A Multi-Source Benchmark Dataset for Deep Learning-Based Cropland Parcel Boundary Extraction

The expertly rectified extraction of cropland parcel boundaries is a critical basis for precision agriculture deployment and land resource management. However, the scarcity of large-scale, high-quality, cross-regional reference datasets limits the generalizability of deep learning models. To address this data bottleneck, we created a standardized, multi-source cropland parcel boundary benchmark dataset. The collection, which covers Europe (Denmark, the Netherlands, Spain) and Asia (Central China), encompasses over 180,000 expert-rectified parcels, with average parcel sizes ranging from 0.05 to 10 hectares and spatial resolutions spanning 0.2 to 10 m. Based on this dataset, we evaluated nine representative deep learning models belonging to three architectural paradigms (edge detection, multi-task semantic segmentation, and instance segmentation) and analyzed the effects of spatial resolution, parcel scale, and topographic relief on extraction accuracy. The experimental results indicate the following: (1) Spatial resolution exerts a non-linear influence on boundary extraction fidelity, where some medium-resolution data conditions (such as 4.8 m imagery in large-scale fields) retained competitive boundary extraction performance for certain models. (2) Even small but regular parcels of merely 0.22 hectares can be extracted with high accuracy, achieving a maximum F1 score of 86.58%. (3) Topographic variations lead to contrasting model performance. In plain and hilly terrain, multi-task semantic segmentation models exhibit obvious advantages. BsiNet yields the highest F1 score of 86.58%, but this value drops to 36.56% in complex mountainous regions. By comparison, edge detection models perform better in mountainous areas, where RCF attains the maximum F1 of 70.85%. In summary, this study establishes a standardized, expertly refined reference parcel dataset and provides an empirical evaluation of how different model architectures perform across varied observational settings, offering an empirical baseline for future model development and evaluation in cropland boundary extraction.