*3.2. Precipitation, Temperature Trends, and SPEI-Based Drought Identification*

All products agreed (at the 0.01 significance level, referring to the *t*-test of significance of linear regression slope) on an upward trend of Tmin of about 0.02–0.03 ◦C per year and of Tmax of about 0.02–0.06 ◦C per year (Figure 4). The annual rainfall sums showed no trend or a declining trend, but none of these were significant at the 0.01 level (Figure 5). The standard deviations of rainfall likewise showed no significant trends (Figure 6). The same applied for seasonal trends (Figure 7). Despite the differences in the precipitation and temperature products, once propagated to the SPEI the differences smoothed out, yet differences in onset, duration, and magnitude of drought remained (Figures S7–S10).

**Figure 4.** Significant annual maximum (Tmax) and minimum (Tmin) temperature trends in the MERRA-2, KMD\_grid, CRU, and JRA-55 gridded products. The black dots represent annual precipitation values with the blue line indicating the linear regression.

**Figure 5.** Non-significant annual precipitation trends in the 10 products. The back dots represent annual precipitation values with the blue line indicating the linear regression.

**Figure 6.** Non-significant annual precipitation standard deviation trends in the 10 products. The black dots represent annual precipitation values with the blue line indicating the linear regression.

Out of the 40 blends, 18 agreed on a statistically significant (at 0.01 level) trend in SPEI of 0.0001 to −0.0098 units month<sup>−</sup>1, suggesting increasing instances of drought occurrence (Figure S7–S10). Those trends were consistent across the temperature products with CHIRPS, GPCP, KMD\_grid, and PERSIANN rainfall, sometimes with ERA-5 and JRA-55, and in one instance with CRU rainfall (Figure S7–S10). Plotting the SPEI mean and standard deviation across the product blends further consolidated the picture (Figure 8). Unambiguous drought years, according to the data products, were 1994, 1996–1997, 1999–2000, 2005–2006, 2009, and 2011. More ambiguous were 1988, 1991–1993, 2001–2004, 2008, 2010, and 2012–2016. Drier conditions in recent years, as suggested by the trend analysis, could be seen in 2005–2006 and 2008–2012, compared to more positive anomalies in the 1990s and early 2000s.

The information from the key informant interviews agrees with all unambiguous droughts in the timespan (2005–2006, 2009, and 2011) and the one year which was unambiguously wet (2007). The interviews also pointed to droughts in 2008, 2010, 2012, and 2014–2015 where the SPEI information based on the different data products was ambiguous. In the other ambiguous years, 2013 and 2016, the key informant interviews pointed to no drought. Hence it would seem that key informants engaged in drought relief on the ground can help resolve the ambiguity resulting from the disagreement between meteorological data products. Their input is thus fundamental for drought identification in regions with scarce ground stations.

**Figure 7.** Non-significant seasonal precipitation trends for the 10 products: (**a**) March–April–May and (**b**) October-November-December. The black dots represent annual precipitation values with the blue line indicating the linear regression.

**Figure 8.** Inter-product SPEI mean (black line) ± 2 standard deviations (gray shading), compared with key informant information from 2005 to 2016 (Red dots). Periods where mean SPEI ± 2 standard deviations were below zero are colored red, those above zero are blue.

#### **4. Discussion**

### *4.1. Uncertainty in Rainfall and Temperature Estimates and Propagation to SPEI*

Reliable assessments of the onset, magnitude, and duration of drought are vital in agro-pastoral ecosystems, not only to understand impacts on livelihoods but also to signal and assess the reliability of responses [2,65]. In the absence of reliable meteorological data as a result of sparse in-situ station density over Kenya [16,32,35] and other African countries, rainfall and temperature data from gridded products can overcome data scarcity for large-scale drought assessment [7,46,66]. These products, however, are subject to uncertainty, including gauge-level measurement errors in the underlying station data, the number and representativeness of the stations used, interpolation steps, structural, parameter, and general input data uncertainties of the meteorological models used [21].

The abundance of gridded data products available thus creates both a challenge and an opportunity for users. Choosing a single product can lead to biased drought estimations as AghaKouchak et al. [67] found out; hence; the use of multiple products in an ensemble approach is preferable [68]. Such an approach will add uncertainty information to the gridded products that can improve decision making in response and management operations [67]. That said, uncertainty in drought magnitude should in no way instill a sense of complacency as increasing extreme events such as droughts over East Africa have already resulted in deterioration of livelihoods and ecosystem integrity [69–71].

