Abstract: ‍ ‍Land cover classification is an important task for remote sensing image interpretation that is employed for detailed terrain and landform analysis and plays a critical role in urban planning， disaster monitoring， and other fields. Synthetic aperture radar （SAR） images are widely used in land cover classification owing to their all-day， all-weather capability， high resolution， and large coverage area. However， SAR images also have limitations： they contain numerous shadows and significant noise due to their unique imaging mechanism. High-resolution SAR images are characterized by complex land cover distributions， large scale variations among targets， and lack of color information. This can easily cause boundary blurring and category confusion. In this paper， we propose a refined land cover classification method guided by coarse-grained category information. This method leverages multi-scale semantic features and coarse category information guidance to enhance the effective semantic feature extraction ability， category discrimination ability， and segmentation performance of a model. To address the issue of numerous shadows and significant noise in SAR images， we propose to employ a residual neural network （ResNet50） for multi-scale feature extraction. The extracted features are processed through a skip-connected decoder combined with an atrous spatial pyramid pooling （ASPP） module， which enhances context modeling capabilities， further preserves edge and structural information， and improves robustness against noise interference. To address the problem of unclear boundaries and category confusion in SAR images， we designed a coarse-grained category guidance module. This module generates dynamic semantic prototypes using class and semantic space features and then weights semantic features to enhance inter-category discrimination. Experiments on the WHU-OPT-SAR dataset show that our algorithm exhibits better semantic structure discrimination in complex scenes while preserving edge clarity. It particularly boosts segmentation accuracy for easily confused categories such as roads and waters.