Yield and yield-reduction in recent drought years in southern Tigray, northern Ethiopia: Implications on food security

The consequences of prolonged precipitation-decient periods are primarily substantial water decit, with the spatial characteristics of a place being (semi)arid, and various socioeconomic factors worsening its impacts and deepening poverty among agrarian communities.This study utilizes a combination of climate, remote sensing and eld survey data from farmers to obtain a rst-hand information on the impacts of recent (2015 and 2017) droughts on crop yield in Raya Azebo and Endamehoni woredas in Ethiopia. Annual rainfall, kiremt rainfall, annual NDVI and Dev-NDVI, kiremt season Dev-NDVI, monthly Dev-NDVI (for June to October), and monthly SPI-1, SPI-3 and SPI-12 (for June to October) were considered as likely factors that could relate with yield and yield loss in the area. Correlation and multiple linear stepwise regression statistical techniques were used to determine drought-yield relationships, and identify more accurate predictors of yield and yield losses in each of the drought years.Results obtained show droughts as having spatiotemporal variations and impacts, with its primary and common reection being reduced vegetation amount, translating to crop failures, food shortages and reduced income of smallholders.Spate irrigation should be further popularized in the low-lying areas of Raya Azebo to augment for future deciencies in the kiremt rainfall. of of The of of negative from the mean NDVI (Kidwell, 1990; The and


Introduction
Thus, in order to detect and estimate droughts and their impacts, some indicators and indexes have been developed over the years. The characteristics of a drought event -onset, intensity, spatial coverage and duration, can be expressed using a drought index. In the opinion of Sivakumar (2011), parameters such as rainfall, runoff, temperature and evapotranspiration should be included in a drought index to enhance its usefulness in describing a drought event. Drought indexes have been generally grouped into meteorological, soil moisture, hydrological, remote sensing and composite/modeled indicators, to correspond with their design and scope of operations. This study is delimited to metrological and remote sensing indicators.
The commonly used meteorological drought indicators include Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI). The SPI is recommended by the World Meteorological Organization WMO (2012) to be used by all National Meteorological and Hydrological Services around the world to characterize meteorological droughts.
Designed by McKee, Doesken, and Kieist (1993) to quantify the magnitude of dry and wet conditions in a desired location, the SPI is effective in drought detection in multiple time-scales, generally between 1 and 48 months. The WMO and GWP (2016) posit that SPI values for 3 months (SPI-3) or less, might be useful for basic drought monitoring; while values for 6 months (SPI-6) or less, are t for monitoring agricultural impacts; and values for 12 months  or longer, are ideal for hydrological impacts. With precipitation being the only input, SPI is de cient when accounting for the temperature component of droughts, as water balance/use are not accounted for. However, SPI is still adopted in this study in spite of the obvious de ciency due to the unavailability of temperature data.
A lot of satellite-based indicators have been developed, with ability to accurately detect droughts over large spatiotemporal resolutions. The Normalized Difference Vegetation Index (NDVI) is a popular drought-monitoring index developed by Kogan (1995), for determination of droughts by the measurement of vegetation vigor and cover in an index ranging from -1 to +1.
However, the NDVI itself needs to be computed vis-à-vis anomalies in order to clearly indicate the presence of drought and its severity. The Dev-NDVI, which is the deviation of NDVI from its long-term mean will easily indicate the magnitude of wetness or dryness of an area. Hence, a negative Dev-NDVI is an indicator of below-normal vegetation condition/health, thus suggesting a drought situation. The magnitude of a drought spell is de ned by the degree of negative departure from the long-term mean NDVI (Kidwell, 1990;Thenkabail et al., 2004). The NDVI and Dev-NDVI are therefore used in this study to obtain a clear perception of agricultural droughts.
Many drought studies have focused on assessment of both meteorological and agricultural droughts in Ethiopia. While Gebrehiwot et al. (2016) undertook a determination of spatiotemporal seasonal agricultural drought in Ethiopia during the 1998 to 2013 cropping seasons; Suryabhagavan (2017) characterized meteorological droughts and climate variability in Ethiopia over three decades (1983 to 2012). The study of Qu et al. (2019) on impacts of drought, includes other countries aside Ethiopia and covering the Horn of Africa, while that of Gidey et al. (2018) characterized the occurrence of drought in northern Ethiopia for 15 years (2001 to 2015). Only the study of Warner and Mann (2018) adopted an integrative approach of the use of satellite (GIS) data with conventional agricultural survey data to assess drought impacts, hence the essence of this study for the area. This study rather than characterize or report drought impacts by use of indicators, or other secondary data, further utilizes the climate and remote sensing data, with eld survey data from farmers to obtain rst-hand information on the impacts of droughts on crop yield, and then determines climatic and vegetation drought indicators which are able to closely predict farmers' yield in the recent drought events (of 2015 and 2017). Therefore, the objective and important contribution of this study to scienti c knowledge of drought is to provide an insight into what drought indicators (meteorological and vegetation/remote sensing) closely relate to and predicts farmers' crop yield, using a small geographical area, with opportunities of further discourse.

