification. To improve recognition quality, an approach based on combining spectrally similar classes and forming alternative classification schemes was proposed. The effectiveness of the method is evaluated using a maximum likelihood statistical classifier and a multilayer perceptron neural network classifier. The results confirm the feasibility of optimizing the structure of the classification scheme and the composition of training samples when solving aerospace environmental monitoring problems.