Effects of variance adjustment techniques and time-invariant transfer functions on heat wave duration indices and other metrics derived from downscaled time-series. Study case: Montreal, Canada

Abstract

Statistical downscaling techniques are often used to generate finer-scale projections of climate variables affected by local-scale processes not resolved by coarse resolution numerical models like global climate models (GCMs). Statistical downscaling models rely on several assumptions in order to produce finer-/local-scale projections of the variable of interest; one of these assumptions is the time-invariance of the relationships between predictors (e.g. coarse resolution GCM output) and the local-scale predictands (e.g. gridded observation-based time-series or weather station observations). However, in the absence of future observations, statistical downscaling studies use historical data to evaluate their models and assume that these historical simulation skills will be retained in the future. In addition, regression-based downscaling models fail to reproduce the observed variance, and hence their projections need to be adjusted accordingly. Two approaches are usually employed to perform this adjustment: randomization and variance inflation. Here, we study the effect of the stationarity assumption when downscaling daily maximum temperatures and using the downscaled information to estimate historical and future metrics like return periods and heat waves durations over Montreal, Canada; and the effect of the two variance adjustment techniques on the historical and future time-series. To do so, we used regional climate model (RCM) output from the Canadian RCM 4.2, as proxies of historical and future local climates, and daily maximum temperatures obtained from the Canadian GCM 3.1. The results show that the root-mean-squared errors between the pseudo-observations and the statistically downscaled time-series (historical and future) varied over time, with higher errors in the future period; and the effects of randomization and variance inflation on the tails of the statistically downscaled time-series.

The stationarity of two statistical downscaling methods for precipitation under different choices of cross‐validation periods

“…The stations where either method passed the stationarity test varied considerably between the precipitation indices on mean, extremes, occurrence, and inter‐annual variability. These results suggest that the downscaled time series of precipitation, and the marginal distribution of the downscaled wet‐day intensities, usually cannot be trusted as a whole, which is consistent with findings that various statistical downscaling methods violated the stationarity assumption for temperature and precipitation in regions of North America (Dixon et al ., ; Gaitán, ; Salvi et al ., ). We used reanalysis data to test the statistical downscaling methods, but if we calibrated quantile‐mapping using GCMs and observations, the non‐stationarity in quantile‐mapping would likely increase due to the temporal misalignment between the low‐frequency oscillations in the GCM and in reality (Maurer et al ., ; Nahar et al ., ).…”

Projections of the future occurrence, distribution, and seasonality of three Vibrio species in the Chesapeake Bay under a high‐emission climate change scenario

“…These are discussed in more depth in Muhling et al . [], but relate primarily to the assumption of stationarity in statistically downscaled projections [e.g., Vrac et al ., ; Gaitan and Cannon , ; Gaitan et al ., ; Gaitan , ; Dixon et al ., ], the simplicity of the water balance model and error from the model trees used to create the spatial fields.…”

“…The use of the statistical framework to derive temperature and salinity in the Chesapeake Bay introduces additional uncertainty to the projections. These are discussed in more depth in Muhling et al [2017], but relate primarily to the assumption of stationarity in statistically downscaled projections [e.g., Vrac et al, 2007;Gaitan and Cannon, 2013;Gaitan et al, 2014;Gaitan, 2016;Dixon et al, 2016], the simplicity of the water balance model and error from the model trees used to create the spatial fields.…”

Can we obtain viable alternatives to Manning’s equation using genetic programming?

“…As the method successfully omitted unrelated variables-like 2 m temperature-from the proposed equations, we conceive that the method could also be used for nonlinear predictor selection, complementing classical approaches like the stepwise selection, often used in conjunction with multiple linear regression (e.g. [47,48] ). The method also provides an alternative to the graphical sensitivity analysis technique by Cannon and McKendry, [49] and to the Bayesian approach used by Robertson and Wang [50] for seasonal streamflow forecasting.…”