A standardized precipitation index (SPI) that uses high-resolution, daily estimates of precipitation from the National Weather Service over the contiguous United States has been developed and is referred to as HRD SPI. There are two different historical distributions computed in the HRD SPI dataset, each with a different combination of normals period (1971–2000 or 1981–2010) and clustering solution of gauge stations. For each historical distribution, the SPI is computed using the NCEP Stage IV and Advanced Hydrologic Prediction Service (AHPS) gridded precipitation datasets for a total of four different HRD SPI products. HRD SPIs are found to correlate strongly with independently produced SPIs over the 10-yr period from 2005 to 2015. The drought-monitoring utility of the HRD SPIs is assessed with case studies of drought in the central and southern United States during 2012 and over the Carolinas during 2007–08. A monthly comparison between HRD SPIs and independently produced SPIs reveals generally strong agreement during both events but weak agreement in areas where radar coverage is poor. For both study regions, HRD SPI is compared with the U.S. Drought Monitor (USDM) to assess the best combination of precipitation input, normals period, and station clustering solution. SPI generated with AHPS precipitation and the 1981–2010 PRISM normals and associated cluster solution is found to best capture the spatial extent and severity of drought conditions indicated by the USDM. This SPI is also able to resolve local variations in drought conditions that are not shown by either the USDM or comparison SPI datasets.
The standardized precipitation index (SPI) represents the amount of precipitation that falls over a given interval of time as standard deviations above or below the historical median, which is centered at zero, for the same interval (Guttman 1999; Edwards and McKee 1997; McKee et al. 1993). The index was proposed by McKee et al. (1993) to improve monitoring capabilities of drought time scale, historical probability, and impacts on water supply. SPI’s multiscalar nature gives it the advantage of being able to recognize simultaneously occurring droughts and wet spells at different time scales. In addition, because it is normalized to a specific location’s historical record, SPI values at different locations, such as in wet and dry climates, can be compared without modification (Edwards and McKee 1997; McKee et al. 1993).
The SPI has been widely used in drought studies within the United States (e.g., Logan et al. 2010; Kangas and Brown 2007; Wu et al. 2007; Hayes et al. 1999; McKee et al. 1995). It has also been applied worldwide to study the frequency and magnitude of specific droughts as well as to understand the climatological characteristics of droughts (e.g., Angelidis et al. 2012; Batisani 2011; Blain 2011; Karavitis et al. 2011; Bonsal and Regier 2007; Patel et al. 2007; Lloyd-Hughes and Saunders 2002; Komuscu 1999).
McEvoy et al. (2012) examined the ability of the SPI and the standardized precipitation evapotranspiration index (SPEI) to monitor hydrological drought in the arid regions of the Great Basin in Nevada and eastern California, the water supply of which primarily comes from runoff of spring snowmelt. Although the SPEI performed slightly better than the SPI, both indices were found to correlate highly with surface water availability. Keyantash and Dracup (2002) evaluated 1-month SPI and several other commonly used drought indices for measuring meteorological drought in two distinctly different precipitation regimes. A set of robust criteria was used to evaluate the indices, and SPI was found to perform well relative to its peers and was recommended for meteorological drought monitoring. Hayes et al. (1999) used the SPI to monitor a severe, 9-month drought in 1996 that affected the Great Plains and Southwest. The SPI was shown to indicate drought onset at least 1 month in advance of the Palmer drought severity index. In addition, the multiscalar nature of the SPI allowed the progression of the drought to be monitored for its duration.
Drought monitoring increasingly takes place on smaller scales, intensifying the need for localized information (e.g., Svoboda et al. 2002). On-the-ground observations of drought-related parameters, such as rainfall, soil moisture, and crop status, are limited by the density of monitoring stations and the ability of people to inspect and track locations that may be experiencing impacts. To help meet the growing need for localized drought-monitoring information, McRoberts and Nielsen-Gammon (2012) developed a computational procedure for calculating SPI that incorporates high-resolution, radar-based estimates of precipitation and evaluated it over Texas; R. V. Cumbie-Ward et al. (2016, unpublished manuscript; see also http://climate.ncsu.edu/spi/AboutThisPage.html) describe the methods to expand a radar-based, high-resolution, and daily updated SPI (HRD SPI) to the contiguous United States.
