Contrasting Berkeley Earth’s New High Resolution Temperature Dataset with ERA5

Surface air temperature is among the most fundamental data for studying historical climate change.  Two major categories of historical surface air temperature data products exist, the observational datasets produced by statistically interpolating direct thermometer measurements from weather stations, ships, and buoys (e.g. NASA, NOAA, HadCRU, and Berkeley Earth products) and the reanalysis datasets (e.g. ERA5, JMA-3Q) that use weather models to merge various data (e.g. pressure, temperature, wind) from surface, atmospheric, and satellite measurements.

Most observational datasets are relatively low resolution, but Berkeley Earth has recently introduced a new high-resolution version (0.25° x 0.25° latitude-longitude gridding, 1850-present).  This presentation will compare and contrast Berkeley Earth’s high resolution temperature dataset with the similarly resolved near-surface temperature component of ERA5 reanalysis (1940-present).  The Berkeley Earth high resolution dataset is an extension of Berkeley Earth’s prior work, and is derived directly from weather station and ocean temperature measurements.  It maintains the substantial efforts to quality control and correct for systematic biases in surface temperature measurements, but adds machine learning and other techniques to improve the spatial resolution and accuracy of interpolated temperature fields.

Reanalysis systems, like ERA5, are a modern marvel that provide spatially complete weather estimates across many variables at relatively high spatial and temporal resolution.  However, to reach their full potential, they require extensive data input streams, with generally greater accuracy in the modern satellite era than the pre-satellite era (e.g. 1940-1970).  This heavy reliance on satellite data also increases the risk of systematic drift in accuracy due to changes in satellite availability or undiagnosed changes in satellite accuracy.  In the specific context of surface air temperature, though ERA5 uses weather station pressure and humidity to directly refine atmospheric conditions, the weather station temperatures are only used indirectly via the estimates of surface/soil conditions.  Because ERA5 is not directly constrained by weather station temperatures, systematic biases of 1-2 °C between measured surface air temperature and reanalysis estimates are common, with larger errors sometimes occurring.

Berkeley Earth and ERA5 are broadly similar, but also exhibit interesting differences.  Predictably, those differences are larger in the pre-satellite era.  In some regions, this gives rise not just to quasi-random variations, but also systematic differences in both seasonality and the apparent global warming trends.  Large differences are more common in regions of limited data (e.g. Antarctica, Greenland, Tibet), but can also occur in other environments.

We will discuss the differences between Berkeley Earth’s new high resolution dataset and ERA5, as well as identify the contexts where we believe observational data sets are likely to be more accurate than reanalysis or vice versa.