Improving Radar-Based Precipitation Nowcasts with Machine Learning Using an Approach Based on Random Forest

A binary classification model is trained by random forest using data from 41 stations in Norway to predict the precipitation in a given hour. The predictors consist of results from radar nowcasts and numerical weather predictions. The results demonstrate that the random forest model can improve the precipitation predictions by the radar nowcasts and the numerical weather predictions. This study clarifies whether certain potential factors related to model training can influence the predictive skill of the random forest method. The results indicate that enforcing a balanced prediction by resampling the training datasets or lowering the threshold probability for classification cannot improve the predictive skill of the random forest model. The study reveals that the predictive skill of the random forest model shows seasonality, but is only weakly influenced by the geographic diversity of the training dataset. Finally, the study shows that the most important predictor is the precipitation predictions by the radar nowcasts followed by the precipitation predictions by the numerical weather predictions. Although meteorological variables other than precipitation are weaker predictors, the results suggest that they can help to reduce the false alarm ratio and to increase the success ratio of the precipitation prediction.