Next Article in Journal
Optimizing Mean Fragment Size Prediction in Rock Blasting: A Synergistic Approach Combining Clustering, Hyperparameter Tuning, and Data Augmentation
Previous Article in Journal
Transition to the New Green Maritime Era—Developing Hybrid Ecological Fuels Using Methanol and Biodiesel—An Experimental Procedure
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Eng
Volume 5
Issue 3
10.3390/eng5030101
eng-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Creating High-Resolution Precipitation and Extreme Precipitation Indices Datasets by Downscaling and Improving on the ERA5 Reanalysis Data over Greece

by
Ntagkounakis Giorgos
1,*,
Panagiotis T. Nastos
1 and
Yiannis Kapsomenakis
2
1
Laboratory of Climatology and Atmospheric Environment, Department of Geology and Geoenvironment, National and Kapodistrian University of Athens, 15771 Athens, Greece
2
Research Center for Atmospheric Physics and Climatology, Academy of Athens, 11521 Athens, Greece
*
Author to whom correspondence should be addressed.
Eng 2024, 5(3), 1885-1904; https://doi.org/10.3390/eng5030101
Submission received: 29 June 2024 / Revised: 9 August 2024 / Accepted: 13 August 2024 / Published: 15 August 2024
(This article belongs to the Section Chemical, Civil and Environmental Engineering)
Browse Figures
Versions Notes

Abstract

:
The aim of this study was to construct a high-resolution (1 km × 1 km) database of precipitation, number of wet days, and number of times precipitation exceeded 10 mm and 20 mm over Greece on a monthly and on an annual basis. In order to achieve this, the ERA5 reanalysis dataset was 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, while as the dependent variable in the training stages, we used 97 precipitation gauges from the Hellenic National Meteorological Service for the period 1980–2010. These stations were also used for validation purposes using a leave-one-out cross-validation methodology. The results of the study showed that the algorithm is able to achieve better R2 and RMSE over the standalone ERA5 dataset over the Greek region. Additionally, the largest improvements were noticed in the wet days and in the precipitation over 10 and 20 mm, where the ERA5 reanalysis dataset overestimates the number of wet days and underestimates precipitation over 10 and 20 mm, while geographically, the ERA5 dataset performs the worst in the island regions of Greece. This indicates that the ERA5 dataset does not simulate the precipitation intensity accurately over the Greek region, and using our methodology, we were able to increase the accuracy and the resolution. Our approach delivers higher-resolution data, which are able to more accurately depict precipitation in the Greek region and are needed for comprehensive climate change hazard identification and analysis.

1. Introduction

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 and 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 (Figure 1). The large mountain ranges give precipitation in the country a distinct longitudinal shift [1]. Additionally, the country experiences intense interseasonal variability [2], which is in line with the Mediterranean climate. The Mediterranean basin experiences extreme precipitation events [3] in the winter and a lot of droughts in the summer [4]. In order to properly model the interseasonal variability as well as extreme precipitation events, Rauscher et al. [5] 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 [6,7]. 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. The difficulty in modelling precipitation with high spatial variability was also noted by Lee et al. [8]. They found that models with coarser resolutions fail to capture the variability of daily precipitation, and additionally, coarser models were not able to simulate precipitation intensity correctly. The different landscapes and microclimates that appear in the Greek region in combination with the high interseasonal variability create a challenging environment that is perfect for testing the efficacy and performance of climate models and downscaling methods.
Climate change is expected to intensify the hydrological cycle and change the spatial distribution of precipitation [9]. According to IPCC reports [10], the Mediterranean is expected to become warmer and dryer by the end of the 21st century. Furthermore, the warming is expected to be more intense in the Mediterranean region than the global mean [11]. In the Greek region, there is a decreasing trend in the precipitation recorded since 1950 [12,13]. Extreme precipitation in the region, on the other hand, in more recent research has revealed an increasing trend [14], which is in line with research conducted in the Mediterranean region [15,16].
Extreme precipitation results in flooding, which causes more economic losses than any other natural disaster. Greece has a long history of flooding [17,18], mainly caused by intense rainstorms, as snowmelt floods are not common in the region. In recent years, deforestation and urbanization have played an important role in increasing the severity and destructive power of floods. The Attica region has been extremely urbanized in the last 70 years, and as a result, it records the majority of the damages from floods [19]. Most of these floods are caused by extreme weather phenomena [20], and their destructive power has been amplified by the intense urbanization of the city [21]. Therefore, changes in extreme precipitation frequency and intensity are very important to policymakers in order to construct mitigation plans. For example, high-resolution precipitation datasets for can inform urban planners about the capacity and the extent to which they must implement stormwater control and other relevant management systems to avoid flooding [22], while the number of wet days alongside precipitation can be used to analyze droughts [23].
The regression kriging method has been used to simulate precipitation before [24,25,26,27]. Recently, the method was paired with more complex algorithms other than just a simple regression [28]. Gradient boosting is usually combined with decision trees to estimate susceptibility or risk [29,30] and with regression trees in order to fill data for meteorological time series [31] or some short-term prediction [32]. More recently, further research has been performed that indicates that gradient boosting can be used to statistically postprocess weather forecast data [33] to achieve better prediction accuracy or process satellite data to estimate total precipitable water [34]. Random forests are also used to predict precipitation [35,36] and temperature [37], indicating their efficacy in predicting climatic variables. Taking into account the previous success of gradient boosting and random forests in simulating precipitation in particular, the aim of this study is to pair the algorithms and utilize ERA5 precipitation data to create a high-resolution precipitation dataset. We believe that pairing two state-of-the-art algorithms along with the newest ERA5 reanalysis dataset will enable us to provide high-resolution data that will provide improvements to the ERA5’s performance over the Greek region. To achieve this, we utilized the land reanalysis, which has a higher resolution, and additionally, we added the AURELHY principal components as independent variables. Furthermore, we also examined extreme precipitation parameters and wet days in order further analyze specific components of precipitation over the region.

2. Materials and Methods

2.1. Data

The study period is 1980–2010, and the primary reason for the chosen period is the fact that the reanalysis datasets were given for this period, and this period is also usually considered a reference period by regional and global climate models. Additionally, there were a lot of meteorological station data for this period, while after 2010, a lot of stations were deprecated due to the financial crisis in Greece. The ERA5 reanalysis data that are used are generated using the same integrated forecasting system that was used in the ERA-Interim reanalysis but correcting the biases that were found. The reanalysis data are given at a 0.1° × 0.1° grid, and a variety of parameters from the ERA5 dataset are used as independent variables in the downscaling algorithm. The newest ERA5 land dataset improved the data resolution, but when validated with real datasets, its main shortcoming is that it overestimates the frequency and duration of precipitation and underestimates its intensity [38,39].
The geographical parameters were generated by a 12 m resolution TanDEM-X Elevation Model, which is a product generated from the TerraSAR-X satellite mission. This raster was upscaled to 1 km × 1 km, and then, the geographical parameters were calculated for each station. We believe that a 1 km resolution is adequate for analyzing the Greek region; therefore, we chose to upscale the raster to a coarser resolution. Future research that aims to apply the methodology to smaller areas and microclimates can utilize higher resolutions. To calculate the distance from shore, lakes’ and rivers’ geographical data were acquired from the government platform https://geodata.gov.gr/ (accessed on 15 May 2024). The daily precipitation totals, the elevation of the precipitation gauges, and their coordinates were provided and quality controlled by the Hellenic National Meteorological Service (HNMS). A total of 97 precipitation gauges (Figure 1) were used, covering most of the Greek mainland as well as the islands. From the gauge data, 77.7% of the stations have more than 50% of the data, 67% have more than 70% of the data, while 40.8% have more than 90% of the data. In our research, we aimed to use the raw gauge data to downscale the reanalysis dataset since we believe they best represent the climate of the region; therefore, no processing was performed to fill in the meteorological time series. The missing data are mostly evenly spread across the region studied, and although the stations do have missing data, they cover a substantial part of the Greek region, which will adequately represent the spatial variability of precipitation, and thus, they can be used as an accurate dependent variable for our analysis.

