We explore the use of three advanced statistical and machine learning methods (a generalized linear model, random forest, and neural network) to predict the occurrence and rain rate distribution of three tropical rain types (deep convective, stratiform, and shallow convective) observed by the radar onboard the GPM satellite over the West Pacific at three-hourly, 0.5-degree resolution. Temperature and moisture profiles from MERRA-2 were used as predictors. All three methods perform reasonably well at predicting the occurrence and rain rate distribution of each rain type. However, none of the methods obviously distinguish themselves from one another and each method still has issues with predicting rain too often and not fully capturing the high end of the rain rate distributions, both of which are common problems in climate models.

Climate model simulations of rainfall in the tropics suffer from pervasive biases, and that can lead to degraded climate simulations in other regions as well. Over the past two decades, high-resolution satellite measurements of tropical rainfall have become available. These data are most commonly used to constrain physics-based climate models by validating statistical properties of rainfall such as means and variances. However, the satellite data contain a wealth of spatiotemporal information on sub-diurnal timescales that can be used to construct predictive models. This study explores the feasibility of predicting rainfall from atmospheric state using a hierarchy of empirical models. Our empirical approach is similar to the physics-based approach in that vertical profiles of atmospheric state at a particular instant of time serve as the predictors, and rainfall over a subsequent time period is the predictand. However, we allow the empirical model to “learn” from data to determine the model parameters. Empirical Orthogonal Function (EOF) decomposition is applied to vertical profiles from NASA MERRA-2 reanalysis to select the dominant predictor modes at analysis time 00 UTC. Rain predictions for the subsequent 6-hour period (00-06 UTC) are separated into different types from TRMM satellite data: stratiform, deep convective, and shallow convective. For each rain type, two generalized linear statistical models (logistic regression for rain occurrence and gamma regression for rain amount) are trained on 2003 data and used to predict during 2004. The results show that the statistical approach can predict spatial patterns and amplitudes of tropical rainfall in the time-averaged sense. The first EOF of humidity and the second EOF of temperature contribute most to prediction. In addition to generalized linear models, other common machine learning techniques (support vector machine and random forest) are compared. Furthermore, marginal nonlinear relationships between predictand and individual predictor are explored via a nonparametric regression technique. Interestingly, incorporating the identified marginal nonlinear relationship into the generalized linear model does not improve the prediction, suggesting that these marginal nonlinear effects are explained by other predictors in the model.

This study further evaluates the modeling approach of Jia et al. (2019) to investigate the potential effects of SST mesoscale variability on the atmospheric dynamics. The approach employs a global atmospheric circulation model coupled to a slab ocean model to produce two ensembles of simulations: one in which the ocean exhibits realistic SST mesoscale variability, and another in which the SST mesoscale variability is suppressed. The latter ensemble is produced by spatially filtering the SST analyses used for the estimation of the oceanic heat flux and the specification of the SST initial condition. The results of the present study, which focuses on the processes of the North Pacific, suggest that while the modeling approach yields the desired SST differences between the two ensembles at the mesoscales, it also introduces SST differences at the large scales that become the primary driver of the large scale differences in the simulated atmospheric flow. Diagnostics based on the eddy kinetic energy indicate that the large scale differences of the atmospheric flow lead to major differences in the dynamics of the jet stream and storm track. Because the large scale SST differences between the two ensembles are primarily driven by the differences between the prescribed estimates of the oceanic heat fluxes, finding a proper pair of those estimates is a necessary condition for the experiment design to detect the atmospheric response to SST mesoscale variability. The paper concludes with proposing a new strategy for the estimation of the oceanic heat fluxes.

We explore the use of three advanced statistical and machine learning methods (a generalized linear model, random forest, and neural network) to predict the occurrence and rain rate distribution of three tropical rain types (deep convective, stratiform, and shallow convective) observed by the radar onboard the GPM satellite over the West Pacific. Three-hourly temperature and moisture fields from MERRA-2 were used as predictors. While all three methods perform reasonably well at predicting the occurrence 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.