Human impacts versus environmental drivers of seagrass loss in the Nansha islands: Evidence from domain-adaptive deep learning

• Physics-aware, domain-adaptive SeagrassMNet achieves 92.6 % IoU. • First multi-temporal Nansha seagrass maps (2019–2023) released. • Seagrass coverage declined 47.7 % from 2019 to 2023 across 30 reefs. • Artificial island construction is the main driver, exceeding SST and sea level. Seagrass meadows are vital blue carbon ecosystems that protect shorelines, support biodiversity and store carbon, but they face accelerating losses from environmental change and human disturbance. Disentangling these drivers is particularly challenging in data-scarce and operationally difficult regions such as the Nansha Islands. This study proposes SeagrassMNet, a physics-informed, domain-adaptive deep model for shallow-water seagrass mapping. The network couples dual attention transformers with adversarial domain adaptation to handle spectral variability and sparse in situ labels. Applied to multi-temporal Sentinel-2 imagery, SeagrassMNet achieved an Intersection over Union (IoU) of 92.56 %, outperforming several state-of-the-art models (UNet++, HRNet, DeepLab v3+, ViT-Adapter-L, SegFormer and Swin Transformer V2). Cross-reef and cross-year tests, together with cross-regional point-based checks on open data (SeagrassSpotter), indicate that the model has promising transferability across broad areas and complex marine environments. The resulting maps provide the first comprehensive record of seagrass distribution for the Nansha Islands and reveal a 47.7 % decline in seagrass cover from 2019 to 2023. Association analyses over this period indicate that environmental factors, particularly sea surface temperature anomalies and sea surface height variability, are systematically correlated with spatial and temporal patterns of seagrass change. Large-scale artificial island construction and coastal modification show even stronger correlations with observed losses, especially where reclamation footprints overlap mapped seagrass meadows, underscoring the need for spatially targeted management in these vulnerable reef systems.