2.2. Methodology

The independent variables used from the ERA5 reanalysis are air temperature, wind speed, evaporation, surface solar radiation, surface pressure, cloud cover, and precipitation. The monthly precipitation total, wet days, and precipitation over 10 and 20 mm were also calculated on a monthly basis using the ERA5 data, and they were used as independent variables when downscaling for these parameters. The climatic elements used from the ERA5 reanalysis for the gauge position were taken from the nearest cell to the gauge. Additionally, the North Atlantic Circulation index (NAOI) was used, which has been proven to influence precipitation in the region [6,40]. The geographical parameters used were taken from the station position or generated by the digital elevation model, including longitude, latitude, elevation, distance from lakes, distance from river, distance from shore, slope, land-to-sea percentage, and the 15 AURELHY PCs. These AUREHLY parameters were calculated by selecting a squared matrix of 11 × 11 DEM grid cells around each point and calculating the elevation differences between the grid and point. This created 121 values to each point in the area studied; then, PCA was applied to the geographical parameters calculated using R “stats” library (version 4.2.2) [41]. In the present study, instead of the 11 × 11 area around the area, the nearest 32 neighbors in the latitudinal direction along with the nearest 46 neighbors in the longitudinal direction were used, making for a total of 1550 points (1550 = (2 × 16 + 1) × (2 × 23 + 1) − 1). After applying PCA, the first 15 principal components were chosen as independent variables for this study. This AYRELHY PC methodology was also used in the study of Mamara et al. [42] in order to generate temperature datasets and in a study by Gofa et al. [43] to generate precipitation datasets in Greece. The AUREHLY methodology and the resulting PCs provides additional information about the terrain of the area studied, which can then be used as independent variables to predict climatic variables that have been proven to be influenced by the geomorphology of the area and the AUREHLY PCs in particular [42,43,44]. The AUREHLY PCs for each month were chosen based on which combination provided the lowest RMSE in the training stages of our model and on guidance from previous studies that used the methodology. The dependent variables studied are total precipitation, wet days, number of days where precipitation exceeds 10 mm, and number of days where precipitation exceeds 20 mm. As wet days, we consider days where precipitation exceeds 1 mm. Additionally, all annual databases were constructed by adding the previously downscaled twelve months.
During the training stage, the independent variables were aggregated and trained against the gauge data. To validate the results, two validation methods are commonly used: One is to randomly split the data and use a smaller part of the dataset to validate the predicted data, and the other validation methodology is leave-one-out cross-validation (LOOCV), where one or more stations are used for validation and the rest for training. In this study, we chose to use the LOOCV methodology since the random-split methodology is subject to the random selection of data, which may influence the final results of the model. In our study, the LOOCV was applied by selecting each station and using it once for validation, effectively training the model as many times as there are stations. We chose the LOOCV methodology since we believe it is the strictest measure of performance since the model is effectively tested against every station where we have data. The maps shown below are the result of training the model with all the stations available, since they represent the model when the greatest amount of data was used. Additionally, maps of the RMSE difference, calculated with the LOOCV methodology, between the predicted and reanalysis data for each station are presented below.
When choosing the different parameters for the downscaling model for each of the different parameters studied (precipitation over 10 mm, 20 mm, wet days, and precipitation amount), an iterative methodology was used. First, the reanalysis parameter studied was trained against the gauge data to construct a model that acted as a “base” model. Every model constructed was then compared to this “base” model. Each of the variables studied was then added, and a model was reconstructed. The RMSE was compared between the original “base” model, and the new ones were constructed after adding each independent variable. The model with the lowest RMSE was chosen as the new “base” model, and the process restarted. In the new “base” model, each independent variable was added, and the performance was compared again. The process was repeated for every month until there was no improvement in RMSE from adding variables, optimizing the performance of the model for the inter-seasonal variability of precipitation. Each metric and model was constructed using the random-split validation ten times, using scikit-learn to split the dataset [45]. The random-split validation was chosen over the LOOCV methodology here in order to save time. The final parameters chosen are presented in Tables S1–S4 in the Supplementary Information section.
The statistical downscaling approach used in the study consists of a regression kriging with a histogram-based gradient boosting regression tree (HBGRT). The regression kriging technique is a spatial interpolation technique used in a variety of geographical problems, where the residuals from a regression model are spatially interpolated using a kriging technique. The regression model chosen was a regression tree, which is essentially a decision tree used for a continuous variable. The gradient boosting technique applied was an ensemble where a regression tree was added to the previous one in order to correct the errors. Finally, the input independent variables were categorized using a histogram in order to then train the regression trees on the most optimal set of input variables. This technique is used for large datasets in order to accelerate the training process while minimizing losses in performance. The whole model used was based on python’s scikit-learn library [45].

3. Results

The results section is categorized according to the variable studied. For each variable, a table is given that shows the improvements that were made in certain metrics (R2, correlation, and RMSE) in the downscaled dataset compared to the standalone ERA5 reanalysis dataset. The metrics were calculated for the downscaled dataset (from here on referenced as HGRP) and the standalone ERA5 reanalysis datasets against data from the gauges, and all formulas used for the calculation of metrics or in our methodology are presented in Formulas (S1)–(S11) in the Supplementary Materials Section. Additionally, maps are given on a seasonal and on an annual basis in order to showcase the differences between the datasets. The maps were constructed by calculating the mean for each dataset for the whole period of 1980–2010. Additionally, for each station, the RMSE was calculated by the LOOCV methodology, and the difference between the reanalysis RMSE and the downscaled RMSE was mapped for each season and on an annual basis. In the maps, positive values represent an improvement in RMSE, while negative values represent a worsening RMSE.

