Leveraging machine learning and deep learning for SAR image classification

The study conducted a comprehensive analysis of contemporary machine learning and deep learning methods for the classification of synthetic aperture radar (SAR) images. The primary objective was to identify architectures and approaches that ensure high classification accuracy while optimising computational efficiency. Particular emphasis was placed on addressing key challenges, including speckle noise, geometric distortions, and the limited availability of labelled data. The research methodology involved a systematic review of the scientific literature from 2015 to 2024 and an analysis of the polarisation characteristics of SAR images using the Copernicus Browser platform. The effectiveness of traditional machine learning methods, such as Support Vector Machines and Random Forest, was evaluated alongside modern deep learning architectures, including ResNet, U-Net, and Vision Transformer. Special attention was given to the impact of adaptive speckle noise filtering using the Lee filter with varying window sizes (3 × 3, 5 × 5, and 7 × 7) on classification performance. The results demonstrated that deep neural networks outperform traditional methods due to their ability to automatical