Author(s): Payam Heidarian; Matteo Benetti; Marco Pilotti; Esmail Ghaemi; Marco Gabella; Alireza Khoshkonesh

Linked Author(s):

Keywords: Rainfall Prediction; Remote Sensing; Machine Learning; Forecast Accuracy; RF; DT; SVM; ANN; ICON

Abstract: Abstract Accurate rainfall prediction enables disaster preparedness, yet numerical weather models often struggle with precision. This study evaluated machine learning (ML) techniques to enhance forecast accuracy from the ICOsahedral Non-hydrostatic (ICON) model, focusing on the Franciacorta region in Northern Italy. External stations supported enhanced interpolation, and both one- and two-hour forecasts were subject to analysis. The methodology encompassed inverse distance weighting interpolation, statistical distribution fitting, and Monte Carlo simulation. The strategic selection of training and testing datasets revealed consistent outperformance of the Random Forest and Decision Tree models for rainfall estimations. Specifically, these models achieved a normalized Root Mean Squared Error (RMSE) reduction of up to 15% in optimal training scenarios. In contrast, Artificial Neural Networks (ANN) demonstrated the highest RMSE, with increases averaging 10%. The accuracy of predictions showed sensitivity to the optimization data. However, specific training data selections improved RMSE by approximately 5%. Furthermore, two-hour-ahead forecasts displayed a 3% improvement in RMSE compared to one-hour-ahead forecasts. These findings highlight the potential of ML, particularly Random Forest and Decision Tree, to refine rainfall predictions in this region. Model performance depended on meticulous training data selection. Future research with larger datasets will refine model optimization and enhance broader applicability.

DOI: https://doi.org/10.64697/978-90-835589-7-4_41WC-P1871-cd

Year: 2025