3.1. Precipitation Total

In the precipitation totals in winter, there are obvious improvements in the resolution of the data, while at the same time, the spatial variability of precipitation also improved. In particular, the model increases the precipitation in the mountainous Peloponnese region and in the western part of mountainous Crete, while at the same time, it decreases precipitation in the Thessaloniki region. From the RMSE differences map (Figure 3), we see that in the mountainous regions of Crete, there are significant improvements in the RMSE, where HGRP adds precipitation. In the Thessaloniki region, there is also an improvement in the RMSE, where the model reduces precipitation. In the mountainous Peloponnese region, there are not as many stations in order to validate the increase in precipitation that HGRP simulates. Additionally, in the RMSE differences map, we can see that there are also improvements in the western part of the Pindos mountain range, where the bulk of precipitation usually occurs in the Greek region. From Figure 2, we can observe that HGRP simulates more precipitation in the northwestern mountain range while also improving on the spatial distribution of the precipitation simulated from the ERA5 reanalysis. In the central part of Greece, there are the most increases of RMSE, where the model reduces precipitation for the region. Additionally, the largest increase in RMSE is recorded in the island of Samos, where HGRP does not seem to have changed the original ERA5 precipitation as much; however, the precipitation simulated in the island is significantly higher than the rest of the islands near Samos in both the ERA5 and our model.
In spring, the model decreases precipitation in the Greek region. More specifically, precipitation is reduced in the mountainous western Greece, in the Katerini region, and in the northern Peloponnese region. In the map of RMSE differences (Figure 3), we observe considerable improvements in the RMSE in those regions, in particular in the northwestern mountainous regions. In the northeastern parts of Greece, precipitation is kept about the same, while the RMSE differences remain mixed. In the islands of the Aegean, precipitation is about the same, with the model increasing precipitation in the islands in the east, in particular Mytilene, Samos, and Rodos. In the Aegean islands, RMSE differences are slightly negative, indicating a slightly worsening performance. Finally, in Crete, the model increases precipitation, particularly in the mountainous western regions. From Figure 3, the results from RMSE differences are mixed; in particular, in the eastern part of the island, a lot of improvements are observed, while in the western part of the island, there are slightly negative values.
In summer, the overall precipitation is also reduced by the model and in particular in the mountainous regions of northwestern Greece and in the Katerini region. In Thrace, precipitation is also reduced but not to the extent that precipitation is reduced in the previously mentioned regions. In the Aegean islands and Crete, the model is in agreement with the ERA5, and precipitation is essentially zero. At this point, it is worth noting that the ERA5 reanalysis dataset and the model perform the worst in these summer months (Table 1), and this can be explained by the overall lack of precipitation in Greece during summer, which makes simulating precipitation very challenging. This is also reflected in the RMSE difference maps, where the results are generally mixed. In the Aegean islands and in the mountainous northwestern regions, there are some slightly negative values, while in the northeastern part of the region, the RMSE suffers most. In the rest of the Greek regions, the values are positive overall. In Table 1, the improvements in R2 and correlation are also presented on a seasonal basis. The model’s outperformance over the ERA5 dataset occurs primarily during winter and autumn, coinciding with periods of heavy precipitation. Conversely, during the typically arid summer months, the model’s output closely resembles the reanalysis data due to the limited rainfall.
In autumn, the results between the HGRP and the ERA5 precipitation totals are more similar than the rest of the seasons. The only region where precipitation is meaningfully increased is the Crete region, where the biggest improvements in the RMSE are also recorded. In the rest of Greece, there are improvements in the spatial distribution of precipitation, in particular in the western Peloponnese region. The RMSE differences remain slightly positive in mainland Greece while slightly negative in the Aegean and Ionian islands.
On an annual basis, the results are similar to the ones recorded in autumn. There are improvements in the spatial distribution of precipitation in the Peloponnese region, and there is a meaningful increase in precipitation in the Crete region. Additionally, the biggest RMSE improvement on an annual basis is also recorded in the Crete region. In the mountainous regions of Pindos in northwestern Greece, the overall RMSE differences are positive, while for the Aegean and Ionian islands, the RMSE differences are mostly slightly negative. Samos remains an outlier, as the precipitation simulated by both the model and the ERA5 reanalysis is obviously significantly different from the rest of the islands, and the RMSE difference is also very different from the rest of the islands.
Although there are significant improvements recorded, because the results are not always uniform in every station, the metrics remain relatively the same compared to the reanalysis dataset (Table 1 and Table S5). The downscaling method seems to perform worst in the summer months while performing the best on an annual basis in every metric studied.

3.2. Wet Days

In Figure 4, when comparing the ERA5 wet days and the wet days from HGRP in all seasons, we can see a reduction in wet days across all areas of Greece, while at the same time, these reductions are translated into improvements in all areas and in all seasons, as we can see from the RMSE differences maps (Figure 5). More specifically, in winter, the pattern that is simulated by the ERA5 remains in the HGRP; however, the amount simulated is massively different. In the northwestern mountain range, the ERA5 dataset reaches the 43-wet-day limit, when in the HGRP, the wet days do not surpass the 33-day threshold. The reduction in wet days in the northwestern mountain range is validated by the RMSE differences maps, where RMSE is reduced in that region. Furthermore, east of the Pindos mountain range, in the greater Larissa and Lamia regions, there are some of the largest RMSE improvements, where HGRP also greatly reduced the number of wet days simulated by the ERA5 reanalysis dataset. Both the ERA5 and the HGRP simulate more wet days in the western part of Crete than the rest of Greece, although the HGRP wet days are much fewer than the ERA5. In that particular region, we can see that there are also very large improvements in the RMSE (Figure 5). In the Aegean and Ionian islands, we can also see that there are improvements in the RMSE across all stations, where the number of wet days is also reduced.
In spring and summer, HGRP seems to keep the spatial distribution of wet days from the ERA5 dataset, but the overall volume of wet days is also massively reduced. In the RMSE differences map, we can see massive improvements in both spring and summer across all stations and areas. In spring, RMSE improves across the northwestern mountain ranges the most, while some slightly negative differences are observed in the eastern part of Crete. In spring, the RMSE differences stay mostly positive across the Aegean and the Ionian islands, while in summer, most differences are positive, but there are a lot of differences that are at zero. This is due to the fact that precipitation is extremely rare; therefore, any difference between HGRP and the ERA5 dataset is extremely small.
In autumn, the ERA5 simulates a large volume of wet days across all mainland Greece, with the exception of Attica. In the Aegean islands, the wet days are noticeably fewer than in the rest of the region, with the exception of the western part of Crete. The bulk of wet days occurs in the northwestern region in both the ERA5 and the HGRP. From Figure 5, we can see that RMSE is improved across all Greece. The largest improvements are recorded in the western part of Crete and the greater Lamia region. These areas also showed very large improvements in the winter RMSE differences.
On an annual basis, the ERA5 dataset, similarly to autumn, simulates a lot of wet days across mainland Greece, while in the Aegean islands, with the exception of Crete, the wet days are noticeably fewer. The HGRP also reduces the volume of wet days simulated by the ERA5, like the rest of the seasons. The spatial distribution of the wet days remains very similar to the ERA5 wet days, but the elevation peaks seem to retain more wet days in HGRP, especially in Crete, where the mountain peaks seem to retain more wet days, which are better defined in HGRP than in the ERA5 reanalysis. The RMSE differences (Figure 5) seem to improve across all Greece. The RMSE is improved the most in the northeastern part of Greece and in the northwestern mountain ranges. There are some slight decreases in RMSE in the eastern part of Crete, where HGRP does not change the number of wet days significantly compared to the reanalysis dataset, especially in comparison to the rest of Greece. In the Aegean and Ionian islands, the RMSE mostly improves but to a lesser extent compared to the rest of Greece. This occurs because the islands also record less precipitation in general.
The downscaling model improves the metrics across all seasons and especially on an annual basis (Table 2 and Table S6). Although the R2 and correlation show small improvements, the RMSE improves dramatically. On an annual basis, the RMSE halves, while it also shows significant improvement in every season and month. In comparison to the precipitation totals, the R2 and correlation stay more uniform across all months and seasons, whereas in Table 1, there is a notable reduction in the summer months. Additionally, it is obvious that the ERA5 reanalysis dataset has an obvious bias towards more wet days.

