Atmospheric Motion Vector Retrieval by Using Deep Learning and Its Assimilation Applications.
Saved in:
| Title: | Atmospheric Motion Vector Retrieval by Using Deep Learning and Its Assimilation Applications. |
|---|---|
| Authors: | Chen, Keyi1 (AUTHOR) cky@cuit.edu.cn, Zhou, Renge1 (AUTHOR), Zhao, Zheng1 (AUTHOR), Li, Xiaoyong1,2 (AUTHOR) |
| Source: | Weather & Forecasting. Feb2026, Vol. 41 Issue 2, p443-464. 22p. |
| Subjects: | Deep learning, Data assimilation, Cyclone forecasting, Wind measurement, Atmospheric circulation, Weather forecasting, Recurrent neural networks, Satellite meteorology |
| Abstract: | This study presents a model that combines a convolutional long short-term memory (Conv-LSTM) neural network with a random sampling transfer learning module to retrieve atmospheric motion vectors (AMVs) at 15 height levels from the images of the infrared and water vapor channels from the Chinese Fengyun-4A and Japanese Himawari geostationary meteorological satellites. The quality of AMVs retrieved by this neural network model is assessed and compared to AMVs obtained through traditional algorithms in assimilating and forecasting experiments for high-impact weather. The study investigates the impacts of assimilating these data types on forecasts of typhoons and the southwest vortex. Results indicate that the neural network model achieves average wind speed errors of 2.3–4.3 m s−1, with correlation coefficients exceeding 0.9 at 15 height levels when being validated against reanalysis data. Compared to traditionally retrieved AMVs, Conv-LSTM AMVs are directly output at preassigned height levels without additional height assignment, provide complete wind fields for specific levels, and improve vertical resolution. Assimilation and forecasting experiments using the WRF data assimilation (WRFDA)-3DVAR system further demonstrate that assimilating AMVs retrieved by the neural network model improve the accuracy of typhoon track, intensity, and precipitation forecasts more effectively than those by traditional algorithms. In the case of southwest vortex forecasts, the key improvements are seen in wind field predictions and the correction of erroneous precipitation locations and magnitudes. Additionally, assimilating AMVs with higher vertical resolution from the neural network model further upgrades forecast performance. Overall, AMVs retrieved by using the neural network algorithm demonstrate significant advantages over traditional methods, offering a novel approach to AMV retrieval and providing new insights for AMV assimilation. Significance Statement: Accurate wind information at various altitudes is essential for weather forecasting, particularly for extreme events such as typhoons and heavy rainfall. This study presents a novel approach to retrieve the atmospheric motion vectors (AMVs) from infrared and water vapor channel images from geostationary satellites using a convolutional long short-term memory (Conv-LSTM) neural network. The proposed model overcomes several limitations of the traditional AMV retrieval methods, including errors in height assignment, vertical resolution, and data gaps. By delivering more accurate and complete wind fields at 15 height levels, the model significantly enhances the assimilation and forecasts of high-impact weather events like typhoons and the southwest vortex. This research not only improves the accuracy of weather predictions but also provides a promising pathway for incorporating deep learning techniques into operational weather forecasting systems, offering fresh insights into AMV retrieval and assimilation. [ABSTRACT FROM AUTHOR] |
| Copyright of Weather & Forecasting is the property of American Meteorological Society and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | This study presents a model that combines a convolutional long short-term memory (Conv-LSTM) neural network with a random sampling transfer learning module to retrieve atmospheric motion vectors (AMVs) at 15 height levels from the images of the infrared and water vapor channels from the Chinese Fengyun-4A and Japanese Himawari geostationary meteorological satellites. The quality of AMVs retrieved by this neural network model is assessed and compared to AMVs obtained through traditional algorithms in assimilating and forecasting experiments for high-impact weather. The study investigates the impacts of assimilating these data types on forecasts of typhoons and the southwest vortex. Results indicate that the neural network model achieves average wind speed errors of 2.3–4.3 m s−1, with correlation coefficients exceeding 0.9 at 15 height levels when being validated against reanalysis data. Compared to traditionally retrieved AMVs, Conv-LSTM AMVs are directly output at preassigned height levels without additional height assignment, provide complete wind fields for specific levels, and improve vertical resolution. Assimilation and forecasting experiments using the WRF data assimilation (WRFDA)-3DVAR system further demonstrate that assimilating AMVs retrieved by the neural network model improve the accuracy of typhoon track, intensity, and precipitation forecasts more effectively than those by traditional algorithms. In the case of southwest vortex forecasts, the key improvements are seen in wind field predictions and the correction of erroneous precipitation locations and magnitudes. Additionally, assimilating AMVs with higher vertical resolution from the neural network model further upgrades forecast performance. Overall, AMVs retrieved by using the neural network algorithm demonstrate significant advantages over traditional methods, offering a novel approach to AMV retrieval and providing new insights for AMV assimilation. Significance Statement: Accurate wind information at various altitudes is essential for weather forecasting, particularly for extreme events such as typhoons and heavy rainfall. This study presents a novel approach to retrieve the atmospheric motion vectors (AMVs) from infrared and water vapor channel images from geostationary satellites using a convolutional long short-term memory (Conv-LSTM) neural network. The proposed model overcomes several limitations of the traditional AMV retrieval methods, including errors in height assignment, vertical resolution, and data gaps. By delivering more accurate and complete wind fields at 15 height levels, the model significantly enhances the assimilation and forecasts of high-impact weather events like typhoons and the southwest vortex. This research not only improves the accuracy of weather predictions but also provides a promising pathway for incorporating deep learning techniques into operational weather forecasting systems, offering fresh insights into AMV retrieval and assimilation. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 08828156 |
| DOI: | 10.1175/WAF-D-24-0213.1 |
