EEDNet: Edge and Edge Direction Network for simple and regular land parcel vectorization

Abstract

Land parcel extraction from remote sensing images plays a crucial role in applications such as agricultural management, yield estimation, and land resource monitoring. These applications depend on vectorized parcels with regular shapes depicted by a limited number of points, making accurate vector-based land extraction results highly important. However, existing methods for land parcel extraction primarily rely on raster-to-vector conversion, transforming pixel segmentation or edge results into vectors. This approach often results in distorted shapes and redundant points. We notice that when humans delineate land parcels, they intuitively identify key points at locations where edge directions change and connect these points sequentially to form vectors. Inspired by this process, we propose Edge and Edge Direction Net (EEDNet) and a novel post-possessing method, which generates parcel polygons as the final output. EEDNet employs a dual-decoder structure for simultaneous learning of parcel edges and their directions. By detecting edges, identifying key nodes through changes in edge directions, and sequentially connecting these nodes under the guidance of edges, EEDNet constructs well-structured parcel polygons, ensuring smooth parcel boundaries and simplified key points. Experimental results show that our method demonstrates the best overall performance across multiple datasets. Specifically, it achieves the highest complete-intersection over union scores of 0.614 on the iFLYTEK dataset, reflecting its ability to balance geometric accuracy and pixel segmentation. Additionally, it records the lowest GTC errors of 0.158 and 0.180 and the lowest GUC errors of 0.077 and 0.101 on the GFDataset and Netherlands datasets, respectively, showcasing its robustness in capturing object-level and geometric features. We release our code at https://github.com/lixianshen20/EEDNet.git.