17:15 〜 18:45
[MGI26-P03] Machine learning based post-processing to enhance daily rainfall forecasts of July 2019-2021 over the Kyushu region focusing on surface conditions at remote moisture sources
キーワード：Kyushu, Rainfall forecast, Moisture sources, FLEXPART, Machine learning, XGBoost
The Kyushu region in Japan is affected by extreme rainfall events during the rainy season, resulting in devastating social and economic damages. A significant increase in the frequency and intensity of such events has challenged accurate rainfall predictions, partially due to anomalies over moisture sources. These force complex changes in the atmosphere, disrupting climatological evaporation, moisture supply, and rainfall patterns. Advanced numerical weather forecasting models utilize data assimilation to reduce errors, but their performance is affected by the amount and completeness of available real-time observations. A machine learning-based method to enhance rainfall forecasts is introduced, which focuses on surface conditions only over remote moisture sources and incomplete data.
The post-processing method is two-fold. First, remote moisture sources associated with the Kyushu region were identified using FLEXPART (FLEXible PARTicle dispersion model), a Lagrangian transport and dispersion model capable of determining particle trajectories and states related to various fields, including but not limited to water cycle studies. Secondly, a machine learning model using XGBoost algorithm was used to establish relationships between moisture source surface parameters and rainfall over the Kyushu region. By applying the post-processing method, which explicitly utilizes surface parameters over moisture sources, a reduction in root mean squared error of up to 37% was achieved for the NCEP GFS (National Centers for Environmental Prediction Global Forecast System) rainfall forecast for the month of July of 2019 to 2021, as an average of lead-times from 1 to 10 days.
FLEXPART was simulated with meteorological data from NCEP GDAS/FNL (Global Data Assimilation System Final). Particles residing over the Kyushu region were tracked backward for up to 10days to identify daily moisture sources. The time duration between moisture originating at a source and reaching the Kyushu region, was considered the lag between the occurrence of a certain surface condition at a source and its respective effect emerges in rainfall over Kyushu. Here, the surface conditions promoting high evaporation are expected to result in higher rainfall. The results from FLEXPART show that during the month of July, the Indian Ocean, South China Sea, Northwestern Pacific Ocean, and the Sea of the Philippines act as major moisture sources for Kyushu. Temperature and wind speed were selected as surface parameters due to their high influence on the evaporation governing the moisture supply. Surface data from NCEP GDAS/FNL reanalysis dataset masked with respective moisture source regions, rainfall forecast from the NCEP GFS, and rainfall observations from IMERG (Integrated Multi-satellitE Retrievals for GPM) were used for XGBoost model, which was evaluated with K-Fold cross-validation. XGBoost algorithm was selected due to its ability to deal with missing data so that changes in moisture source regions can be effectively handled.
As a conclusion, utilizing explicit surface data extraction from remote moisture sources led to increased accuracy in rainfall forecasts, surpassing the accuracy achieved without explicit data extraction, affirming the effectiveness of focusing on remote moisture sources.
The post-processing method is two-fold. First, remote moisture sources associated with the Kyushu region were identified using FLEXPART (FLEXible PARTicle dispersion model), a Lagrangian transport and dispersion model capable of determining particle trajectories and states related to various fields, including but not limited to water cycle studies. Secondly, a machine learning model using XGBoost algorithm was used to establish relationships between moisture source surface parameters and rainfall over the Kyushu region. By applying the post-processing method, which explicitly utilizes surface parameters over moisture sources, a reduction in root mean squared error of up to 37% was achieved for the NCEP GFS (National Centers for Environmental Prediction Global Forecast System) rainfall forecast for the month of July of 2019 to 2021, as an average of lead-times from 1 to 10 days.
FLEXPART was simulated with meteorological data from NCEP GDAS/FNL (Global Data Assimilation System Final). Particles residing over the Kyushu region were tracked backward for up to 10days to identify daily moisture sources. The time duration between moisture originating at a source and reaching the Kyushu region, was considered the lag between the occurrence of a certain surface condition at a source and its respective effect emerges in rainfall over Kyushu. Here, the surface conditions promoting high evaporation are expected to result in higher rainfall. The results from FLEXPART show that during the month of July, the Indian Ocean, South China Sea, Northwestern Pacific Ocean, and the Sea of the Philippines act as major moisture sources for Kyushu. Temperature and wind speed were selected as surface parameters due to their high influence on the evaporation governing the moisture supply. Surface data from NCEP GDAS/FNL reanalysis dataset masked with respective moisture source regions, rainfall forecast from the NCEP GFS, and rainfall observations from IMERG (Integrated Multi-satellitE Retrievals for GPM) were used for XGBoost model, which was evaluated with K-Fold cross-validation. XGBoost algorithm was selected due to its ability to deal with missing data so that changes in moisture source regions can be effectively handled.
As a conclusion, utilizing explicit surface data extraction from remote moisture sources led to increased accuracy in rainfall forecasts, surpassing the accuracy achieved without explicit data extraction, affirming the effectiveness of focusing on remote moisture sources.