Drought indicators can help to detect, assess, and reduce impacts of drought. However, existing indicators often have deficiencies that limit their effectiveness, such as statistical inconsistency, noncomparability, arbitrary metrics, and lack of historic context. Further, indicators selected for drought plans may be only marginally useful, and relatively little prior work has investigated ways to design operationally practical indicators. This study devises a generalizable approach, based on feedback from users, to develop and evaluate indicators for decision-making. This approach employs a percentile-based framework that offers clarity, consistency, and comparability among different indicators, drought levels, time periods, and spatial scales. In addition, it characterizes the evolution of droughts and quantifies their severity, duration, and frequency. User preferences are incorporated into the framework’s parameters, which include percentile thresholds for drought onset and recovery, severity levels, anomalies, and consecutive time periods for triggering. To illustrate the approach and decision-making implications, the framework is applied to California Climate Division 2 and is used with decision-makers, water managers, and other participants in the National Integrated Drought Information System (NIDIS) California Pilot. Stakeholders report that the framework provides an easily understood and beneficial way to assess and communicate drought conditions, validly compare multiple indicators across different locations and time scales, quantify risks relative to historic droughts, and determine indicators that would be valuable for decision-making.
Drought causes substantial impacts throughout the United States (NCDC 2013). It can develop quickly or insidiously and elude characterization until damage has already occurred. A tenet of drought preparedness is that sound indicators can help decision-makers and the public to detect and assess drought conditions and take actions to reduce impacts (NIDIS 2006).
Indicators are variables to define and characterize drought conditions (e.g., Redmond 1991; Steinemann et al. 2005). They are the link between data and decision-making, often a component of drought plans. Indicators can help answer the following stakeholder questions: How do we know it’s a drought? How severe is it? When should we take action? How do we know we’re out of a drought?
Indicators and indicator systems can present a number of limitations for decision-making. For instance, indicators are often based on metrics that are subjective, statistically inconsistent across different temporal and spatial scales, noncomparable with other indicators, and disconnected from relative risk or historic conditions.
In addition, indicators in plans are typically selected in an ad hoc and arbitrary way, without understanding or testing their performance. Decision-makers often lament the lack of guidance for developing and evaluating indicators. For example, a recent survey of 19 drought managers from western states found that 16 of the 18 states with drought plans did not use their indicators, citing them as “not at all useful,” “ignored,” “a guess,” “borrowed from another state,” and “never tested” (Steinemann 2014).
Despite the importance of indicators for drought preparedness and response, relatively little prior work has investigated ways to develop indicators and to evaluate their usefulness for decision-making. As Redmond (2002) notes, indicators often escape into widespread usage without being evaluated, despite underlying scientific and operational problems. Redmond (1991) lists desirable properties of indicators, such as detailed understanding of caveats and availability of data. Keyantash and Dracup (2002), based on Redmond (1991), offer six criteria (robustness, tractability, transparency, sophistication, extendability, and dimensionality) and assigned weights and values to indicators in a study of Oregon. Steinemann et al. (2005), based on a review of drought plans and interviews with water officials, provide a set of criteria for indicators and triggers, including suitability for drought type, data availability and consistency, clarity and validity, temporal and spatial sensitivity and specificity, statistical consistency, and links with drought management goals.
In a quantitative approach, the stochastic characteristics of indicators, such as their frequency, duration, and transition probabilities between severity levels, can be examined to evaluate performance (e.g., Steinemann 2003; Paulo and Pereira 2008). In a link with decision-making, drought indicators can be evaluated as to when they would have provided early warning, or comported with drought decisions in actual drought situations or during virtual drought exercises (e.g., Steinemann and Cavalcanti 2006; Shukla et al. 2011). Other studies (e.g., Mo and Lettenmaier 2014; Hao and AghaKouchak 2014; Anderson et al. 2013) have evaluated drought indicators according to how well they correlate with the U.S. Drought Monitor (Svoboda et al. 2002).
This article presents a generalizable approach for developing, selecting, and evaluating drought indicators, using an adaptable framework to address decision-making needs. The approach is designed to be useful in characterizing the evolution of droughts and in quantifying their severity, duration, frequency, and spatial extent. As an illustration, the framework is applied to the analysis of drought conditions in a climatic division in California, and feedback is obtained from water managers, drought decision-makers, and other stakeholders throughout the state.
