Precipitable water vapor fusion based on a generalized regression neural network

Water vapor plays an important role in Earth’s weather and climate processes and energy transfer. Plenty of techniques have developed to monitor precipitable water vapor (PWV), but joint use of different techniques has some problems, including systematic biases, different spatiotemporal coverages and resolutions among different datasets. To address the above problems and improve the data utilization, we propose to use a generalized regression neural network (GRNN) to fuse PWVs from Global Navigation Satellite System (GNSS), Moderate-Resolution Imaging Spectroradiometer (MODIS), and European Centre for Medium‐Range Weather Forecasts Reanalysis 5 (ERA5). The core idea of this method is to use the high-quality GNSS PWV to calibrate and optimize the relatively low-quality MODIS and ERA5 PWV through the constructed GRNNs. Using the proposed method, we generated more than 400 PWV maps that combine GNSS, MODIS, and ERA5 PWVs in North America in 2018. Results show that the overall bias, standard deviation (STD), and root-mean-square (RMS) error are 0.0 mm, 2.1 mm, and 2.2 mm for the improved MODIS PWV, and 0.0 mm, 1.6 mm, and 1.6 mm for the improved ERA5 PWV. Compared to the original MODIS and ERA5 PWV, the total improvements are 37.1% and 15.8% in terms of RMS. The RMS improvements are mainly contributed from the calibration of bias for the MODIS PWV and optimization for the ERA5 PWV. It also demonstrates that the original MODIS PWV tends to be greater than the GNSS PWV while the ERA5 PWV has very small biases. After calibration and optimization, the correlation coefficients between the modified PWV and the GNSS PWV are 0.96 for the MODIS PWV and 0.98 for the ERA5 PWV. The proposed method also diminishes the temporal and spatial variations in accuracy, generating homogeneous PWV products. Since the biases among the three datasets are well removed and data accuracies are improved to the same level, they are thus easily fused and jointly used.