A revision of the Combined Drought Indicator (CDI) as part of the European Drought Observatory (EDO)

Abstract. Building on almost ten years of expertise and operational application of the Combined Drought Indicator (CDI), which is operationally implemented within the European Commission’s European Drought Observatory (EDO) for the purposes of early warning and monitoring of agricultural droughts in Europe, this paper proposes a revised version of the index. The CDI conceptualizes drought as a cascade process, where a precipitation shortage (WATCH stage) develops into a soil water deficit (WARNING stage), which in turn leads to stress for vegetation (ALERT stage). The main goal of the revised CDI proposed here, is to improve the indicator’s performance for those events that are currently not reliably represented, without drastically altering the modelling framework. This is achieved by means of two main modifications: (a) use of the previously occurring CDI value to improve the temporal consistency of the timeseries, (b) introduction of two temporary classes – namely, soil moisture and vegetation greenness – to avoid brief discontinuities in a stage. The efficacy of the modifications is tested by comparing the performances of the revised and currently implemented versions of the indicator, for actual drought events in Europe during the last 20 years. The revised CDI reliably reproduces the evolution of major droughts, out-performing the current version of the indicator, especially for long-lasting events. Since the revised CDI does not need supplementary input datasets, it is suitable for operational implementation within the EDO drought monitoring system.



Soil Moisture 123
The soil moisture anomaly index (zSM) is computed using the modelled soil moisture output of the evolution of a drought event is conceptualized by a "cause-effect" relationship, assuming that a 156 shortage in precipitation leads to a soil moisture deficit, culminating in reduced vegetation 157 productivity. In its original form, data for the variables zSPI, zSM and zFAPAR (see above) are used 158 to characterize three stages of an idealized agricultural drought: 159 • "WATCH", in which the precipitation is below normal (zSPI = 1), and an early warning signal 160 of a potential drought affecting agriculture can be observed; 161 • "WARNING", when a precipitation deficit propagates in the hydrological cycle and affects 162 soil water content (zSPI = 1 & zSM < -1). 163 • "ALERT", when the effects of drought become visible as vegetation stress (zSPI = 1 & During the operational implementation of the indicator, two additional recovery stages were 166 introduced (see https://edo.jrc.ec.europa.eu/documents/factsheets/), aimed at better capturing 167 the "fade-out" phase of a drought, namely the "PARTIAL RECOVERY" and "FULL RECOVERY" 168 stages. In both stages, the previous month's zSPI (zSPI m-1 ) is introduced to account for the 169 preceding conditions: 170 • "PARTIAL RECOVERY": zSPI returns to normal values even if vegetation is still negatively 171 affected (zSPI m-1 = 1 & zSPI = 0 & zFAPAR < -1). 172 • "FULL RECOVERY": Both precipitation and FAPAR return to normal conditions (zSPI m-1 = 1 & 173 This operational implementation of the index is the one commonly referred to in the 175 scientific and technical drought literature when CDI is described. 176 The CDI modelling framework described above is summarised in Fig. 1, where the different 177 stages of CDI (from WATCH to FULL RECOVERY) are depicted according to the eight cases that can 178 be obtained by combining the two possible binary states for each of the three main variables (zSPI, 179 zSM, zFAPAR), as well as a function of zSPI m-1 . 180 Due to its operational status, the maps of the CDI that are currently available in EDO are 181 always processed using data available up to the release date of a new map. For this reason, some 182 inconsistencies in the reference baseline and actual data (e.g. FAPAR data source) are present in 183 this operational dataset. For the present study, a self-consistent dataset has been produced by re-184 computing the CDI with the best data available at the end of 2018. This dataset (referred to here 185 as CDI-v1) is consists of 648 dekadal maps at 5-km spatial resolution, from January 2001 to 186 December 2018. In order to compute the CDI at this spatial resolution, the original data for zSPI 187 and zFAPAR were initially resampled over the zSM grid, using the nearest neighbour and spatial 188 average procedure, respectively.