2. Background and motivation
Drought indicators are typically based on meteorological and hydrologic variables, such as precipitation, streamflow, soil moisture, reservoir storage, and groundwater levels. An indicator can combine several variables into a single index value, such as the Palmer drought severity index (PDSI).
Drought levels are categories of severity of drought, with nomenclature such as “mild, moderate, severe, or extreme drought” or “level 1, level 2, or level 3 drought.” Drought triggers are specific indicator values or thresholds that define a drought level (such as “severe drought is between the 6th and 10th percentile”), when it begins and ends, and can be associated with the timing of drought responses.
Indicators and their drought levels not comparable with other drought indicators. Drought plans typically list multiple indicators, with values for each indicator corresponding to different levels of drought. However, these levels may not be statistically consistent among the indicators. While at any given time different indicators could understandably reflect different severity levels of drought, the inconsistency is the definition of a level in terms of probability of occurrence. The broader point is that if the likelihood, rarity, or extremity of conditions should be a factor in defining drought, then the severity level should relate to a statistically consistent percentile.
Indicators not statistically consistent across different temporal and spatial scales. Even for an individual indicator, a specific value or drought level may have a different probability of occurrence, depending on time and location. For example, the PDSI value of −4.0, corresponding to “extreme drought,” occurs less than 1% of the time in southeastern California in July and more than 10% of the time in central California in July (Guttman et al. 1992). Even statistically based indicators such as “percent of normal” can have a different probability of occurrence, depending on location and time.
Indicator values and levels not scientifically or operationally meaningful. Indicators are often based on scales that are arbitrarily defined and that may be difficult to interpret directly. To quote a drought manager, “What does a −1.5 Palmer index really mean?” Moreover, the decision of how rare conditions need to be in order to define drought is relegated to severity levels predefined by these scales. For instance, the definition of a so-called severe drought would occur 6.7% of the time for the standardized precipitation index (SPI), 14% on average for the surface water supply index (SWSI), and 10% on average for the PDSI (Steinemann 2003).
Indicators not adequately specified for drought onset and drought recovery. Drought onset and recovery triggers need to be associated with an indicator being below/above a certain threshold (e.g., a percentile or drought level) for a certain period of time (e.g., consecutive months) in order to invoke/revoke that level. However, drought plans often neglect to specify the threshold, the time period for the trigger, or both. Further, drought plans may mention onset but not recovery, or assume that onset triggers can be reversed to function as recovery triggers, even though goals and criteria for going into a drought versus getting out of a drought may differ.
To better understand the types of indicators that would be useful to decision-makers, a series of workshops were conducted with over 75 stakeholders (including water managers, water users, drought officials, and other decision-makers) in the National Integrated Drought Information System (NIDIS) California Pilot activities during the period of 2012–14 (NIDIS 2006, 2014). Stakeholders identified a set of attributes deemed desirable for indicators:1 1) statistically consistent and comparable terms (“all in terms of percentiles; that would be really useful”), 2) relative to historic drought conditions (“so we can get a sense for how severe this is, relative to the past”), 3) individual and separate (“rather than preaggregated into an index”) yet possible to combine (“with user-defined weights”), 4) easy to understand and implement (“related to familiar concepts, such as return period”), 5) able to represent a range of conditions (“from dry to wet”), and 6) relevant across different time scales and spatial scales (“means the same thing [in terms of likelihood], regardless of the time or place”; “possible to scale up and scale down”).
Motivated by the limitations identified in drought plans and desirable attributes of indicators as identified by stakeholders, our approach adopts a convention of placing indicators in terms of percentiles. A percentile-based approach not only meets the set of attributes listed above, but can also be used as the basis for determining objectively defined drought level thresholds, according to decision-maker needs (see, e.g., Quiring 2009; Steinemann and Cavalcanti 2006; Steinemann 2003; Carrão et al. 2014; Mizzell et al. 2010).
