Accurate parameter estimation is critical to the reliability of process-based hydrologic and water quality models such as the Soil and Water Assessment Tool (SWAT). These models are extensively used to simulate watershed responses to changes in land use, management practices, and climate conditions. This study presents a deep learning-based workflow that leverages two-dimensional Convolutional Neural Networks (2D CNNs) for inverse modeling and parameter estimation within the SWAT framework. The approach was applied to the East Fork Shoal Creek (EFS) watershed, a subbasin of the Kaskaskia River watershed in Illinois, USA. CNN models were trained on large datasets of SWAT-simulated watershed outputs—including streamflow, sediment, total nitrogen, and total phosphorus—generated through Sobol sampling of parameter combinations. A comprehensive hyperparameter tuning process was conducted, examining variations in filter size, kernel size, pool strategy, learning rate, dropout rate, number of epochs, and batch size. The resulting CNN architectures were optimized to simulate all key watershed responses effectively, capturing complex spatial and temporal dynamics. The CNN-based approach achieved predictive accuracy comparable to or better than conventional tools like SWAT-CUP, with the EFS-CNN showing strong KGE (0.52–0.80) and PBIAS (5.09–26.72) across flow, sediment, nitrogen, and phosphorus. Transfer learning was effective, as EFS-trained CNNs performed equally well on the neighboring Lost Creek watershed within the Kaskaskia River basin. These findings underscore the potential of deep learning techniques to improve parameter estimation, facilitate efficient model transferability in watershed modeling, and help mitigate equifinality. The proposed approach significantly reduces computational demands and improves scalability, offering a robust pathway for advancing process-based hydrologic modeling in data-rich and regional-scale applications.