A Structural Knowledge-Guided Deep Learning for Retrieving Precipitable Water Vapor From Satellite Remote Sensing Observations

Abstract

Near-infrared (NIR) satellite remote sensing provides global precipitable water vapor (PWV) observations at ~1-km spatial resolution; however, its applicability is severely constrained under cloudy-sky conditions, where a substantial proportion of retrievals become unusable due to cloud contamination. To address this limitation, this study proposes a structural knowledge-guided deep learning (SKDL) framework for satellite-based PWV retrieval with enhanced accuracy and spatial continuity. The framework incorporates structural knowledge learned from boosting models into a deep neural network (DNN) via a knowledge distillation mechanism, thereby improving the representation of complex, nonlinear feature interactions. Furthermore, cloud properties are explicitly introduced as new variables to account for cloud-induced radiative transfer effects. The proposed method is evaluated against ERA5 reanalysis and global navigation satellite system (GNSS) observations. ERA5-based spatial validation shows that SKDL-LightGBM (SKDL-LGB) achieves a mean absolute error (MAE) of 1.6 mm and a structural similarity index (SSIM) of 0.90. GNSS-based station validation further indicates that SKDL-LGB) attains a root-mean-square error (RMSE) of 2.15 mm under all-weather conditions, outperforming standalone boosting and deep learning models. These findings demonstrate that integrating distilled structural knowledge and cloud information substantially improves NIR-based PWV retrieval performance, particularly in cloudy-sky conditions. The proposed framework offers a robust and scalable solution for generating high-accuracy, spatially continuous PWV products.