RECONSTRUCTION OF GEOMETRIC MODELS OF OBJECTS FROMSATELLITE IMAGES BASED ON ARTIFICIAL NEURAL NETWORKS

Abstract

Segmentation of natural objects, such as soil and vegetation types, is possible using methods that operate on the spectral brightness of individual pixels, determined by phytocenological, humification and other mechanisms of their homeostasis. Segmentation of objects of artificial nature, such as houses, railways and highways, requires knowledge about a large number of spectral parameters, the structure of the object and its surroundings. Modern approaches to extracting three-dimensional information about man-made objects can be conditionally divided into traditional (based on classical processing methods) and neural network (data mining). The capabilities of the first methods have exhausted themselves under the conditions of the increasingly complex structure of the earth's surface and the need for automation of data processing. The second group of methods is being actively developed at the present time, its possibilities are unlimited and their quality is directly related to the volume of the training sample. The article discusses a method for constructing three-dimensional models of earth surface objects, based on a two-stage (integral and local analysis) and multiscale approach and assuming that the initial data is pre-processed by the method of increasing the resolution, which is characterized by a representative decrease in the resolution of the training sample images.