Adapting to Climate Change with Machine Learning: The Robustness of Downscaled Precipitation in Local Impact Analysis

Abstract

The skill, assumptions, and uncertainty of machine learning techniques (MLTs) for downscaling global climate model’s precipitation to the local level in Bolivia were assessed. For that, an ensemble of 20 global climate models (GCMs) from CMIP6, with random forest (RF) and support vector machine (SVM) techniques, was used on four zones (highlands, Andean slopes, Amazon lowlands, and Chaco lowlands). The downscaled series’ skill was evaluated in terms of relative errors. The uncertainty was analyzed through variance decomposition. In most cases, MLTs’ skill was adequate, with relative errors less than 50%. Moreover, RF tended to outperform SVM. Robust (weak) stationary (perfect prognosis) assumptions were found in the highlands and Andean slopes. The weakness was attributed to topographical complexity. The downscaling methods were shown to be the dominant source of uncertainties. This analysis allowed the derivation of robust future projections, showing higher annual rainfall, shorter dry spell duration, and more frequent but less intense high rainfall events in the highlands. Apart from the dry spell’s duration, a similar pattern was found for the Andean slopes. A decrease in annual rainfall was projected in the Amazon lowlands and an increase in the Chaco lowlands.