Towards pan-European drought risk maps: quantifying the link between drought indices and reported drought impacts

Drought in Europe is a hazard with a wide range of transboundary, environmental and socio-economic impacts on various sectors including agriculture, energy production, public water supply and water quality. Despite the apparent importance of this natural hazard, observed pan-European drought impacts have not yet been quantitatively related to the most important climatological drivers to map drought risk on a continental scale. This contribution approaches the issue by quantitatively assessing the likelihood of drought impact occurrence as a function of the standardized precipitation evapotranspiration index for four European macro regions using logistic regression. The resulting models allow mapping the sector-specific likelihood of drought impact occurrence for specific index levels. For the most severe drought conditions the maps suggest the highest risk of impact occurrence for ‘Water Quality’ in Maritime Europe, followed by ‘Agriculture & Livestock Farming’ in Western Mediterranean Europe and ‘Energy & Industry’ in Maritime Europe. Merely impacts on ‘Public Water Supply’ result in overall lower risk estimates. The work suggests that modeling and mapping for North- and Southeastern Europe requires further enhancement to the impact database in these regions. Such maps may become an essential component of drought risk management to foster resilience for this hazard at large scale.


Introduction
Drought is a natural hazard known to be very difficult to grasp and its general characteristics, creeping onset, long lasting duration, large spatial extent and crossboundary effects have hindered scientists and practitioners to precisely define the hazard (Wilhite et al 2007, EEA 2009). Independent of its definition, drought can cause a variety of direct and indirect, negative (and sometimes positive) impacts. Even though the majority of drought impact research and public recognition focuses on the agricultural sector, drought has more damage potential. Its multifaceted character affects a variety of environmental and socioeconomic systems that can be classified into a number of different categories (Stahl et al 2012).
For the last three decades, the European Commission estimated that drought has caused over 100 billion Euros of losses to the European Union members (EC 2007b), has covered about 37% of Europe's surface, and has affected more than 100 million inhabitants (Kossida et al 2012). Despite some controversy over global drought trends (Dai 2012, Sheffield et al 2012there is medium confidence that Southern Europe has experienced more intense and longer droughts (IPCC 2012). Furthermore, there is medium confidence that drought will intensify for the Mediterranean and Central European region due to climate change (IPCC 2012). Hence, increasing resilience to this hazard through an appropriate drought risk assessment at different scales is a valid concern and important to address. In Europe, an assessment at the pan-European scale, i.e. to support EU level policy development and cross boundary communication is still pending.
The risk of natural disasters in a very general sense is a combination of hazard and vulnerability Any further distribution of this work must maintain attribution to the author (s) and the title of the work, journal citation and DOI. (IPCC 2007). Commonly, the drought hazard is described by one or a set of drought indicators and for Europe the standardized precipitation and evapotranspiration index (SPEI) (Vicente-Serrano et al 2010a) has become popular in recent years (Potop 2011, López-Moreno et al 2013, Spinoni et al 2013. But, even though more than 100 drought indicators are known (Zargar et al 2011) and have been compared in a number of studies, clear guidance on their usage is still lacking. One reason for this is that most indices describe general anomalies of meteorological conditions, but only few drought indices were developed with or have been tested against observed drought impact data. Vulnerability to drought is typically estimated by a combination of relevant, subjectively weighted vulnerability factors (Kumar 2008, Jordaan 2012, Fu et al 2013, Sreedhar et al 2013. This approach requires explicit but difficult to obtain information on physical, ecological and socioeconomic parameters. Although drought impacts are symptoms of vulnerability (Knutson et al 1998), the majority of current approaches do not consider past drought impact reports to estimate vulnerability, and only a few studies have validated their approaches using historical drought impact observations (Aggett 2012, Naumann et al 2013, Karavitis et al 2014).
The presented work takes advantage of a new database of reported drought impacts in Europe to derive a first-order estimate of pan-European drought risk. The employed approach estimates the regional likelihood of drought impact occurrence (LIO) as a function of a drought index using a statistical model. The objective of the effort is to provide insights to sector specific differences in drought risk on a pan-European scale. As mapping is the appropriate method to communicate complex spatial and temporal information (WHO 2014), the study focused on an approach that allows mapping the LIO specific hazard levels.

