Detection of inhomogeneities in precipitation time series in Portugal using direct sequential simulation

• We investigate a geostatistical approach for the homogenisation of climate data.
• We identify irregularities and fill in missing data in precipitation series.
• A higher number of simulations enhance the detection of inhomogeneities.
• Recommended settings for the presented method are derived from sensitivity analyses.

Climate data homogenisation is of major importance in climate change monitoring, validation of weather forecasting, general circulation and regional atmospheric models, modelling of erosion, drought monitoring, among other studies of hydrological and environmental impacts. The reason is that non-climate factors can cause time series discontinuities which may hide the true climatic signal and patterns, thus potentially bias the conclusions of those studies. In the last two decades, many methods have been developed to identify and remove these inhomogeneities. One of those is based on a geostatistical simulation technique (DSS — direct sequential simulation), where local probability density functions (pdfs) are calculated at candidate monitoring stations using spatial and temporal neighbouring observations, which then are used for the detection of inhomogeneities. Such approach has been previously applied to detect inhomogeneities in four precipitation series (wet day count) from a network with 66 monitoring stations located in the southern region of Portugal (1980–2001). That study revealed promising results and the potential advantages of geostatistical techniques for inhomogeneity detection in climate time series. This work extends the case study presented before and investigates the application of the geostatistical stochastic approach to ten precipitation series that were previously classified as inhomogeneous by one of six absolute homogeneity tests (Mann–Kendall, Wald–Wolfowitz runs, Von Neumann ratio, Pettitt, Buishand range test, and standard normal homogeneity test (SNHT) for a single break). Moreover, a sensitivity analysis is performed to investigate the number of simulated realisations which should be used to infer the local pdfs with more accuracy. Accordingly, the number of simulations per iteration was increased from 50 to 500, which resulted in a more representative local pdf. As in the previous study, the results are compared with those from the SNHT, Pettitt and Buishand range tests, which were applied to composite (ratio) reference series. The geostatistical procedure also allowed us to fill in missing values in the climate data series. Finally, based on several experiments aimed at providing a sensitivity analysis of the procedure, a set of default and recommended settings is provided, which will help other users to apply this method.