Answers are grounded in this paper and cite their sources.
Accurate extraction of coastal aquaculture facilities from remote sensing imagery remains challenging because these facilities exhibit irregular spatial distributions, substantial scale variations, and weak contrast against complex marine backgrounds, often with indistinct boundaries. To address these challenges, this study proposes FCS-UNet, an enhanced U-Net framework incorporating three components: Dual Hybrid Attention (DHA), Multi-Scale Convolutional Attention Module (MSCAM), and Progressive Stacked Feature Refinement (PSFR). The contribution of this study is the integration of adapted attention and multi-scale modules with a bottleneck block of stacked depthwise-separable convolutions, evaluated by facility type on a new dataset. Experimental results show that FCS-UNet achieves an Accuracy of 93.64 ± 0.42%, a Dice coefficient of 95.25 ± 0.24%, and an IoU of 91.01 ± 0.47%, improving the U-Net baseline by 3.68, 2.75, and 5.02 percentage points, respectively. Dongwuyang experiments and ablation studies further evaluate the model’s performance. Furthermore, facility-type-stratified evaluations demonstrate that FCS-UNet achieves the highest Dice and IoU for both raft and cage facilities in both study areas. Under a random image-level split within a single coastal region, these results demonstrate the effectiveness of FCS-UNet for coastal aquaculture facility extraction.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: