Transcript – Validating Rainfall Parameterizations in Climate Models Using Predictive Empirical Analysis

In this study, we used statistics and machine learning methods to study rainfall. One of the biggest problems in climate modelling is that models do not simulate rain very well, as we’ve seen figure one in some places, it rains too much in some places at Rangitoto. One way to fix this is to improve the simulation of rainfall using better perambulations of rainfall.

We want to construct perambulations based on data, so we use satellite data from the GPS satellite and divide terrain into three types: stratifying convective and shallow convective. You focus on the tropical Pacific Ocean and we get to use one year of data 2017 to train our models and another year of data 2018 to validate the models. We use three different types of models, a generalized linear model, which is a statistical model, a random forest model, which is a decision tree model and a neural network. A deep fit for what neural network with five layers of history are hidden. We also compare the three models the climate model, the community atmospheric model. As we see in table one, all three models predict the three different types of rain quite well. But what of more interest to us is prediction of extreme rainfall or high rain rates as shown in figure for the x axis shows the rain rates and millimeters, but are the y axis shows the probability density and each of the three panels corresponds to the three rain types. The solid line is observations, it’s what we see is that the glam and RF do not predict high rate rain rates very well, but the neural network predicts the high rain rates quite well. So, you can simulate weather and climate extremes much better than the other models. So, what we learn from the study is that the machine learning approaches can help address some of the biases that we see in climate models. Thank you.