Quantile-based bias-correction of extreme rainfall: Pros & cons of popular methods for climate signal preservation

Abstract

Bias correction is a common practice in climate sciences. However, bias-corrected climate projections do not necessarily preserve signals in moments and quantiles compared to raw climate models. Focusing on extreme rainfall and Depth-Frequency curves, the goal of this paper is to demonstrate the efficacy of three popular techniques in preserving signals in the first- and second-order moments and in a selection of quantiles. With this aim, a thorough sensitivity analysis is undertaken and a real-world application leveraging a multi-model EURO-CORDEX ensemble showcases the findings. The target techniques are Quantile-Quantile Downscaling (QQD), Detrended Quantile Mapping (DetQM), and Quantile Delta Mapping (QDM). Results highlight that QQD shows significant errors in the preservation of signals in the mean and percentiles; DetQM and QQD show errors in the percentiles; QDM in the standard deviation. Errors depend not only on the bias correction technique, but also on the magnitude and accordance of the bias and the signal. The main implications are: i) a climate projection having a certain bias and signal, bias-corrected with different methods, provides different extremes; ii) climate projections having the same signal, bias-corrected with the same method, provide different extremes according to the bias magnitude; iii) physically consistent combinations of bias and signal (as those experienced in the real-world application) provide for a large uncertainty range associated to the final, bias-corrected moments and Depth-Frequency curves.