The abovementioned method to generate HRD SPI is described briefly here. Historical distribution parameters of the Pearson type-III distribution are determined using accumulated precipitation data for a variety of time scales from stations in the National Weather Service (NWS) Cooperative Observer Program (COOP). These stations are grouped into homogeneous regions by following the method of Hosking and Wallis (1997) to provide more accurate estimates of the historical frequency distribution. The scale and shape parameters of the Pearson type-III distribution determined for each station are normalized by the location parameter and then interpolated to the Hydrologic Rainfall Analysis Project (HRAP) grid. PRISM normals are also reprojected and aggregated to the HRAP grid and are used as the location parameter of the historical distribution at each grid point. Gridded precipitation analyses from the Advanced Hydrologic Prediction Service (AHPS) and National Centers for Environmental Prediction (NCEP) Stage IV precipitation products are accumulated over time scales of interest and combined with PRISM normals data to determine the fraction of normal precipitation. The normalized scale and shape parameters are used to determine the cumulative probability corresponding to the estimated precipitation amount, and this probability is transformed using an inverse normal function to obtain the SPI.
Because both AHPS and NCEP Stage IV precipitation data are available at daily intervals and at roughly 4-km resolution, this method allows for the production of daily, high-resolution maps of SPI for various time scales. Although the AHPS and NCEP Stage IV precipitation datasets are similar, spot comparisons do show differences. Using both to calculate SPI allows for a comparison between different precipitation inputs. The HRD SPI is calculated using two different PRISM normals datasets (1971–2000 and 1981–2010) from the PRISM Climate Group (PRISM Climate Group 2014; Daly et al. 2008), as well as distribution parameters calculated from two different clusterings of NWS COOP sites, each based on either the 1971–2000 (cluster set 1) or 1981–2010 (cluster set 2) normals period, in addition to the two different high-resolution precipitation datasets. HRD SPI are generated using the 1971–2000 PRISM normals and associated cluster solution (cluster set 1) for consistency with McRoberts and Nielsen-Gammon (2012) while developing the method for calculating the HRD SPI. The 1981–2010 PRISM normals and association cluster solution (cluster set 2) are used to calculate HRD SPI that reflect the most recent normals period.
Names are assigned to HRD SPI generated with each input data combination (Table 1) and will be used throughout this paper. In brief, HRD SPI calculated using AHPS precipitation is referred to as AHPS-SPI, while HRD SPI calculated using NCEP Stage IV precipitation is referred to as NCEP-SPI. In addition, HRD SPI calculated using the 1971–2000 PRISM normals and cluster set 1 has a subscript of “1” (i.e., AHPS-SPI1 and NCEP-SPI1), and HRD SPI calculated using 1981–2010 PRISM normals and cluster set 2 has a subscript of “2” (i.e., AHPS-SPI2 and NCEP-SPI2).
The HRD SPI dataset described here is updated each day for the contiguous United States and is made publicly available from the State Climate Office of North Carolina (http://www.climate.ncsu.edu/drought). Grids can be viewed online with a map tool or can be downloaded in GIS-ready formats. Through our engagement with national drought-monitoring efforts, we know that HRD SPI are already being used by U.S. Drought Monitor (USDM) authors for their weekly assessments as well as by several state programs, including Vermont, Missouri, Texas, and North Carolina.
McRoberts and Nielsen-Gammon (2012) previously evaluated the ability of the prototype high-resolution SPI, referred to as MPEDE-SPI [where MPEDE represents multisensor precipitation estimates (MPE) drought estimator], to measure drought conditions during the 2008–09 Texas drought. They used a comparison with the station-based SPI from the High Plains Regional Climate Center’s Applied Climate Information System and reports from county extension agents. Comparisons revealed that the MPEDE-SPI and station-based SPI agreed well, with instances of larger differences seemingly tied to areas where radars consistently under- or overestimated precipitation. The advanced resolution gained by the MPEDE-SPI allowed for a better estimate of conditions in areas where no surface gauges were present, as validated by county extension reports. The primary potential problem noted with the MPEDE-SPI was spatial discontinuities in drought depictions that resulted from radar biases, particularly at longer durations (McRoberts and Nielsen-Gammon 2012). A comparison with climate-division SPI from the National Centers for Environmental Information (NCEI; known as the National Climatic Data Center at the time of the drought) over a 3-yr period also revealed that the two were in strong agreement over most of Texas.
The ability of the above-described HRD SPI to monitor drought is assessed here using two drought case studies: a flash drought over the central Great Plains in 2012 (GREATPLAINS12) and a more slowly evolving drought over the Carolinas in 2007–08 (CAROLINA07). We chose these two events for comparison because they were during the recent period when HRD SPI is available, they covered the range of drought severities (D0–D4 on the USDM’s scale), they represented droughts with different spatial extents and temporal variations for drought onset and amelioration, and they covered geographies with varying radar coverages.
