Human influence on climate detectable in the late 19th century

Abstract

The physics of the heat-trapping properties of CO[Formula: see text] were established in the mid-19th century, as fossil fuel burning rapidly increased atmospheric CO[Formula: see text] levels. To date, however, research has not probed when climate change could have been detected if scientists in the 19th century had the current models and observing network. We consider this question in a thought experiment with state-of-the-art climate models. We assume that the capability to make accurate measurements of atmospheric temperature changes existed in 1860, and then apply a standard "fingerprint" method to determine the time at which a human-caused climate change signal was first detectable. Pronounced cooling of the mid- to upper stratosphere, mainly driven by anthropogenic increases in carbon dioxide, would have been identifiable with high confidence by approximately 1885, before the advent of gas-powered cars. These results arise from the favorable signal-to-noise characteristics of the mid- to upper stratosphere, where the signal of human-caused cooling is large and the pattern of this cooling differs markedly from patterns of intrinsic variability. Even if our monitoring capability in 1860 had not been global, and high-quality stratospheric temperature measurements existed for Northern Hemisphere mid-latitudes only, it still would have been feasible to detect human-caused stratospheric cooling by 1894, only 34 y after the assumed start of climate monitoring. Our study provides strong evidence that a discernible human influence on atmospheric temperature has likely existed for over 130 y.

Keywords: climate change detection and attribution; satellite temperature data; stratospheric temperature.

Fig. 1.

Observed and simulated changes in global-mean annual-mean temperature in six atmospheric layers. Results are synthetic and observed temperatures averaged over the layers sampled by the primary satellite instruments used to measure atmospheric temperatures: the SSU (panels A–C) (21) and the MSU/AMSU. MSU/AMSU provides lower stratospheric temperature (TLS), total tropospheric temperature (TTT), and lower tropospheric temperature (TLT) (panels D–F, respectively) (22). Approximate altitude ranges of the weighting functions for these layers are given in brackets (27, 28). Results are anomalies relative to climatological means over 1986 to 2024. Model simulations are for the period 1860 to 2024 and are from nine different CMIP6 models and 32 realizations of historical climate change (see Materials and Methods and SI Appendix). Vertical green lines are the start dates of major volcanic eruptions: KR $=$ Krakatoa; SM $=$ Santa Maria; NO $=$ Novarupta; AG $=$ Agung; FU $=$ Fuego; EC $=$ El Chichón; PI $=$ Pinatubo. The vertical brown line marked HT denotes the unusual Hunga Tonga eruption. Unlike the other eruptions shown in the figure, Hunga Tonga injected a large amount of water vapor into the stratosphere and mesosphere (29). Forcing by “small” early 21st century eruptions was included (up until December 2014) in the CMIP6 historical runs, but the start dates of these eruptions are not shown here (30). The red lines are least-squares linear trends in the satellite temperature data.

Fig. 2.

Model fingerprints of atmospheric temperature change in response to combined anthropogenic and natural external forcing. Fingerprints were calculated for four vertical domains: SSU+MSU (panels A, B, and C), SSU (panels D, E, and F), MSU (panels G, H, and I), and TROP (panels J, K, and L). Results are for three different periods: 1860–1882 (panels A, D, G, and J), 1920–1959 (panels B, E, H, and K), and 1994–2024 (panels C, F, I, and L). Each fingerprint pattern, $F(x)$, is the first EOF of the CMIP6 multimodel average zonal-mean annual-mean atmospheric temperature changes (see Materials and Methods and SI Appendix). The choice of the 1860–1882 and 1994–2024 periods reduces the influence of the large Krakatoa and Pinatubo eruptions (which occurred in 1883 and 1991, respectively). The period 1920–1959 has few major eruptions (Fig. 1). All results are for global geographical coverage. The variance explained by each EOF is given in brackets. EOF loadings are plotted at the approximate heights of the maxima of each SSU and MSU weighting function, indicated by dotted horizontal blue lines. Results were smoothly interpolated in the vertical.

