Noise-Resistant Land Cover Classification of SAR Images Based on Automatic Sample Extraction

Synthetic Aperture Radar (SAR) can provide stable multi-temporal data for land cover classification (LCC). Nonetheless, existing SAR sample extraction methods are costly, time-consuming, and lack effective temporal-spatial contextual feature representations, leading to suboptimal performance of deep learning algorithms in SAR LCC. To address these issues, this paper proposes a Speckle-robust LCC method for SAR images based on automatic sample extraction. Spatial consistency analysis and morphological processing are performed using multi-temporal historical land cover data, followed by the design of the confidence index (CI) in combination with SAR temporal characteristics for automatic sample extraction. Subsequently, a classification model combining weakly supervised learning and feature fusion is developed to enhance the model's noise resistance and temporal feature extraction capability. Experimental results in Kaifeng City and Yueyang City demonstrate the effectiveness of our method, achieving an Overall Accuracy (OA) exceeding 90% and providing an efficient, automated solution for SAR image LCC.