Two operationally produced drought products that are available on a monthly time step over the central Great Plains are used for the initial case-study comparison with the HRD SPI: climate-division SPI from the NCEI (Vose et al. 2014) and a gridded SPI from the Western Regional Climate Center’s WestWide Drought Tracker (Western Regional Climate Center 2014). Over the Carolinas, the HRD SPI is also compared with SPI produced from the dynamic drought index toolkit (DDIT), which is only produced over North and South Carolina. Other operational SPI products were considered but were ultimately excluded from this analysis. These include daily gridded SPI from NOAA’s Climate Prediction Center, which is not archived, and daily station-based SPI from the High Plains Regional Climate Center, with which McRoberts and Nielsen-Gammon (2012) visually compared the prototype MPEDE-SPI but which does not have publicly available GIS-format data. For both case studies, the HRD SPI is also compared with USDM maps.
Climate-division SPI produced by NCEI (NCEI-SPI) uses monthly climate-division precipitation data with a reference period of 1931–90 and the Pearson type-III distribution (see the NCDC document ftp://ftp.ncdc.noaa.gov/pub/data/cirs/climdiv/divisional-readme.txt and the NOAA computer program code SPICOMPUTE by N. Guttman: http://www1.ncdc.noaa.gov/pub/data/software/palmer/spi.f). To calculate monthly climate-division precipitation, station data are interpolated by using a climatologically aided approach that minimizes errors due to geography, and the estimates at grid points within each climate division are averaged using areal weights (ftp://ftp.ncdc.noaa.gov/pub/data/cirs/climdiv/divisional-readme.txt Vose et al. 2014). Station data come from the Global Historical Climatology Network, which includes data from a variety of networks, including NWS COOP.
The SPI from the Western Regional Climate Center’s WestWide Drought Tracker (WWDT-SPI) was developed to provide high-resolution (approximately 4 km) drought monitoring in the western half of the conterminous United States, where terrain can have large effects on temperature and precipitation and where there are significant data limitations, both in the density of surface stations as well as in the coverage provided by remotely sensed data, such as radar-derived precipitation estimates (Western Regional Climate Center 2014). The WWDT-SPI is based on monthly data from the PRISM Climate Group (PRISM Climate Group 2014; Daly et al. 1994) and uses an incomplete beta distribution determined over the period of 1895–2010.
The Carolinas DDIT was developed shortly after the worst drought up to that time affected the Carolinas from 1998 to 2002. It was developed to address several monitoring deficiencies, including the need for drought indices with localized detail (Carbone et al. 2008). The DDIT generates weekly and monthly 1-, 3-, 6-, 9-, 12-, and 24-month SPI data calculated for individual stations and aggregated to a variety of spatial scales to meet different user needs (Dow et al. 2009; Carbone et al. 2008). DDIT-SPI is generated at monthly and weekly time scales using daily precipitation data since 1950 from 238 NWS weather stations in the Carolinas (Carbone et al. 2008). Following calculation of SPI at each station, values are interpolated to a 4-km grid using inverse distance weighting, allowing for aggregations to a variety of spatial units, such as climate divisions or counties (Dow et al. 2009; Carbone et al. 2008).
d. U.S. Drought Monitor
The USDM is a weekly product that attempts to depict, on a single map and with a simple scale, the current spatial extent and severity of drought across the United States (Svoboda et al. 2002). This weekly product is created through expert assessment of a suite of drought indicators, atmospheric and hydrological observations, and reports of on-the-ground conditions and impacts, all from a variety of sources and at various spatial scales, including point, climate division, and gridded. Subjective assessments of these sources, as well as of local data and reports, by experts at federal, regional, state, and local levels across the country are incorporated into the USDM process (Svoboda et al. 2004). The five USDM categories are based on the percentile chance of occurrence in any given year out of a 100-yr period (Svoboda et al. 2002). The objective indicators used in the USDM process are transformed to their historical frequency of occurrence for the location and time of year in question to allow for a better comparison with the USDM categories (Table 2).
a. Full-period SPI comparison
HRD SPIs are initially compared with WWDT-SPI and NCEI-SPI for the approximately 10-yr period for which HRD SPI are available, ending with December of 2015. Only WWDT-SPI and NCEI-SPI are used for this comparison because they are both available over the contiguous United States. Because both HRD SPI and WWDT-SPI are gridded, Pearson’s correlation coefficient is calculated for each grid point and then is averaged over the full grid for each time scale. NCEI-SPI is available at the climate-division scale; therefore, each HRD SPI version is first aggregated up to climate divisions before the Pearson correlation coefficient is calculated. As with the WWDT-SPI comparison, correlations between NCEI-SPI and each HRD SPI version are averaged over all climate divisions for each time scale. Since NCEI-SPI is only available for 1-, 2-, 3-, 6-, 9-, 12-, and 24-month durations, the comparison between NCEI-SPI and HRD SPI is for these time scales only. The comparison between WWDT-SPI and HRD SPI is for all 14 time scales for which the HRD SPI are available.
