A refined edge-aware convolutional neural networks for agricultural parcel delineation

• A novel CNN model was developed for agricultural parcel delineation. • Four model innovations were integrated to enhance edge-awareness. • A post-processing workflow was proposed to optimize parcel shape and geometry. • The proposed model outperformed representative deep learning models. • The model demonstrated transfer mapping capability using limited samples. Accurate delineation of agricultural parcels is crucial for ensuring food security and implementing precision agriculture practices. However, one of the primary challenges in parcel delineation revolves around the accurate extraction of parcel edges due to the diverse morphologies of these parcels. Current deep learning models for agricultural parcel delineation often produce unclosed or incomplete edges, highlighting the need to enhance the edge-awareness of the model. In this study, we proposed a refined edge-aware U-Net (REAUNet) to extract accurate agricultural parcel boundaries from Sentinel-2 images and GF-2 images. REAUNet integrates four innovative components for edge enhancement, i.e., the deep supervision, the edge detection block, the dual attention block and the refine module. The edge detection block and the dual attention block strengthen the ability of capturing spatial contextual information, and the deep supervision constrains parcel delineation across multiple scales, and the refine module further refines the parcel edges. Together, these components enhance the edge-awareness of the model, resulting in more accurate agricultural parcel delineation. From the ablation experiment, REAUNet with all four components reached the higher accuracy than the scenarios that only three out of four components were retained, which suggested that the inclusion of all four components was essential. The comparative evaluation showed that our proposed model surpassed the representative deep learning models for both thematic and geomantic accuracy. REAUNet demonstrated an increase of 2.0 % in F1-score and a decrease of 3.4 % in GTC (Global Total-Classification) compared to the SEANet, and exhibited a 0.8 % increase in IoU and a 2 % decrease in GTC relative to the MPSPNet. We also assessed the capability of the REAUNet model transferred for a new geographic region, and reported that a pre-train REAUNet model with only a small number of local samples could achieve a satisfactory accuracy. To conclude, the newly proposed REAUNet model reaches a high mapping accuracy for delineating agricultural parcels from the high-resolution images. With the powerful edge-aware capability, REAUNet can effectively utilize limited training samples to transfer a deep learning model for a new region, facilitating the generation of agricultural parcel products and a large-scale agricultural parcel mapping in practical applications.