Materials And Methods
Page 4/22 The study area This study was implemented in two purposively selected woredas (districts) in the southern Tigray region, northern Ethiopia. Tigray has been previously established as one of the most drought-vulnerable regions in Ethiopia. Located between 12°38'44" -12°57'10" N Latitude and 39°27'18'' -39°55'56'' E Longitude, with an altitude ranging from 1109 to 3760 m above sea level (m.a.s.l.). The study area includes one district of mid and highland agroecology (Endamehoni) and another woreda (Raya Azebo) of lowland agroecology. Nine (9) tabias/kabeles (villages) are included in the portion selected for this study (Figure 1).

Meteorological data
The Climate Hazards group InfraRed Precipitation with Stations (CHIRPS) grid precipitation dataset from 2015 to 2018, obtained from (ftp://ftp.chg.ucsb.edu/pub/org/chg/products/CHIRPS-2.0/) was used for this study. The CHIRPS is the combination of a "high-resolution climatology, time-varying cold cloud duration precipitation estimates, and in situ precipitation estimates", with high research quality and very low errors Shukla et al., 2017). The high reliability, spatial spread and accuracy of the CHIRPS over the point-based data makes it preferred for this study. Also, Bayissa (2018) and Gebrechorkos et al. (2018) have earlier confirmed the accuracy of the CHIRPS data for use in studies on Ethiopia. The rainfall variables used in the study include the annual and main (kiremt) season (JJAS) rainfall.

Satellite images data
The Moderate Resolution Imaging Spectroradiometer (MODIS) MOD13Q1 composite images with 250-m resolution data obtained from the NASA Land Processes Distributed Active Archive Center (LP DAAC),

USGS/Earth
Resources Observation and Science (EROS) Center (https://lpdaac.usgs.gov/data _access/data_pool), was used for this study. In the opinion of (Wu et al., 2015) MODIS data products are accurate and appropriate for drought monitoring. Following download, the images were re-projected and resampled from a sinusoidal to geographic projection (WGS84); and then NDVI values were extracted from the re-projected (*.tif) images using R scripts. To obtain the conventional NDVI values ranging from -1 to +1, the extracted values were rescaled with the factor 0.0001. The NDVI variables used in this study includes the Annual NDVI, from which the Dev-NDVI of the main cropping (kiremt) season, and the monthly Dev-NDVI of June to October.
Farmers' yield (survey) data A one-time direct interview of 34 farmers was used for this study. These 34 farm owners (sorghum = 16; barley = 18) are randomly selected, and are those with high yield recall from memory. While more farmers were interviewed as they are always found in groups, only those who met our criteria of crop type and recent crop yield memory, were included the study, with their farmlands measured with the use of IPAQ, for appropriate relation of the yield to specific farmlands. Two cereal crops -barley and sorghum were selected in the highland and lowland respectively, due to their widespread importance as dominant crops and staple foods in the area, and also to improve the accuracy of yield recall from farmers' memory.
Page 5/22 Data analyses Data pre-processing procedures The analysis of historical meteorological droughts was done using the Standardized Precipitation Index (SPI) in R software version 3.4 with the aid of suitable scripts and packages. The SPI was calculated on three timescales: 1-, 3-and 12-month (SPI-1, SPI-3, and SPI-12). To obtain agricultural droughts Dev-NDVI was calculated using the equation (1).
Where the NDVI i is the NDVI value for month 'i'; NDVI mean-m is the long-term mean NDVI for the same month m over the spatiotemporal period.
Three operations were implemented to get three variables of yield data for the analyses.
1. Using the raw yield data obtained from farmers, conversion to yield per hectare was done by dividing the actual (raw) yield value by the area of land and multiplying the result by 10,000 (which is the area of a hectare). Hence the formula: Where: Actual yield = raw yield values obtained from farmer; Area(m 2 ) = total area of farmer's plot in m 2 ; 10,000 = total meter squares (m 2 ) in a hectare 1. The yield per hectare (YpHa) data was used to compute a Standardized variable of crop yield (SCY) for each of the sampled Tabias within the study area, using equation (3).
The SCY was then be used to obtain the crop yield loss ratio (YLR) of the area, using equation (4). The formulae applied to obtain the SCY and the YLR has been used in a recent study of Elhag & Zhang (2018) to monitor the impact of drought on crop yield elsewhere. Where: Yj is the crop yield in j Year of a Tabia; Ȳ is the average, and σ is the standard deviation of crop yield during the period under review.

