Some Pitfalls in Statistical Downscaling of Future Climate

Abstract

Statistical downscaling (SD) is commonly used to provide information for the assessment of climate change impacts. Using as input the output from large-scale dynamical climate models and observation-based data products, SD aims to provide a finer grain of detail and to mitigate systematic biases. It is generally recognized as providing added value. However, one of the key assumptions of SD is that the relationships used to train the method during a historical period are unchanged in the future, in the face of climate change. The validity of this assumption is typically quite difficult to assess in the normal course of analysis, as observations of future climate are lacking. We approach this problem using a “perfect model” experimental design in which high-resolution dynamical climate model output is used as a surrogate for both past and future observations. We find that while SD in general adds considerable value, in certain well-defined circumstances it can produce highly erroneous results. Furthermore, the breakdown of SD in these contexts could not be foreshadowed during the typical course of evaluation based on only available historical data. We diagnose and explain the reasons for these failures in terms of physical, statistical, and methodological causes. These findings highlight the need for caution in the use of statistically downscaled products and the need for further research to consider other hitherto unknown pitfalls, perhaps utilizing more advanced perfect model designs than the one we have employed.

Evaluation of some distributional downscaling methods as applied to daily maximum temperature with emphasis on extremes

“…For the GCM, CAT 5 and the tails have similar seasonal cycles of MAE with larger values during a broad plateau from April to September indicating that during this time area‐averages (GCM) are less representative than point values (observations). This plateau can be explained largely by two factors: the “mountain snow effect” in spring and the “coastal effect” during summer (Lanzante et al ., ). Both of these effects result from the larger footprint of the GCM compared to the observations.…”

“…In the presence of a warming signal the most challenging values to downscale occur in the right tail in the form of “novel” values. The downscaling “mountain snow effect” occurs when snow is present in the historical training period but absent in the future application period (Lanzante et al ., ). But since right tail values occur mostly during snow‐free conditions in both the training sample as well as in the future this mitigates the “mountain snow effect,” thus degrading tail values much less compared to the rest of the distribution than would normally be expected.…”

“…In examining the seasonality of performance of the downscaling methods we found generally larger errors (MAE) in the GCM as well as the downscaling methods during the warm season. This seasonality is largely due to factors identified previously (Lanzante et al ., ) that are linked to spatial resolution differences between the GCM (coarse) and observations (fine). The spring skill minimum, referred to as the “mountain snow” effect, is related to a non‐stationarity arising from the disappearance of snow during spring in a warmed climate, and is a particular challenge for methods that utilize a change‐factor approach.…”

“…(), and following Lanzante et al . () we refer to it as bias correction quantile mapping (BCQM) since the term “quantile mapping” is often used more generically in reference to a variety of SD techniques from the distributional class. Mathematically it can be expressed as: where F represents the cumulative distribution function (CDF) and F −1 its inverse.…”

“…Although there is no unique way to classify SD methods Lanzante et al . () propose four classes: direct transfer function, distributional mapping, spatial mapping, and weather generators. …”