Abstract: Timely and accurate acquisition of rice planting information is of great significance for national food security. Optical remote sensing has served as a primary means for extracting information on rice planting areas. However, it is challenging to acquire optical remote sensing data in regions with frequent rainfall and cloud cover. In contrast, although microwave radar data offer subtle spectral information, this limitation can be effectively overcome by multi-temporal synthetic aperture radar (SAR) data. Additionally, the increasingly mature deep learning technology further provides a solid foundation for extracting information about rice planting areas. Building upon HRNet-a high-resolution deep learning network, this study proposed an improved network model termed AFFCA-HRNet, where AFFCA denotes the adaptive feature fusion and channel-spatial attention. The HRNet was improved as follows. First, a channel attention mechanism was constructed to perform feature operations on the components obtained through max and average pooling. Second, the improved channel attention module was embedded into the HRNet, forming a residual structure in conjunction with the original convolutional layer, thereby generating a feature extraction module with adaptive fusion. Based on the multi-temporal SAR dataset from Sentinel-1, the AFFCA-HRNet model was employed to extract information on rice planting areas in Wangcheng District, Changsha City, Hunan Province. The results demonstrate that by leveraging high-level semantic and spatial features, the AFFCA-HRNet model achieved an overall accuracy of 94.9% and a Kappa coefficient of 0.938 9 in information extraction, outperforming the widely utilized semantic segmentation network Deeplabv3. The proposed model, by integrating deep learning with multi-temporal microwave radar data, offers an effective solution for rice classification in regions exposed to frequent rainfall and cloud cover, holding great application potential.