Data and methods
Spatial domain and resolution Historical drought impacts are commonly reported for local administrative entities. However, all European countries are heterogeneous in their administrative units regarding spatial extent and size. To overcome this challenge, the European Union has created a standardized hierarchical geocode that refers to country-specific administration units, the Nomenclature of Units for Territorial Statistics (NUTS) (European Commission 2005). Non-physical data in Europe is mostly referenced to these NUTS regions at three different levels (NUTS-1 to NUTS-3). However, even these NUTS-regions are spatially not comparable in size. Hence, for this study we composed spatial units ('NUTS-combo level', supplements 1, figure S1) as a combination of NUTS level polygons in different countries that approximately match a reference size of Belgium as an optimal tradeoff between spatial resolution and coverage. Further, to account for the large climatic differences in the region under investigation, we used a classification that divides Europe into climatologically comparable macro regions (figure 2, left) adapted from the European and Mediterranean Plant Protection Organization (Bouma 2005). For our purposes we associated the Eastern Mediterranean countries to Southeastern Europe due the spatial extents of major droughts events in the past.

Drought impact reports
Reports on drought impacts in Europe are numerous and available in various sources and formats. Within the EU FP-7 project Drought R&SPI (eu-drought.org) a unique standardized and categorized collection of drought impact reports was archived in the European Drought Impact Report Inventory (EDII) (Stahl et al 2012). An impact is hereby defined as a negative consequence of drought for environment, society or economy. Reports on drought impacts are spatially referenced either to their respective NUTS region or to locations such as rivers and lakes. Reports with only vague locational information are assigned to the next higher administrative level to ensure correct reference. Due to this generalization impacts may appear to cover a larger area then they actually did. As a temporal orientation, entries have a date stamp in form of either an exact date, a season or the year of occurrence. All impact reports are categorized into one of fifteen impact categories (sectors) and into a more detailed specification of impact types. Full documentations and the data themselves are accessible at http://www. geo.uio.no/edc/droughtdb/.
While the database continues to grow, this study used the content of July 2014, at which time the EDII database contained over 2500 reported drought impacts. As a compromise between data availability and currentness of data, the period of 1970-2012 was selected. For this time frame 1509 reported drought impacts were registered in the EDII database for the NUTS-combo regions that comprise Maritime Europe: 56 were available for Northeast Europe, 464 for Southeast Europe and 347 for Western-Mediterranean Europe. The majority of these reported impacts relate to the following well-known major drought events: 1975-1976West-Central Europe, 1991 Impact reports in four sectors (referenced as 'impact category' within the EDII database) with a sufficient sample size in terms of quantity and spatial and temporal distribution were selected for analysis in this study. These sectors are: (1) 'Agriculture & Livestock Farming', (2) 'Energy & Industry', (3) 'Public Water Supply' and (4) 'Water Quality'. Entries in these categories were first re-assigned from their original reporting level to the 'NUTS-combo level' (figure 1). While Scandinavia, France and Northeastern Europe have only few entries, Central Europe, Southern and Western Iberia, and the Balkan-regions have many entries in these chosen categories. The category 'Agriculture & Livestock Farming' has the highest pan-European coverage. Impacts on 'Energy & Industry' are least represented and mainly available in Central Europe, Southeast Europe and the Iberian Peninsula. Entries on 'Public Water Supply' are stronger represented and additionally cover Britain and Southeast Europe with a maximum of reported occurrence in Southern England and Bulgaria. The distribution of  entries for 'Water Quality' again are mainly from Central Europe, the Iberian Peninsula and the southeastern Balkan.
To develop a suitable regional statistical sample for each of the four European macro-regions: Northeastern, Maritime, Western Mediterranean, and Southwestern Europe (figure 2, left), several possibilities were considered: number of reports per year, classified severity of impacts, or occurrence of one or more impact. The former two were biased to reporting differences across Europe (Stahl et al 2012). Therefore, binary datasets, i.e. 'impact' or 'no impact', were created for 1970-2012, indicating years with drought impact occurrence in a particular category and in the respective NUTS-combo polygons. For temporal comparability, occurrences of multiyear-drought impacts were assigned to each applicable year. Seasonal and short-term information were generalized to the year of occurrence. Finally, NUTS-region records were pooled for each macro-region. Figure 2 shows the time series of impact occurrence for each macro-region and sector. 'Agriculture & Livestock Farming' and 'Public Water Supply' are also the categories with the most impacted years. Northeast Europe has overall only few entries. The Western-Mediterranean region has entries for two distinct events, the '1991-95' and '2004-7 Figure 2 also shows that the report sources are generally diverse in all regions, but differ somewhat in the proportions: whereas Maritime Europe, Northeastern Europe and the Western Mediterranean are dominated by academic and governmental work, more drought impact reports in Southeastern Europe stem from non-governmental reports and the media (newspapers, world wide web).

