Exploration of Deep-Learning-Based Error-Correction Methods for Meteorological Remote-Sensing Data: A Case Study of Atmospheric Motion Vectors

Abstract

Meteorological satellite remote sensing is important for numerical weather forecasts, but its accuracy is affected by many things during observation and retrieval, showing that it can be improved. As a standard way to measure wind from space, atmospheric motion vectors (AMVs) are used. They are separate pieces of information spread out in the troposphere, which gives them more depth than regular surface or sea surface wind measurements. This makes rectifying problems more difficult. For error correction, this research builds a deep-learning model that is specific to AMVs. The outcomes show that AMV observational errors are greatly reduced after correction. The root mean square error (RMSE) drops by almost 40% compared to ERA5 true values. Among these, the optimization of solar observation errors exceeds 40%; the discrepancies at varying atmospheric pressure altitudes are notably improved; the degree of optimization for data with low QI coefficients is substantial; and there remains potential for enhancement in data with high QI coefficients. Furthermore, there has been a significant enhancement in the consistency coefficient of the wind’s physical properties. In the assimilation forecasting experiments, the corrected AMV data demonstrated superior forecasting performance. With more training, the model can fix things better, and the changes it makes last for a long time. The results show that it is possible and useful to use deep learning to fix errors in meteorological remote-sensing data.

Keywords

• deep learning
• atmospheric motion vectors
• error correction
• data assimilation
• meteorological remote sensing