Item

Abstract

Meteorological variables required for driving snow energy-balance models are now widely available at ~1km scales from operational convection-permitting forecast models but not at ~100 m snowdrift-permitting scales. Simple statistical approaches to downscaling with lapse rates or linear regression are unreliable for variables such as precipitation rate and wind speed that do not have straightforward relationships with elevation, but more reliable downscaling with atmospheric dynamics models is computationally expensive. We explore the application of convolutional neural networks as a tool for downscaling that is more informative than simple statistical methods and faster than dynamical downscaling. Applications are illustrated by modelling late-lying snowdrifts in the Cairngorm Mountains of Scotland.