3.3. Number of Days Precipitation Exceeds 10 mm

In the number of days that precipitation exceeds 10 mm (from here on P10), shown in Figure 6, we can observe that HGRP changes the spatial distribution of P10 significantly. In winter, the ERA5 reanalysis simulates most of P10 over the northwestern mountainous regions and in the western part of the Peloponnese region. The HGRP adds a lot of P10 over the western part of the Peloponnese region and in Crete; in the RMSE differences maps (Figure 7), we can see that in those particular areas, some of the largest improvements are recorded. In Crete in particular, very large improvements are observed across the whole island, and some of the largest improvements in all of Greece are observed in the mountainous regions of the island. In the Ionian and Aegean islands, the HGRP added some events in limited areas; however, from Figure 7, we observe that the overall results are mixed in terms of RMSE improvements. In the Attica and Evia regions, in Figure 6, we observe an increase in P10 and a change in the spatial distribution, which is translated into positive RMSE differences (Figure 7). In northern Greece and in the greater Larissa and Lamia regions, RMSE differences remain negative, and HGRP performs the worst. The changes in P10 made by HGRP are not as large as the ones mentioned above, and the RMSE differences are rather small, as they remain less than one in most stations.
In spring, the overall pattern of P10 is retained in the HGRP dataset; however, the greater resolution aids in capturing the Greek orography better. In Crete and the Aegean islands, HGRP seems to increase the events simulated by the ERA5 dataset. In Figure 7, we observe that the RMSE differences in those regions are positive, and in Crete in particular, we see the best improvements. In the Peloponnese, Attica, and northwestern Greece, the RMSE differences are positive, while in the greater Larissa and Thessaloniki regions, RMSE differences are mostly negative. In these particular areas, although there are no significant changes in the volume of P10, there are changes in the resolution of the data simulated.
In summer, it is worth noting that days where precipitation exceeds 10 mm are extremely rare and mostly occur in the northern mountainous regions of Greece. From Figure 6, we can see that HGRP reduces P10 in the Katerini and the northeastern part of Greece; however, the station coverage in those areas is not sufficient to make a conclusion regarding the improvement in RMSE. In the rest of the mainland Greek region, there are improvements in the spatial distribution of P10. In the Peloponnese region, we can see that the P10 generated by HGRP is more spread out compared to the standalone ERA5 dataset. In the Peloponnese region, there are also major improvements in RMSE (Figure 7). In the Aegean islands, the ERA5 dataset and HGRP are in agreement, as P10 is essentially zero in most occasions in those areas.
In autumn, the changes in the spatial distribution made by the HGRP are similar to the ones recorded in winter. HGRP increases P10 in the western Peloponnese region and in Crete, and in those regions, we also observe large RMSE improvements. Increases are also observed in the Katerini region, and there is a better spatial distribution in HGRP in northern Greece. These improvements are also observed in the RMSE differences map, where in most of northern Greece, RMSE improves in the HGRP. In the greater Attica and Evia regions, we observe an increase in P10 in the HGRP, and the RMSE differences are also mostly positive in the region. In the Aegean islands, HGRP slightly increases P10, with the RMSE also improving, as shown in in Figure 7.
On an annual basis, the changes made by HGRP are mostly similar to the ones recorded in autumn and winter, which is to be expected since P10 occurs the most in those seasons. From Figure 6, we can see that HGRP adds P10 in the western Peloponnese region and in Crete. In those particular areas, the biggest improvements of RMSE are recorded, similarly to autumn and winter. In the northwestern part of Greece, large improvements in RMSE are also recorded. These improvements are due to the higher spatial resolution of the HGRP and the better spatial distribution achieved. In Attica and Evia, there is also an increase in P10, which leads to improvements in the RMSE in those areas. In the Aegean and Ionian islands, the HGRP adds P10 in most islands and especially in the eastern part of the Aegean. In those areas, there are also positive improvements in Figure 7. In northern Greece and the greater Larissa and Lamia regions, there are fewer changes made by the HGRP, and this is also evident by the very small changes in RMSE recorded in those areas, as shown in Figure 7.
The number of days precipitation exceeds 10 mm also shows improvements in the metrics (Table 3); however, these improvements are not as numerous as the ones for the number of wet days. The improvements are mainly to the RMSE and less so in R2 and correlation. The reanalysis dataset here does not have an obvious bias towards more precipitation over 10 mm, in contrast to the number of wet days, where there was an overall bias towards more wet days. Rather, it seems to lack P10 in particular regions, which is further solidified by the maps of the RMSE difference, where the mountainous regions of the Peloponnese, Crete, and western Greece record most of the improvements (Figure 7).

