Machine Learning Downscaling of Regional Precipitation Patterns in Mountainous Terrain

Eszter Jonsdottir, Andrew Charles Meltzer, Joao Bakker

Zenodo (CERN European Organization for Nuclear Research) · 24 December 2022

In one sentence. CNN downscales 25 km regional climate precipitation to 1 km in mountain terrain, cutting RMSE 34% versus raw output and 12% versus quantile mapping.

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.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.