Improving the accuracy and spatial resolution of precipitable water vapor dataset using a neural network-based downscaling method

Abstract

Precipitable water vapor (PWV) is a crucial variable in water and energy transfers between the surface and atmosphere, and it is sensitive to climate and environmental changes. Among various PWV monitoring techniques, the European Centre for Medium-Range Weather Forecasts Reanalysis 5 (ERA5)-derived PWV, with a spatial resolution of 0.25 × 0.25°, has excellent spatiotemporal continuity. However, its accuracy has relatively large uncertainties, and its spatial resolution is inadequate for small regions such as the Tibetan Plateau (TP). Therefore, the aim of this study was to propose machine learning-based modification and downscaling methods to improve the accuracy and spatial resolution of ERA5-derived PWV. First, a modification model based on a back propagation neural network (BPNN) was proposed to improve the accuracy of ERA5-derived PWV. The results showed that the root mean square error (RMSE) of ERA5-derived PWV decreased from 2.83 mm to 2.24 mm in China, i.e., an improvement of 20.8%. In the TP region, the RMSE decreased from 2.92 mm to 1.96 mm, and the improvement was 32.9%. Subsequently, a BPNN-based downscaling model was established using the modified ERA5-derived PWV to generate PWV with a 6-hourly, 0.1° × 0.1° spatiotemporal resolution in the TP region. Compared with global navigation satellite system-derived PWV, the RMSE of the generated PWV was 2.13 mm. The spatial distribution of BPNN-derived PWV based on the downscaling method exhibited suitable stability in the TP region, indicating that the proposed method could significantly improve the accuracy and spatial resolution of ERA5-derived PWV in the TP region.