Open AccessSEMI-SUPERVISED SPECTRAL-TEXTURE IMAGE CLASSIFICATION ALGORITHM
Author(s) -
S. A. Rylov
Publication year - 2019
Publication title -
interèkspo geo-sibirʹ
Language(s) - English
Resource type - Journals
ISSN - 2618-981X
DOI - 10.33764/2618-981x-2019-4-1-37-43
Subject(s) - artificial intelligence , pattern recognition (psychology) , multispectral image , sample (material) , computer science , segmentation , texture (cosmology) , co training , contextual image classification , image (mathematics) , computer vision , semi supervised learning , chemistry , chromatography
When classifying satellite images, training sample often turns out to be unrepresentative. This leads to low segmentation quality. In such conditions, it is advisable to use semi-supervised classification methods, which simultaneously utilize both training sample and unclassified data. At the same time, high resolution satellite images are characterized by high interclass heterogeneity of spectral characteristics, which demands to take spatial information into account. We propose a new semi-supervised classification algorithm for multispectral images, that utilizes both spectral and texture features. The use of the semi-supervised concept allows improving the classification quality when the amount of training sample is small. The results of experiments on model and satellite images confirming the effectiveness of the proposed algorithm are given.