A Dual-Stage Reconstruction and Optimization Deep Learning Framework for Generating High-Precision Seamless Precipitable Water Vapor Across the Mainland United States

Precipitable water vapor (PWV) is critical to global climate dynamics, the terrestrial water cycle, and extreme weather events. However, current Moderate Resolution Imaging Spectroradiometer (MODIS) near-infrared (NIR) PWV products suffer from low accuracy, vulnerability to clouds, and incomplete spatiotemporal coverage. Therefore, we propose a dual-stage reconstruction and optimization deep learning framework that fuses MODIS NIR remote sensing data with Global Navigation Satellite System (GNSS) observations to generate high-precision, large-scale, and spatially continuous PWV. Firstly, a cascaded spatiotemporal generative adversarial network (CSTGAN) was developed to reconstruct PWV in cloud-contaminated regions using the clear MODIS PWV, generating seamless reconstructed PWV. Secondly, an optimization model based on a random forest method was applied to correct the reconstructed PWV by GNSS-derived PWV, yielding final enhanced accuracy PWV . The CSTGAN integrates spatial and temporal reconstruction modules within a generative adversarial network architecture, enabling a cascade learning mechanism to capture complex spatiotemporal patterns. The study focuses on the mainland United States from 2018 to 2023. During reconstruction, the CSTGAN achieves a Pearson correlation coefficient (R) exceeding 0.98 and a root mean square error (RMSE) below 1.39 mm on the test sets under 70% artificial masking rates when comparing the masked regions of reconstructed dataset with their clear-sky MODIS counterparts. After optimization, the final seamless PWV exhibit an R of 0.97 and RMSE of 2.97 mm against independent GNSS validation data, representing a 44.5% RMSE reduction compared to the reconstructed PWV product. This approach overcomes the critical limitation of MODIS NIR PWV unavailability under cloud conditions, significantly enhancing the utility of satellite-derived PWV. The resulting 1km resolution, seamless PWV advance hydrological cycle analysis, climate monitoring, meteorological retrievals, and atmospheric delay correction, offering transformative potential for Earth system science research and applications.