Analysis of Modern Methods for Recognizing Small Aerial Objects Based on Machine Learning

In recent years, the use of small unmanned aerial vehicles (UAVs) has increased, necessitating their accurate recognition against objects with similar radar characteristics, primarily birds. Despite the development of machine learning methods, the problem of reliably extracting UAV features remains relevant. The aim of this paper is to comparatively analyze existing domestic and international approaches to recognizing small aerial objects based on machine learning algorithms. This paper presents the results of evaluating the effectiveness of traditional methods (support vector machines, random forests), deep neural networks, and transformer architectures. The potential for extracting informative features from micro-Doppler time-frequency spectrograms obtained in various frequency ranges is explored. Convolutional neural networks and transformers are found to achieve recognition accuracy of up to 97 % with low noise levels; however, their use is associated with significant computational costs and the need for large volumes of training data. The article highlights the potential of hybrid neural network architectures, which integrate the advantages of various methods, to improve the accuracy and robustness of analysis. The results have practical implications for developing effective UAV recognition algorithms in real- time environments with limited computing resources. The findings expand existing understanding of the potential and limitations of modern machine learning algorithms in radar systems.