The Chesapeake Bay Total Maximum Daily Load (TMDL) is evaluated largely through watershed model load accounting, yet whether modeled reductions are realized as in-bay nutrient recovery remains uncertain, particularly for phosphorus retained in legacy sediments. We developed PBGFormer, a physics-biogeochemical-guided Transformer that combines Landsat 7/8 reflectance with learnable river advection-diffusion and turbidity-gated sediment-interaction priors, and generated 30 m total nitrogen (TN) and total phosphorus (TP) maps of Chesapeake Bay for 2000–2025 using DINEOF gap filling. The model achieved a validation accuracy with an R2 value of 0.830 for TN and an R2 value of 0.691 for TP. The satellite record revealed a pronounced model-observation divergence. Chesapeake Bay Program CAST accounting indicates TP load reductions of 22.0% by 2024, corresponding to 92% of the official phosphorus reduction goal, whereas two independent observations showed much smaller realized changes: USGS RIM flow-normalized TP loads declined by 6.9%, and satellite-derived bay TP concentrations declined by 5.5%. Thus, phosphorus accounting was approximately 3.6-fold more optimistic than observed recovery. Spatially, 90.8% of bay waters showed accelerated post-TMDL TN decline, but only 38.4% showed accelerated TP decline; 54.1% exhibited nitrogen-only recovery. This asymmetric response was not attributable to worsening hypoxia: bottom waters warmed significantly (+0.38 to +0.45 °C decade⁻¹), but official hypoxic-volume metrics were flat to declining from 2000 to 2025. Extrapolation further exposed a phosphorus recovery debt: CAST trajectories imply near-term TP load attainment around 2026, whereas satellite concentration trends imply roughly five additional decades of in-bay recovery, and long-term RIM TP loads are statistically flat. These results show that watershed load accounting can substantially overstate phosphorus recovery and that legacy/internal phosphorus imposes a structural ceiling on external-load-based restoration.

暂无