The SPEI
For this study the SPEI (Vicente-Serrano et al 2010a) was selected as drought hazard indicator. As an evolution of the well-established standardized precipitation index (SPI) (McKee et al 1993) the SPEI was developed to not only account for water input through precipitation, but also for water losses through evaporation. In contrast to SPI, which has a limited interpretability in dry regions (Wu et al 2007), SPEI was specifically developed for the semi-arid environments in Southern Europe (Beguería et al 2010, Vicente-Serrano et al 2010a, 2010b. SPEI has been shown empirically to be a better suited predictor for a number of environmental variables than SPI, including river flow (Lorenzo-Lacruz et al 2010, López-Moreno et al 2013). SPEI can be derived from widely available meteorological observations and has been used for drought quantification in the recent literature (Beguería et al 2010, Lorenzo-Lacruz et al 2010, Wang et al 2014. For this study SPEI was derived from the E-OBS (version 9) data for the period 1970-2012, which provide estimates of daily precipitation and temperature interpolated from station data to a 0.25°grid (Haylock et al 2008). The SPEI was calculated following the recommendations of Stagge et al (2014), using the Hargreaves method to estimate potential evapotranspiration (Hargreaves 1994) and the generalized extreme value distribution for standardization (Stagge et al in revision a).
To adapt to the annual resolution of the impact data, the SPEI was calculated for a time scale of twelve months for the month of December for each year from 1970 to 2012 for all grid cells. Then, the mean SPEI-12 Dec value of all grid cells within each NUTS-combo region was extracted. The resulting annual drought indicator, the regional mean of SPEI-12 Dec will be referred to as 'SPEI' throughout this article. Figure 2 shows the time series of this SPEI for all NUTS-combo regions within each of the macro-regions. The variability of the NUTS-combo SPEI values within each macro region can be high (figure 2).

Modeling the likelihood of drought impact occurrence
The aim of this study is to provide quantitative insights into the relation between observed drought impacts and the SPEI. To this end, we follow the approach of previous assessments (Gudmundsson et al 2014, Stagge et al in revision b), which related drought impact occurrence to drought indicators using binary logistic regression. Logistic regression predicts the likelihood of drought impact occurrence, LIO as where the left hand side of the equation is known as the logit transformation. The model parameters α and β are estimated using standard regression techniques within the framework of generalized linear models (Harrel 2001, Venables and Ripley 2002, Zuur et al 2009. The LIO is hence a measure for the probability of drought impact occurrence, which is dependent on the drought hazard indicator (here SPEI). With this probabilistic model, the occurrence of drought impacts cannot not directly be predicted as 'impact' or 'no impact', but, the likelihood of drought impact occurrence gives estimates in a range from zero (0% probability of impact occurrence) to one (100% probability of impact occurrence).
In this study we sampled pairs of the binary response variable (i.e. the drought impact occurrence series) and the SPEI values of all NUTS-combo regions and pooled them into one sample for each macroregion (omitting NUTS-combo regions without any reported impact in the category). Due to the data sampling strategy, as well as the fact that droughts are by definition rare events the number of impact occurrences compared to the number of no-impact occurrences is generally low. However, in most cases, the distributions of impact and no impact occurrence along the predictor variable SPEI are fairly well separated (figure 3, box plots). The logistic regression models were then fitted for each region and each impact category. Only models for which to SPEI was found to be a significant predictor (p β < 0.01) were retained or further analysis and application to predict and map the LIO for selected SPEI values.
Following a previous study (Gudmundsson et al 2014) model performance was also assessed using the area under the ROCs (receiver operating characteristics) curve, A ROC , which allows to quantify the skill of probabilistic models (Mason andGraham 2002, Wilks 2011). Any A ROC > 0.5 indicates that decisions of the resulting model will be on average superior to random guessing and A ROC = 1.0 indicates a perfect model (see supplement 2, figure S2 for more detailed information).

