Research on paddy field segmentation methods based on high-resolution remote sensing images

To address the challenge of accurate paddy field extraction in rice cultivation areas, this study proposes an optimized method that integrates multi-dimensional spectral and temporal features derived from remote sensing imagery. Based on the Transformer framework, a multi-dimensional feature set combining multispectral bands and spectral indices was constructed, and a temporal phenology-differential attention encoder was designed to explicitly model the spectral evolution of rice at key growth stages. Experiments were conducted using fused Gaofen-2 and Sentinel-2 imagery. The optimized model achieved an precision of 94.23% and a mean Intersection over Union IoU (mIoU) of 87.24%, representing a significant improvement over the baseline model. The results indicate that the method effectively improves segmentation performance for fragmented paddy fields with narrow bunds and complex boundaries, offering a reliable and efficient solution for automated, precise paddy field extraction.