Target Detection and Classification from SAR Images for Military Applications

Synthetic Aperture Radar (SAR) imagery is an essential tool for contemporary military operations, enabling day-and-night, all-weather surveillance and reconnaissance. Despite its importance, however, target detection and classification are still difficult because of noise, variability in the targets, and background clutter. To overcome these difficulties, this research proposes a hybrid approach that combines denoising methods with deep learning-based object detection to improve SAR image analysis. Different filtering methods such as Mean, Median, Frost, Kuan, GMAP, Non-Local Means (NLM), and Denoising Autoencoder (DAE) were evaluated. The results show that the best noise removal performance was produced by DAE with an MSE of 46.00, SSIM of 0.88, and PSNR of 24.48 dB, proving its efficiency in maintaining significant image details and removing noise efficiently. The preprocessed images were applied to YOLOv5 and YOLOv8 models for object detection. From the results, YOLOv8 performed better than YOLOv5, with 89.5% Precision, 87.8% Recall, 88.7% Accuracy, and F1 Score of 88.64, versus YOLOv5's 85.7% Precision, 83.9% Recall, 84.84% Accuracy, and F1 Score of 84.79. In addition, CCE loss was smaller in YOLOv8 (0.092) compared to YOLOv5 (0.136), validating its better detection performance, particularly for camouflaged and low-contrast targets of the military.