Reliable regional high-resolution data for precipitation and extreme precipitation have become increasingly more important since they are essential for assessing the accuracy and correcting the biases of Global and Regional Climate models which in turn help us make more accurate predictions about the future. These future predictions are essential for Climate Change Risk & Vulnerability assessments which are legally required according to the new EU Taxonomy Regulation and will be used to make important decisions by businesses and law makers. Precipitation modelling and downscaling techniques are used in order to achieve higher resolution datasets which are useful in predicting precipitation extremes and droughts. The frequency and intensity of precipitation is important to policymakers because it disrupts farming and causes natural hazards like floods and mudslides. Moreover, higher resolution data can also be used in hydrological and other engineering models to make predictions about the future frequency and intensity of climate-change related hazards.

In the Greek region precipitation and extreme precipitation are very difficult parameters to simulate due to their distribution, and in Greece’s case due to their rarity in certain months.  The Greek region despite the large number of islands is dominated by mountains in the Greek mainland. The large mountain ranges give precipitation in the country a distinct longitudinal shift. Additionally, the country experiences intense interseasonal variability which is in line with the Mediterranean climate. The Mediterranean basin experiences extreme precipitation events in the winter and a lot of droughts in the summer. In order to properly model the interseasonal variability as well as extreme precipitation events previous papers found that when resolution is increased the accuracy of the prediction also increases in the greater European region. In their study they also found that dry days and precipitation extremes are affected by the resolution of the models to a greater extent than standalone precipitation. This is extremely important for the Greek region because in Summer precipitation is extremely rare and the Greek region records a large number of consecutive dry days. In Winter on the other hand mainland Greece records a large number of extreme precipitation events, with most of them occurring in western Greece. Additionally in Greece precipitation exhibits very high spatial variability as documented by a wealth of research on this subject. This makes precipitation modelling in Greece quite hard, because of the high spatial resolution needed in order to adequately describe the precipitation variability in the area. Models with coarser resolutions fail to capture the variability of daily precipitation and additionally coarser models weren’t able to simulate precipitation intensity correctly. The different landscapes and microclimates that appear in the Greek region in combination with the high inter-seasonal variability create a challenging environment which is perfect for testing the efficacy and performance of climate models and downscaling methods.

 The aim of this study is to construct a high resolution (1km x 1km) database of precipitation, number of wet days and number of times precipitation exceeded 10mm and 20mm over Greece on a monthly and on an annual basis. In order to achieve this the ERA5 reanalysis dataset is downscaled using regression kriging with Histogram-Based Gradient Boosting Regression Trees. The independent variables used are spatial parameters derived from a high resolution digital elevation model and a selection of ERA5 reanalysis data. 

Our results confirmed biases that were also observed in previous papers whereby the ERA5 reanalysis overestimates the frequency of precipitation and underestimates its intensity. In our research we found that the number of wet days simulated by the ERA5 data was very inflated, while precipitation exceeding 10 and in particular 20 mm was understated. More specifically, in the Precipitation Totals the main improvements came from the increased resolution and an improvement on the spatial distribution of precipitation. In contrast, in the number of wet days and the number of times precipitation exceeded 10 mm there were large improvements in the metrics studied. In the number of wet days the RMSE halved on an annual basis with additional large reductions on a monthly basis. This was achieved by reducing the number of wet days simulated by the ERA5 dataset. On the number of days where precipitation exceeded 10mm there were improvements in both the metrics studied and the geographical distribution of the events. Finally, on the number of days where precipitation exceeded 20mm there were smaller improvements in the metrics because the occurrence of such events is very rare. However, it is safe to assume that P20 was underestimated in the ERA5 reanalysis and HGRP was able to improve its accuracy.

The largest improvements geographically were recorded in the region of Crete, where we found that the ERA5 reanalysis dataset underestimate every variable studied, with the exception of wet days. Next the mountainous regions of Peloponnese also recorded large improvement with the smallest improvements occurring in Western Greece. At this point however it is important to note that the gauge dataset used had a large number of stations in the mountainous regions of Crete, therefore in future research it would be helpful if more stations could be added in the mountainous regions of western Greece in particular, where there is also higher elevation and the bulk of precipitation occurs.

Overall the main driver of the variables studied continues to be the ERA5 variables however with our downscaling methodology we were able to achieve significant improvements in the metrics when compared to the standalone ERA5 dataset. The largest improvements were recorded in the wet days and the number of days where precipitation exceeded 10mm, while there were smaller to no improvements in the precipitation total. Finally, we can confidently conclude the algorithm tested seems to be a good fit for creating precipitation datasets.