A method to infer the observation time of a station at annual resolution is discussed and applied to stations in the U.S. Historical Climatology Network (HCN). The procedure is based on a tendency for the percentiles of the monthly distribution of positive day-to-day maximum temperature changes (i.e. day n+1 > day n) to exceed the corresponding absolute percentiles of the distribution of negative day-to-day changes at afternoon stations. Similarly, absolute percentiles of negative day-to-day minimum temperature change tend to exceed the corresponding positive interdiurnal changes at morning observation sites. Equal percentiles are generally found at stations that use a midnight observation hour. Based on annual and seasonal summations of these monthly percentile differences, discriminant functions are developed that are capable of differentiating between afternoon, morning and midnight observation schedules.
Across the majority of the United States, this procedure correctly classifies observation time in over 90% of the station-years tested. This high rate of classification success allows discrepancies with observation time metadata to be readily identified. Application of the procedure to U.S. historical climate network stations indicates the presence of inconsistent observation times in the corresponding metadata file. An examination of maximum and minimum temperatures from consecutive days at these sites and corresponding observations from adjacent stations reporting midnight, afternoon and morning observation hours, further confirms the presence of errors in the observation time documentation