11:35 AM - 11:55 AM
[U12-03] AI-based Meteorological State Retrieval with Uncertainty Quantification from Satellite Observation
★Invited Papers
Keywords:Machine learning, air temperature, moisture, geostationary satellite, uncertainty quantification
The artificial intelligence/machine learning (AI/ML) have presented new capabilities for representing the statistical relationship between multispectral images and atmospheric state utilizing rich information contained in high-dimensional, spatial/spectral satellite measurement data. Here, we report how current AI/ML techniques are useful in meteorological applications by demonstrating capabilities and limitations. Particularly, we apply deep neural networks (DNNs) to measurements by geostationary satellite, which is attractive for weather monitoring of severe weather and forecasting because the observation is spatially and temporally uniform and frequent.
Measurements of water vapor over oceanic regions are highly demanding because heavy rainfall in coastal areas is often associated with a massive amount of water vapor transport in the middle/lower troposphere over the ocean. The infrared measurement generally loses its sensitivity to the lower parts of thick clouds, so that a three-dimensional (3D) profile of temperature and humidity by hyperspectral infrared sounders such as the Atmospheric InfraRed Sounder (AIRS) is primarily limited to clear-sky atmospheric volume. We recently proposed a convolutional deep neural network (DNN)-based all-weather atmospheric sounding method (hereafter, HimMet) of air temperature and relative humidity with uncertainty estimate from measurements taken from a multispectral imager onboard the geostationary satellite Himawari-8. As the HimMet utilizes spatial features (2D) in the image in addition to multispectral features, this is a new technique of imager sounding. After a pretraining of DNN in a self-supervised learning way, the DNN was fitted to radiosonde measurements acquired for two years through maximum likelihood estimation to jointly diagnose uncertainty estimates, assuming the probabilistic distribution of error as Gaussian. Because the radiosonde stations distributes sparsely and nonuniformly, the training is performed by the weighted-average loss function with large weights around radiosonde stations. It is shown that highly accurate temperature and humidity retrieval was obtained under all-weather conditions owing to the use of spatial and multispectral features. Validation using the ERA5 reanalysis as a reference revealed that overall RMSE was 1.2 (1.6) K for air temperature and 10.2% (10.4%) for the relative humidity over the ocean (land). Humidity estimation by HimMet was remarkably more accurate than that of the AIRS standard product when testing was done for mostly clear-sky portions of the atmosphere. Validation using radiosonde measurements for 2 years under all-weather conditions yielded an RMSE of 1.43 K for air temperature and 11.8% for relative humidity. The validation also showed that the uncertainty estimate is quantitatively reasonable. Although slight underestimation of error standard deviation prevails, it still provided a good diagnostics of estimation uncertainty under varying conditions including cloud effects, land/ocean type, and pressure level. We also found apparent hot/cold bias when the air temperature is abnormally low/high and wet/dry bias when the relative humidity is close to 0%/100%. It should be noted that these biases cannot be captured by the present method of uncertainty quantification.
This kind of AI/ML method is promising and useful in resolving several problems because of their capabilities of rapid-processing. Important applications include atmospheric stability diagnostics, early convective initiation detection and severe weather preconditioning, and the assimilation of meteorological states from geostationary satellites into numerical weather prediction.
Measurements of water vapor over oceanic regions are highly demanding because heavy rainfall in coastal areas is often associated with a massive amount of water vapor transport in the middle/lower troposphere over the ocean. The infrared measurement generally loses its sensitivity to the lower parts of thick clouds, so that a three-dimensional (3D) profile of temperature and humidity by hyperspectral infrared sounders such as the Atmospheric InfraRed Sounder (AIRS) is primarily limited to clear-sky atmospheric volume. We recently proposed a convolutional deep neural network (DNN)-based all-weather atmospheric sounding method (hereafter, HimMet) of air temperature and relative humidity with uncertainty estimate from measurements taken from a multispectral imager onboard the geostationary satellite Himawari-8. As the HimMet utilizes spatial features (2D) in the image in addition to multispectral features, this is a new technique of imager sounding. After a pretraining of DNN in a self-supervised learning way, the DNN was fitted to radiosonde measurements acquired for two years through maximum likelihood estimation to jointly diagnose uncertainty estimates, assuming the probabilistic distribution of error as Gaussian. Because the radiosonde stations distributes sparsely and nonuniformly, the training is performed by the weighted-average loss function with large weights around radiosonde stations. It is shown that highly accurate temperature and humidity retrieval was obtained under all-weather conditions owing to the use of spatial and multispectral features. Validation using the ERA5 reanalysis as a reference revealed that overall RMSE was 1.2 (1.6) K for air temperature and 10.2% (10.4%) for the relative humidity over the ocean (land). Humidity estimation by HimMet was remarkably more accurate than that of the AIRS standard product when testing was done for mostly clear-sky portions of the atmosphere. Validation using radiosonde measurements for 2 years under all-weather conditions yielded an RMSE of 1.43 K for air temperature and 11.8% for relative humidity. The validation also showed that the uncertainty estimate is quantitatively reasonable. Although slight underestimation of error standard deviation prevails, it still provided a good diagnostics of estimation uncertainty under varying conditions including cloud effects, land/ocean type, and pressure level. We also found apparent hot/cold bias when the air temperature is abnormally low/high and wet/dry bias when the relative humidity is close to 0%/100%. It should be noted that these biases cannot be captured by the present method of uncertainty quantification.
This kind of AI/ML method is promising and useful in resolving several problems because of their capabilities of rapid-processing. Important applications include atmospheric stability diagnostics, early convective initiation detection and severe weather preconditioning, and the assimilation of meteorological states from geostationary satellites into numerical weather prediction.