b. SPI comparison for individual case studies
For the two case studies, objective comparisons are made between SPI datasets on the climate-division scale to allow for a triplicate comparison with NCEI, which is available at this aggregation. The purpose of this analysis is to make an objective, quantitative assessment of the HRD SPI dataset’s quality during each drought event. HRD SPI and WWDT-SPI grids are averaged over climate divisions within each case-study domain (Fig. 1), where each grid point receives equal weight. GREATPLAINS12 uses SPIs for each month in 2012, and CAROLINA07 uses SPIs for each month in 2007–08.
SPI data are grouped by duration, and Pearson’s correlation coefficients and root-mean-square differences (RMSD) are calculated for each possible comparison over all months and climate divisions. These values are then averaged over all durations to obtain an average spatial and temporal correlation and RMSD between each SPI combination. Last, the spatial pattern of the association of the SPI datasets is estimated by averaging the correlation coefficients across all durations for each climate division. These are calculated for each of the climate divisions in the study area (Fig. 1); SPI comparison values for each climate division are obtained by averaging the correlation coefficients and RMSDs over all durations.
c. Objective U.S. Drought Monitor comparison
An objective comparison with USDM maps is performed in each case study. For GREATPLAINS12, which is performed first, the HRD SPI input dataset combination and time scale that have the highest correlation and smallest RMSDs with the USDM is determined. A similar comparison is analyzed for CAROLINA07 but only using the version of HRD SPI that performed best in GREATPLAINS12.
To compare the HRD SPI and USDM, SPI grids are converted to USDM categories with the percentiles approach used by the USDM and shown in Table 2. USDM shapefiles are converted to rasters with the same gridcell size as the SPI files, resulting in two spatially and temporally comparable datasets. RMSDs are calculated between the USDM and each gridded SPI (converted to USDM categories) over an N-week period to assess how each HRD SPI variety and duration match the USDM. A single RMSD between the USDM and HRD SPI is calculated by combining data over all grid cells and durations.
d. Visual analysis with USDM
A visual analysis of the HRD SPI to the USDM and other SPIs is performed to evaluate whether HRD SPI provides relevant, localized information needed for drought monitoring. This qualitative analysis allows for a comprehensive assessment of the value of the HRD SPI relative to independent SPIs.
a. Full-period SPI comparison
Pearson’s correlation coefficients calculated between gridded WWDT-SPI and each HRD SPI version reveal that the two SPI analyses are strongly correlated over the full period that HRD SPI data are available (Table 3). The same is true for the climate-division comparison between NCEI-SPI and each HRD SPI version (Table 4). In each case, HRD SPI versions using AHPS precipitation have stronger correlations. In addition, the AHPS-SPI1 version has the highest correlation coefficients, but only by approximately 0.01 when compared with correlations for AHPS-SPI2. This may be because AHPS precipitation is more similar to the precipitation used in the PRISM grids and NCEI climate-division precipitation estimates. Also, WWDT-SPI uses the full PRISM period of record (since 1895) to calculate historical distributions, whereas NCEI-SPI uses the period of 1931–90. The distribution parameters resulting from the 1971–2000 PRISM normals and cluster set 1 may be more similar to those used in the WWDT-SPI and NCEI-SPI, leading to slightly higher correlations. Overall, these results indicate that each HRD SPI variety agrees strongly with independent SPI analyses.
The central United States experienced a rapid onset of drought conditions during the summer of 2012. In the winter and spring of 2012, Texas, New Mexico, Oklahoma, and southern Kansas were in the midst of a drought that had begun the previous year (Hoerling et al. 2013a). Portions of Iowa and Nebraska were also experiencing conditions that were abnormally dry to drought level in the spring of 2012, but the rest of this region was largely drought free. In May, June, and July the central United States experienced strong heat waves that, when combined with reduced precipitation and agricultural demand for water, led to a rapid onset of drought conditions (Hoerling et al. 2013b, 2014). The May–August 2012 period saw the most severe precipitation deficits over the central Great Plains since record keeping began in 1895 (Hoerling et al. 2013b, 2014). By late August 2012, over 75% of the United States was experiencing drought or abnormally dry conditions, with nearly the entire GREATPLAINS12 study region experiencing extreme or exceptional drought, the two most severe categories used by the U.S. Drought Monitor. The primary cause for the rapid onset of drought in the summer of 2012 was a reduction in atmospheric moisture transport into the Great Plains from the Gulf of Mexico, which is typically a major source of moisture for the region (Hoerling et al. 2013b, 2014).
