Statistical and Machine Learning Methods Applied to the Prediction of Tropical Rainfall

Large geographical biases exist in climate model simulations of tropical rainfall, such as an overly persistent double ITCZ, and the rain rate distribution of most climate models is far different than observed, with too much weak rain and not enough heavy rain. We explore the use of three advanced statistical and machine learning methods (a generalized linear model, random forest, and feedforward neural network) to predict the frequency occurrence, rain rate distribution, and spatial pattern of three tropical rain types (deep convective, stratiform, and shallow convective) observed by the radar onboard the GPM satellite over the tropical Pacific at 0.5° horizontal resolution. Three-hourly temperature and moisture fields from MERRA-2 were used as predictors. While all three methods perform reasonably well at predicting the frequency occurrence and spatial pattern of each rain type, the neural network is the only method able to produce rain rate distributions similar to observations, especially for the top 5-10% of observed values. However, the neural network took the most effort to train and has a relatively high root mean square error, suggesting that it sometimes assigns high rain rates to situations that in reality produce much weaker rain rates.