3.4. Number of Days Precipitation Exceeds 20 mm

Firstly, days where precipitation exceed 20 mm (from here on P20) are very rare in the Greek region, and they mostly occur in the winter and autumn months, while in summer in particular, they are very rare. In winter, as shown in Figure 8, the ERA5 simulates almost all P20 in northwestern Greece and to a smaller extent in the Peloponnese and the northeastern Greek region. In the HGRP, P20 displays a greater spatial extent in all of Greece, and there is an increase in P20 events over the Peloponnese and Crete in particular. In these areas, there is also the largest improvement in RMSE, as we can see in Figure 9. In Attica and the Evia regions, there is an increase in P20 in the HGRP dataset, which improves RMSE across those areas. In the Aegean and Ionian islands, HGRP increases P20 across most islands. Samos is again an outlier, recording much higher amounts of P20 in the ERA5 reanalysis than the rest of the islands. These changes mostly result in improvements in RMSE in the Ionian and Aegean islands.
In spring, HGRP is very different from the standalone ERA5 reanalysis dataset. The events in HGRP are much more spread out than in the ERA5 reanalysis, where most of the events are simulated in the northwestern mountainous regions. In HGRP, the P20 in the northwestern mountainous regions is more defined throughout most of the region, which does improve RMSE in the southern part of the Pindos mountain range (Figure 9). In the Peloponnese region, there is also an increase in P20 in HGRP, especially in areas with higher elevation in the region, which improves RMSE in the area. In Crete, the standalone ERA5 dataset does not capture the orography of the island and, obviously, does not simulate enough P20 in the island, whereas in the HGRP, there is an increase in P20, particularly in higher-elevation areas, which improves RMSE in the area (Figure 9). Additionally, HGRP increases P20 in the northeastern part of Greece and also improves on the spatial distribution and capturing the underlying elevation of the region better. As seen in Figure 9, there is an improvement in the RMSE in the HGRP in northeastern Greece from the changes made by the HGRP. In the Aegean and Ionian islands, HGRP adds P20 in most areas, which translates into mostly positive RMSE changes. In summer, HGRP follows the pattern of the rest of the seasons, where P20 is more spread out in the Greek region; however, in summer, P20 events are almost nonexistent; therefore, this is a mistake. This is further confirmed by the RMSE differences, which are overwhelmingly negative in summer, making this the only season and parameter where HGRP does not improve on the standalone ERA5 dataset.
In autumn, similarly to winter, most P20 simulated by the ERA5 reanalysis dataset is centered around the northwestern mountainous regions and to a lesser extent in the Peloponnese region. In the HGRP, there is an increase in P20 in the Peloponnese region and Crete, which improves RMSE significantly (Figure 9). In Attica and Evia, there is an increase in P20 in the HGRP and a greater spatial distribution of the events, and this is validated by an improvement in the RMSE metrics in these regions. Additional increases of P20 are also recorded in the northeastern part of Greece, where the RMSE differences remain mixed, and in the Aegean and Ionian islands, where RMSE mostly improves.
On an annual basis, the changes made by HGRP are similar to winter and autumn, which is to be expected as most P20 happens in those seasons, similarly to P10. HGRP increases P20 in the greater Greek region, with the most substantial increases happening in the western part of the mountainous regions of Peloponnese, in Crete, and in north eastern Greece. In the Peloponnese region, the changes result in improvements in RMSE; similarly, in the mountainous regions of Crete, the RMSE improves the most in Greece. By contrast, in the northeastern part of Greece, the metrics deteriorate by the increase in P20 in the region. In Attica and Evia, there is also an increase in P20, which results in positive changes in RMSE. In the Ionian and Aegean islands, there are mostly increases by the HGRP, which improves the metrics in those regions. Samos remains an outlier, where the ERA5 dataset simulates much more P20 in the island when compared to the rest of the region, and this is transferred over to the HGRP.
In Table 4, there are improvements in all metrics across all seasons except summer. Obviously, the metrics recorded in P20 are worse than the rest of the parameters studied because these events are very rare in the Greek region. It is safe to conclude that the ERA5 reanalysis simulates a lot fewer days where precipitation exceeds 20 mm, and this is improved by the HGRP. From Figure 8, we observe that in order to improve the results of the ERA5, HGRP increases P20 in all seasons, and unfortunately, this is not correct in summer, which is why it is the only month where the model underperforms the ERA5 dataset. Additionally, in the rest of the seasons, there are significant improvements in the spatial resolution of the ERA5 dataset, which are not adequately reflected in the metrics.

4. Discussion

Firstly, we can confidently conclude that this research further validates the results from Wu et al. [38] and Jiang et al. [39], where they found that the ERA5 overestimates the frequency and duration of precipitation and underestimates its intensity. With our methodology, we were able to significantly improve the accuracy of the predictions by both improving the values simulated by the model and their geographical distribution. The largest improvements were definitely in the number of wet days and the number of days where precipitation exceeded 10 and 20 mm. In contrast, the total precipitation did not exhibit any large improvement in the metrics studied; instead, most of the improvement came from the higher resolution achieved by the downscaling model and the improved geographical distribution of the precipitation.
More specifically, in the number of wet days, there is a wide difference between the simulated and the real values. The main improvement in the number of wet days was the reduction in RMSE and less so in the R2. This occurred because the geographical distribution of wet days simulated by the ERA5 is correct, but the quantity of wet days simulated is extremely inflated. In the number of days precipitation exceeded 10 mm, the improvements were smaller compared to the wet days and less conclusive in terms of the bias in the model. The downscaling method does seem to increase the quantity of the events over the total of the Greek region and, in particular, western Greece and Crete. Furthermore, the ERA5 reanalysis significantly underestimates the number of days precipitation exceeded 20 mm. Although with our downscaling methodology, we were able to achieve significant improvements, the annual R2 achieved by our model is still only 0.31 compared to the 0.24 R2 of the ERA5 reanalysis. In comparison, the downscaled annual R2 is 0.45 and 0.58, and the ERA5 reanalysis R2 is 0.38 and 0.56 for number of days precipitation exceeded 10 mm and wet days, respectively. These results indicate that, ultimately, the main force for the downscaling method is still the ERA5 data and that the rarer and more extreme a parameter is, the harder it is to simulate in general.
In terms of geographical distribution, we can see that the ERA5 is not able to correctly simulate the precipitation that occurs in the mountainous regions of the Peloponnese and Crete. In Crete in particular, the ERA5 reanalysis underestimated every variable studied, with the exception of the number of wet days. It is important to note that the gauge dataset we used does have a lot of high-elevation stations in the mountainous regions of Crete, which could be one of the reasons that the differences are so pronounced in the region. However, from the maps of the RMSE difference, we can see that some of the biggest improvements occur in that region. The overall results are also further validated by the R2 differences maps (Figures S1–S4) presented in the Supplementary Information section and the metrics presented in Tables S5–S8. In the rest of Greece, although there is an adequate number of stations covering the whole region, we could not include any very high-elevation stations in the mountainous areas of western Greece, which could influence the overall results since the bulk of precipitation in Greece occurs in the mountainous regions of western Greece. If such data do become available, they could be the subject of future research.
Overall, the improvements in the metrics can be explained by the increased resolution of our dataset, which greatly influences the performance of ERA5 in the oceanic regions of Greece since in Table 5, we can observe that the bulk of improvements happen at islands’ stations. The increased resolution allows the model to better depict the geography of the islands, which greatly increases the accuracy of the predictions made. Achieving high-accuracy results for the maritime regions of Greece is especially important since they are highly water-stressed areas, which are expected to be greatly impacted by future temperature increases.

