A deep learning method for cultivated land parcels (CLPs) delineation from high-resolution remote sensing images with high-generalization capability

Accurate cultivated land parcels (CLPs) information is essential for precision agriculture. Deep learning methods have shown great potential in CLPs delineation but face challenges in detection accuracy, generalization capability, and parcel optimization quality. This study addresses these challenges by developing a high-generalization multi-task detection network coupled with a specialized parcel optimization step. Our detection network integrates boundary and region tasks and design distinct decoders for each task, employing performance-enhancing modules as well as more balanced training strategies to achieve both accurate semantic recognition and fine-grained boundary depiction. To improve network's ability to train more generalized models, our study identifies the variations in image hue, landscape surroundings, and boundary granularity as the key factors contributing to generalization degradation and employ color space augmentation and attention mechanisms on spatial and hierarchy to enhance the generalization. Additionally, the parcel optimization step repairs long-distance boundary breaks and performs object-level fusion of delineated regions and boundaries, resulting in more independent and regular CLP results. Our method was trained and validated on GaoFen-1 images from four diverse regions in China, demonstrating high delineation accuracy. It also maintained stable spatiotemporal generalization across different times and regions. Comprehensive ablation and comparative experiments confirmed the rationale behind our model improvements and demonstrated our method's effectiveness over existing single-task models (SegNet, MPSPNet, DeeplabV3+, U-Net, ResU-Net, R2U-Net), and recent multi-task models (ResUNet-a, BSiNet, SEANet).