In the current study, uncertainty manifests itself in differences between the data values of gridded meteorological products, with annual minimum and maximum temperature varying less than rainfall. The temporal pattern of the Tmax and Tmin input was also more similar across products than that of rainfall. The variation of SPEI across data blends therefore predominantly reflects the variation of the rainfall data. Plotting the SPEI ensemble mean ±2 standard deviations identified periods of unambiguous dry and wet years, while ambiguous periods could be resolved by information from key informants engaged in drought relief on the ground. It should be noted that the uniform weighting of SPEI ensemble members neglects the similarity between some of the data blends, as they use similar data and assumptions, which are, however, hard to disentangle and quantify in an alternative weighting scheme. As such, we could not authoritatively pick out a superior data product but observed the similarity in detecting drier years. Drought occurrence was thereby much less ambiguous than drought severity.

#### *4.2. Annual and Seasonal Trends*

By comparing 10 precipitation products, we found no evidence of a statistically significant trend (although there could be a trend), neither in annual rainfall nor seasonal rainfall totals, nor annual standard deviations. This finding is in contrast with the declining rainfall trend over East Africa reported by [36,38,71,72] and [11]. It is also in contrast with the key informant information that the March-April-May (MAM) rain season, being the longer of the two seasons and essential in the farming calendar, has demonstrated unreliability in recent years. Agricultural water demand is likely rising considering the growing population [20,37], nevertheless declining length of the March-April-May season could be the principal factor of increasing water scarcity, rather than burgeoning anthropogenic water needs. Since rain-fed agriculture is the primary source of livelihoods in the study area and the primary contributor to the economy [36,37], a decrease of rainfall in the long season and a general shortening of the season is a major concern [73]. However, the reported unreliability of the March-April-May season in recent years could also be reflective of generally drier soil conditions in response to the positive temperature trend which we did find across all data products, or changes in sub-seasonal rainfall timing that are not visible as a trend in annual standard deviations. Both would propagate to lower SPEI values, which in our case and for most products agree with an increase in drought instances in recent years.

The absence of evidence of a significant trend in the shorter October-November-December (OND) rain season in our case (Figure 7) differs from recent studies over

Kenya [38]. The key informants and the Kenyan Government GoK [74], however, support our finding by mentioning that the shorter OND season has shown more reliability in supporting farming compared to the longer MAM season. This is manifested by greater seasonal rainfall averages in the OND season in most products (Table S2). The OND season, however, shows greater variation than the MAM season (Table S2) as also reported by [75]. The MAM, especially due to its lower variability, thus remains important for agroecosystem productivity in the region, with a likely atmospheric teleconnection with the OND as shown by [71]. The MAM season plays a primary role in the farming calendar of the study area, accounting for about 30% of crop productivity, and supporting cultivation of staple pulses such as pigeon peas and green grams [74].

With regard to temperature, all data products compared in this study agreed on positive trends in min/max temperatures. While the products were in greater agreement about the magnitude of the Tmin trend, the Tmax trend magnitude varied more between products. This agrees with findings over Kenya by Ayugi and Tan [46] who found increasing trends of min/max temperatures, and Camberlin [76] who similarly reports a marked warming in the Horn of Africa. Ayugi and Tan [46] found warm days to be increasing and cold nights to be decreasing, as well as summer days to be increasing, over Kenya, confirming the picture of rising temperatures.

### *4.3. Anomalies, Drought Identification, and the Value of Triangulation*

The 10 different precipitation products compared in this study generally agreed on years with negative rainfall anomalies. However, the products disagreed considerably on the magnitudes of those anomalies. The anomalies, seen in Figure 2, demonstrate the prevailing inter-annual variability in the study area [75]. The anomalies propagated to droughts of varying magnitude, confirmed by unanimously negative SPEI values or key informants in 27% of the 30 years. However, in 1988, 1991–1993, and 2001–2004 there was disagreement between the products and the key informant information did not reach that far back.

The 2010–2011 period is widely reported as the worst drought in a 60-year span in the Horn of Africa [11,71,77] which is confirmed by the key informants for the study area but unanimously confirmed by the SPEI products only for 2011. While in most years the multi-product approach allows us to robustly identify drought and get a handle on the uncertainty in drought magnitude, from 2008 onwards, the greater disagreement between the data products, both in terms of SPEI direction and magnitude, highlights the potential of information from actors engaged in drought relief in the region. Our key informants worked in disaster risk management, food security, water storage/harvesting and climate change resilience building, i.e., sectors that are sensitive to drought conditions. These experts' inputs are therefore viewed as important in the continued assessment and response to droughts with their observations contributing to resolving ambiguity.

