Decoding the Surface Temperature Record

Abstract:

Historical temperature observations from surface stations have been recorded using a variety of units and levels of precision, with metadata that are often incomplete. As a result, the amount of rounding applied to these observations is generally unknown, posing a challenge to statistical methods that are sensitive to the use of discrete data. Methods used to infer distributional changes often assume that data are continuously distributed and can only be reliably applied when the specific discreteness of each sample is known. We present a new technique, termed `precision-decoding,' that identifies the original precision and units of time series data. Applying it to the GHCND database, we identify temporal and spatial patterns in the precision and units used by surface stations. We show that many archived values have been offset from the original observations due to double-rounding in the presence of conversion between Fahrenheit and Celsius, and provide additional metrics to identify stations in need of further quality control. While the discreteness of the data is unlikely to have influenced global mean temperature trends, we show that it can affect higher-order moments of the temperature distribution such as the variance or skewness, and that it can alter the apparent frequency of record-breaking events.