ГІБРИДНА СИСТЕМА РОЗПІЗНАВАННЯ СУПУТНИКОВИХ ЗОБРАЖЕНЬ З НЕЙРОМЕРЕЖЕВИХ ЕКСТРАКТОРОМ ТА ІНФОРМАЦІЙНО ЕКСТРЕМАЛЬНИМ КЛАСИФІКАТОРОМ

Context. The study solves the relevant task of developing an interpretable and adaptive recognition system for semantic segmentation of satellite imagery by integrating neural network feature extractors with an information-extreme classifier.Objective. To improve the accuracy of satellite land cover classification by developing a hybrid machine learning model that combines a deep convolutional neural network for extracting informative features with an information-extreme classifier, enabling the construction of highly reliable decision rules even in the presence of overlapping recognition classes in the feature space.Method. A hybrid model is proposed that combines efficient spatial feature extraction using a convolutional neural network (CNN) with an information-extreme intelligent (IEI) technology for data analysis, based on maximizing the information capacity of the recognition system during machine learning. For feature aggregation, GlobalAveragePooling is applied instead of the classical Flatten operation. Additionally, regularization techniques such as weight decay and cyclical learning rate scheduling are implemented. The optimization of IEI model parameters is carried out using a modified Kullback information criterion, interpreted as a measure of recognition class diversity.Results. The developed model achieves high classification accuracy (95%) on the test set and demonstrates stable performance, with improved efficiency of the neural feature extractor due to a reduced number of training epochs enabled by regularization techniques. As a result of the information-extreme machine learning process, the optimal geometric parameters of the hyperspherical recognition class containers were determined, allowing the construction of highly reliable decision rules even under conditions of recognition class overlap in the feature space.Conclusions. The proposed hybrid model enables the construction of highly reliable decision rules through information-extreme machine learning, even in cases of a priori fuzzy partitioning of recognition classes in the feature space, based on the input training matrix formed during feature extraction.