Multi-Scale Bilateral Spatial Direction-Aware Network for Cropland Extraction Based on Remote Sensing Images

The information of cropland is obtained efficiently and accurately as the basis for achieving precision agriculture (PA). As the boundary between cropland and other types of land in remote sensing images with different resolutions is fuzzy, the characteristics of cropland are easily confused when extracting cropland, resulting in inaccurate identification and extraction of cropland under large and complex backgrounds and rough localization of marginal areas. We proposed a two-path multiscale attention self-supervised network with the perception of four directions in pixel space, called the multiscale bilateral spatial direction-aware network (MBSDANet), to solve these problems and improve the model’s ability to extract cropland in small samples. One path extracts attentional feature maps by spatial directions to preserve detailed direction-aware information and generate high-resolution features; the other path obtains local-to-global information through pyramid pooling and attention awareness to capture dense multiscale cropland features to separate targets in complex contexts. The features of the two branches at different levels are fused by weighting the multi-aware information. Triangular self-supervised and boundary-aware losses are used to achieve fine segmentation and extraction of cropland in small samples. We tested the extraction method on cropland in Denmark and the Hebei Province of China, demonstrating its effectiveness and generalization. Compared to other neural network models, MBSDANet achieves better accuracy with a precision of 0.9481, an IoU of 0.8937, and an F1 score of 0.9438.