1) Comparisons between SPIs
Both AHPS-SPI1 and AHPS-SPI2 correlate more strongly with WWDT- and NCEI-SPIs than either NCEP-SPI1 or NCEP-SPI2 (Table 5), suggesting the AHPS precipitation product is more similar to the precipitation values used by the WWDT- and NCEI-SPIs. In contrast, little difference is observed in correlation coefficients between SPIs generated with the same precipitation but different PRISM normals and cluster solutions. The HRD SPIs have much weaker correlations with NCEI-SPI (averaging 0.39) than the WWDT-SPI (averaging 0.75). For a comparison, the correlations between the WWDT-SPI and NCEI-SPIs were calculated and were found to be similar to those between the NCEI-SPI and HRD SPIs, averaging 0.39.
RMSDs between SPI datasets show similar patterns to the correlation analysis (Table 6). HRD SPI using AHPS precipitation generally has lower RMSDs with comparison SPIs than HRD SPI using NCEP Stage IV precipitation. In addition, average RMSDs between the HRD SPIs and the WWDT-SPI is 0.62, as compared with 1.02 with NCEI-SPI.
When broken down by climate divisions, the HRD SPIs are strongly correlated with WWDT-SPI over the northeast portion of the study area (0.75 ≤ r ≤ 1.0, with slightly weaker correlations in central southern Colorado, New Mexico, and westernmost Texas; Fig. 2a). NCEI-SPI and HRD SPIs also generally correlate more strongly in the northeastern part of the study region, but overall correlations are much lower, and even negative, in some southwestern climate divisions (Fig. 2b). These negative correlations could be because of lack of input data for the AHPS product. AHPS may not be picking up precipitation that surface gauges are, and this could be resulting in a decrease in HRD SPI but an increase in station-based SPIs. For comparison, the average climate-division correlations are calculated between WWDT-SPI and NCEI-SPI, revealing a similar pattern to the NCEI-SPI and HRD SPI comparison (Fig. 2c). This result suggests the influence of the use of PRISM data in both the WWDT and HRD SPIs.
2) Objective comparison with the USDM
The period of study for the comparison with the USDM is from 1 May 2012 through 28 August 2012, which was characterized by a rapid onset of drought conditions throughout the central Great Plains region, with ongoing drought conditions in the southern and western portions of the 10-state study area.
RMSDs between the HRD SPIs and the USDM are shown in Table 7. These are used as a measure of the relative agreement between each HRD SPI version and time scale and the USDM. Shorter time scales generally have lower RMSDs, indicating that these durations more closely match the conditions that were indicated by the USDM. Considering the rapid onset of drought conditions during this time, this result is not unexpected. The 1-month duration is an exception to this pattern and has RMSDs that are more similar to those of longer durations. This time scale is the most sensitive to precipitation, or a lack thereof, and can consequently change dramatically from one week to the next. In contrast, the USDM by design tends to depict a slower evolution of drought conditions, likely leading to higher RMSDs with the 1-month HRD SPI. When averaged over all time scales for each input dataset combination, the AHPS-SPI2 has the lowest RMSD. Therefore, all subsequent analyses, including the CAROLINA07 case study, will focus on the AHPS-SPI version only.
Analysis of the HRD SPI durations with the lowest RMSD (Fig. 3) shows that shorter durations (of roughly ≤5 months) have the lowest RMSDs over the areas that experienced shorter-term drought conditions, whereas longer durations (of roughly ≥12 months) generally have the lowest RMSDs in areas that were mostly experiencing ongoing drought conditions that had begun much earlier. This result suggests that the HRD SPI is capable of recognizing the different time scales of the droughts indicated by the USDM for this period.
3) Visual analysis with the USDM
(i) Localized drought-monitoring information
Over the course of July of 2012, the severity of drought throughout the central Great Plains intensified; most of the area was experiencing severe and extreme drought, with a few pockets of exceptional drought, by 31 July 2012 (Fig. 4a). The 3-month AHPS-SPI2 (Fig. 4b) indicates widespread severe and extreme drought, with small pockets of exceptional drought, throughout this area, lining up with the areas indicated by the USDM. The WWDT-SPI also indicates significant amounts of severe and extreme drought (Fig. 4c), but these are not as pervasive as the AHPS-SPI2 shows. This is especially evident over southern Oklahoma into Arkansas where the WWDT-SPI has areas of moderate or no drought while the AHPS-SPI2 shows pockets of extreme and severe drought.
