Development of a lightweight degradation-aware framework for SAR land-cover classification and pastoral ecosystem monitoring

This study evaluates the robustness of a lightweight deep learning framework for Synthetic Aperture Radar (SAR) land-cover classification under simulated airborne radar imaging degradations. A MobileNetV3-CBAM model was trained on 8,000 SAR image patches representing four land-cover classes: grassland, bareland, agriculture, and urban land. To simulate controlled variations in SAR image quality associated with airborne radar acquisition conditions, degradations related to speckle noise and spatial resolution were applied to the test data. A degradation-aware fine-tuning strategy was then introduced to improve model robustness under these challenging conditions. The baseline MobileNetV3-CBAM model achieved a macro-F1 score of 0.9409 on clean SAR images. After degradation-aware fine-tuning, the macro-F1 score increased to 0.9624 on clean images and showed substantial improvements under degraded conditions. In particular, the macro-F1 score improved from 0.2808 to 0.9413 under severe spatial resolution degradation and from 0.5069 to 0.9261 under severe speckle noise degradation. Furthermore, the Robustness Index (RI) increased from 0.5436 for the baseline model to 0.9854 after fine-tuning, indicating a marked enhancement in the model’s ability to maintain classification performance under degraded imaging conditions. A comparison with the baseline MobileNetV3 architecture was also conducted to examine whether the CBAM attention mechanism provided additional benefits under the evaluated degradation-aware conditions. The final MobileNetV3-CBAM model contains 1.27 million trainable parameters and has a model size of 4.96 MB, making it suitable for lightweight SAR image classification applications. These results demonstrate that degradation-aware fine-tuning substantially enhances the robustness of SAR land-cover classification and provides a practical framework for evaluating model sensitivity to controlled SAR image-quality degradations relevant to airborne pastoral ecosystem monitoring.