Quantifying spatial correlations and trends in rainfall data under a Mediterranean climate using non-parametric statistical and machine learning approaches

Numerous studies have highlighted the rapid pace of climate change in South Western Australia and quantifying hydrological shifts may provide insights into climate impacts in other Mediterranean regions. Identifying connections in data from spatially distributed rainfall stations using different interpolation approaches, can fill gaps in historical datasets and allow effective installation of new rain gauges based on optimal density and location. Different insights were revealed by using multiple approaches across the time series: the Seasonal Mann-Kendall Test, the Mann-Kendall Test under the Scaling Hypothesis (MKLTP), the Lanzante’s test (LAT) as single change point detection tests, the E-Agglomerative (ECP) and E-Divisive (EDP) change detection algorithms as multiple change point detection tests. Twenty seven loading factors (e.g. seasonal and annual Sen’s slope, the year and number of change points) were calculated from daily rainfall data collected over 100 years (1920-2019) from 107 meteorological stations in South Western Australia. The results illustrated that the rate of rainfall fluctuation in terms of Sen’s slope varied from -2.6 mm yr^-1 in coastal areas to 0.9 mm yr^-1 in inland areas. The scaling trend analysis identified that 53% of the stations were effected by long-term persistence in the wet season, in contrast to only 20% in the dry season. The single change point methods identified a change in the 1940s-1950s and the multiple change point methods identified two changes, in the 1940s and 2000s. The spatial correlation of stations were also mapped using an unsupervised machine learning approach (K-Means), the Multiscale Bootstrap Resampling (MBR), and loading factors, into three optimal clusters, indicating that rainfall in the coastal areas continue to decline, whereas rainfall in the inland areas has increased over the previous 100 years. These statistical and machine learning approaches are effective in identifying spatial and temporal variability in climate change trends.