Correlating yield, climatic and drought-indicator variables
To obtain the relationship between drought and crop yield in the study area, YpHa, SCY and YLR for 2015 and 2017 were correlated with annual rainfall; kiremt rainfall; annual NDVI; annual Dev-NDVI; kiremt Dev-NDVI; monthly Dev-NDVI (for June to October); and monthly SPI-1, SPI-3 and SPI-12 (for June to October). The use of all these variables in measuring the relationship is to determine which factor is most related to the declared farmers' yield and losses for the years under review.
The analysis of the relationship between variables listed above with YpHa, SCY and YLR was done in three batches. Firstly, the relationships across the entire study area was assessed. Next, a split woreda-based analysis was conducted to ascertain the relationships based on the two different woredas (Endamehoni and Raya Azebo) selected for the study. The purpose of the split-site analysis is to account for inherent variations expected in rainfall patterns, crop yield and crop loss in the two agroecologically dissimilar districts (woredas).  (Table 1). Regressing yield, climatic and drought-indicator variables Multiple linear stepwise regression was implemented in SPSS to obtain the most significant and relevant predictors for both crop yield and yield loss in the drought years of 2015 and 2017. The stepwise regression is an approach to selecting a subset of effects for a regression model and has been widely used in literature, with several authors preferring its application where little theory to guide the selection of terms for a model exists; or where there is the focus on predictors which interactively seem to provide a good fit; and where the performance of a model is to be improved by reducing the variance from unnecessary terms (Bachmair et al., 2018;Blauhut et al., 2016;Jones et al., 2011;Yahaya and Timothy, 2015;Zhu et al., 2018). The stepwise regression also exempts variables of high multicollinearity (variance inflation factor -VIF) in the selection of the most relevant predictors, hence its adoption in this study. The step-wise multiple regression was therefore conducted using the threshold of probability of the predictor to enter as <=0.050 and the probability of the predictor to be removed as >=0.100.

drought and crop yield: relationship and climatic/vegetative predictors
The results are presented in three subsections to represent the lowland case study (Raya Azebo), the highland case study (Endamehoni), and the entire study area. These sections correspond to the three-batch analyses implemented to unravel the relationship of drought with crop yield, and arriving at key drought indicators that more accurately predicted the crop yield in the event of drought. Figure 2 contains key spatial variables used to obtain the results.

Raya Azebo (lowland)
For the sampled area in Raya Azebo, only annual rainfall, values show a strong positive significant correlation with SCY and YpHa, while June and October SPI-1 indicate a strong negative association with SCY and YpHa (Supplementary Table S2). This implies that higher annual rainfall translates to higher crop yield while higher meteorological drought in June and September led to lower crop yield. Kiremt season rainfall also had a strong positive significant relationship with YpHa. Other notable variables indicating a moderate positive relationship with YpHa and SCY are annual Dev-NDVI, kiremt Dev-NDVI; while a moderate negative relationship exists between July and August SPI-1, June and July SPI-1; June and October SPI-3 (Supplementary Table S2). Annual rainfall and kiremt season rainfall had a strong negative relationship with YLR. Hence, the lower the (annual and kiremt) rainfall, the higher the yield loss. On the other hand, annual Dev-NDVI, June and October SPI-1 are strongly positively and significantly associated with YLR. Hence, higher deviation from NDVI and higher meteorological drought reflected in higher crop losses. Kiremt Dev-NDVI, June Dev-NDVI, July SPI-1, June and October SPI-3, and June SPI-12 were moderately positively correlated with YLR Annual rainfall and July SPI-3 are the most significant predictors of SCY in Raya Azebo. The two-step regression (Supplementary Table S7) results show that annual rainfall alone explains 38% of the variation in the standardized crop yield while and the combination of both predictors -annual rainfall and July SPI-3 for the year explain 56.3% of the variation in SCY for Raya Azebo (Table 2). Similar to the SCY, the YpHa is significantly predicted by annual rainfall and July SPI-3 with the coefficient of determination (r 2 ) indicating that 35.7% of the variations in the yield per hectare obtained in the study area is attributed to annual rainfall alone, while the combination of annual rainfall and July SPI-3 explains 53.3% of the variations (

