On the Use of Knowledge‐Informed Machine Learning and Multisource Data for Spatially Explicit Estimation of Irrigation Water Withdrawal

Abstract Irrigation plays a crucial role in the earth system, yet our understanding of irrigation water withdrawal (IWW) remains limited due to the scarcity of spatially explicit data. While process‐based models and remote sensing can bridge this data gap, their estimates often fail to capture real IWW and are associated with large uncertainties. Here, we present a knowledge‐informed, explainable machine learning framework that combines random forest (RF) with Shapley additive explanations to generate spatially explicit IWW estimates across China. Our framework incorporates irrigation domain knowledge, state‐of‐the‐art irrigated cropland maps, and various socioeconomic, hydroclimatic, and auxiliary factors. RF shows reasonable performance in spatial and temporal cross‐validation, achieving a coefficient of determination exceeding 0.85 and a root mean square error below 0.45 km 3 /year when evaluated against held‐out prefecture‐level data. The predictions of IWW depth are primarily driven by geographic and knowledge‐based predictors, most of which exhibit nonlinear and non‐monotonic impacts on model outputs. By integrating the RF model with a temporal downscaling approach, we develop a new gridded IWW product for China (named CIWW1km), which provides monthly IWW depth and volume at 1 km resolution from 2000 to 2020. CIWW1km aligns closely with prefecture‐level IWW reports and explains over 85% of the variance in independent IWW observations (i.e., data excluded from training) across over 150 basins and counties. It highlights a rapid increase in IWW in China's arid zone, driven by irrigated area expansions. CIWW1km outperforms existing products and is well‐suited for hydrological and climate studies, and water‐food nexus analyses.