Deep information-extreme machine learning for autonomous UAV based on decursive data structure for semantic segmentation of digital image of a region

• Published in: Radìoelektronnì ì komp'ûternì sistemi (Online) Vol. 2025; no. 3; pp. 122 - 136
• Main Authors: Cheranovskyi, Valerii, Myronenko, Mykyta, Kovalevskyi, Serhii, Kraskovskyi, Roman, Otroshchenko, Mykhailo
• Format: Journal Article
• Language: English
• Published: National Aerospace University «Kharkiv Aviation Institute 10.09.2025
• Subjects: autonomous uav; decursive binary tree; digital image of the region; information criterion; information-extreme machine learning; optimization
• ISSN: 1814-4225; 2663-2012; 2663-2012
• DOI: 10.32620/reks.2025.3.09

The subject of the research is functional categorical models of deep information-extreme machine learning based on linear and hierarchical data structures, methods for optimizing machine learning parameters based on information criteria and constructing a decursive binary data tree for a given alphabet of recognition classes. The aim of the research is to improve the accuracy of machine learning for an autonomous UAV for semantic segmentation of a digital image of a region obtained via an optoelectronic observation channel. This goal is achieved by developing a method of deep information-extreme machine learning for an on-board recognition system of an autonomous UAV using a decursive binary data structure. A new method of deep information-extreme machine learning for autonomous UAVs has been developed, based on a hierarchical data structure in the form of a decursive binary tree. The novelty of the method lies in the maximization of the average interclass code distance within a given dimensionality of the Hamming feature space by optimizing the selection level of coordinates of statistically averaged binary realizations of the recognition classes. At the same time, the level of depth of information-extreme machine learning according to the principle of deferred decisions is determined by the number of parameters of the system's functioning that are optimized according to the information criterion. This approach, unlike neural-like structures, provides flexibility for the onboard recognition system during retraining in the event of an expansion of the recognition class alphabet. The Kullback-Leibler information measure modified by the authors serves as a criterion for optimizing machine learning parameters. In addition, the proposed method involves the transformation of the input training matrix into a working binary matrix specified in the Hamming space, which in the process of machine learning adapts to its maximum accuracy. Results: Based on the results of deep information-extreme machine learning, error-free decision rules based on the training matrix were constructed within the framework of a geometric approach. It is shown that the accuracy of the deep information-extreme machine learning is affected by the sequence of optimization of the parameters of the recognition system. The results of functional testing and cross-validation have confirmed the high accuracy of information-extreme machine learning for an autonomous UAV, as demonstrated by semantic segmentation of a digital image of a region. Conclusions: For the first time, a method of deep information-extreme machine learning based on a hierarchical data structure in the form of a decursive binary tree has been developed, which, unlike the known ones, additionally optimizes the level of selection for coordinates of binary averaged vectors of recognition features.