Bias correction of global ensemble precipitation forecasts by Random Forest method

• Published in: Earth science informatics 2021-06, Vol.14 (2), p.677-689
• Main Authors: Zarei, Morteza, Najarchi, Mohsen, Mastouri, Reza
• Format: Article
• Language: English
• Subjects: Algorithms; Basins; Bias; Decision trees; Earth and Environmental Science; Earth Sciences; Earth System Sciences; Ensemble forecasting; Ensemble precipitation; Information Systems Applications (incl.Internet); Machine learning; Mathematical models; Meteorological satellites; Numerical models; Ontology; Precipitation; Precipitation forecasting; Research Article; Simulation and Modeling; Space Exploration and Astronautics; Space Sciences (including Extraterrestrial Physics; Weather forecasting
• ISSN: 1865-0473; 1865-0481
• DOI: 10.1007/s12145-021-00577-7

One of the most important topics in operational applications of precipitation forecasts is their improvement by bias correction methods. In this study, the ensemble precipitation forecasts of six numerical models from the TIGGE (THORPEX Interactive Grand Global Ensemble) database, associated with four basins in Iran for 2008–2018, were extracted and bias-corrected by the Quantile Mapping (QM) and Random Forest (RF) methods. Random Forest is a supervised machine learning algorithm made of an ensemble of decision trees. The results in all four basins demonstrated that most models had better skills in forecasting precipitation depth after bias correction using the RF method, compared to using the QM method and raw forecasts. In the dichotomous evaluation for 5-mm and 25-mm precipitation thresholds, all models improved their performance after bias correction. However, the QM performed slightly better than the RF. In probabilistic evaluations, significant improvements were observed after bias correction using the RF method, compared to using the QM in the models, and the reliability diagrams of the bias-corrected forecasts by the RF concentrated around the 1:1 line in all four basins. In seasonal evaluation, models had better probabilistic forecasts in autumn and winter than in spring and summer, and showed better scores in the lower tercile category than in the middle and upper tercile categories. In general, the improvement of a model’s performance after bias correction with the Random Forest method shows the importance of this method for operational application.