We develop a framework to define and assess characteristics of different indicators and systems of defining drought and an approach for analyzing the indicators. This framework and approach are generalizable, focusing on the quantitative performance of indicators and decision-making implications. Because drought is widely acknowledged as complex, and drought situations vary, this approach is not intended to opine which specific indicators are “best,” but rather, to provide the analytic foundations for users to develop, evaluate, and choose indicators that would be most suitable for their purposes.
This article provides three main contributions: 1) an approach to characterize, analyze, and compare, in statistically comparable terms, any number of indicators, temporal scales, and spatial scales; 2) a way to assess drought frequency, duration, and severity and to compare current conditions with historic droughts; and 3) a framework that allows users to specify values for key parameters determining drought, such as percentile thresholds for drought onset and recovery for one or more drought levels, trigger specifications of consecutive time periods, anomaly time scales, as well as to customize indicators for drought determinations. The percentile-based approach developed here offers a consistent quantitative treatment of indicators that can be used together with stakeholder input to help them determine which indicators might provide the most value for drought preparedness and response.
First, we develop and detail characteristics of drought indicators and triggers for the framework. The framework is adaptable and transferable and can be applied to any location and at any temporal and spatial scale. Second, we analyze a range of values for each characteristic and assess the results, using an application for California. We select values that are typical and based on stakeholder input, recognizing that other values could also be explored. Third, we apply the framework to the reconstruction and analysis of historic conditions in a region of California. Fourth, we use these analyses together with stakeholders in California NIDIS activities to understand decision-making implications and obtain user feedback on the framework and evaluation approach.
a. Framework for drought indicators and triggers
We detail the indicator and trigger characteristics that define the general framework. For each characteristic, we also provide examples of values that will be explored for the application to California.
Variable for indicator, such as precipitation, streamflow, groundwater levels, reservoir levels, etc. We selected precipitation for this application because of its relevance to many stakeholders and the availability of a long-term dataset. The framework permits the comparison and inclusion of any number of indicators, so additional indicators could be analyzed along with or instead of the precipitation indicators.
Metric for indicator, such as percentiles, percent of normal, raw values, etc. We selected percentiles, for the reasons detailed earlier. Indicators can be transformed to percentiles by fitting a distribution to the data, developing an empirical cumulative distribution function, or using other probability estimators. Trigger values are associated with percentiles, which can also define drought levels. Other indicators, if included, would also be analyzed in terms of percentiles.
Spatial scale for indicator, such as climate division, hydrologic basin, etc. For our example of California, we selected Climate Division 2 (CD2), the Sacramento drainage basin,2 because of its importance to a range of stakeholders in the state. The spatial scale for the indicator may differ from the spatial scale for the decision that is triggered by the indicator. For instance, a climate division indicator in one part of the state may be used to trigger a county-level response in another part of the state.
Temporal scale for indicator, such as month, week, etc. This is the basic time period for analyzing the indicator, even though the trigger may involve multiple consecutive time periods. For precipitation, we selected one month as the unit of data aggregation, at different anomaly time scales. While precipitation data are available at shorter time scales, the choice of 1 month as the basic unit represents period of interest to a range of stakeholders in California.
Drought severity levels. Percentile thresholds can be used to designate levels of drought (and range from dry to wet) conditions. In this case, we selected thresholds based on commonly used levels in drought plans: drought level 1 = 20%–35%; drought level 2 = 10%–20%; drought level 3 = 5%–10%; drought level 4 = 2%–5%; and drought level 5 <2%.
Percentile threshold for drought onset PONSET. This is the percentile level below which drought onset (or a certain level of drought) will be defined, given the indicator is at/below that threshold for the designated prior time period. We examined PONSET for the drought levels of <35%, <20%, and <10%.
Consecutive time periods for drought onset NMONSET. This is the prior time period that the indicator needs to be at/below the percentile level of PONSET for drought onset to be invoked. We looked at NMONSET (NM = number of months) ranging from 1 to 6 consecutive months.
Percentile threshold for drought recovery PRECOV. This is the percentile level above which the drought recovery (for a certain level) is defined, given the indicator is at/above that threshold for the designated prior time period. We examined PRECOV for the drought levels of >20% and >35%.
