Anomaly Detection in SAR Images Using a Hybrid Self-Supervised Approach

Synthetic Aperture Radar (SAR) imagery, due to its capability for imaging under diverse weather and illumination conditions, has become one of the key tools for environmental monitoring, change detection, and both military and civilian applications. However, the presence of speckle noise and statistical heterogeneity of the background poses significant challenges to the anomaly detection process. In this study, a simple and computationally efficient approach based on noise reduction using a median filter, adaptive thresholding, and anomaly map generation is proposed. The method calculates statistical indices (mean and standard deviation) and defines a sensitivity threshold to effectively distinguish anomalous regions from complex backgrounds. For evaluation, the TenGeoP-SARwv dataset, containing more than 37,000 Sentinel-1 SAR images, was utilized. Experiments conducted on five sample images showed that the mean intensity (μ) ranged from 13,763 to 25,487, the standard deviation (σ) from 2,329 to 3,621, and the adaptive threshold (τ) from 18,749 to 32,730. These values demonstrate the adaptability of the proposed method to diverse statistical conditions. The final results indicate that the percentage of detected anomalous regions across the images varied between 2.44% and 3.25%. Notably, Image 4, which exhibited the lowest mean intensity, produced the highest anomaly percentage (3.25%), reflecting the algorithm’s capability to reveal unusual patterns even in low-backscatter data. Conversely, higher-intensity images, such as Image 3, showed lower anomaly ratios, highlighting the robustness of the method varying conditions. Visual analysis of the anomaly maps and histogram plots confirmed the quantitative findings, showing that the proposed approach can accurately highlight specific regions without generating widespread false detections. These characteristics make it an efficient tool for applications such as environmental monitoring, oil spill detection, natural resource management, and surveillance systems. Furthermore, the proposed approach can serve as a foundation for developing more advanced machine learning and deep learning-based methods in the future.