Using Visualization of Semi-Supervised Learning to Study Drivers of Distinct Atmospheric River Conditions Historically and in GCMs

Abstract:

Deep learning is a powerful tool to develop insights into the drivers of atmospheric rivers (ARs), the major source of extreme precipitation in regions including western North America. Despite advances in understanding ARs, predictability of timing and landfall details is illusive. A semi-supervised learning approach is taken to explore the relationship between large-scale meteorological patterns (LSMPs) on AR days in western North America and the features of the jet stream that drive them. Using reanalysis, we focus on the historical period and first select a broad set of days with AR conditions, using a version of the algorithm described in Goldenson et al. (2018). This algorithm uses feature detection methods to identify contiguous regions that meet empirical thresholds for quantity and spatial dimension of landfalling column-integrated water vapor features in the North Pacific. By limiting the dataset to those days with AR conditions, we reduce the dimensionality of the problem to better distinguish the variability of LSMPs on such days. To do so, a self-organizing map is used to group days with atmospheric river conditions into nodes that share common features in related meteorological fields -- simultaneously considering column water vapor, 700mb level zonal winds and 500mb heights in the NE Pacific as input fields. Finally, these LSMPs, labeled 1 through 6, are used as the labels in a neural net for classification based on 850mb zonal winds across the whole North Pacific. We use layerwise relevance propagation to visualize the regions the neural net relies upon to predict a particular LSMP, and make inferences about the aspects of the jet that are driving these variations between LSMPs. LSMPs for some of the sets of AR conditions display higher potential for predictability than others, for example the LSMP to which more of the weaker events are assigned is also the most internally variable and least predictable. For those LSMPs with better predictability, the neural net tends to rely on the jet exit region. Finally we evaluate GCMs for their ability to capture the character and frequency of the large-scale meteorological patterns, and use the neural net defined based on historical data to explain how the GCM biases and projected changes to the eddy-driven jet relate to their projected AR conditions.