Long-term temperature data from individual climate stations almost always suffer from inhomogeneities, owing to non-climatic factors. These include sudden changes in station location, instruments, thermometer housing, observing time, or algorithms to calculate daily means; and gradual changes arising from instrumental drifts or from changes in the environment due to urban development or land use. Most abrupt changes tend to produce random effects on regional and global trends, and instrument drifts are corrected by routine thermometer calibration. However, changes in observation time (Vose et al., 2004) and urban development are likely to produce widespread systematic biases; for example, relocation may be to a cooler site out of town (Böhm et al., 2001). Urbanisation usually produces warming, although examples exist of cooling in arid areas where irrigation effects dominate.
When dates for discontinuities are known, a widely used approach is to compare the data for a target station with neighbouring sites, and the change in the temperature data due to the non-climatic change can be calculated and applied to the pre-move data to account for the change, if the discontinuity is statistically significant. However, often the change is not documented, and its date must be determined by statistical tests. The procedure moves through the time series checking the data before and after each value in the time series (Easterling and Peterson, 1995; Vincent, 1998; Menne and Williams, 2005): this works for monthly or longer means, but not daily values owing to greater noise at weather timescales. An extensive review is given by Aguilar et al. (2003).
The impact of random discontinuities on area-averaged values typically becomes smaller as the area or region becomes larger, and is negligible on hemispheric scales (Easterling et al., 1996). However, trends averaged over small regions, in particular, may be biased by systematic heterogeneities in the data (Böhm et al., 2001), and the impact of non-random discontinuities can be important even with large averaging areas. The time-of-observation bias documented by Karl et al. (1986) shows a significant impact even with time series derived for the entire contiguous United States. Adjustments for this problem remove an artificial cooling that occurs due to a switch from afternoon to morning observation times for the U.S. Cooperative Observer Network (Vose et al., 2004).
Estimates of urban impacts on temperature data have included approaches such as linear regression against population (Karl et al., 1988), and analysis of differences between urban and rural sites defined by vegetation (Gallo et al., 2002) or night lights (Peterson, 2003) as seen from satellites. Urbanisation impacts on global and hemispheric temperature trends (Karl et al., 1988; Jones et al., 1990; Easterling et al., 1997; Peterson, 2003; Parker, 2004, 2006) have been found to be small. Furthermore, once the landscape around a station becomes urbanized, long-term trends for that station are consistent with nearby rural stations (Böhm, 1998; Easterling et al., 2005, Peterson and Owen, 2005). However, individual stations may suffer marked biases and require treatment on a case-by-case basis (e.g., Davey and Pielke, 2005); the influence of urban development and other heterogeneities on temperature depends on local geography and climate so that adjustment algorithms developed for one region may not be applicable in other parts of the world (Hansen et al., 2001; Peterson, 2003).
Homogenization of daily temperature series requires much more metadata than monthly assessment (see the extensive discussion in Camuffo and Jones, 2002) and only a few series can be classed as totally homogeneous. Daily minima and maxima, and consequently also DTR and analysis of extremes, are particularly sensitive to non-climatic heterogeneities, including changes in height above ground, housing and ventilation of instruments (Auer et al., 2001; Brunet et al., 2006). The ongoing automation of measuring networks is typically accompanied by a change from large and unventilated screens to small and continuously ventilated ones. Assessment of potential homogeneity problems in a network of 60 daily maximum and minimum temperature series, for Europe for the 20th century by Wijngaard et al. (2003), suggests that 94% of series should be classed as of doubtful homogeneity. The percent of doubtful series reduces to 61% when considering 158 series for 1946–1999. Vincent et al. (2002) in a Canadian study of over 200 daily temperature series, develop daily adjustments by smooth interpolation of monthly adjustments. But a new technique adjusts higher order daily statistics (Della Marta and Wanner, 2006).