5. Conclusions

The goal of this study was to create high-resolution (1 km × 1 km) monthly databases for precipitation totals, number of wet days, and number of days precipitation exceeded 10 and 20 mm using regression kriging with a histogram-based gradient boosting regression tree. In order to achieve this, we used climatic data from the newest land-based reanalysis dataset, geospatial variables from a high-resolution digital elevation model, the AUREHLY principal components, and the North Atlantic Circulation Index as the independent variables. As dependent variables, we used 97 precipitation gauges from the Hellenic National Meteorological Service for the period 1980–2010. In order to compare the results between the standalone ERA5 dataset and our downscaling methodology, we used an iterative LOOCV cross-validation. The downscaling was carried out on a monthly basis, where both the gauge data and the ERA5 data were aggregated on a monthly basis and then downscaled.
Our results confirmed biases that were also observed in previous papers [38,39], 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. One of the reasons behind these biases may be because of the coarse resolution of the model, which does not manage to capture the intense geographical variation that precipitation exhibits in the Greek region. With our methodology and the higher resolution we achieved, we managed to correct some of these biases, especially in the Greek islands, which recorded most of the increases in accuracy (Table 5). More specifically, in the precipitation totals, the main improvements came from the increased resolution and an improvement in the spatial distribution of precipitation, while in the island stations, the metrics were improved with RMSE decreasing by 7.7% while remaining mostly the same on the rest of Greece. 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 and seasonal basis. This was achieved by reducing the number of wet days simulated by the ERA5 dataset. On the number of days where precipitation exceeded 10 mm, there were improvements in both the metrics studied and the geographical distribution of the events. Finally, on the number of days where precipitation exceeded 20 mm, 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 differences in the variables studied can be attributed first to the coarse resolution of the ERA5 dataset, but additionally, the geographical variability of a small and complex region like Greece poses unique challenges in simulating its climate variable that cannot be addressed by more generalized models made in order to depict the European and global climate. In future versions of the ERA dataset, where higher resolution is achieved, we could expect more accurate results.
The largest improvements geographically were recorded in the region of Crete, where we found that the ERA5 reanalysis dataset underestimated every variable studied, with the exception of wet days. Next, the mountainous regions of the Peloponnese also recorded a 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 where 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 10 mm, while there were smaller to no improvements in the precipitation total. The improvements mainly occurred by better distribution of precipitation, wet days, P10, and P20 in the Greek region, which can be attributed to the better resolution that we were able to achieve when compared to the standalone dataset as well as improving on the wet days bias that was clearly exhibited by the ERA5 dataset. The improvements made in relation to P10 and P20 indicate that extreme precipitation indices provided by the standalone ERA5 dataset should be adjusted to correct the biases of the dataset before being used for hazard assessments like flooding, droughts, landslide susceptibility, etc. Caution is also warranted when using the ERA5 for analysis in small island regions since its resolution may not be adequate depending on the size of the region. Our research improves upon the ERA5 dataset by offering new insights on its accuracy in Greece, while we can also confidently conclude the algorithm tested seems to be a good fit for creating precipitation datasets.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/eng5030101/s1, Table S1: Parameters used for downscaling Precipitation; Table S2: Parameters Used for Downscaling Number of Wet Days; Table S3: Parameters Used for Downscaling Number of Days with Precipitation over 10 mm; Table S4: Parameters Used for Downscaling Number of Days with Precipitation over 20 mm; Table S5: Monthly precipitation total metrics; Table S6: Monthly number of wet days metrics; Table S7: Monthly number of days where precipitation exceeds 10 mm metrics; Table S8: Monthly number of days where precipitation exceeds 20 mm metrics; Figure S1: Maps of precipitation total R2 difference for each station; Figure S2: Maps of number of wet days R2 difference for each station; Figure S3: Maps of number of days where precipitation exceeds 10 mm R2 difference for each station; Figure S4: Maps of number of days where precipitation exceeds 20 mm R2 difference for each station; Formula (S1): Root Mean Square Error; Formula (S2): Pearson correlation formula; Formula (S3): R2 formula; Formula (S4)–(S8): Gradient Boosting Algorithm. Formula S9: Random Forest Regression Formula; Formula (S10) and (S11): Principal Component Analysis.

Author Contributions

Conceptualization, N.G.; Methodology, N.G., P.T.N. and Y.K.; Writing—original draft, N.G.; Writing—review & editing, P.T.N. and Y.K.; Visualization, N.G.; Supervision, P.T.N. and Y.K.; Project administration, P.T.N. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The gauge data used in this study were provided by the Hellenic National Meteorological Service, the reanalysis data by the ECMWF, the digital elevation model by TerraSAR-X satellite mission and the shoreline, and the lake and river datasets by the https://geodata.gov.gr/ platform (accessed on 15 May 2024). The datasets generated during the current study are available from the corresponding author on reasonable request.

