Fine-Grained Extraction of Coastal Aquaculture Ponds From Remote Sensing Images Using an Edge-Supervised Multi-task Neural Network

Precise monitoring and extraction of coastal aquaculture ponds are crucial for ecological conservation and the sustainable use of coastal wetlands. However, traditional image analysis techniques often struggle to efficiently extract aquaculture ponds in the complex, heterogeneous environments of coastal wetlands. While deep learning approaches have enhanced automation, existing models are limited by their reliance on single-feature attention, which fails to capture the global structural characteristics of aquaculture ponds, such as boundaries and internal spatial relationships, thus restricting their generalization capability. To address these issues, this paper introduces a novel Multi-task Aquaculture pond Extraction Network (MAENet). This network facilitates the supervised extraction process by modeling pond boundaries and the distances from internal points to these boundaries. It notably enhances performance in complex environments and significantly boosts generalization capabilities by learning global structural features. Firstly, a shared encoder-decoder architecture was constructed, leveraging large kernel depthwise separable convolution and residual optimization, thereby enhancing both local and global feature representations. Then, a selective task fusion module was developed to optimize the allocation of shared feature space, effectively alleviating the negative transfer problem of multi-task architecture. Finally, the adoption of an adaptively optimized multi-task loss function obviates the necessity for manual allocation of weights. Ablation and comparative experiments using GF2 imagery from the Qingdao coastal wetlands show that MAENet outperforms existing methods in complex scenarios. Furthermore, transfer experiments with JL1 imagery from the Jiangmen and Yantai demonstrate the strong generalization capability of the proposed method across different environments.