Evaluation of multiple bias-adjustment methods for estimating heat stress conditions in southern South America

Regional Climate Models (RCMs) are valuable tools capable of providing finer-scale climate information, which is particularly relevant in regions like southern South America (SA), where the complex topography and the land-coast contrast strongly influence climate. Despite this, RCMs present systematic errors that need to be corrected for their proper use in impact studies, especially those relying on climate impact indices exceeding specific thresholds, such as heat-stress conditions. In these cases, bias adjustment (BA) methods are commonly used. These methods link climate model historical simulations and observations through the calibration of transfer functions that are subsequently applied to adjust systematic errors in the simulated distribution. In this study, different BA methods were evaluated for southeastern South America (SESA) with a special focus on the estimation of multivariate heat-stress indices, namely the wet bulb temperature and a simplified version of the wet bulb globe temperature. Both indices are based on temperature and humidity variables. The BA methods were calibrated using the historical CORDEX-CORE RCM simulations for the SA domain and the MSWX high-resolution observation dataset. The assessment accounted for: a) two adjustment strategies for estimating the bias-corrected indices (direct and indirect); b) comparison of univariate and multivariate BA methods; c) evaluation of trend-preserving and non-trend-preserving methods. In all cases, BA methods were trained and validated with a cross-validation scheme in the austral summer season during the historical period.  

Results show that under the indirect approach (i.e. adjusting individual variables involved in the indices calculation), all univariate methods presented similar performance, with no remarkable differences between trend- and non-trend-preserving methods. Notwithstanding, in this set up, the multivariate method considerably improved the representation of the thermal indices. This improvement was evident for the RegCM4.7 simulations, where the calculation of the indices using the individually adjusted variables amplified the errors. The lowest biases were found under the direct approach (i.e. adjusting indices directly),  although performance among methods varied depending on the heat stress index analyzed. 

Overall, this study provides insight into the suitability of the BA methods for estimating multivariate thermal indices and paves the way for future assessments of heat stress conditions over SA.

Acknowledgement: A.C. acknowledges support from project PROTECT (PID2023-149997OA-I00) funded by MICIU/AEI/10.13039/501100011033 and ERDF A way of making Europe.