Hydro-learning: Integrating AI with SWAT for multi-resolution flow and sediment prediction in data-scarce basins

One of the main challenges in predicting hydrological processes like streamflow and sediment yield in ungauged catchments is the lack of calibration data. Approaches such as regionalization, parameter transmission, and remote sensing can overcome a lack of calibration data. To address this, a novel hybrid framework was developed by integrating the physically based SWAT (Soil and Water Assessment Tool) model with deep learning (CNN-BiGRU-Attention). Termed Hydro-Learning, the framework transfers predictive capabilities from a small gauged basin (Bakerman Watershed, BW) to a larger, partially ungauged basin (Khalilkan Watershed, KW), which includes the fully ungauged sub-basin AKW. Model training was based on SWAT outputs from BW and tested across three distinct spatial resolution scenarios: high resolution (12.5 m DEM/10 m LULC), medium resolution (30 m DEM/100 m LULC), and coarse resolution (250 m DEM/500 m LULC). The model had excellent predictive weather when comparing with the gauged data for BW or even when compare the AI results with the outputs originates from the SWAT models within KW and AKW. Recorded AIC of 12160 and the highest R 2 and NSE of 0.98 alongside RMSE of 1.49 m 3 /sec for the stream flow in the KW, performance over sediment was similarly good, with an R 2 and NSE maxima of 0.83 with RMSE of 146,567.66 tons. The model was also really robust and worked very well for the AKW, providing great results for streamflow and solid results for sediment despite the lack of available data for the calibration. Streamflow predictions remained consistently strong across scenarios, while sediment yield was highly sensitive to input resolution, with high-resolution inputs yielding the best results. This hybrid approach offers a viable solution for hydrological forecasting in data-scarce regions and may serve as a more practical alternative to conventional regionalization or recalibration methods. The two-stage framework demonstrates high generalization ability, which provides the promise for hydrological forecasting at global scale.