Extracting agricultural parcel boundaries from high spatial resolution remote sensing images based on a multi-task deep learning network with boundary enhancement mechanism

Extracting agricultural parcel boundaries from remote sensing images based on deep learning methods is currently the most promising method. Due to the diversity of agricultural types and limitations of deep learning networks, parcel boundaries extracted by edge detection deep learning networks lack target specificity, and parcel boundaries extracted by semantic segmentation deep learning networks generally causes a loss of precise boundary information. Based on the principles of multi-task deep learning networks PsiNet, BsiNet, and ResUNet-a, we constructed a parallel multi-task deep learning network capable of simultaneously performing edge detection and semantic segmentation for extracting agricultural parcel boundaries from high-resolution remote sensing images. To improve the effectiveness of parcel boundary recognition, the edge enhancement concept from the context-aware tracing strategy (CATS) is introduced into the multi-task parallel deep learning network, which enhanced the continuity and effectiveness of parcel boundary recognition and accelerated network training convergence. Finally, we conducted experiments on three different study areas with different remote sensing data sources and agricultural parcel types. The experimental results demonstrate that the method we proposed achieved better continuity in identifying parcel boundaries and improved the recognition of boundaries between neighboring parcels with strong similarities in remote sensing features such as texture and spectral characteristics.