A study on loss function against data imbalance in deep learning correction of precipitation forecasts

Abstract

Sample imbalance has always been an important issue to be solved in heavy precipitation forecasts. Introducing appropriate constraints into the loss function of a machine learning model can mitigate the negative impact of sample imbalance on model training. This study explores the role of Dice loss function in dealing with sample imbalance and verifies its application on the correction of heavy precipitation forecasts using the U-Net neural network. On this basis, the application of ordinal classification in forecasts correction is further verified. The results show that the concept of Dice loss is highly similar to that of threat score, which can suppress the negative impact of sample imbalance on heavy precipitation forecasts by ignoring most of the events that are insensitive to heavy precipitation forecasts. Model correction experiments further indicate that the correction models trained with weighted cross entropy, differentiable threat score loss and Dice loss (WCE, DTS and Dice models) improve the precipitation forecasts of ECMWF (European Centre for Medium-Range Weather Forecasts) model. For the forecasts of all precipitation levels, Dice model effectively suppresses false alarms while improving hits, whereas WCE model greatly promotes false alarms while significantly improving hits. In contrast, DTS model significantly suppresses large precipitation false alarms while improving its hits, but causes substantial small precipitation false alarms while significantly promoting its hits. As a result, for large precipitation forecasts, both Dice and DTS models perform significantly better than WCE model, while for small precipitation forecasts, Dice model performs better than DTS and WCE models. Furthermore, the correction model combining ordinal classification and Dice loss exhibits the best forecasting skill. This model further suppresses the false alarms of large precipitation while improving its hits.