Complex agricultural parcel boundary extraction through semantic segmentation network with dual-task constraints

Agricultural parcel boundary extraction from remote sensing images facilitates efficient land management and production prediction, thereby contributing to global food security. Various semantic segmentation neural networks have gained success in handling large parcels with regular boundaries. However, they encounter difficulty on land with blurred boundaries, diverse textures, and irregular shapes. We propose a semantic segmentation network named DDM-Net with three techniques to handle these issues. First, we introduce a dual-feature attention (DFA) module between the encoder and decoder. The DFA leverages both category and multiscale information to mitigate the loss of detailed features caused by multilayer convolution and pooling, thereby significantly enhancing the recognition of blurred boundaries. Second, we design a new multifeature extraction and fusion module to suppress the influence of irrelevant texture information, which enables the network to effectively reduce the extraction of erroneous boundaries. Third, we architect a dual-task learning framework where the primary task focuses on boundary extraction, whereas an auxiliary task dedicated to mask prediction provides global perspectives. This dual-task approach comprehensively captures parcel features, resulting in superior boundary continuity. Our network is pre-trained on a simple parcels dataset from Xinjiang and then transferred to handle a complex parcels dataset from Sichuan and Chongqing. Experimental results demonstrate that DDM-Net outperforms current state-of-the-art networks, achieving improvements of 2.87% (Sichuan), 3.67% (Chongqing), and 0.84% (Xinjiang) in the Qual on three different datasets. Our study is expected to facilitate further agricultural research based on agricultural parcel boundary extraction.