Results
For Northeast Europe, the lack of sufficient data prevented robust model identification and therefore had to be excluded from further analysis. For all other macro regions and impact categories the SPEI was found to be a significant predictor and models could be fitted with low standard errors for both coefficients (figure 3). All final model confidence intervals show   LIO (0.49) for all models at SPEI = −3. The model for 'Water Quality' for Southeast Europe has a smaller A ROC (0.7). Compared to the other regions, its LIO curve has a negative shift with a low increase starting around SPEI = −1.5.
The maps in figure 4 project selected points of the modeled LIO curves in figure 3 onto maps to facilitate regional comparisons. The spatial visualization shows the resulting LIO, which we interpret as a drought risk for each impact category and for five different drought hazard levels (SPEI = −1, −1.5, −2, −2.5 and −3). While the LIOs for all impact categories and macro regions are low for SPEI = −1, they then increase differently with hazard intensification, although only barely for −1.5. For all impact categories the risk maps show regional differences mainly for the most severe hazard conditions. The application of the models for 'Agriculture & Livestock Farming' and 'Water Quality' in general result in the highest risks for a given hazard level, followed by 'Energy & Industry', and finally 'Public Water Supply'. For 'Agriculture and Livestock Farming', the risk of drought impacts is highest in Western Mediterranean Europe, reaching a likelihood of 46%, whereas Maritime and Southeast Europe have lower LIOs. In contrast to that, the highest risk for impacts on the sector of 'Energy & Industry' is predicted for Maritime Europe (34%); the lowest risk for southeastern Europe (15%). No regional differences in drought risk could be identified for 'Public Water Supply'; all models predict comparably low LIO (∼20%) under most severe hazard conditions. The impact category of 'Water Quality' shows the overall highest and most diverse risk of impact occurrence during drought. For the most severe SPEI condition, for Maritime Europe a LIO of ∼50% was modeled, while for the Western Mediterranean a LIO of 34% and for Southeast Europe an even lower risk was modeled.

