Research on Rice Identification Methods in Mountainous Regions

Rice is a vital staple crop in China, and accurate identification of rice planting areas is essential for agricultural management and food security. However, mountainous regions pose significant challenges due to cloud contamination, limited use of multi-dimensional features, and low classification accuracy. This study proposed a novel rice identification method based on Graph Convolutional Networks (GCN), integrating multi-source remote sensing data tailored for complex terrains. A coarse-to-fine cloud removal strategy was developed by fusing SAR imagery with temporally adjacent optical images, achieving high accuracy (RMSE=0.0391, SAM=0.0729, mSSIM=0.9221, CC=0.9537), and providing reliable cloud-free data. A comprehensive feature library, including spectral, texture, polarization, and terrain attributes, was constructed and optimized, yielding 19 key features that enhanced classification performance. The GCN model trained with this feature set achieved 98.3% overall accuracy in Huoshan County, Dabie Mountains, with 96.8% agreement with statistical yearbook data. Ablation experiments showed terrain features significantly improved performance in complex topography. Compared with SVM, RF, and U-Net models, the proposed method outperformed in accuracy, spatial adaptability, sample efficiency, and computational cost, demonstrating its effectiveness and potential for high-precision rice mapping in mountainous environments.