These inputs are particularly valuable in drought assessments for relatively constrained spatial extents, as informant data on droughts can be assumed to cover the entire study area. For large-area drought estimations covering larger regions or featuring more localized droughts, spatially explicit information on the location and extent of informant activities must be collected during interviews and integrated into the verification of the drought occurrence estimation. The involvement of key informant observations and meteorological data covers the blind spots of the respective category.

According to the EM-DAT global disasters database EM-DAT [78], the year 2010 experienced large-scale drought conditions in the coastal, northern-most, and north-eastern locations. Our analysis suggests that the 2010–2011 drought conditions had existed already since 2008 and continued until 2012, even though the year 2010 showed wetter conditions in some of the products, as also confirmed by [70]. The effects of the severe 2011 drought might have carried over to 2012, with SPEI showing no sign of relief, although the actual magnitude of SPEI is ambiguous in that year. The effects of this prolonged drought period were devastating among the households largely dependent on rainfed agriculture. Essential sectors such as energy, which is largely hydro-based, were negatively impacted across East Africa [2,6]. In Kenya, a total of 3.75 million persons, primarily in the north and parts of the south-east, were affected by the resulting food shortage according to the global record of mass disaster occurrence [78]. The drought period 2005–2006, confirmed by most products, was followed by wetter conditions in 2007, which exacerbated impacts. As [18,36,77] discuss, livelihoods and natural ecosystems across East Africa were severely impacted by the drought and, as Nicholson [70] reiterates, subsequent flash floods. The drought conditions seem to have commenced in 2004 and peaked in 2006, a classic demonstration of the evolving nature of the hazard [4,22].

A case of disagreement between the SPEI blends are the years 2014–2015, which were confirmed as drought years by actors engaged in drought relief in the area and the EM-DAT database. EM-DAT mentions the year 2014 with only a few areas in northern and north-eastern Kenya affected. In this light, south-eastern Kenya, including Kitui West, might have seen milder drought conditions. The National Drought Management Authority of the Kenyan government (NDMA) reports that in 2013–2014, during the OND, the greater Kitui region experienced moderate drought conditions and instances of decline in crop production and crop failure [79,80]. Triangulation of the SPEI calculations with qualitative information on the ground showed its greatest value here. The qualitative input effectively resolved the ambiguity between the data products. However, the qualitative data, too, have the potential for errors, including false recollections, difficulties in estimating the length of a drought and distinguishing trends and extremes, influences of recent events and media attention on past occurrences, and willfully biased responses with the aim to attract funding by exaggerating the severity of the drought situation [81]. On their own, the qualitative data lack information on drought magnitude and timing, which is something that the SPEI analysis can provide, albeit with uncertainty.

#### **5. Conclusions**

We revealed uncertainties related to the choice of rainfall and temperature products for the calculation of SPEI in the context of identifying past drought conditions in the semi-arid Kitui West area of Kitui County, south-east Kenya. We thereby complement existing studies with a demonstration of the variation of data products and the resulting SPEI calculations at the sub-national scale, which is relevant for assessing drought impacts on agriculture-based livelihoods. In an attempt to resolve the ambiguity in drought identification resulting from the differences in products, we assessed the value of complementing the SPEI analysis with key informant interviews, effectively demonstrating the added value of triangulation.

We observed that blends of 10 gridded rainfall and four gridded temperature products unanimously identified years experiencing drought conditions amidst a few variations. Moreover, 18 of the 40 SPEI combinations, revealed trends towards drier conditions, statistically significant at the 0.01 level. Using the ensemble of gridded meteorological data blends in the calculation of drought indices, the SPEI in this study, facilitated greater understanding of the uncertainties in onset, duration, and magnitude of past droughts. These uncertainties were driven more by the variation between rainfall products than temperature products in our case. Understanding past droughts is important to study their social-ecological impacts and assess the adequacy of responses in the future. Our study thus holds an important lesson for studies of past droughts: using any one of the available data products would risk severely misrepresenting drought characteristics and perhaps instituting erroneous responses. It is similarly important to bear in mind that, in the absence of a dense ground-station network, there is no benchmark dataset against which the individual data products can be assessed. Searching for a "best" product is thus not viable, and the value of these products can only be realized in an ensemble as we have revealed.

