How Reliable Are Automated Homogenization Methods for Early Climate Records?

Early instrumental observations are fundamental to understanding historical climate variability, but long-term instrumental records remain scarce outside Europe and North America. This data gap limits our ability to assess climate change over decades globally and highlights the urgent need to rescue and harmonize historical observations from data-poor regions. The Global Early Instrumental Monthly Meteorological Multivariable Database (HCLIM) provides a unique resource for rescued early observations, but the systematic evaluation of automated homogenization methods for such data remains limited.

In this study, we assess the performance of two widely used automated homogenization techniques, CLIMATOL and BART (Bayesian Analysis of Records), applied to early instrumental temperature series from HCLIM. The performance of the methods is evaluated based on data storage in preprocessed datasets, breakpoint characteristics in homogenized series, and consistency with the 20CRv3 reanalysis. French and South Asian station networks are used as representative examples of dense and sparse networks, respectively.

Our results show minor structural differences in preprocessing across methods, with BART retaining fewer but longer and more consistent records (80%) compared to CLIMATOL (96%). BART detected approximately eight times more breakpoints than the other methods, indicating higher sensitivity to inhomogeneities. Strong regional and temporal variation in breakpoint detection was observed. Comparisons with 20CRv3 show that BART generally shows the lowest deviations and highest consistency, especially in dense station networks such as France. In contrast, CLIMATOL shows mixed performance depending on network density and local climate variability. Larger deviations are found in sparse regions such as Southeast Asia, highlighting the strong influence of station density on homogenization quality.

These findings demonstrate that automated homogenization methods differ significantly in data storage, breakpoint sensitivity, and reconstruction capability. Overall, BART shows the highest accuracy, while CLIMATOL shows balanced performance. This study contributes to best practices for applying automated homogenization to salvaged historical data and supports ongoing work to expand climate records in data-sparse regions, improve the basis for climate model validation, reanalyses, and assessment of historical climate extremes relevant to adaptation planning.