Machine Learning Downscaling of Regional Precipitation Patterns in Mountainous Terrain

Abstract

Coarse-resolution global and regional climate model output systematically misrepresents precipitation in
mountainous terrain, smoothing the sharp windward-leeward gradients and elevation-dependent intensification that
govern water availability for millions of people living downstream of mountain catchments. We trained a
convolutional super-resolution network to downscale 25 km regional climate model precipitation fields to a 1 km grid
over a 180 km by 140 km domain spanning the windward and leeward slopes of a mid-latitude coastal mountain
range, using a 34-year (1988-2021) archive of daily precipitation from 96 gauge stations and a terrain-corrected
radar mosaic as training and verification reference, and benchmarked the network against quantile mapping and
random forest baselines applied to the same predictor fields. The CNN incorporated static elevation, slope and
aspect layers alongside the coarse precipitation, wind and moisture flux predictors, and was trained with a hybrid
loss combining mean absolute error on wet-day accumulation with a categorical term penalising missed wet-day
occurrence. Across withheld stations, the CNN reduced domain-mean daily precipitation root-mean-square error by
34% relative to the uncorrected regional climate model and by 12% relative to the quantile-mapping baseline, with
the largest gains concentrated on windward slopes above 1200 m, where the uncorrected model underestimated
winter accumulation by a median of 28%. Skill degraded, however, at the highest, most poorly gauged elevations
(above 2000 m), where training data were sparsest and the network's reconstructed orographic enhancement was
only weakly constrained; in this band the random forest baseline, despite lower overall skill, produced more
physically consistent elevation gradients than the CNN. Leeward rain-shadow precipitation was reproduced with
good fidelity by all three methods once the coarse-model dry bias was removed. These results indicate that
CNN-based downscaling can meaningfully correct the orographic precipitation deficit endemic to coarse climate
model output, but that its advantage over simpler statistical methods is elevation-dependent and should not be
assumed to extend uniformly into sparsely observed high-elevation terrain without independent verification.