Data-driven hydrological modeling has seen rapid development in recent years owing to its flexibility to approximate the complex relationships between driving forces and hydrological fluxes. However, traditional data-driven models typically cannot simultaneously capture the processes that pose both chronic and acute impacts on streamflow, thus impeding further inference. Therefore, this study presents a baseflow-filtered hydrological inference model to gain insights into hydrological processes in irrigated watersheds. The proposed model starts with separating the streamflow process into two sub-processes using a process-based baseflow separation method. Each sub-process is simulated through a new interpretable data-driven model. The resulting hydrological inferences facilitate the identification of the dominant factors influencing flows in saturated and unsaturated zones. The proposed model is applied to three irrigated watersheds, and the evaluation metrics show that the proposed model outperforms two conventional data-driven models. Our findings reveal that predictors associated with air temperature and long-term (i.e., monthly) irrigation are mainly responsible for characterizing baseflow dynamics, while precipitation and short-term (i.e., semi-weekly or weekly) irrigation are primarily responsible for describing overland flow and interflow dynamics. The fidelity of the derived hydrological inference is further demonstrated through sensitivity analysis. The results show that the relative importance of predictors not only reflects their significance on model performance, but also influence the changes on streamflow.

Abstract Machine learning (ML) models have been widely used for hydrological simulation. Previous studies have reported that conventional ML models fail to accurately simulate extreme flows which are crucial for design flood estimation and associated risk analysis in the context of climate change. Therefore, this study proposes a joint probabilistic rainfall‐runoff model (JPRR) for improving high‐to‐extreme flow projection. With the aid of paired copula constructions, bootstrap aggregation, and multi‐model ensemble approaches, the proposed model is able to effectively characterize the dependence relationships of predictors (i.e., precipitation time series with different moving sums) with various probability distributions. JPRR has been applied to four pristine basins in China, representing different climate zones and landscapes. The results reveal that JPRR significantly outperforms three well‐known ML models (i.e., random forest, artificial neural networks, and long short‐term memory) in high‐to‐extreme flow simulations. In JPRR, the copulas exhibiting the right tail dependence play a more important role in streamflow simulations at mountainous basins. Moreover, a significant difference in streamflow projections (from 2030 to 2099) derived from JPRR and benchmark models imply that flood risks from conventional ML models may be underestimated under changing climatic conditions.