Discussion
The database on reported drought impacts tapped for this study is a new information source. So far it covers only a fraction of the NUTS-combo regions and years as some countries have not yet been covered well by the overall search for impact reports. In the derived binary variable of impact occurrence, no impact can have two reasons: 1st there was no impact and 2nd no report was found due to sampling focus or local reporting traditions. In addition, the content of the EDII is also somewhat biased in space and time with an overall increasing trend in the number of reports for more recent events. Such biases will reduce as the database will grow. Despite these uncertainties and limitations, the logistic regression models used to predict the likelihood of impact occurrence as a function of SPEI were found to be significant for all four impact categories for three of the four macro regions. The identified models thus allow a quantitative assessment and visualization of regional differences in drought risk across Europe as desired e.g. by Kossida et al 2012.
'Agriculture & Livestock Farming' are most likely to be impacted by light to moderate drought already, which corresponds to the commonly assumed impact propagation from a meteorological drought to agricultural drought and later to hydrology and economy (Wilhite and Glantz 1985). Our models estimated the risk of impact occurrence in this sector to be highest in the Western Mediterranean region (figure 4). As the Western Mediterranean and Maritime regions are well covered by reported impacts we expect these result to be representative. The presence of drought risk all over Europe shows that drought is not only a Mediterranean problem (Kossida et al 2012) but also that the higher risk for the Mediterranean corresponds to previous assessments (EC 2007a(EC , 2008. For the other two regions the results may be more biased to the few subregions for which reports were available. The chosen drought indicator, the SPEI-12, may also not be ideal for rainfed and irrigated agriculture alike. Stagge et al (in revision b) found shorter lag times to be relevant predictors for example for agricultural impacts in Slovenia. Overall, the result may also be influenced by timelier and more consistent impact reporting for the Agriculture and Water Supply sectors than for other sectors (e.g. Ding et al 2010). Impacts on agricultural production experience high public recognition and tend to be well documented in the media, governmental reports and scientific research. Furthermore, the European Union pays compensation for losses greater than 30%, a serious reason for an early impactreporting by affected farmers. Information sources of the impact reports are mainly based on NGO-reports and media, with potentially higher uncertainty in the latter.
'Energy & Industry' are likely to be impacted only for more severe droughts. The highest risk for impacts on this sector are detected for Maritime Europe which corresponds to the high water abstraction rates for energy production (EEA 2012a) for this macro region. Our models indicate that Southeast Europe has a generally lower likelihood for drought impacts. Rather than indicating a high resilience of this region this result may be influenced by the spatial and temporal distribution of data and droughts are of low relevance for the sector. EDII-data are mainly from the Western Balkan countries, and only start from the year 2000. Hence not all climatic characteristics and past impacts will be captured in the model.
In comparison to the other impact categories, 'Public Water Supply' was modeled to be at lowest risk for similar SPEI values with perhaps the most similar LIO estimates for the three macro regions for which data was available. The chosen twelve month timescale of the standardized drought indicator is likely a good indicator for water supply impacts (Amor et al 2009). As EDII data is distributed quite well over space and time the somewhat lower risk for all macro regions might reflect a generally high awareness and level of water resources planning and management.
Impacts on 'Water Quality' show a very different pattern. The category incorporates impacts on several sectors and details of impacts may hence differ regionally. Rather than a high vulnerability, the high risk for impacts on 'Water Quality' for Maritime Europe may instead reflect a more expanded water quality monitoring and controlling network for several sectors. Government reports were ample in these regions (figure 2). However, it may also reflect the commonly less than good ecological status of water bodies in Maritime-Europe (EEA 2012b). The very low risk for Southeastern Europe may also be caused by the reporting tradition and sources or by the same regional bias as for 'Energy & Industry', or alternatively by a lower relevance and awareness of water quality issues. In this region, the results would benefit from an improved impact database.
Some of the decisions for spatial and temporal sampling also affect the results. As the SPEI is a standardized variable, it normally corresponds to a drought with the same occurrence frequency everywhere. Hence the risk maps are a first step towards the kind of mapping as for example done to display the risk of damage for a 100 year flood. The use of the mean SPEI, the annual time scale and the pooling of the samples into regional models, however, limits the sharpness of the distinction between times with or without impacts and the corresponding drought hazard indicator. As data availability increases, the analyses could be repeated at smaller spatial units and higher temporal resolution. Stagge et al (in revision b) tested such modeling for a few countries in Europe with promising results and Gudmundsson et al 2014 tested a similar method for a drought impact with underlying quantitative data (area burned by wildfire).
Ideally, drought risk mapping should explain differences caused by different vulnerability and exposure to drought. To date most published approaches classify drought risk by a combination of a hazard-and a vulnerability index, where the latter is predominantly obtained as a combination of subjectively weighted vulnerability factors (Iglesias et al 2009, Jordaan 2012, Sreedhar et al 2013. Several approaches validate their results by past impacts (Aggett 2012, Naumann et al 2013, Karavitis et al 2014), but without directly statistically linking them to past impacts. Considering the terminology on Disaster Risk Reduction by UNISDR (2009); a hazard is a climatological phenomenon that causes impacts. But the question whether we can already talk about a drought risk if there is no related impact appears to be unresolved, particularly for the use of pre-categorization and fixed risk thresholds. This study defined drought by its impacts: if a system was impacted by a drought it was vulnerable to this drought. The proposed approach thus promotes a paradigm shift to using past drought impacts as a proxy for impact category specific vulnerabilities without the need of subjective weighting procedures.

Conclusion
In contrast to flood risk (e.g. Van Alphen et al 2009, Alfieri et al 2014), drought risk has not yet been assessed on a pan-European scale, mainly because quantitative information on vulnerability and damages is difficult to obtain. To bridge this gap, the presented study tested the potential of a data-driven approach for modeling the likelihood of drought impact occurrence, derived from the EDII database , as a function of a meteorological drought indicator across Europe. The resulting drought risk maps for impacts in four selected categories illustrate some differences among the European macro regions in the sensitivity to the annual drought hazard levels investigated.
For the most severe droughts in particular, it is shown that drought risk in Europe varies for different impact categories and regions. The highest risk of impact occurrence was detected for 'Water Quality' in Maritime Europe, followed by 'Agriculture & Livestock Farming' in Western Mediterranean Europe and'Energy & Industry' for Maritime Europe. Merely impacts on 'Public Water Supply' result in overall lower risk estimates. These results suggest that drought risk should be analyzed sector specific.
Future work could improve the specificity of such drought risk maps by separating into climatically more homogeneous regions and by testing other drought hazard indicators, such as the ones used by the European Drought Observatory (http://edo.jrc.ec.europa. eu/) or identified in other studies to be relevant for monitoring purposes in different regions. Furthermore, the reported impact data within the EDII database is still biased in a number of ways, including regionally varying sampling density and preference to some particular impact categories. The work proves, however, that there is potential in the use of impact reports towards the description of the risk of a complex natural hazard such as drought, for which vulnerabilities and damages are difficult to quantify otherwise. Drought risk maps may then become an essential component of drought risk management to foster resilience to this hazard at transboundary scales.