Acknowledgments

We would like to thank Hellenic National Meteorological Service for providing us with precipitation gauge data to use in our study.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Sindosi, O.A.; Bartzokas, A.; Kotroni, V.; Lagouvardos, K. Influence of orography on precipitation amount and distribution in NW Greece; a case study. Atmos. Res. 2015, 152, 105–122. [Google Scholar] [CrossRef]
  2. Metaxas, D.A.; Philandras, C.M.; Nastos, P.T.; Repapis, C.C. Variability of precipitation pattern in Greece during the year. Fresenius Environ. Bull. 1999, 8, 001–006. [Google Scholar]
  3. Toreti, A.; Xoplaki, E.; Maraun, D.; Kuglitsch, F.G.; Wanner, H.; Luterbacher, J. Characterisation of extreme winter precipitation in Mediterranean coastal sites and associated anomalous atmospheric circulation patterns. Nat. Hazards Earth Syst. Sci. 2010, 10, 1037–1050. [Google Scholar] [CrossRef]
  4. Hoerling, M.; Eischeid, J.; Perlwitz, J.; Quan, X.; Zhang, T.; Pegion, P. On the increased frequency of Mediterranean drought. J. Clim. 2012, 25, 2146–2161. [Google Scholar] [CrossRef]
  5. Rauscher, S.A.; O’Brien, T.A.; Piani, C.; Coppola, E.; Giorgi, F.; Collins, W.D.; Lawston, P.M. A multimodel intercomparison of resolution effects on precipitation: Simulations and theory. Clim. Dyn. 2016, 47, 2205–2218. [Google Scholar] [CrossRef]
  6. Feidas, H.; Noulopoulou, C.; Makrogiannis, T.; Bora-Senta, E. Trend analysis of precipitation time series in Greece and their relationship with circulation using surface and satellite data: 1955–2001. Theor. Appl. Climatol. 2007, 87, 155–177. [Google Scholar] [CrossRef]
  7. Nastos, P.T.; Politi, N.; Kapsomenakis, J. Spatial and temporal variability of the Aridity Index in Greece. Atmos. Res. 2013, 119, 140–152. [Google Scholar] [CrossRef]
  8. Lee, M.H.; Im, E.S.; Bae, D.H. Impact of the spatial variability of daily precipitation on hydrological projections: A comparison of GCM-and RCM-driven cases in the Han River basin, Korea. Hydrol. Process. 2019, 33, 2240–2257. [Google Scholar] [CrossRef]
  9. Donat, M.G.; Lowry, A.L.; Alexander, L.V.; O’Gorman, P.A.; Maher, N. More extreme precipitation in the world’s dry and wet regions. Nat. Clim. Chang. 2016, 6, 508–513. [Google Scholar] [CrossRef]
  10. Stocker, T.F.; Qin, D.; Plattner, G.K.; Tignor, M.M.M.B.; Allen, S.K.; Boschung, J.; Nauels, A.; Xia, Y.; Bex, V.; Midgley, P.M. Contribution of working group I to the fifth assessment report of the intergovernmental panel on climate change. In Climate Change 2013: The Physical Science Basis; Cambridge University Press: Cambridge, UK; New York, NY, USA, 2013; Volume 5, pp. 1–1552. [Google Scholar]
  11. Giorgi, F.; Lionello, P. Climate change projections for the Mediterranean region. Glob. Planet. Chang. 2008, 63, 90–104. [Google Scholar] [CrossRef]
  12. Stefanidis, S.; Stathis, D. Spatial and temporal rainfall variability over the Mountainous Central Pindus (Greece). Climate 2018, 6, 75. [Google Scholar] [CrossRef]
  13. Markonis, Y.; Batelis, S.C.; Dimakos, Y.; Moschou, E.; Koutsoyiannis, D. Temporal and spatial variability of rainfall over Greece. Theor. Appl. Climatol. 2017, 130, 217–232. [Google Scholar] [CrossRef]
  14. Tzanis, C.G.; Pak, A.N.; Koutsogiannis, I.; Philippopoulos, K. Climatology of Extreme Precipitation from Observational Records in Greece. Environ. Sci. Proc. 2022, 19, 51. [Google Scholar] [CrossRef]
  15. Partal, T.; Kahya, E. Trend analysis in Turkish precipitation data. Hydrol. Process. 2006, 20, 2011–2026. [Google Scholar] [CrossRef]
  16. Ruiz-Leo, A.M.; Hernández, E.; Queralt, S.; Maqueda, G. Convective and stratiform precipitation trends in the Spanish Mediterranean coast. Atmos. Res. 2013, 119, 46–55. [Google Scholar] [CrossRef]
  17. Angelakis, A.N.; Antoniou, G.; Voudouris, K.; Kazakis, N.; Dalezios, N.; Dercas, N. History of floods in Greece: Causes and measures for protection. Nat. Hazards 2020, 101, 833–852. [Google Scholar] [CrossRef]
  18. Diakakis, M.; Mavroulis, S.; Deligiannakis, G. Floods in Greece, a statistical and spatial approach. Nat. Hazards 2012, 62, 485–500. [Google Scholar] [CrossRef]
  19. Mimikou, M.; Koutsoyiannis, D. Extreme floods in Greece: The case of 1994. In Proceedings of the US-ITALY Research Workshop on the Hydrometeorology, Impacts, and Management of Extreme Floods, Perugia, Italy, 13–17 November 1995. [Google Scholar]
  20. Mimikou, M.; Baltas, E.; Varanou, E. A Study of Extreme Storm Events in the Greater Athens Area, Greece; IAHS-AISH Publ.: Wallingford, CT, USA, 2002; pp. 161–165. [Google Scholar]
  21. Bathrellos, G.D.; Karymbalis, E.; Skilodimou, H.D.; Gaki-Papanastassiou, K.; Baltas, E.A. Urban flood hazard assessment in the basin of Athens Metropolitan city, Greece. Environ. Earth Sci. 2016, 75, 319. [Google Scholar] [CrossRef]
  22. Chen, Y.; Samuelson, H.W.; Tong, Z. Integrated design workflow and a new tool for urban rainwater management. J. Environ. Manag. 2016, 180, 45–51. [Google Scholar] [CrossRef]
  23. Funk, C.; Harrison, L.; Alexander, L.; Peterson, P.; Behrangi, A.; Husak, G. Exploring trends in wet-season precipitation and drought indices in wet, humid and dry regions. Environ. Res. Lett. 2019, 14, 115002. [Google Scholar] [CrossRef]
  24. Bajat, B.; Pejović, M.; Luković, J.; Manojlović, P.; Ducić, V.; Mustafić, S. Mapping average annual precipitation in Serbia (1961–1990) by using regression kriging. Theor. Appl. Climatol. 2013, 112, 1–13. [Google Scholar] [CrossRef]
  25. Teng, H.; Shi, Z.; Ma, Z.; Li, Y. Estimating spatially downscaled rainfall by regression kriging using TRMM precipitation and elevation in Zhejiang Province, southeast China. Int. J. Remote Sens. 2014, 35, 7775–7794. [Google Scholar] [CrossRef]
  26. Paparrizos, S.; Maris, F.; Matzarakis, A. Integrated analysis of present and future responses of precipitation over selected Greek areas with different climate conditions. Atmos. Res. 2016, 169, 199–208. [Google Scholar] [CrossRef]
  27. Nastos, P.T.; Kapsomenakis, J.; Philandras, K.M. Evaluation of the TRMM 3B43 gridded precipitation estimates over Greece. Atmos. Res. 2016, 169, 497–514. [Google Scholar] [CrossRef]
  28. Seo, Y.; Kim, S.; Singh, V.P. Estimating spatial precipitation using regression kriging and artificial neural network residual kriging (RKNNRK) hybrid approach. Water Resour. Manag. 2015, 29, 2189–2204. [Google Scholar] [CrossRef]
  29. Song, Y.; Niu, R.; Xu, S.; Ye, R.; Peng, L.; Guo, T.; Li, S.; Chen, T. Landslide susceptibility mapping based on weighted gradient boosting decision tree in Wanzhou section of the Three Gorges Reservoir Area (China). ISPRS Int. J. Geo-Inf. 2018, 8, 4. [Google Scholar] [CrossRef]
  30. Tien Bui, D.; Ho, T.C.; Pradhan, B.; Pham, B.T.; Nhu, V.H.; Revhaug, I. GIS-based modeling of rainfall-induced landslides using data mining-based functional trees classifier with AdaBoost, Bagging, and MultiBoost ensemble frameworks. Environ. Earth Sci. 2016, 75, 1101. [Google Scholar] [CrossRef]
  31. Körner, P.; Kronenberg, R.; Genzel, S.; Bernhofer, C. Introducing Gradient Boosting as a universal gap filling tool for meteorological time series. Meteorol. Z. 2018, 27, 369. [Google Scholar] [CrossRef]
  32. Liao, S.; Liu, Z.; Liu, B.; Cheng, C.; Jin, X.; Zhao, Z. Multistep-ahead daily inflow forecasting using the ERA-Interim reanalysis data set based on gradient-boosting regression trees. Hydrol. Earth Syst. Sci. 2020, 24, 2343–2363. [Google Scholar] [CrossRef]