Consecutive time periods for drought recovery NMRECOV. This is the prior time period that the indicator needs to be at/above the percentile level of PRECOV for drought recovery to be invoked. We looked at NMONSET ranging from 1 to 6 consecutive months.
Anomaly time scale for indicator, such as 1, 3, 6, 9, 12, 18, or 24 months. This represents the deficit (or surplus) across different time scales, which can reflect different aspects of drought and its effects. We note the time scale for an indicator may differ from the time scale for a trigger. For example, for a drought onset defined by a 3-month precipitation anomaly below 25% for two consecutive months, the time scale for the indicator is the three prior months, and the time scale for the trigger is two consecutive months. We also note that the 3-month anomaly for two consecutive months differs from a 6-month anomaly for one month.
b. Data and analysis
State climate division monthly precipitations were obtained from the NOAA/NCDC website (ftp.ncdc.noaa.gov/pub/data/cirs). This dataset is based on daily observations of the Global Historical Climatology Network (GHCN) and extends from 1895 to present (dataset updated monthly). Divisional values were derived from area-weighted averages of gridpoint estimates interpolated from available station data. Station data were gridded using climatologically aided interpolation to minimize biases arising from topographic and network variability. A more complete description of the methods used to construct this dataset is contained in Vose et al. (2014).
Precipitation over several time scales ranging from 1 to 24 months were first calculated using the monthly NCDC climatic division data. Next, cumulative probabilities (i.e., percentiles) were determined using a method similar to that used by McKee et al. (1993), from a derived gamma distribution described below, to calculate SPI values. The only difference from McKee et al. (1993) is that, in the present study, the resulting statistics are left as cumulative probabilities rather than converted to standard z scores. The source code for the calculations was obtained from the Colorado Climate Center (http://ccc.atmos.colostate.edu/standardizedprecipitation.php).
Precipitation amounts are first fitted to a gamma distribution using the base time period of interest; in this case, we used the period 1900–2014. A separate distribution is used for each calendar month or ending month in the case of time scale anomalies greater than 1 month. The parameters describing the gamma distribution are then used to calculate the cumulative probability of each precipitation value using a modified incomplete gamma function to allow cases with zero precipitation, H(x) = q + (1 − q)G(x), where G(x) is the incomplete gamma function and q is the probability of zero precipitation.
4. Results and discussion
We reconstruct historic conditions in California CD2, according to precipitation indicators based on percentiles, for a range of time scale anomalies from January 1900 to September 2014.
We then examine a range of values for the characteristics of percentile threshold for onset and recovery, number of consecutive months for onset and recovery, and time scale of anomaly. With these variations, we assess the implications for drought duration (number of months in drought, onset to recovery), total number of drought months, frequency (number of droughts during a time period), severity (percentile value of an indicator), and the timing of months in drought. This approach for determination of drought can apply to either a drought event or a specific drought level within that event.
A drought (or drought level) is defined to start in month M1 when the precipitation probability is less than PONSET for NMONSET consecutive months (ending in month M1). A drought (or drought level) is defined to end in month M2 when the precipitation probability is greater than PRECOV for NMRECOV consecutive months (ending in month M2). The number of droughts and the average drought duration, from January 1900 to September 2014, for combinations of PONSET, PRECOV, NMONSET, and NMRECOV for California CD2 are calculated and presented in Tables 1–8.
As the anomaly time scale increases, the number of droughts during the period of record generally decreases, while the average drought duration increases, for each set of PONSET and PRECOV (see Tables 1–3). As the number of consecutive months (i.e., NM) increases, the number of droughts during the period of record decreases, while the average drought duration increases, for each set of PONSET and PRECOV. As PONSET increases (from more to less severe), the number of droughts increases for all values of NM and anomalies. But, average duration decreases for longer anomalies and shorter NM and increases for shorter anomalies and longer NM. As PRECOV increases (from more to less severe), the number of droughts generally decreases, while average duration increases for all values of NM and anomalies.
