DUO CHAN , Geoffrey Gebbie, Peter Huybers

Various observational estimates of historical land surface air temperature (LSAT) trends differ on account of differences in corrections. Relative to the most-recent estimate provided by NOAA’s Global Historical Climatology Network Monthly Version 4 (GHCNm4), an estimate by Berkeley Earth is 0.02°C warmer and one by the Climate Research Unit (CRUTEM5) is 0.14°C warmer between 1880–1940. Such systematic offsets can arise in LSAT records as a result of poorly-documented changes in measurement characteristics, including changes in instrumentation and movement of stations, as well as how these breakpoints are corrected for across different estimates. Building on an existing pair-wise homogenization algorithm (PHA0 applied in GHCNmV4), we propose a revised version (PHA1) that accounts for autocorrelation in climate variables and iteratively operates to adjust breakpoints. Tests on synthetic data generated by adding breakpoints to CMIP6 simulations and realizations from a Gaussian process indicate that PHA1 outperforms PHA0 in identifying small breaks and recovering accurate climate trends. Applied to unhomogenized station temperatures compiled within GHCNmV4, PHA1 is shown to detect breakpoints that correspond with available station metadata. Uncertainties associated with PHA1 are estimated by randomly perturbing algorithmic parameters. The continental mean temperature warming found using PHA1 is consistent with that of Berkeley Earth to within estimated uncertainties, despite using a different homogenization approach. Relative to unhomogenized data, the PHA1 homogenization increases 1880–2022 temperature trend by 0.18°C per century, with a 95% confidence interval of 0.11–0.24°C per century, leading to a continental mean temperature warming of 1.74°C between 1880–1889 and 2012–2021 with a 95% confidence interval of 1.63–1.90°C.