An Adaptive Canny Edge detection architecture with dual attention for precise segmentation of individual aquaculture ponds

Aquaculture mapping is crucial for effective monitoring and managing aquaculture resources. However, accurately identifying and geotargeting individual aquaculture ponds (IAPs) from medium-resolution remote sensing imagery presents significant challenges due to varying pond sizes, complex boundary conditions, and environmental factors. In this study, we proposed ACEU-Net (Adaptive Canny Edge-detection UNet), a novel deep learning architecture designed specifically for IAPs segmentation from Sentinel-2 imagery. The ACEU-Net incorporates three key components: an Adaptive Canny Edge detection module that dynamically adjusts threshold parameters through backpropagation, an Edge-aware Dual Attention Module (EDAM) that integrates channel and spatial attention mechanisms with edge information, and a Pyramid Scene Parsing (PSP) pooling module that enhances multi-scale contextual feature extraction. Experimental validation conducted across three towns in Guangdong Province, China, demonstrates that ACEU-Net consistently outperforms state-of-the-art segmentation models, achieving average performance metrics of 0.905 (F1-score), 0.825 (IoU), 0.968 (OA), and 0.885 (MCC). Ablation studies confirm the effectiveness of each proposed module. The superior performance of ACEU-Net in delineating complex pond boundaries under varying environmental conditions positions it as a valuable tool for sustainable aquaculture management and environmental monitoring.