Addressing the Small Aquaculture Pond Mapping Challenge: A Water Signal Attention-Guided Network Using PlanetScope Imagery

Precise, fine-scale mapping of aquaculture ponds (APs) is the technical foundation for refined aquaculture management and environmental regulatory compliance. Despite advancements, current remote sensing workflows often struggle to resolve small-scale APs (<1 ha) or delineate boundaries in dense clusters due to low spectral contrast. To address these challenges, we propose a Water Signal Attention-Guided Network (WSAG-Net), a fine-scale and automated approach for AP mapping using PlanetScope imagery. WSAG-Net incorporates weakly supervised water segmentation into the attention learning process, guiding the model to prioritize water regions. A dedicated joint loss function is employed to jointly optimize the auxiliary water segmentation and the main AP extraction, ensuring that water signal prior knowledge is embedded into shared feature representations. This design enhances discriminative semantic learning and improves the robustness of AP extraction. Tested on PlanetScope imagery, WSAG-Net achieved a Frequency-Weighted Intersection over Union (FWIoU) of 91.09% and an Overall Accuracy (OA) of 95.25%, outperforming all baseline models in both boundary delineation and the identification of small, clustered APs (<1 ha). Furthermore, compared to existing public AP datasets, our method substantially reduces the omission of small APs (<1 ha). This study addresses the persistent difficulty of delineating densely clustered small APs, presenting a practical and transferable framework for fine-scale AP inventory and compliance monitoring.