With PONSET and PRECOV constant, the average drought duration increases as the time scale anomaly increases from 1 to 12 months for all values of NMONSET and NMRECOV (Tables 4, 5). But the frequency of droughts decreases for NMONSET = 1 or 2 months and increases for NMONSET = 3, 4, 5, or 6 months, as the anomaly increases from 1 to 12 months. A reason is that shorter-term anomalies tend to be more oscillatory, making it difficult to stay below a threshold for multiple consecutive months.
When the number of consecutive months is increased and/or the anomaly time scale is decreased, for PONSET = PRECOV = 20%, the total number of drought months during the period decreases, in general. But for NM = 1 month, the number of drought months does not change appreciably for the range of anomaly scales examined here (see Table 6).
For lower values (more severe) of PONSET, the number of droughts and the average duration does not vary appreciably with PRECOV. For higher values (less severe) of PONSET, the number of droughts increases and average duration decreases as PRECOV decreases (more severe). For lower values of PRECOV, the number of droughts increases and average duration decreases as PONSET increases. For higher values of PRECOV, the number of droughts increases as PONSET increases; however, the average drought duration remains relatively constant across all values of PONSET examined (see Tables 7, 8).
To illustrate this approach and results, we depict how this framework would have classified monthly drought severity according to precipitation percentiles, across a range of anomaly time scales, in California CD2 from January 1900 to September 2014 (Fig. 1). For any given month, the specific percentiles (drought severity levels) across different time scale anomalies can also be extracted (see Table 9).
We explore specific values of PONSET = PRECOV = 20% and with NMONSET = NMRECOV = 3 months to assess which months would have been classified as drought (noted in asterisks in Fig. 2). These values were selected to exemplify the analysis, based on the results above, as they would capture the main drought events during this time period examined (California Department of Water Resources 2000, 2012).3
5. Stakeholder feedback
This approach for developing and evaluating indicators was used in NIDIS drought preparedness workshops, presentations, and interactions with over 75 stakeholders throughout the state of California during the period of 2012–14. Reports from stakeholders include feedback such as “This is really useful—it shows drought in ways that make sense”; “Now I see what the Governor means by driest year in recorded state history”; “It gives decision-makers what they need”; and “We’d like to use [this approach for indicators] throughout the state.”
In particular, decision-makers reported that this approach enables them to 1) assess precipitation and other indicator deficits at different time scales, 2) place current conditions in historical context, 3) validly compare multiple indicators and view them all in one place, 4) determine which indicators would have been most useful for decision-making, 5) choose indicators that are relevant and practical, 6) quantify and compare risks, and 7) communicate drought conditions in ways that are meaningful and understandable to the public.
The analyses reported above, together with feedback from interactions with stakeholders, produced the following observations and conclusions.
Decision-makers often have different criteria for the indicators and desirable performance characteristics. For instance, a 1- or 3-month precipitation anomaly is important to a dryland farmer, whereas an 18- or 24-month precipitation anomaly is important to a water manager with a multiyear storage reservoir system. Also, one decision-maker would want to implement water use restrictions as soon as drought conditions start developing (e.g., 3- or 6-month anomalies, lower values of NMONSET and less severe values of PONSET), but another decision-maker would want to wait until drought conditions have definitively developed before implementing restrictions (e.g., 9- or 12-month anomalies, higher values of NMONSET and more severe values of PONSET).
Decision-makers also noted that a shorter-term anomaly tends to be more responsive to changes (and more likely to provide an element of early warning but also false alarm of onset and false assurance of recovery) and a longer-term anomaly tends to be more stable and persistent (and more likely to show definitive onset and recovery but also slower to respond). While the criterion of needing to be above/below a threshold for a number of consecutive months can help to make the shorter anomalies more stable and definitive for decision-making, it may be difficult to meet that criterion because of “jumping around between levels.”
The determination of PONSET and PRECOV often involves an important policy decision: How extreme or rare do conditions need to be in order to declare a drought level or event? And how much do conditions need to recover in order to rescind the declaration? Stakeholders noted that this approach provides an explicit way to specify the desired percentile thresholds associated with PONSET and PRECOV. This overcomes a common problem with other indicator systems where this decision is implicitly made by drought indices or indicators that have predetermined values for the drought levels and their corresponding probabilities.