Fig. 3.

Influence of the “start date” for detecting human-caused atmospheric temperature signals on signals and S/N ratios. Results are for global geographical coverage and four vertical domains (SSU$+$MSU, SSU, MSU, and TROP). The four rows correspond to monitoring start dates of 1860 (panels A, B), 1900 (panels C, D), 1920 (panels E, F), and 1986 (panels G, H). The signals (left column) were calculated by projecting latitude-height temperature data from 32 individual CMIP6 HIST_ext realizations onto the fingerprint $F(x)$ (Fig. 2). Trends of increasing length L years were computed from the resulting projection time series. CMIP6 control run temperature data were then projected onto $F(x)$, yielding the projection time series $N_{\mathit{ctl}}(t)$. The noise ${\sigma }_{\mathit{ctl}}(L)$ is estimated by fitting nonoverlapping $L$-year trends to $N_{\mathit{ctl}}(t)$ and calculating the SD of the $L$-year trend distribution. The S/N ratio (right column) is the $L$-year signal trend divided by the respective values of ${\sigma }_{\mathit{ctl}}(L)$. For each start date, $L$ varies from $5,6,\dots L_{max}$ years, with $L_{max}=39$ years for 1986 and 40 years for all other start dates. The horizontal black line in panels B, D, F, and H is the 1% significance level.

Fig. 4.

Sensitivity of anthropogenic fingerprint detection time, $t_d$, to the selected start date for monitoring atmospheric temperature change. Values of $t_d$ were calculated using the S/N ratios from the right-hand column of Fig. 3. The four rows provide (from top to bottom) the detection times for monitoring start dates of 1860, 1900, 1920, and 1986. For each assumed start date, values of $t_d$ are given for the four vertical domains (SSU$+$MSU, SSU, MSU, and TROP), and for four geographical domains: global (90^°N–90^°S), NH (0^°–90^°N), SH (90^°S–0^°), and NH mid-latitudes (40^°N–55^°N). Fingerprint detection times in individual model HIST_ext realizations are indicated by smaller filled colored circles. Larger filled colored circles denote the multimodel average $t_d$ value, which is only plotted if the fingerprint is detectable at the 1% significance level within 40 y after the 1860, 1900, and 1920 start dates (or within 39 y after the 1986 start date) in at least 10 HIST_ext realizations. Fingerprint detection times in satellite data are represented by unfilled color symbols.

Fig. 5.

Comparison of the natural internal variability of global-mean monthly-mean atmospheric temperature in six different atmospheric layers: SSU3, SSU2, SSU1, TLS, TTT, and TLT (panels A–F). Model estimates of internal variability are from two sources: preindustrial control runs (CTL) and externally forced HIST_ext simulations from which the multimodel average temperature time series (MMA) was subtracted. A regression-based approach was used to remove the MMA from observed temperature time series (54). Temperatures from the CTL simulations and the “MMA-removed” HIST_ext realizations and satellite temperature data were digitally smoothed to compare variability on timescales of approximately 10y (SI Appendix). The statistic plotted is ${\sigma }_{\text{FILT}}$, the temporal SD of the filtered data. All ${\text{HIST}}_{\mathit{ext}}$ and observed time series were linearly detrended after removal of the MMA. The 432-mo analysis period in the ${\text{HIST}}_{\mathit{ext}}$ realizations and satellite data (January 1986 to December 2021) excludes the January 2022 Hunga Tonga eruption, which is not included in the ${\text{HIST}}_{\mathit{ext}}$ simulations but had a large effect on observed stratospheric temperature (SI Appendix, Fig. S6). The CTL run distributions of ${\sigma }_{\text{FILT}}$ are based on analysis of over 180 nonoverlapping 432-mo segments of $T_{\mathit{ctl}}(t)$ data.