An ensemble approach to SPEI could not, however, identify all droughts unanimously in our case, using an ensemble of 10 rainfall products times four temperature products over the Kitui West area in south-east Kenya. This ambiguity could only be resolved with the information from 14 key informants engaged in disaster relief on the ground. Our study thus demonstrates the value of triangulating quantitative drought analysis with qualitative data. The qualitative data alone, in turn, would miss information on drought onset, duration, and magnitude; this is what the ensemble approach to SPEI provides, albeit with uncertainty. It is thus the juxtaposition of both types of data that is most fruitful.

Engaging organizations involved in disaster relief locally in drought identification will also strengthen their role in the region. Since drought is a gradually evolving phenomenon with long-lasting socio-economic impacts, there is need to develop and/or intensify integrated interventions and capacity building where affected communities are actively engaged at sub-national levels. The evolving and complex dry conditions accompanied by uncertainty are a challenge for the relatively recently devolved Kitui County administration, which has the mandate to coordinate multistakeholder risk management strategies at county-level. Such management strategies and collaborative networks should be flexible to detect, track, and respond effectively to various unique drought episodes. Effective responses include enhancement of government, private sector, and community-based disaster relief systems, targeting, for example, crop diversification with cultivation of drought resistant varieties as championed by the Kenya Red Cross [82]. An ensemble approach to SPEI will provide the necessary quantitative basis for these policies, while the experience of community, regional and national organizations will help resolve data ambiguities as well as strengthen the implementation of national policies.

Appreciating uncertainties in drought characteristics should in no way distract from decisive action to mitigate the impacts of droughts, improve disaster relief, and strengthen adaptive capacity, because extreme events such as droughts have been increasing over East Africa and have already resulted in deterioration of livelihoods and ecosystem integrity. While there is likely spatial variation over the region, we confirmed a statistically significant trend towards increasingly drier conditions also for Kitui West with just over half of the SPEI ensemble members. This trend was partly driven by a significant increase of minimum and maximum temperature over time in all data products, while negative annual and seasonal rainfall trends in some of the products could not be proven statistically significant. Beyond the temperature, and therefore evapotranspiration, effect, it will be worth investigating next how the timing and sub-annual variation of rainfall propagates into negative SPEI values, i.e., drier conditions. Such an analysis should go beyond trends in annual standard deviations of rainfall, which in our case did not turn out significantly either.

**Supplementary Materials:** The following are available online at https://www.mdpi.com/article/ 10.3390/w13243611/s1, Figure S1: Illustration of the areal weighting approach, Figure S2: Various grid resolutions of the data products used and their contribution to the areal average, Figures S3 and S4: Correlation matrices between the weighted gridded rainfall products and KMD Gridded data and the weighted gridded Min/Max temperature and the KMD Gridded data, Figure S5: Rainfall anomalies zoomed in the respective decades over the study period ; zoom1, 1987–1996, zoom2, 1997–2006, and zoom3, 2007–2016, Figure S6: Cumulative negative SPEI among the 40 combinations, Figures S7–S10: SPEI outputs using CRU, MERRA-2, JRA-55, and KMD\_grid Tmax/Tmin with the 10 rainfall products, with linear trend superimposed., Table S1: KMD-Grid, CRU, MERRA-2, and JRA-55 Tmax/Tmin Statistics, Table S2: Seasonal (MAM and OND) and annual precipitation statistics,. Equation (S1): Steps used in weighting of respective Gridded products., Breakdown S1: Guiding questions used in key informant interviews.

**Author Contributions:** Conceptualization, methodology, investigation and project administration P.B. and T.K.; Software, F.B.; Validation, P.B., F.B. and T.K.; Formal analysis, F.B.; Resources, P.B. and T.K.; Data curation, P.B. and F.B.; Writing-original draft preparation, P.B. and T.K.; Writing—review and editing, P.B., T.K. and P.R.; Visualization, F.B, T.K. and P.B.; Supervision, T.K. and P.R.; Funding acquisition, T.K. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research including the APC, was funded by Geo.X Research Network for Geosciences in Berlin and Potsdam, grant number SO\_087\_GeoX.

**Data Availability Statement:** The data and code used in the article can be accessed as detailed under, Table 1, the corresponding author can readily provide further clarification on a need basis.

**Acknowledgments:** We are grateful to the Kenya Meteorological Department for providing the requested gridded and ground station data. Further, we are equally thankful to the various gridded data providers as accessed in the respective interfaces and the input of Key informants.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

#### **References**