As shown in Fig. 4, the WWDT-SPI indicates that most of Nebraska, with the exception of the southeastern corner, is experiencing exceptional drought. In contrast, the AHPS-SPI2 indicates an area of severe drought, with a few pockets of extreme drought throughout most of Nebraska, while also indicating moderate to severe drought conditions in southeastern Nebraska. When compared with the USDM, the AHPS-SPI2 clearly captures the severity of conditions throughout most of the state better than does the WWDT-SPI. In contrast, in western Kansas, the Oklahoma Panhandle, and northern Texas, the WWDT-SPI captures the areas of extreme and exceptional drought indicated by the USDM that the AHPS-SPI2 shows as being in moderate to extreme drought.
The similarities between the AHPS-SPI2 and the WWDT-SPI suggest that the information provided by the AHPS-SPI2 for drought monitoring is similar to independently generated SPI products. The major advantage of the AHPS-SPI2 is its incorporation of bias-corrected radar-derived precipitation data, which can provide high-resolution estimates of precipitation in areas that lack surface observations. Inspection of USDM maps reveals clear instances of broad agreement between the AHPS-SPI2 and the USDM, but the spatial pattern in drought severity is more localized and detailed in the AHPS-SPI2 than in the USDM. Even though AHPS precipitation data are examined while creating the USDM, the USDM does not depict the same spatial variability suggested by it. One possibility for this is a tendency to aggregate large areas as experiencing the same conditions, especially if there are no on-the-ground observations to suggest otherwise. In addition, there are other factors, such as high temperatures and stress to a variety of sectors, that influence the decision to depict a particular location or region as experiencing a specific severity of drought, regardless of whether it has received the same amount of precipitation as its neighbors.
(ii) Issues with HRD SPI
The HRD SPI may have an advantage over independent SPI analyses where radar coverage is good, but where coverage is poor this SPI could have biases that make it a less-useful product. As an example, on 1 May 2012, the 8-month AHPS-SPI2 (Fig. 5a) indicates pockets of extreme and severe drought in northern and southern Missouri that are not indicated by the USDM map for the same date (Fig. 5c). These areas line up well with areas of known poor radar coverage (Fig. 5b). Further muddling the issue is that the 8-month WWDT-SPI indicates some dryness (Fig. 5d) in southern Iowa, along the border with northern Missouri, but this does not line up with the area indicated by either the USDM or the AHPS-SPI2. At different time scales these areas are not as obviously different in the AHPS-SPI2 maps, suggesting that it takes a longer duration of time for these errors resulting from radar coverage to accumulate. A local knowledge of radar errors in these regions would provide valuable information about the limits of the AHPS-SPI2, as well as all other HRD SPI varieties, improving its utility for drought monitoring. In addition, AHPS precipitation estimates, and the resulting SPI, would be improved by utilizing more quality-assured surface gauge precipitation measurements, if available, for bias correction in these areas.
A second case study is performed to evaluate the gridded SPI product over the Carolinas during 2007–08, when both states experienced widespread drought. Drought onset began rapidly from the late spring to early summer of 2007, on the heels of a dry winter, as demand for water increased with the start of the growing season. Record-setting heat in the summer of 2007 combined with an absence of tropical systems in the summer and autumn of 2007 contributed to widespread drought expansion across both states, despite minor events that brought sporadic and localized improvement. By 25 December 2007, roughly one-half of the Carolinas was experiencing exceptional drought, and all of North Carolina and most of South Carolina were in some stage of drought (http://droughtmonitor.unl.edu/MapsAndData/WeeklyComparison.aspx). Dry conditions persisted into 2008, with another dry start to summer. Tropical rains brought relief to parts of the Carolinas in August and early September of 2008, but widespread improvement was absent in the Carolinas until the winter of 2008/09, when precipitation ranged from normal to above normal (http://droughtmonitor.unl.edu/MapsAndData/WeeklyComparison.aspx).
The evolution of conditions in North and South Carolina was monitored by using a variety of data and drought indices, including SPI. Two resources for SPI available at the time of this drought are the NCEI and the Carolinas DDIT. The research presented in this section provides an evaluation of the AHPS-SPI2 over the Carolinas during 2007–08 through a comparison with independently generated SPI from three sources: NCEI, WWDT, and DDIT. SPI from NCEI and DDIT represent the types of available data at the time of the 2007–08 drought, and the WWDT-SPI is used to allow comparison with another gridded product. USDM maps are also incorporated as “ground truth” for quantifying the ability of each SPI dataset to measure the spatial extent and severity of drought conditions in 2007–08.
