5:15 PM - 7:15 PM
[ACG41-P05] Preliminary Validation of the GSMaP Precipitation Gap-Filling Algorithm Using Machine Learning
Keywords:Global Precipitation Measurement, Machine Learning, Satellite Observation
The Global Satellite Mapping of Precipitation (GSMaP) utilizes high-frequency infrared (IR) cloud observations from geostationary meteorological satellites (GEO) to fill the observational gaps in the precipitation network provided by passive microwave (PMW) radiometers onboard multiple polar-orbiting satellites. Specifically, cloud motion vectors are calculated from continuous GEO observations to predict the movement of precipitation clouds observed by PMW at previous and later times. Additionally, to account for changes in precipitation intensity due to cloud development and dissipation during movement, precipitation intensity is adjusted based on the cloud-top temperature after advection. However, the precipitation intensity adjustment method based on cloud-top temperature has significant errors for orographic precipitation and cirrus clouds that do not accompany strong precipitation.
Nguyen et al. (2020) reported that the accuracy of IR-based precipitation estimation can be significantly improved by using machine learning to classify precipitation cloud types in advance and constructing individual brightness temperature–precipitation intensity fitting functions for each cloud type. In this study, we attempted to dramatically enhance the precipitation estimation accuracy of GSMaP by incorporating this machine-learning-based cloud classification approach into the GSMaP precipitation interpolation algorithm.
The methodology of the previous study is based on the PERSIANN-Cloud Classification System (CCS) developed by Houg et al. (2004). First, cloud patches are detected using an IR brightness temperature threshold, and spatial distribution parameters of IR brightness temperature within each cloud patch are extracted as features. A self-organizing map is then created by vectorizing these features, enabling visualization of precipitation clouds with similar characteristics. In this study, we constructed a self-organizing map for precipitation cloud patches over Japan using multi-band IR observations from Himawari-8. As a result, we obtained consistent findings with previous studies, which indicated that the temperature gradient around the minimum brightness temperature is effective in classifying orographic precipitation. Furthermore, while the previous study used only a single IR band, we found that the cloud optical thickness index from split-window observations and the brightness temperature difference in the water vapor band are effective indicators for classifying orographic precipitation.
By reconstructing brightness temperature–precipitation intensity fitting functions for each classified precipitation type, we achieved a substantial reduction in precipitation intensity estimation errors. When applying this type-specific fitting approach to a precipitation case along the southern coast of Honshu, the underestimation of orographic precipitation occurring on mountain slopes was significantly reduced.
This study was supported by the Japan Aerospace Exploration Agency (JAXA) as part of the collaborative research project "Development of High-Resolution GSMaP Algorithm."
Nguyen et al. (2020) reported that the accuracy of IR-based precipitation estimation can be significantly improved by using machine learning to classify precipitation cloud types in advance and constructing individual brightness temperature–precipitation intensity fitting functions for each cloud type. In this study, we attempted to dramatically enhance the precipitation estimation accuracy of GSMaP by incorporating this machine-learning-based cloud classification approach into the GSMaP precipitation interpolation algorithm.
The methodology of the previous study is based on the PERSIANN-Cloud Classification System (CCS) developed by Houg et al. (2004). First, cloud patches are detected using an IR brightness temperature threshold, and spatial distribution parameters of IR brightness temperature within each cloud patch are extracted as features. A self-organizing map is then created by vectorizing these features, enabling visualization of precipitation clouds with similar characteristics. In this study, we constructed a self-organizing map for precipitation cloud patches over Japan using multi-band IR observations from Himawari-8. As a result, we obtained consistent findings with previous studies, which indicated that the temperature gradient around the minimum brightness temperature is effective in classifying orographic precipitation. Furthermore, while the previous study used only a single IR band, we found that the cloud optical thickness index from split-window observations and the brightness temperature difference in the water vapor band are effective indicators for classifying orographic precipitation.
By reconstructing brightness temperature–precipitation intensity fitting functions for each classified precipitation type, we achieved a substantial reduction in precipitation intensity estimation errors. When applying this type-specific fitting approach to a precipitation case along the southern coast of Honshu, the underestimation of orographic precipitation occurring on mountain slopes was significantly reduced.
This study was supported by the Japan Aerospace Exploration Agency (JAXA) as part of the collaborative research project "Development of High-Resolution GSMaP Algorithm."