Accurately mapping coastal aquaculture facilities using very high-resolution (VHR) satellite imagery is crucial for marine spatial planning; however, this process remains challenging due to the high density of facilities, ambiguous boundaries, and significant multiscale variations. This study proposes an improved YOLOv8-seg framework that integrates Efficient Multi-scale Attention (EMA), weighted BiFPN feature fusion, and the Wise-IoU v3 loss for segmenting joint cages and rafts in Sandu Bay, China. A comprehensive ablation design (8 architectures × 2 losses) quantifies the effects of individual modules and their interactions. The optimal model (EMA + BiFPN + WIoU v3) achieved a masked mAP50 of 0.778 and a prediction R² of 0.988 for cage and raft facility areas, both based on ground-truth annotations. To assess operational scalability, the VHR-based segmentation workflow was extended to Sentinel-2 bay-wide grid density maps across nine dates (2019–2026), identifying 576 persistent core aquaculture cells and significant core-periphery spatial structures. Combining the derived density products with CMEMS water quality products revealed a strong negative correlation between aquaculture facility density and KD490 (Spearman’s ρ = −0.358), while the overall density–CHL relationship was weak, and spatial heterogeneity in CHL variance partitioning was stronger than the density effect. In-situ 4-hour measurements (10,684 records from six sites) were used to perform a first-order cross-validation of satellite CHL and SST products, and NH₃-N was suggested as a complementary indicator of aquaculture-related nutrients, although caution is warranted due to the limited number of monitoring stations. These results demonstrate an end-to-end Earth observation workflow—ranging from VHR instance segmentation to regional aquaculture density products and multisensor environmental association analysis—providing a reusable framework for operational monitoring of coastal aquaculture.