A CNN-Transformer Hybrid Network With Boundary Guidance for Mapping Cropland Field Parcels From High-Resolution Remote Sensing Imagery

Accurate mapping of cropland field parcels (CFPs) is essential for efficient agricultural production and management. However, CFPs in high-resolution remote sensing imagery exhibit substantial variability in size, shape, and spatial distribution, making it challenging to accurately capture their features using only local or global information, which limits the mapping precision. To address this, we propose a convolutional neural network (CNN)-transformer hybrid network with boundary guidance (CTHBNet) for effective CFP mapping from high-resolution imagery. CTHBNet comprises three key modules: 1) a CNN-transformer fusion encoder that integrates local and global features to enhance recognition of multi-scale parcels; 2) a hierarchical information fusion decoder that progressively restores parcel shapes to delineate complete boundaries; and 3) a boundary-guided feature enhancement module that refines the boundary clarity and reduces parcel adhesion. Moreover, a multi-task learning strategy jointly optimizes parcel extent, boundary, and distance features to improve the mapping precision. We evaluated CTHBNet on GaoFen-2 imagery across four distinct agricultural regions in China. The results showed that CTHBNet can achieve an overall accuracy of over 92%, an F1-score exceeding 82%, and a mean intersection over union exceeding 84%, demonstrating its effectiveness in CFP mapping. Comparative experiments confirmed the superior accuracy and boundary delineation of CTHBNet, and ablation experiments validated the effectiveness of each proposed module. Furthermore, experiments on five global regions and multiple public benchmarks further demonstrated its robust generalization capability. In particular, by fusing the complementary strengths of CNN and transformer models, CTHBNet significantly improves the representation of complex parcel boundaries and diverse shapes, thereby enhancing its adaptability to heterogeneous cropland structures and increasing the overall mapping precision.