Page 39

et al., 1997). At the same time, historical sea surface temperature data continue to be digitized to fill in gaps in the available in situ data (e.g., Woodruff et al., 1998). While new approaches to derive surface temperature over land using satellite data are under development (e.g., Basist et al., 1998), the satellite data these approaches require may only go back to 1987.

The possibility, indeed probability, of erroneous data is addressed by every major data set compiler as part of the quality control effort (e.g., Jones et al., 1999; Peterson et al., 1998c). While all erroneous data points cannot be removed from a data set without the risk of removing a great deal of good data as well, biases due to large isolated errors can be eliminated. Biases due to discontinuities in the observing network are a much more difficult problem to resolve. However, a great deal of work on homogeneity problems has been done over the past decade or more, as summarized in a recent review of homogeneity research (Peterson et al., 1998b). This work attempts to estimate the magnitude of the bias caused by random station moves, installation of new instrumentation, and changes in observing practices such as changing the time of observation of maximum/minimum thermometers from late afternoon to early morning. Once the magnitude of the bias is determined, the data can be adjusted to account for these inhomogeneities.

A more difficult problem is assessing the impact of small, gradual changes in the observing network. Urbanization (and land-use changes in general) and the resultant urban warming is the most commonly cited example of this type of problem. Recent efforts to assess this bias focus on identifying which stations are rural and which are urban, using map-based metadata12 (Peterson and Vose, 1997) or night-lights derived metadata (Owen et al., 1998). Long-term global temperature trends calculated both from the full land surface network, and from rural stations only, turn out to be very similar (differing by about 0.05 °C per 100 years), despite some differences in regions sampled (e.g., India has few long-term rural stations) (Peterson et al., 1999).

The uneven spatial distribution of in situ data, and the change in their distribution over time, can also potentially create biases. Some of the approaches to addressing this problem are: (a) acquiring more data through digitization of historical records, (b) improving internationalcontinue

12 Metadata (or data about data), in this context, is information that describes the environment in which a measurement is made and/or the methods and/or tools used to make the measurement.