EAMNet: a dual-decoder network with edge-semantics synergy for agricultural parcel extraction from remote sensing images

The extraction of agricultural parcels from remote sensing imagery is essential for monitoring and managing agricultural resources. However, existing deep learning methods often face challenges in accurately extracting agricultural parcels and delineating their boundaries due to blurred boundaries, diverse textures, and complex morphologies. We propose EAMNet, a multitask neural network developed to address these challenges through three key innovations. First, we integrate a coordinate attention module between the encoder and main decoder to enhance feature map representations, particularly improving the network’s ability to recognize blurred boundaries. Second, we develop an FE module at the output of the main decoder to extract and fuse multiscale features, enabling the network to more effectively handle diverse textures and complex morphologies. Finally, we propose a multitask network where mask mapping functions as the primary task, whereas boundary mapping and distance mapping act as auxiliary tasks. These tasks collaboratively optimize the network’s performance by improving boundary accuracy and spatial understanding. We evaluate EAMNet on three datasets: one GaoFen-2 (GF-2) image in Shandong, China; another GF-2 image in Sichuan, China; and a Sentinel-2 image in the Netherlands. Experimental results demonstrate that EAMNet outperforms current state-of-the-art networks, achieving improvements of 0.77%, 2.17%, and 2.28% in intersection over union across three different datasets. EAMNet’s high precision and generalization capability make it a promising tool for agricultural parcel extraction in precision agriculture.