There is increasing interest in the use of stochastic models to represent single-site and spatial–temporal rainfall for applications ranging from flood design to the downscaling of outputs from gridded numerical climate models. While single-site modelling focuses attention on the temporal structure (modelling of daily rainfall commonly involves a two-stage simulation of rainfall occurrence and rainfall amounts), spatial modelling provides a powerful tool to identify spatial anomalies. To fit such models, long records of high-resolution rainfall data are required. In this paper, we consider some aspects of the data that should be examined carefully in order to avoid severe biases in the representation of spatial variability. Problems may arise, for example, because of the replacement of some rain gauges or changes in the way measurements are recorded. In the UK, a primary source of daily or subdaily observational rainfall sequences for research use is the UK Meteorological Office (UKMO). Two sets of UKMO daily data are used here to illustrate some of the issues, and to reveal problems that may be widespread in UK rainfall data. Some practical suggestions for identifying and remedying such problems are given.
The data sets are from two regions: the Blackwater catchment (about 50 × 40 km2) and northeast Lancashire (about 60 × 40 km2). It shows their locations. Each data set consists of daily rainfall time series from a network of gauges. Typically, rain gauge data in the UK are accompanied by documentation providing some basic information about the site network. All the gauges are identified with UKMO station numbers (six digits) and positioned in national grid coordinates. There are 34 unique sequences in the Blackwater data set and 39 in the Northeast Lancashire data set, of which five sequences in each data set are recorded by automatic gauges. The longest time spans of sequences are from 1908 to 2000 for the Blackwater data set and from 1961 to 2000 for the Northeast Lancashire data set, respectively.
Even in the best-regulated systems, discrepancies in record can occur, and preliminary checks on record lengths and completeness are needed. For example, in our experience there is occasionally disagreement between the supporting documentation and the data themselves; some nominally different sequences are actually identical during their overlapping periods; sometimes abnormal records are mixed into the daily sequences (e.g. a monthly total appears on the last day of a month); etc.
It can be useful to distinguish between automatic gauges and those that are read manually. Most automatic gauges are of ‘tipping bucket’ type (for a description of the tipping bucket mechanism; for these gauges, daily rainfall totals are usually derived by counting the number of bucket tips in a day and converting this to a rainfall amount (Wood et al. 2000). This procedure can result in discretization errors, particularly during periods of light rain. Apparent differences between records at nearby gauges may therefore be due simply to the different gauge types (see ‘Differences owing to recording mechanisms’ for an example of this). In general, automatic gauges can be identified by the presence of ‘AUTO’ or ‘LOGGER STA’ in their names.
Changes in data resolution
From a careful check of the data sets themselves (rather than from any official documentation), a substantial change in the resolution of the recordings can be seen. Before the 1970s the resolution was 0.3 mm. In the early 1970s the resolution was improved to 0.1 mm for most of the sites (except for some tipping bucket gauges for which the highest resolution is 0.2 mm). This change had little effect on recorded rainfall amounts, but resulted in a substantial increase in recorded rainfall occurrences. Here we select the three longest sequences from the Blackwater data set to illustrate the effect of resolution change. We calculated the annual amounts and numbers of wet days for the recorded sequences, and then recalculated these quantities applying a threshold of 0.3 mm, to reproduce the effect of recording to a resolution of 0.3 mm over the whole time period.
It shows the annual amounts and the numbers of wet days for the three sequences. The solid lines were calculated from the original sequences, whereas the dots were calculated with a threshold of 0.3 mm applied. In the left plot (for the amounts) the dots are more or less superimposed on the solid lines throughout the whole period, whereas in the right plot (for the occurrences) the dotted sequences are noticeably lower than the solid lines from around 1970. From 1975 onwards, on average over the three sequences, the annual amounts decrease by only 2.1 mm when a threshold is applied, whereas the annual numbers of wet days decrease on average by about 15 days (or around 10%).
The improvement in recording resolution may thus introduce an illusory increasing trend in rainfall occurrences. This will be illustrated in detail in ‘Model-based checks’.
No Comments, Comment or Ping