Quantifying Anthropogenic Contributions to Extreme Temperature Changes: Conventional and Bayesian Approaches

Abstract

In this thesis, we rigorously quantify the contributions of greenhouse gas (GHG), aerosol (AA), and natural (NAT) forcing on the intensity of extreme temperature changes (annual minima of day/night time temperature: TXn, TNn; annual maxima of day/night time temperature: TXx, TNx) from frequentist and probabilistic perspectives. In addition, more reliable future projections of extreme temperature are provided based on the evaluation of model performances in capturing the observed trends. Although many previous detection and attribution studies have found strong evidence of human influences on extreme temperature changes, quantification of individual external forcings (greenhouse and anthropogenic aerosol forcings) has not been well assessed owing to data limitations. For this reason, we first examine observed extreme temperature changes during 1951-2015 and verify individual contributions of external forcings at global, continental, and subcontinental scales based on the optimal fingerprint method. The results represent that all observed extreme temperature indices gradually increase at global and continental scales. The cause of observed warming is found to be anthropogenic forcings by conducting a two-signal analysis in which observed changes are regressed onto anthropogenic forcing (ANT) and NAT, consistent with previous findings. In addition, new anthropogenic evidence is detected over Europe (TXn) and South America (TNx, TXx) due to extended observational records and improved spatial coverage. Generally, the regional scale has a low signal-to-noise ratio because of high internal variability. Nevertheless, more than 60% of subregions show human influences on all extreme temperature indices except for TXn (45%). Three-signal analysis in which observation is regressed onto GHG, AA, and NAT reveals that GHG detection occurs across global and all continental regions. The global and continental GHG results resemble ANT detection calculated from two-signal analysis, indicating that observed warming is mainly due to greenhouse gas-only forcing. To quantify contributions of individual forcing, attributable trends calculated by multiplying scaling factors to corresponding recent trends are analyzed in TXx which provided GHG and AA detection simultaneously across most continental domains. Most of the observed warming is explained by GHG, which is offset by the aerosol cooling effect, implying that more future warming is expected when anthropogenic aerosol mitigation policies are realized. Although the optimal fingerprint method was elaborately developed, it is vulnerable to strong multicollinearity in independent variables. In this context, as an alternative detection method, a new Bayesian attribution approach was devised, which can separate individual forcing influences. For the additional ANT and NAT analysis, ALL detection must be preceded. When logarithm Bayes factors of ALL against NAT are equal to or greater than unity (‘substantial’ evidence), additional ANT contribution is defined. It means that improved ALL detection compared to NAT is due to the ANT influence under linear additivity assumption (ALL=ANT+NAT). Similarly, additional NAT contribution is estimated by comparing Bayes factors between ALL and ANT (ln BALL,ANT ≥ 1). The Bayesian attribution method provides similar ANT and NAT results to the optimal fingerprint method in warm extremes even at a regional scale. Additional GHG and AA contributions are obtained in the same way with the ANT precondition. When ANT is closer to observations compared to AA or GHG (i.e. ln B ≥ 1), additional GHG or AA detection is defined, respectively. It indicates that improved ANT detection is due to additional GHG or AA contributions assuming that ANT is a linear combination of GHG and AA. The Bayesian method shows that GHG is the main cause of observed extreme temperature changes at global and continental scales. Furthermore, AA influences are robustly detected over Europe and Asia in warm extremes. In cold extremes, GHG detection obtained from the Bayesian method is more frequent than optimal fingerprint at a regional scale. It is presumably because the Bayesian method is not affected by multicollinearity which degenerates signals in the optimal fingerprint. Based on an improved understanding of observed extreme temperature changes through detection analyses, future changes in global land extreme temperatures are constrained. After selecting a period in which AA and NAT influences are negligible (i.e. greenhouse gas forcing is the most dominant factor in observed warming), we extend ALL forced model samples into four SSP scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5) based on conventional and Bayesian attribution results for each model. The constraint method based on optimal fingerprint adjusts future projections of extreme temperature by multiplying scaling factors. It provides a dramatic decrease of 5-95% range as well as median value especially in warm extremes after the mid-21st century, consistent with previous constraining studies for mean temperatures. On the contrary, TNn shows more warming than raw projections, reflecting model underestimation in the observed trend. The Bayesian-based method constrains future projection by assigning higher weights to models which show better agreement with observations than other models. Although the resulting 5-95% ranges are also reduced in the Bayesian constraint, they are generally less than the optimal fingerprint-based results. The median does not change much after Bayesian adjustment in all extreme temperatures except for some decreases in medians in high emission scenarios (SSP3-7.0, SSP5-8.5) in the late 21st century. These differences in constrained projections between the two methods are due to different ways of adjusting individual models: scaling up or down model projections (in an absolute way) or giving weights to models (in a relative way).