5:15 PM - 7:15 PM
[ACG41-P17] Application of AI for Atmospheric Motion Vector Estimation Using Geostationary Meteorological Satellite Images
Keywords:Earth Observation Satellite, Atmospheric Motion Vector, Himawari Geostationary Meteorological Satellite , Convolutional Neural Network
In recent years, the increasing severity of meteorological disasters due to climate change demands more accurate numerical weather prediction (NWP). Among the initial conditions used in NWP, wind is one of the variables that lacks sufficient accuracy. Wind speed and direction vary with altitude. Obtaining global wind distribution is crucial for NWP. Radiosondes provide us highly accurate wind data, however most observatories are located on land, and their coverage over the ocean are highly limited. Therefore, satellite observations are indispensable for global coverage. Multiple-layer vector wind, called the atmospheric motion vector (AMV), can be retrieved from cloud and water vapor motions derived from geostationary or polar-orbit satellite images. AMV estimates face challenges related to height assignment. To improve the accuracy of height assignment, research and development utilizing AI technology have become increasingly active.
The objective of this study is to address the current limitations of AMV by developing an AI model that utilizes deep learning to improve AMV estimation from satellite images. To develop the AI model for AMV estimation, we integrated radiosonde observation data provided by the University of Wyoming and Himawari geostationary meteorological satellite data. The radiosonde data include latitude, longitude, pressure altitude, wind direction, and wind speed, while the Himawari data consist of cloud and water vapor images along with AMV (hereafter referred to Himawari AMVs). A dataset consisting of 7,357 observations collected from 16 observatories across Japan in 2023 was created by combining the two data sources. By augmenting the provided images through rotation and flipping, the dataset size was expanded from 7,357 to 58,856 samples. The training data consisted of data from 15 observatories excluding the Wakkanai observatory, while the validation data included 5,528 samples from the Wakkanai observatory. The input data is created from two consecutive satellite images of 33×33 pixels, and the ground truth was the radiosonde wind data. Evaluation metrics were defined as follows: cosine similarity (C) representing wind direction accuracy, length similarity (L) representing wind speed accuracy, and an overall score (= C/2 + L/2, with a maximum value of 1). Separate CNN models were used to estimate wind direction and wind speed. Further details of the CNN models will be provided in the presentation date.
Preliminary experimental results showed that the AI-driven AMV wind direction achieved high accuracy, with a large proportion of data exceeding C > 0.9, surpassing the accuracy of Himawari AMV. On the other hand, wind speed estimation showed a lower proportion of data exceeding L > 0.9, indicating that further improvements to the CNN model are necessary. The preliminary results suggest that AI technology has the potential to improve the quality of AMV data. We have plans to integrate altitude data from the Earth observation satellite Aeolus and develop AI models tailored to different altitude ranges to address the challenges identified in the initial experiments.
The objective of this study is to address the current limitations of AMV by developing an AI model that utilizes deep learning to improve AMV estimation from satellite images. To develop the AI model for AMV estimation, we integrated radiosonde observation data provided by the University of Wyoming and Himawari geostationary meteorological satellite data. The radiosonde data include latitude, longitude, pressure altitude, wind direction, and wind speed, while the Himawari data consist of cloud and water vapor images along with AMV (hereafter referred to Himawari AMVs). A dataset consisting of 7,357 observations collected from 16 observatories across Japan in 2023 was created by combining the two data sources. By augmenting the provided images through rotation and flipping, the dataset size was expanded from 7,357 to 58,856 samples. The training data consisted of data from 15 observatories excluding the Wakkanai observatory, while the validation data included 5,528 samples from the Wakkanai observatory. The input data is created from two consecutive satellite images of 33×33 pixels, and the ground truth was the radiosonde wind data. Evaluation metrics were defined as follows: cosine similarity (C) representing wind direction accuracy, length similarity (L) representing wind speed accuracy, and an overall score (= C/2 + L/2, with a maximum value of 1). Separate CNN models were used to estimate wind direction and wind speed. Further details of the CNN models will be provided in the presentation date.
Preliminary experimental results showed that the AI-driven AMV wind direction achieved high accuracy, with a large proportion of data exceeding C > 0.9, surpassing the accuracy of Himawari AMV. On the other hand, wind speed estimation showed a lower proportion of data exceeding L > 0.9, indicating that further improvements to the CNN model are necessary. The preliminary results suggest that AI technology has the potential to improve the quality of AMV data. We have plans to integrate altitude data from the Earth observation satellite Aeolus and develop AI models tailored to different altitude ranges to address the challenges identified in the initial experiments.