  33. Velthoen, J.; Dombry, C.; Cai, J.J.; Engelke, S. Gradient boosting for extreme quantile regression. Extremes 2023, 26, 639–667. [Google Scholar] [CrossRef]
  34. He, X.; Chaney, N.W.; Schleiss, M.; Sheffield, J. Spatial downscaling of precipitation using adaptable random forests. Water Resour. Res. 2016, 52, 8217–8237. [Google Scholar] [CrossRef]
  35. Lee, Y.; Han, D.; Ahn, M.H.; Im, J.; Lee, S.J. Retrieval of total precipitable water from Himawari-8 AHI data: A comparison of random forest, extreme gradient boosting, and deep neural network. Remote Sens. 2019, 11, 1741. [Google Scholar] [CrossRef]
  36. Yan, X.; Chen, H.; Tian, B.; Sheng, S.; Wang, J.; Kim, J.S. A downscaling–merging scheme for improving daily spatial precipitation estimates based on random forest and cokriging. Remote Sens. 2021, 13, 2040. [Google Scholar] [CrossRef]
  37. Pang, B.; Yue, J.; Zhao, G.; Xu, Z. Statistical downscaling of temperature with the random forest model. Adv. Meteorol. 2017, 2017, 7265178. [Google Scholar] [CrossRef]
  38. Wu, G.; Qin, S.; Mao, Y.; Ma, Z.; Shi, C. Validation of precipitation events in ERA5 to gauge observations during warm seasons over eastern China. J. Hydrometeorol. 2022, 23, 807–822. [Google Scholar] [CrossRef]
  39. Jiang, Y.; Yang, K.; Shao, C.; Zhou, X.; Zhao, L.; Chen, Y.; Wu, H. A downscaling approach for constructing high-resolution precipitation dataset over the Tibetan Plateau from ERA5 reanalysis. Atmos. Res. 2021, 256, 105574. [Google Scholar] [CrossRef]
  40. Kalimeris, A.; Ranieri, E.; Founda, D.; Norrant, C. Variability modes of precipitation along a Central Mediterranean area and their relations with ENSO, NAO, and other climatic patterns. Atmos. Res. 2017, 198, 56–80. [Google Scholar] [CrossRef]
  41. R Core Team. R: A Language and Environment for Statistical Computing, version 4.2.2; R Foundation for Statistical Computing: Vienna, Austria, 2022; Available online: https://www.R-project.org/ (accessed on 31 October 2022).
  42. Mamara, A.; Anadranistakis, M.; Argiriou, A.A.; Szentimrey, T.; Kovacs, T.; Bezes, A.; Bihari, Z. High resolution air temperature climatology for Greece for the period 1971–2000. Meteorol. Appl. 2017, 24, 191–205. [Google Scholar] [CrossRef]
  43. Gofa, F.; Mamara, A.; Anadranistakis, M.; Flocas, H. Developing gridded climate data sets of precipitation for Greece based on homogenized time series. Climate 2019, 7, 68. [Google Scholar] [CrossRef]
  44. Hiebl, J.; Frei, C. Daily temperature grids for Austria since 1961—Concept, creation and applicability. Theor. Appl. Climatol. 2016, 124, 161–178. [Google Scholar] [CrossRef]
  45. Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 2011, 12, 2825–2830. [Google Scholar]
Eng 05 00101 g001
Figure 1. Digital Elevation Mode and Stations used in this study.
Figure 1. Digital Elevation Mode and Stations used in this study.
Eng 05 00101 g001
Eng 05 00101 g002
Figure 2. Maps for mean ERA5 and predicted precipitation.
Figure 2. Maps for mean ERA5 and predicted precipitation.
Eng 05 00101 g002
Eng 05 00101 g003
Figure 3. Maps of precipitation total RMSE difference for each station.
Figure 3. Maps of precipitation total RMSE difference for each station.
Eng 05 00101 g003
Eng 05 00101 g004
Figure 4. Maps for mean ERA5 and predicted number of wet days.
Figure 4. Maps for mean ERA5 and predicted number of wet days.
Eng 05 00101 g004
Eng 05 00101 g005
Figure 5. Maps of number of wet days RMSE difference for each station.
Figure 5. Maps of number of wet days RMSE difference for each station.
Eng 05 00101 g005
Eng 05 00101 g006
Figure 6. Maps for mean ERA5 and predicted number of days where precipitation exceeds 10 mm.
Figure 6. Maps for mean ERA5 and predicted number of days where precipitation exceeds 10 mm.
Eng 05 00101 g006
Eng 05 00101 g007
Figure 7. Maps of number of days where precipitation exceeds 10 mm RMSE difference for each station.
Figure 7. Maps of number of days where precipitation exceeds 10 mm RMSE difference for each station.
Eng 05 00101 g007
Eng 05 00101 g008
Figure 8. Maps for mean ERA5 and predicted number of days where precipitation exceeds 20 mm.
Figure 8. Maps for mean ERA5 and predicted number of days where precipitation exceeds 20 mm.
Eng 05 00101 g008
Eng 05 00101 g009
Figure 9. Maps of number of days where precipitation exceeds 20 mm RMSE difference for each station.
Figure 9. Maps of number of days where precipitation exceeds 20 mm RMSE difference for each station.
Eng 05 00101 g009
Table 1. Seasonal precipitation total metrics.
Table 1. Seasonal precipitation total metrics.
SeasonR2CorrelationRMSE
ModelReanalysisModelReanalysisModelReanalysis
Winter0.450.400.670.63125.41126.52
Spring0.440.450.670.6761.7364.63
Summer0.450.520.670.7234.0331.68
Autumn0.460.460.680.6888.3289.89
Annual0.420.400.650.63226.32233.99
Table 2. Seasonal number of wet days metrics.
Table 2. Seasonal number of wet days metrics.
SeasonR2CorrelationRMSE
ModelReanalysisModelReanalysisModelReanalysis
Winter0.600.600.780.785.9411.19
Spring0.630.600.790.774.049.88
Summer0.710.680.840.832.255.10
Autumn0.560.510.750.714.037.60
Annual0.580.560.760.7511.8329.50
Table 3. Seasonal number of days where precipitation exceeds 10 mm metrics.
Table 3. Seasonal number of days where precipitation exceeds 10 mm metrics.
SeasonR2CorrelationRMSE
ModelReanalysisModelReanalysisModelReanalysis
Winter0.530.370.730.613.634.63
Spring0.360.340.600.582.422.59
Summer0.370.320.600.561.241.34
Autumn0.450.420.670.652.602.98
Annual0.450.380.670.627.048.30
Table 4. Seasonal number of days where precipitation exceeds 20 mm metrics.
Table 4. Seasonal number of days where precipitation exceeds 20 mm metrics.
SeasonR2CorrelationRMSE
ModelReanalysisModelReanalysisModelReanalysis
Winter0.340.190.580.432.633.23
Spring0.210.170.460.421.501.64
Summer0.020.090.140.300.790.79
Autumn0.380.340.610.581.832.11
Annual0.310.240.560.484.685.64
Table 5. Metrics for annual precipitation totals, annual number of wet days, annual number of P10, and annual number of P20 for the mainland region of Greece and the island region.
Table 5. Metrics for annual precipitation totals, annual number of wet days, annual number of P10, and annual number of P20 for the mainland region of Greece and the island region.
AreaR2CorrelationRMSE
ModelReanalysisModelReanalysisModelReanalysis
Annual Precipitation Total
Continental0.400.450.630.67220.71219.47
Oceanic0.440.370.660.61234.27253.69
Annual Number of Wet Days
Continental0.570.550.760.7412.1332.11
Oceanic0.530.530.730.7311.3725.22
Annual Number of P10
Continental0.410.450.640.677.437.65
Oceanic0.510.300.720.556.429.15
Annual Number of P20
Continental0.270.290.520.544.745.04
Oceanic0.370.190.610.434.586.41
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Giorgos, N.; Nastos, P.T.; Kapsomenakis, Y. Creating High-Resolution Precipitation and Extreme Precipitation Indices Datasets by Downscaling and Improving on the ERA5 Reanalysis Data over Greece. Eng 2024, 5, 1885-1904. https://doi.org/10.3390/eng5030101

AMA Style

Giorgos N, Nastos PT, Kapsomenakis Y. Creating High-Resolution Precipitation and Extreme Precipitation Indices Datasets by Downscaling and Improving on the ERA5 Reanalysis Data over Greece. Eng. 2024; 5(3):1885-1904. https://doi.org/10.3390/eng5030101

Chicago/Turabian Style

Giorgos, Ntagkounakis, Panagiotis T. Nastos, and Yiannis Kapsomenakis. 2024. "Creating High-Resolution Precipitation and Extreme Precipitation Indices Datasets by Downscaling and Improving on the ERA5 Reanalysis Data over Greece" Eng 5, no. 3: 1885-1904. https://doi.org/10.3390/eng5030101

Article Metrics

Back to TopTop