1) Drought-index comparison
Monthly climate-division SPI for 1-, 3-, 6-, 9-, 12-, and 24-month durations over the Carolinas were obtained from NCEI and DDIT, where the choice of SPI durations is based on availability of DDIT-SPI. In a similar way, monthly AHPS-SPI2 and WWDT-SPI were obtained over the Carolinas for 2007–08 and were averaged over each climate division, with each grid cell receiving equal weight.
Correlation coefficients between the AHPS-SPI2 and DDIT-SPI (0.88), NCEI-SPI (0.90), and WWDT-SPI (0.90) suggest that the AHPS-SPI2 is strongly linearly correlated with each of the comparison SPIs at the climate-division level (Table 8). Correlations are stronger for durations of ≤12 months, with average magnitudes ranging from 0.88 to 0.97, but are lower for the 24-month duration, with average magnitudes that range from 0.77 to 0.79. For comparison, average correlations between the DDIT-SPI and the WWDT- and NCEI-SPIs are 0.93 and 0.92, respectively, which values are only slightly higher than corresponding correlation coefficients between the AHPS-SPI2 and WWDT- and NCEI-SPIs.
RMSDs agree with the results of the correlation analysis, being smallest at shorter durations, where values are less than or equal to 0.5, and largest for the 24-month duration (Table 9). In addition, the RMSDs between AHPS-SPI2 and WWDT- and NCEI-SPIs are nearly identical to the RMSDs between the DDIT-SPI and the WWDT- and NCEI-SPIs, suggesting that the AHPS- and DDIT-SPIs both agree about as well with these two comparison SPIs when averaged to climate divisions. In contrast, the WWDT- and NCEI-SPIs are nearly perfectly linearly correlated, with an average correlation coefficient of 0.99, and have the most similar magnitudes of the four comparison datasets, as indicated by average RMSD of 0.17.
2) USDM comparison
RMSDs computed between weekly USDM maps and gridded AHPS-SPI2 in 2007–08 are lowest for the 1-month duration and increase with the length of the duration (Table 10). This result suggests that the 1-month AHPS-SPI2 duration best matches the severity and extent of drought conditions indicated by the USDM in the Carolinas during 2007–08. Visual comparison of the different SPIs with the USDM over the Carolinas in 2007–08 reveals an overall consistent pattern in drought depiction, with subtle differences in the severity or spatial distribution.
From mid-August through September of 2007, hot temperatures and decreased precipitation led to drought expansion and degradation across the Carolinas. Tropical Storm Gabrielle brought slight relief to parts of eastern North Carolina in early September of 2007, but it was not enough to remove abnormal dryness and drought from the region. By the beginning of October, over one-half of the area of the Carolinas was experiencing severe, extreme, or exceptional drought, according to the 2 October 2007 USDM map (Fig. 6e).
The 3-month AHPS-SPI2 for 2 October (Fig. 6a) indicates that exceptional drought (D4) is less widespread, more spotty, and mostly absent from the southern piedmont climate division of North Carolina—an overall less-severe indication than the USDM depiction. The 3-month NCEI- and WWDT-SPIs for 30 September (Figs. 6b,c), in contrast, indicate that exceptional drought is present throughout central North Carolina into much of South Carolina, which is a greater spatial extent than the USDM’s depiction. Further complicating the picture is that the 3-month DDIT-SPI for 30 September (Fig. 6c) also indicates that sites across central North Carolina and South Carolina are in exceptional drought (D4), with bordering sites experiencing less-severe dryness. On an individual basis, the sites in this region agree with the severity indicated by the AHPS-SPI2 for the same locations. An interpolation between these sites would result in a depiction of exceptional drought throughout the central Carolinas, missing the localized variations that the AHPS-SPI2 captures, while still suggesting less of a spatial extent to these conditions than does the WWDT- or NCEI-SPI. Numerous sites across the Carolinas do not have SPI calculated because of insufficient data, making between-station estimates of drought conditions problematic and highlighting a drawback of interpolating station data, a technique on which the WWDT-SPI, the NCEI-SPI, and the larger spatial aggregation units of the DDIT-SPI all rely.
5. Summary discussion and conclusions
HRD SPIs correlate strongly with independently produced WWDT-SPI and NCEI-SPI over the contiguous United States and over the approximately 10-yr period for which HRD SPI are available. Comparison of HRD SPIs with SPIs generated by WWDT and NCEI averaged over climate divisions in the central Great Plains in 2012 reveals strong correlations and low RMSDs with WWDT-SPI but weaker correlations and higher RMSDs with NCEI-SPI. Use of PRISM data in both the HRD SPI and WWDT-SPI may be a large part of the reason for their similarity. Comparisons between the WWDT-SPI and NCEI-SPI over GREATPLAINS12 also revealed lower correlations. Even though the HRD SPI only uses PRISM for the location of the historical distribution and the WWDT-SPI uses a different probability distribution function, the similarities between the two SPI products are clear and can likely be attributed to this common dataset.
