Disentangling Regional Drivers of Top Antarctic Snowfall Days With a Convolutional Neural Network

Abstract

Snowfall is the primary contributor to Antarctic surface mass balance. Identifying regional-scale mechanisms that drive heavy snowfall provides context for changes in Antarctic surface mass balance in a warmer climate. We compare drivers of top snowfall days across five Antarctic regions using machine learning and traditional synoptic diagnostics. A convolutional neural network identifies top snow days with an accuracy of 92%–94% per region when trained on just atmospheric moisture and low-level meridional wind, highlighting the importance of atmospheric river-like structures to top Antarctic snowfall days. The network's skill depends mainly on low-level wind in East Antarctica and atmospheric moisture in West Antarctica, suggesting that dynamic processes are comparatively more important in driving East Antarctic top snowfall days. We leverage the quasi-geostrophic omega equation to identify mechanisms for ascent and snowfall production, and we find that East Antarctic top snowfall days feature stronger synoptic-scale forcing for ascent compared to West Antarctica.