Regularized linear regression to estimate the spatial sensitivity governing the pattern effect, comparative analysis to contemporary methods, and observational applications

Abstract

How the spatially varying temperature field affects global radiation (i.e., the "pattern effect") is crucial to understanding how sensitive Earth's temperature is to anthropogenic forcing. We capture this phenomenon in a sensitivity map using regularized linear regression. When trained on 1,000 simulated years in a climate model, the resulting sensitivity maps are consistently able to explain over 75% of the variance in net top-of-atmosphere radiation in an out-of-sample internal variability test. However, when the training data are constricted to 24 years to mirror the length of available observations, that value ranges between 0% and 75% with a median of 50%. This implies that 24-year observational sensitivity maps produced by our method carry significant uncertainty. Tested against the forced climate response in an RCP 8.5 simulation, the ideal 1,000-year training case captures ~75% of the forced response magnitude, while sensitivity maps derived from 24-year periods are unreliable for projecting the warming scenario. Acknowledging the implication that our results depend highly on the particular behavior of the last two decades, we present the first physically interpretable radiative feedback sensitivity maps derived entirely from observations. We then unify several alternative methods under a common training and testing procedure. These methods all generate predictive frameworks from internal variability, except for an included Green's function. The latter approach was the primary method used to generate pattern effect sensitivity maps prior to the methods discussed in this thesis, so it grounds our comparative analysis to the current state-of-the-field. All methods match or improve upon the Green's function's ability to predict internal variability, but vary widely in their ability to predict a step forcing 4xCO2 warming simulation.