The HRD SPI generated using AHPS precipitation and the 1981–2010 PRISM normals and associated cluster solution (AHPS-SPI2) is found to have the lowest RMSDs with the USDM over the central Great Plains during the 1 May–28 August 2012 period of rapid expansion, suggesting that this input-dataset combination is the best for monitoring drought. Only this HRD SPI variety is considered for further analysis.
Visual comparison of AHPS-SPI2 with WWDT-SPI and the USDM over the central Great Plains reveals that the AHPS-SPI2 is capable of capturing localized detail that is also reflected in the USDM but not in the WWDT-SPI. This spatial detail is largely a result of the incorporation of radar-derived precipitation into the AHPS precipitation product, giving it the ability to capture the hit-or-miss nature of convective precipitation over the summer and allowing for a much more localized depiction of drought conditions in the resulting SPI. Even though AHPS precipitation data are incorporated into the USDM process, the USDM maps do not always show this same level of variation. Several reasons exist for this fact; in determining areas that are experiencing a given drought severity, large areas may be aggregated together for ease of communication. In addition, if no on-the-ground observations support local-scale variations, it is likely that drought severity will be interpolated between points.
The AHPS-SPI2 generally compares the worst to the USDM and WWDT-SPI in areas with poor radar coverage, where underestimates in the AHPS precipitation product can translate to more severe drought indicated by the AHPS-SPI2. In particular, areas in the Rockies have greater disparities between the different products. This is likely because the complex, mountainous terrain causes radar beam blockages, the distances between radars is so great that precipitation is missed, and there are fewer surface stations to correct the bias in radar estimates of precipitation. The influence of the radars on the AHPS-SPI2, and all HRD SPI varieties, should be considered when using this product in areas with poor radar coverage, especially in future long-term drought events. Future bias corrections to the input gridded precipitation datasets, from incorporation of more quality-assured surface gauge observations or from improved bias-correction functions, may help with this issue.
Climate-division comparisons during the Carolinas 2007–08 drought of AHPS-SPI2 with SPI generated by DDIT, WWDT, and NCEI suggest that the AHPS-SPI2 is able to provide information that is similar to that of independently generated SPI. Visual comparisons with independent SPIs and the USDM highlight the AHPS-SPI2’s unique ability to indicate conditions in which no or few surface gauges are located, as well as its high spatial resolution (roughly 4 km), both of which are valuable additions to drought monitoring. There are still advantages to utilizing a station-based SPI, such as the DDIT-SPI, since it represents more accurate values at the individual station locations. As monitoring tools, both have merit, and the operational use of one over the other is likely to be driven by a user’s needs for localized or more regional detail.
A reduction in precipitation is a driving force behind drought, but it is not the only factor that contributes to a drought’s severity or duration. Both the 2012 central Great Plains drought and the 2007–08 Carolinas drought were exacerbated by strong heat waves, which, when combined with growing-season demands, led to numerous impacts that would not necessarily be expected solely on the basis of the precipitation deficit. For these reasons, it is not expected that the HRD SPI, or any SPI, will be able to completely match USDM-indicated drought severity and extent 100% of the time. However, the multiscalar nature of the SPI combined with the increased precipitation information provided by the NCEP Stage IV and AHPS precipitation products should allow for a better estimation of those areas that are experiencing precipitation deficits that may lead to drought impacts. A multiscalar drought index that incorporates a temperature component, such as the standardized precipitation evapotranspiration index, and that is calculated with the same high-resolution data inputs that are used for the HRD SPI, may be a more robust drought index, especially during heat-exacerbated drought events.
Funding support for this research was by U.S. Department of Agriculture National Institute of Food and Agriculture Grant 2011-67019-20042 and by NOAA through the regional Carolinas Integrated Sciences and Assessments program. The statements, findings, and conclusions are those of the authors and do not necessarily reflect the views of NOAA, the Department of Commerce, or the U.S. Department of Agriculture. The HRD SPI product and research described here would not be possible without the freely available AHPS precipitation, NCEP Stage IV, and NWS COOP data from NOAA, the monthly PRISM normals data from the PRISM Climate Group, or the drought-index information from the dynamic drought-index tool from the South Carolina Department of Natural Resources. The authors also thank the reviewers for their insights and suggestions, which have greatly improved this article.