E2EVAP: End-to-end vectorization of smallholder agricultural parcel boundaries from high-resolution remote sensing imagery

Rapid and accurate agricultural parcel mapping from high-resolution remote sensing imagery is fundamental to precision agriculture for smallholder farming systems. However, due to the narrow and small-size parcels, and the significant spatio-spectral variability, the existing two-stage segmentation methods cannot extract individual parcels automatically. In this article, the end-to-end vectorization of smallholder agricultural parcel boundaries (E2EVAP) framework is proposed for extracting the vertices of each parcel boundary individually in smallholder farming regions, where the semantic-contour interaction and topological loss through hierarchical instance representation (called "channel to instance") are designed for aggregating the foreground features and jointly establishing the topological relationship between instances to alleviate the topological overlap between parcel objects. Vertex shift correction based on the deep attention corner snake module guided by the parcel boundaries is also proposed to adaptively correct the boundary vertex location shift of large-scale irregular parcels. The comprehensive experimental results obtained on the iFLYTEK public agricultural parcel dataset confirm that E2EVAP shows a superior performance (with a mask mAP of 0.335 and a boundary mAP of 0.201), compared with the pixel-based (such as Mask R-CNN, HTC, and SOLOv2) and contour-based (such as DeepSnake and E2EC) benchmark methods. We believe that E2EVAP has the potential to be widely used for accurate vectorization mapping of agricultural parcels in smallholder farming areas such as South Africa and southern China. The code of E2EVAP is at https://github.com/YangPanHZAU/E2EVAP.