Abstract Satellite‐derived precipitable water vapor (PWV) products are widely used for meteorological and climatological research, yet the relative contributions of environmental and observational variables to product uncertainties remain unquantified. Here, we present an explainable machine‐learning framework that integrates extreme gradient boosting with SHapley Additive exPlanations (SHAP) to establish a data‐driven error budget for satellite PWV product bias. The model accurately characterizes nonlinear relationships between product bias and multiple influencing factors, with an R2 of 0.70, an RMSE of 2.19 mm, and an MAE of 1.57 mm. The SHAP‐based error budget reveals that background atmospheric moisture is the dominant contributor (29.69%), followed by surface temperature (14.27%), latitude (11.35%), and longitude (9.34%). Moving beyond previous qualitative evaluations, this framework provides a quantitative decomposition of satellite PWV product bias into multiple environmental and observational variables. The approach can be adapted to other evaluation metrics and extended to other satellite‐derived water vapor and geophysical products.

Read original article