A machine learning approach to model over ocean tropical cyclone precipitation

Abstract

Extreme rainfall found in tropical-cyclones (TCs) is a risk for human life and property in many low to mid latitude regions. Probabilistic modeling of TC rainfall in risk assessment and forecasting can be computational expensive, and existing models are largely unable to model key rainfall asymmetries such as rain-bands and extra-tropical transition. Here, a machine learning-based framework is developed to model over-water TC rainfall for the North Atlantic basin. First, a catalog of high-resolution TC precipitation simulations for 26 historical events is assembled for the North Atlantic basin using the Weather Research and Forecasting (WRF) Model. The simulated spatial distribution of rainfall for these historical events are then decomposed via principal component analysis (PCA), and quantile regression forest models (QRF) are trained to predict the conditional distributions of the first five principal component (PC) weights. Conditional distributions of rain rate levels are estimated separately using historical satellite data and a QRF model. With these models, probabilistic predictions of rainfall maps can be made given a set of storm characteristics and local environmental conditions. The model is able to capture storm total rainfall compared to satellite observations with a correlation coefficient of 0.96 and r-squared value of 0.93. Additionally, the model shows good accuracy in modeling hourly total rainfall compared to satellite observations. Rain rate maps predicted by the model are also compared to historical satellite observations and to the WRF simulations during cross-validation, and the spatial distribution of estimates captures rainfall variability consistent with TC rain-bands, wavenumber asymmetries and possibly extra-tropical transition.