Comprehensive Analysis of a Parcel-Level Crop Mapping (PARCM) Framework From the Perspective of Spatial Heterogeneity and Spatiotemporal Transferability

Efficient mapping of spatial extent and crop attributes of agricultural parcels is essential for agricultural management, policy making, and food security. Currently, parcel-level crop maps are primarily created using modal classes of pixel-level thematic maps or statistical properties of per-parcel features. However, these methods assume that a parcel contains one crop type, reducing accuracy in multicrop parcels, and focus on local and season-specific crops, overlooking spatiotemporal variations, especially in smallholder landscapes. In this study, we developed a parcel-level crop mapping framework, named PARCM, and comprehensively analyzed its performance regarding spatial heterogeneity and spatiotemporal transferability. For PARCM, a rule-based fusion method was designed to integrate two core procedures: pixel-level crop mapping (PIXCM) using random forest and agricultural parcel extraction using a modified recurrent residual U-Net. PIXCM-based crop maps aid in generating parcel sub-boundaries and final crop maps. We mapped major autumn-harvested crops in 17 heterogeneous agricultural landscapes across Northeastern and Central China using 10-m and 2-m resolution remote sensing imagery. Results demonstrated that PARCM outperformed PIXCM and a parcel feature average-based method, achieving an average overall accuracy (OA) and Kappa above 0.85. PARCM performed well in regions with fewer crop types (low compositional heterogeneity) and more uniform distribution (low configurational heterogeneity), with OA exceeding 90%. It also achieved strong seasonal transferability (average OA = 89.60% and Kappa = 0.82) and spatial transferability (OA > 88.0% and Kappa > 0.80). Our framework can maintain the spatial consistency of crops and parcels while adapting to varying spatiotemporal complexities, providing a strong solution for parcel-level crop mapping.