Investigating the uncertainty in gridded in situ climate data for the UK

The Met Office has been generating gridded climate datasets from UK in situ observations for over 25 years. The current version of these gridded data is known as the HadUK-Grid dataset – it includes monthly, seasonal, annual and long-term average values for 11 variables at 1km resolution. Three of these variables are also available at daily resolution. These datasets have a variety of applications including climate monitoring, model calibration and model validation.

Although the software, data sources, file formats, grid resolution and distribution methods have all changed since we first started producing these datasets, the techniques used to create the grids from in situ observations have remained largely the same. The observations are first detrended, either by converting the values to anomalies from the long-term average, or by using regression analysis to model the dependency on geospatial variables such as terrain elevation or proximity to the coast. The anomaly values or regression residuals are then interpolated to the target grid points using inverse-distance weighted averaging.

Here we present a review of the uncertainties in our gridded data. The aims of the analysis are threefold – to update the information we provide to users regarding the quality of our gridded products, to better understand the strengths and weaknesses of our gridding methods, and to investigate the efficiency of the quality control tests in our gridding software.

A leave-one-out cross-validation analysis has been carried out for the majority of our archive of gridded data. Time series graphs will be presented showing how the RMS error varies through the data record for each variable. These graphs show clear trends, seasonal cycles and outliers. Case studies have been used to understand the causes of some of these features.

Based on this analysis we have investigated aspects of our gridding techniques and quality control methods which have the potential to be improved. Preliminary results are presented which show how different gridding methods affect the cross-validation results and we draw some initial conclusions regarding the impact on gridding uncertainties.