Precipitable water vapor fusion of MODIS and ERA5 based on convolutional neural network

Sensing precipitable water vapor (PWV) in the earth’s atmosphere is of significant importance for contributing to severe weather event monitoring and forecasting. PWV can be measured and retrieved with various techniques of different accuracies and spatial and temporal resolutions. In this study, we aim to achieve PWV estimates of high accuracy and resolutions by fusing the moderate resolution imaging spectroradiometer (MODIS) and the fifth generation of the European Centre for Medium-Range Forecasts (ECMWF) global reanalyzes (ERA5) with convolutional neural network (CNN). The region is focused on the west coast of America, and the experimental duration lasts for three years, from 2019 to 2021. The fused PWV values reveal a good agreement with the global navigation satellite system (GNSS) PWV, showing a mean absolute error (MAE) of 0.9 mm and a root mean square error (RMSE) of 1.3 mm. The fused PWV demonstrates significant improvement in accuracy compared to MODIS PWV estimates, with the MAE and RMSE reduced by 75.3% and 72.4%. Meanwhile, they also outperform the ERA5 PWV, revealing decreases of 34.6% in RMSE and 35.2% in MAE, respectively. Besides, the multilayer perceptron algorithm is also applied in PWV fusion as a comparison, which reveals a worse performance than the CNN fusion model. Furthermore, the fused PWV is less affected by seasonal variations and can provide more detailed spatial features compared to MODIS and ERA5 PWV. The proposed approach contributes to exploiting the full potential of MODIS and ERA5 products and offers promising potential for meteorological applications.