Forecasting rainfall events based on zenith wet delay time series utilizing eXtreme Gradient Boosting (XGBoost)

Accurate rainfall prediction is vital for mitigating flood and storm disasters as well as for planning agricultural activities and water resources management. GNSS observations enable the estimation of atmospheric water vapor content through the Zenith Wet Delay (ZWD) value, where previous studies indicate a strong correlation between the ZWD-derived indicators and rainfall events. However, specifying these indicators is challenging due to the spatial variability of precipitation and the location of GNSS stations. While many studies have integrated meteorological parameters with GNSS-derived Zenith Total Delay (ZTD) values to enhance prediction accuracy, the scarcity of meteorological instruments at GNSS stations remains a limitation. In this study, we employed ZWD-derived features and utilized the eXtreme Gradient Boosting (XGBoost) classification method to predict rainfall events. Ten parameters (including station latitude, longitude, elevation, ZWD monthly anomaly, ZWD slope, ZWD maximum, maximum ZWD derivative, month, hour, and precipitation flag) were used as features in the input layer of the considered XGBoost model. For training, data from 40 GNSS stations spanning five consecutive years (2016 to 2020) in the eastern United States of America were analyzed to derive the required features from 4-hour ZWD time series. To evaluate the proposed method, estimated rainfall was compared with the observations of weather stations during 2021. Furthermore, the results of five GNSS stations (not included in the training) were compared with the regional rainfall events of 2016 to 2021. Our results indicate that the proposed method achieves a mean True Forecast Rate (TFR) and a mean False Forecast Rate (FFR) of approximately 0.75 and 0.15, respectively, demonstrating performance comparable to studies incorporating meteorological parameters.