Abstract: The soil and water assessment tool (SWAT) has been widely used in distributed-parameter hydrological models in order to simulate agricultural non-point source pollution (NPS) and mitigation strategies. Although the SWAT has presented significant effectiveness in the humid and semi-humid areas, great difficulties still remain in the dry regions. Particularly, the essential parameters are often required to calibrate under water-limited settings. Some challenges are also attributed to the model's reduced depiction of the hydrological response, including the soil moisture dynamics and surface runoff production. In this study, a hybrid framework was established to integrate the "physics-constrained + data-driven" approaches. The SWAT model was also combined with a deep learning architecture, namely a convolutional neural network paired with a long short-term memory network (CNN-LSTM). An enhanced particle swarm optimization technique (IPSO) was included in this linked framework to optimize three elements: the CNN-LSTM network architecture, its hyperparameters, as well as the daily fusion weight that was employed to linearly amalgamate SWAT and CNN-LSTM outputs. Furthermore, the IPSO assessed the error between the fused prediction series and observed pollutant concentrations at each iteration using adaptive inertia weighting and stochastic perturbations. Thereafter, the particle placements were dynamically adjusted in the parameter space. The SWAT and CNN-LSTM models were then used to rotate daily between major (master) and secondary (auxiliary) roles. The error correction markedly diminished the bias in real time. The individual performance of the SWAT and CNN-LSTM was initially evaluated to compare the combined SWAT-IPSO-CNN-LSTM system on daily timescales. Subsequently, the efficacy of IPSO after hyperparameter optimization was assessed against nine metaheuristic algorithms. The Ningxia section of the Yellow River basin was selected as the research region. A multi-scale investigation was implemented on the watershed nitrogen and phosphorus dynamics. Total nitrogen (TN) and total phosphorus (TP) loads were simulated, and then the enhancement of the coupled model was simulated. The results show that better performance was achieved over the individual SWAT configuration. In TN simulations, the coefficient of determination (R²), and the Nash–Sutcliffe efficiency (NSE) rose by 14.1% and 14.5%, respectively, while the percent bias (PBIAS) decreased by 38.6%; In TP simulations, the R² and NSE were enhanced by 10.7% and 12.0%, respectively, while there was a 65.3% reduction in the PBIAS. Moreover, the centered root-mean-square error (CRMSE) decreased from 0.233 (TN) and 0.192 (TP) in the uncoupled SWAT simulations to 0.117 and 0.127, respectively, in the linked framework, indicating a significant reduction in the prediction error. A spatiotemporal study of the coupled model demonstrated that the peak TN (57% of annual maximum) and TP (68%) loads on the wet season were dominated by the synergistic impacts of the intense precipitation and fertilizer application. The NPS levels were relatively slightly reduced in the dry season, due mainly to the compensatory winter irrigation. At the sub-watershed level, the southern areas exhibited the pollution peaks that linked to the rainfall-induced runoff, whereas the northern irrigation districts shared a bimodal pollution pattern mostly driven by intensive agriculture practices. Specifically, the irrigation district of the Yellow River diversion accounted for 31%-37% of the TN and TP loads among river segments; The largest pollutant export per unit area was observed in the Hongliugou and Kushuihe segments, due to the concentrated agropastoral activities. The coupled SWAT-IPSO-CNN-LSTM technique can be expected to efficiently mitigate the calibration uncertainty in a dry watershed. Its adaptive error-correction can accurately and rapidly forecast the daily nutrient pollution. Thus, this hybrid framework can enhance the technological assistance for the water environment in drought-prone areas.