The revised version of CDI, as proposed here (CDI-v2) 190
In order to better understand the modifications to the CDI that are proposed here, two case 191 studies where CDI-v1 was not able to capture in full the evolution of the drought, are first 192 reported.
in the context of a nested approach, since each successive stage is contained within the definition 198 of the previous one. This is exemplified by the inclusive nature of the calculation (see above, 199 where "&" is used in the definition of the classes). This approach can lead to abrupt breaks in 200 tracking a drought event, when a substantial temporal shift among the three quantities can be 201 observed. 202 For example, the plots in Fig. 2 report the timeseries of SPI-3 (upper panel), zSM (middle 203 panel) and zFAPAR (lower panel) for a year that includes a drought event in Spain. Dotted vertical 204 lines demarcate the full span of the drought event. At the top of each plot, a box demarcates the 205 period when the stage-specific conditions for WATCH, WARNING and ALERT are met. By an a 206 posteriori analysis of the event, it is easy to assess a desirable sequence of stages for each dekad, 207 as reported in the bottom part of the lower plot (i.e. the ideal outcome of a revised CDI, CDI-v2 208 ideally). However, from the actual sequence of CDI values (CDI-v1) it can be seen that the event is 209 interrupted in the middle of the soil moisture deficit period due to the return of precipitation to 210 normal conditions. 211 A second example is shown in Fig. 3 for a drought event in France, where the timeseries of 212 SPI-3, zSM and zFAPAR suggests an extensive period of soil moisture deficit following a 213 https://doi.org/10.5194/nhess-2020-204 Preprint. Discussion started: 7 August 2020 c Author(s) 2020. CC BY 4.0 License. precipitation deficit, that caused a short period of FAPAR anomalies. Even if two periods meeting 214 the requirement for a WARNING and an ALERT status are observed (see boxes at the top of the 215 middle and lower panels, respectively), a temporary return above the thresholds is observed (for 216 one or two dekads) in both zSM and zFAPAR timeseries. In an a posteriori analysis, a single 217 continuous ALERT period would have been likely detected (see ideal CDI sequence at the bottom 218 of the Figure). CDI-v1 instead treats those gaps as interruptions, causing a "back-and-forth" 219 transition between the ALERT and WARNING stages. 220 This behaviour is in contrast to the "cause-effect" principle on which the indicator is based, 221 and even if this occurrence cannot be always avoided in real case studies, it should be kept to a 222 minimum. It is worth noting how, also in this second case, according to CDI-v1 the event stops well 223 before the end of the soil moisture deficit, due to the return of precipitation to normal conditions 224 (SPI-3 > -1). 225 The two examples reported above highlight the main drawbacks of the current operational 226 version of the CDI, which can be summarized as follow: 227

•
Lack of a proper cascade process in favour of a nested approach, which can cause an early 228 interruption in drought events in case of notable shifts between timeseries; 229 • absence of check on the possible small gaps within a stage, which can lead to 230 inconsistencies in the temporal sequence and quick alternation of different stages. 231 The revised version of the CDI that is proposed here (i.e. CDI-v2 from hereafter) addresses 232 these two key issues by introducing two principal modifications: 233 • Set-up different rules to ensure temporal continuity based on the previous dekad's CDI 234 (CDI d-1 ) rather than the preceding SPI (SPI m-1 ); 235 • adding a second set of threshold values to detect both temporary gaps within a stage, and 236 https://doi.org/10.5194/nhess-2020-204 Preprint. Discussion started: 7 August 2020 c Author(s) 2020. CC BY 4.0 License.
the "fade-out" phase of a drought. 237 These modifications are implemented according to the scheme depicted in Fig. 4, where the 238 upper part of the Table is Fig. 6 were extrapolated from the detailed reports of EDO for the most recent drought 287

events. 288
In all the cases studied, the start of the drought event coincides for the two versions of the 289 indicator (CDI-v1 and CDI-v2), as is to be expected given the analogous conditions adopted to 290 define a new event. Over some sites, the two versions do not differ substantially, as in the case of 291 Wattisham and Magdeburg (Fig. 5), and Silkeborg and Poznan (Fig. 6), where only minor signs of 292 the issues highlighted in Figs. 2 and 3 can be observed. In those study sites, the temporal evolution 293 of the droughts appears to be well reproduced by both versions of the indicator, with the start-, 294 peak-and end-dates consistent with the scientific literature for the events (Buras et al., 2020;Ciais 295 et al., 2005;Hanel et al., 2018;Rebetez et al., 2006). are rather similar to that depicted in Fig. 2, with a long period of soil water deficit and plant water 301 stress during the whole dry season following a rainfall deficit early in spring and a hot and dry 302 summers. In these cases, the new version of the index seems capable to capture those instances 303 when a drought is prolonged by higher than normal evaporative demand even after the rainfall 304 returns to normal. Considering the well documented severity of those droughts (Garcia-Herrera et 305 al., 2007;MeteoAM, 2007;Spinoni et al., 2015), the behaviour of CDI-v2 seems much more in line 306 with the expected evolution of the droughts. 307 Finally, for some study cases -specifically Deols (2011 drought), Strasbourg (2015 drought) retroactive application of the revised indicator to past drought events, without the need for 381 additional inputs or changes in the underlying datasets. For similar reasons, the three main stages 382 of drought (i.e. "WATCH", "WARNING" and "ALERT"), which were originally defined in Sepulcre-383 On a general level, it is apparent that both the point-scale timeseries and the spatial maps 405 obtained with the new version of the indicator, better approximate the expected spatiotemporal 406 characteristics of a drought event, with a more realistic succession of the "WATCH", "WARNING" 407 and "ALERT stages", and a large spatial consistency in the modelled patterns. In addition, in spite 408 of the improved performance of the revised version of the CDI, the "look and feel" of the indicator 409 are not substantially altered. Given the well established and wide community of users of the 410 current version of the CDI that is implemented in EDO, this is a key consideration that can ensure a 411 smooth transition to the operational use within EDO, of the revised version of the CDI that is 412 proposed here.