Rainfall nowcasting by integrating radar and rain gauge data with machine learning for Ischia Island, Italy

Abstract

Study region Ischia Island in southern Italy. Study focus: This study investigates the efficacy of a Random Forest (RF) regression model for short-term rainfall nowcasting by integrating rainfall radar data and rain gauge measurements to enhance predictive accuracy at rain gauge stations. The hyperparameters of the RF regression model have been adjusted to ensure reliable rainfall nowcasting with lead times suitable for early warning purposes. K-Fold Cross-Validation has been adopted to minimize overfitting, and the model has been trained for four operational scenarios, offering valuable insights into optimizing the RF model's performance. New hydrological insights for the region: The results reveal that the combination of rain gauge data and radar data (divided into East and West sectors) ensures the best performance, improving the RMSE by approximately 13 % compared to using rain gauge data alone. The model effectively captures rainfall patterns and demonstrates robust predictive capability up to a 120-minute lead time. The optimal lag times for the rain gauge and rainfall radar data are 0, −10, and −20 min, and 0, −10, −20, and −30 min, respectively. Additionally, results indicate that contribution of the rain gauge data surpasses that of the rainfall radar data and East rainfall radar data plays a more significant role in prediction than data from the West. • Rainfall nowcasting via machine learning (Random Forest). • Setting model structure to maximize nowcasting accuracy. • Integration of radar and rain gauge data for training the model. • Analysing the feature importance of several types of data. • Evaluation of model performance for several lead times.

Topics & Concepts