To obtain accurate rainfall rate data covering a wide area, it is necessary to generate composite rainfall rate data from multiple radars because single radar observation has a limited range, accuracy deteriorated area due to rainfall attenuation, and missing area due to shielding by buildings and topography. Therefore, a composite method of rainfall rates for the MP-PAWR with other radars is needed. As conventional radars in Japan, there are X-band and C-band parabolic multi-parameter radars that compose eXtended RAdar Information Network (XRAIN). The objective of this study is to generate composite rainfall rate data from MP-PAWR data and XRAIN composite data which is more accurate than both. In this study, assuming that the raw observation data of all radars that compose XRAIN are not available by an MP-PAWR operator, we propose a method to selectively composite MP-PAWR data and XRAIN composite data on the Cartesian coordinate axes of latitude and longitude. The selection process of the proposed method used a gradient boosting decision tree (GBDT) model, one of the machine learning models, that estimates whether MP-PAWR data or XRAIN composite data is closer to ground rainfall rate in real-time.
For training and evaluation of the GBDT model, we used the data obtained by the MP-PAWR installed in Saitama University, actual XRAIN composite data, the numerical weather forecast data, and the data obtained by rain gauges composing Automated Meteorological Data Acquisition System (AMeDAS). The training condition of the GBDT model was the following. The training data consists of 375 hours from July to October 2019 – 2022, including various rainfall cases. The training label represents MP-PAWR data or XRAIN composite data, whichever has the smaller absolute error relative to observed 1-hour rainfall by the rain gauges. Also, to obtain a GBDT model with a high generalization performance for various rainfall cases, a total of 17 input features were used, including features that represent MP-PAWR and XRAIN observation conditions, the difference of rainfall rates between MP-PAWR data and XRAIN composite data, and numerical weather forecast data. As 1-hour rainfall observed by rain gauges was used to create the training label, each feature was pre-processed to match the sampling frequency of the rain gauges, e.g., by taking a 1-hour average. The evaluation was carried out under operational assumptions. Specifically, composite rainfall rate data were generated using features obtained every minute, and 1-hour average values of composite rainfall rate data were evaluated using observed 1-hour rainfall by rain gauges as true values.
Figure 1 shows the ratio of Root Mean Squared Error (RMSE) for MP-PAWR data, the composite rainfall rate data generated by the maximum method (Max.) which selects the maximum value from MP-PAWR data and XRAIN composite data, and the composite rainfall rate data generated by the proposed method, to the RMSE of XRAIN composite data for three cases: convective rainfall, typhoon, and stratiform rainfall. Here, the maximum method was evaluated as a reference of a simple and typical composite method. As shown in Figure 1, the RMSE of the composite rainfall rate generated by the maximum method worsen compared to XRAIN composite data for the convective rainfall case and stratiform rainfall case. On the other hand, the proposed method can generate more accurate composite rainfall rate data than XRAIN composite data in all cases. Specifically, the composite rainfall rate data generated by the proposed method improves RMSE by 13% for the convective rainfall case, 3% for the typhoon case, and 18% for the stratiform rainfall case, respectively, compared to XRAIN composite data.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner