The High-Accuracy InSAR PWV Retrieval Methods Based on Machine-Learning Fusing Multisource Data

Abstract

High-accuracy, high-spatial-resolution precipitable water vapor (PWV) is essential for severe weather observation and prediction. Although interferometric synthetic aperture radar (InSAR) can provide PWV information at fine spatial resolution, its retrieval accuracy is poor. Most existing methods optimize InSAR PWV using global navigation satellite system or ERA5 PWV. However, performance improvements remain limited, and effectiveness in regions with complex topography still requires further validation. Therefore, Kunming and its surrounding areas in Yunnan Province, China, with complex topography, are selected as the study area. First, differential zenith tropospheric delay (ΔZTD) was extracted from the Sentinel-1A time series using small baseline subset-InSAR. Then, two models were established: the direct fusion mapping model (DFMM) and the stagewise correction refinement model (SCRM) with physical processes. Both models employed a backpropagation neural network to retrieve ΔPWV by integrating InSAR ΔZTD, continuous operational reference system (CORS) ZTD, CORS PWV, ERA5 PWV, surface pressure, air temperature, and geopotential. Finally, absolute PWV was recovered by adding ERA5 PWV at the reference epoch to the retrieved ΔPWV. The results indicate that, compared with the original InSAR ΔPWV with a root-mean-square error (RMSE) of 11.72 mm, the RMSE decreased to 2.29 mm for DFMM and 2.02 mm for SCRM. Both models’ RMSE improvement rates exceed 80%. In the time series, the retrieved DFMM and SCRM PWV show high consistency with the CORS PWV, with R values exceeding 0.94. Spatially, the retrieved DFMM PWV and SCRM PWV show an R of 0.97 with ERA5 PWV and reveal more detailed PWV information. The proposed models achieve high spatial resolution and high accuracy in InSAR PWV retrieval over complex topography, providing high-spatial-resolution PWV datasets to support refined meteorological analysis and related applications.