Parcel-scale crop planting structure extraction combining time-series of Sentinel-1 and Sentinel-2 data based on a semantic edge-aware multi-task neural network

This study proposes a parcel-scale remote sensing methodology for agricultural cultivation pattern extraction to address the classification inaccuracies observed in previous crop planting structure identification studies. Specifically, a novel semantic edge-aware multi-task neural network (SEANet) is first employed to delineate agricultural land parcel boundaries precisely. Subsequently, the Global Separability Index (GSI) was used to optimize feature selection. Ultimately, a parcel-scale random forest (RF) model was developed to enable accurate crop type classification and spatial delineation of crop planting structures. Results show: 1) SEANet achieves accurate extraction of parcel boundaries, extracting a total of 175,000 parcels, and demonstrates a relative error of 6.98% compared to statistical yearbook data. The proposed method effectively addresses issues of rough and incomplete boundaries; 2) The GSI results indicate that VH polarization features, the normalized difference red-edge1 index (NDRE1), and texture features exhibit robust crop discrimination capabilities. The critical periods for crop identification are May and July, with key spectral bands in B2∼B5, and B12 of the Sentinel-2 satellite; 3) The developed parcel-scale RF model effectively addressed the challenges of mixed pixels and salt-and-pepper noise, with an overall accuracy (OA) of 94.86% and a Kappa coefficient of 0.93, which demonstrated excellent classification performance.