Assessing Uncertainty in Deep Learning Techniques that Identify Atmospheric Rivers in Climate Simulations

Atmospheric rivers (ARs) can be the difference between CA facing drought or hurricane-level storms. ARs are a form of extreme weather defined as long, narrow columns of moisture which transport water vapor outside the tropics. When they make landfall, they release the vapor as rain or snow. Convolutional neural networks (CNNs), a machine learning technique that uses filters to recognize features, are the leading computer vision mechanism for classifying multichannel images. CNNs have been proven to be effective in identifying extreme weather events in climate simulation output (Liu et. al. 2016, ABDA'16, http://bit.ly/2hlrFNV). Here, we compare three different CNN architectures, tuned with different hyperparameters and training schemes. We compare two-layer, three-layer, four-layer, and sixteen-layer CNNs' ability to recognize ARs in Community Atmospheric Model version 5 output, and we explore the ability of data augmentation and pre-trained models to increase the accuracy of the classifier. Because pre-training the model with regular images (i.e. benches, stoves, and dogs) yielded the highest accuracy rate, this strategy, also known as transfer learning, may be vital in future scientific CNNs, which likely will not have access to a large labelled training dataset. By choosing the most effective CNN architecture, climate scientists can build an accurate historical database of ARs, which can be used to develop a predictive understanding of these phenomena.