Key points are not available for this paper at this time.
The Fen River Basin (FRB), a critical ecological corridor in China's Yellow River Basin, faces escalating water-security challenges under climate change and intensive human activities. Pressures include concentrated precipitation patterns, severe agricultural non-point source pollution (contributing >60 % nitrogen loads), groundwater overdraft, and increasing river flow interruptions (67 days/year in 2020), demanding integrated solutions aligned with Sustainable Development Goal 6 (SDG6). We propose a physics-embedded deep learning (PIDL) paradigm with bidirectional coupling between mechanistic and data-driven engines: 1) SWAT-modeled soil water stress index (SWSI) and groundwater depth are embedded into CNN-Attention-BiLSTM inputs to enforce physical plausibility; 2) Deep learning prediction errors dynamically update SWAT parameters (e. g. , SOLK, CHN2) via Bayesian inversion. NSGA-II multi-objective optimization generates management strategies, validated through Monte Carlo simulations and ecological feasibility checks. The coupled framework outperformed standalone models in spatio-temporal accuracy: Runoff prediction: R 2 = 0. 94, RMSE = 0. 12 mm/d (37 % improvement vs. unidirectional coupling) ; Pollution load error reduced by 14. 3 % (hotspot identification accuracy: ±1. 5 km) ; Ecological flow compliance reached 92 % (vs. 69 % baseline). NSGA-II-optimized strategies achieved synergistic benefits: drip irrigation (65 % coverage, 12 % groundwater reduction), vegetative buffers (50m width, 31 % nitrogen load reduction), and dynamic ecological flows (dry season: 15 m 3 /s; wet season: 25 m 3 /s). Monte Carlo confirmed robustness (±11 % fluctuation). As implemented by Shanxi Water Resources Department, the PIDL framework enables cross-scale water governance (18–25 % systemic efficiency gain), balancing allocation, pollution control, and ecological restoration. Its "monitor-simulate-optimize-validate" architecture provides a replicable pathway for SDG-oriented management in semi-arid basins. • 10-m Sentinel-2 NDVI-SWAT fusion. Pinpointed agricultural pollution hotspots (F1: 0. 79, +31 % accuracy) in semi-arid basins. • PIDL with bidirectional coupling SWAT→DL physical inputs + Bayesian-updated SWAT parameters (Runoff R 2 = 0. 94, RMSE↓37 %). • Radar precipitation assimilation CMPA data + attention weights reduced flood peak errors by 42 % and captured drought-flood lags. • Satellite-optimized water governance NSGA-II strategies (drip irrigation, buffers, eco-flows) adopted provincially, boosting efficiency 18–25 %.
Liu et al. (Thu,) studied this question.