In light of the main statewide drought events during the last century, the analysis in Fig. 2 shows how the multiyear events encompassed the range of time scale anomalies and how indicators at short- and midrange anomalies time scales have earlier onset and recovery, higher frequency, and shorter duration than the longer-term anomalies.
This leads to a point made by stakeholders, that is, a short-term recovery does not necessarily mean the drought is over, and the longer-term deficits can persist. An example is the drought of 1987–92, where March 1991 was exceptionally wet and temporarily reduced critical water demands, but not enough to end the 5 years to date of drought. Stakeholders also appreciated that extracting percentiles (Table 9) enabled them to compare specific severity levels among droughts. For instance, while 1976–77 is remembered as a major drought event, they saw that the 2014 drought has been the most severe to date (at precipitation anomaly time scales below 24 months). These examples point to the importance of multiple indicators at different time scales to characterize complexities of drought, as this approach is designed to facilitate.
Criteria for drought onset and recovery can also include multiple conditions on these multiple indicators, as well as different combination algorithms. For example, as suggested by one decision-maker, given a set of four indicators (3-, 6-, 9-, 12-month precipitation anomalies), drought onset can require that two of four meet the conditions and drought recovery require that three of four meet the conditions, and at higher values of NMRECOV and PRECOV to “be more conservative.”
Also, a drought index can be developed from multiple individual indicators with user-defined weights, with a new probability distribution (and percentiles) determined for the data stream resulting for the index. As stakeholders noted, a problem with using indices (such as the PDSI, SWSI, or Drought Monitor) is that they embed subjectivity in ways not apparent to or changeable by decision-makers, such as the choice, weighting, and time scale of individual indicators to create the overall index value.
While percentiles meet a range of user-defined needs for indicators and provide statistical comparability, they are not intended as a panacea. For instance, a series of summer months with low-percentile precipitation may not signal drought, if summer normals are close to zero anyway and contribute a negligible amount to annual totals. Thus, other types of indicator metrics, for example, conveying information on amount of deficit, can be important for decision-makers.
In future activities and extensions to this work, regionally specific indicators are being developed for NIDIS pilot activities in the state of California, along with a web-based indicator system. This system is intended for use in stakeholder workshops, involving activities such as virtual drought exercises and retrospective drought analyses, to identify the specific indicators that would be most useful for drought assessment and decision-making.
This article reported on an approach to define, assess, and communicate drought conditions in ways that are practical to decision-makers and the public and that address common deficiencies with drought indicators. It offers an adaptable framework to determine drought according to different indicator variables, spatial scales, temporal scales, severity levels, anomalies, percentile thresholds for drought onset and recovery, and consecutive time periods for triggering. The characteristics of this framework for drought indicators have been designed, based on broad stakeholder input, to provide the following benefits: 1) a useful and understandable way to view drought; 2) a statistically consistent and equitable basis to define and assess indicators, triggers, and drought levels; 3) a valid comparison across any number of indicators, time scales, and locations; 4) clear and intuitive indicators and triggers that relate to other operational concepts, such as return period, and that place current conditions in historical context; 5) a degree of flexibility to incorporate decision-maker preferences in drought determinations; and 6) a quantitative basis to assess and communicate drought conditions and risks in ways that are meaningful to the public. Overall, this framework offers an approach to develop and select drought indicators that can be practical and valuable for decision-making.
This study received support from the California–Nevada Applications Program (CNAP), a Regional Integrated Sciences and Assessment (RISA) program, via National Oceanic and Atmospheric Administration Grants NA11OAR4310150 and NA13OAR4310172, and from the National Integrated Drought Information System (NIDIS). We thank Shraddhanand Shukla, Amy Davis, and three reviewers for their helpful comments on this manuscript.
Current affiliation: Melbourne School of Engineering, The University of Melbourne, Melbourne, Victoria, Australia.
Direct statements from individual stakeholders are provided in quotation marks.
State drought events, based on water years (from 1 October to 30 September), are reported by the California Department of Water Resources as the following: 1912–13, 1918–20, 1923–26, 1928–35, 1947–50, 1959–62, 1976–77, 1987–92, 2000–02, 2007–09, and from 2014 to the present.