Generative Modeling of Future Precipitation Patterns

Neural networks are currently under utilized in climate science. This is unfortunate given that they have potential to learn the patterns underlying climate change from many types of long term data. We use PredNet, a convolutional recurrent neural network made to predict the next frame in a video, with an input of 33 years of United States precipitation “images” from 1981 to 2013 to predict that of 2014 (1). While our results were less revealing than was desired, we illustrate that next frame video prediction has the possibility to be used to great effect in field of climate science in the future.