Endamehoni (highland)
There are slightly different results obtained in the analyses of data in this area. There is a strong negative significant relationship between kiremt Dev-NDVI, June Dev-NDVI, July and September SPI-1 and SPI-3; and July September and October SPI-12, all with YpHa (Supplementary Table S3). The SCY had a strong negative significant relationship with kiremt Dev-NDVI, September SPI-1; July and September SPI-3; and July, September and October SPI-12. Additionally, it is worthy of mention that the correlation between kiremt season rainfall, August SPI-1, June, August and October SPI-3, and August SPI-12 with YpHa and SCY were moderately negative, though not significant. Therefore, higher wetness led to lower yield in this woreda. The crop yield loss was strongly and positively significantly related with Annual Dev-NDVI, June Dev-NDVI, August and September SPI-1; June to October SPI-3, and August to October SPI-12. Furthermore, kiremt season rainfall, July SPI-1 and July SPI-12 showed a moderate positive relationship with the YLR but not significant. Higher rainfall reflected in higher yield loss in this area and higher Dev-NDVI did not imply higher yield as crop losses were recorded due to excessive rainfall, as confirmed by data obtained during farmers' survey (Supplementary Table S3).
October SPI-12 is the most significant predictor of SCY in Endamehoni for 2015. The coefficient of determination (r 2 ) show that October SPI-12 explain 43.7% of the variation in standardized crop yield in the area for the year under review (  Table S1). Hence, higher SPI-12 drought in July and September led to higher crop losses, whereas October drought had no major impact on crop losses.
Kiremt season rainfall, June Dev-NDVI and annual NDVI are the key predictors for SCY in the entire study area. From the three-step regression, kiremt rainfall alone explains 39.8% of the variation in the standardized crop yield; Kiremt rainfall and June Dev-NDVI explain 50.4%; and the combination of all three predictors-kiremt rainfall, June Dev-NDVI and the annual NDVI for the year explain 57.5% of the variation in crop yield in the study area. The key predictors for YLR are kiremt rainfall and June Dev-NDVI. The two-step regression shows that kiremt rainfall alone explains 51.3% of the variation in yield loss ration; while the combination of kiremt rainfall and June Dev-NDVI explain 66.9% (Table 2). Some additional variables that are excluded from the stepwise threshold, though with significant predictive possibilities but high multicollinearity, include the September SPI-3 (p=.047) and October SPI-12 (p=.046). From a one-step regression analysis (Supplementary Table S7), annual rainfall is the most significant predictor of YpHa. The coefficient of determination (r 2 ) obtained shows that 40.3% of the variations in the yield per hectare obtained in the entire study area is attributed to annual Rainfall.  had better and improved yield in this season. As above, the results are presented in three subsections to represent the lowland case study (Raya Azebo), the highland case study (Endamehoni), and the entire study area. These sections also correspond to the three-batch analyses implemented to unravel the relationship of drought with crop yield, and arriving at key drought indicators that more accurately predicted the crop yield in the drought events.

Raya Azebo (lowland)
There is a strong positive significant relationship between annual NDVI with YpHa and SCY. A positive moderate not-significant relationship also exists between kiremt Dev-NDVI, YpHa and SCY (Supplementary  Table S5). June to September Dev-NDVI values show a moderate negative not-significant relationship with YpHa and SCY. Crop loss relates moderately with kiremt season rainfall, June and July Dev-NDVI, August SPI-1; July to September SPI-12 values. All these are not significant in their relationship. However, June SPI-3 has a strong negative significant relationship with YLR while October SPI-12 has a strong positive relationship with YLR, which is also significant (Supplementary Table S5).
The results of a seven-step multiple regression (Supplementary  (Table 3). The coefficient of determination (r 2 ) obtained shows that 37.1% of the variations in the yield loss ratio obtained in the study area in 2017 is attributed to the kiremt rainfall values (Table 3).

Endamehoni (highland)
Key results found for this area indicate that YpHa and SCY are moderately positively correlated with Annual and June Dev-NDVI (Supplementary Table S6). The SCY however shows a strong positive relationship with kiremt Dev-NDVI, August and September Dev-NDVI. The YLR had a strong negative significant relationship with kiremt Dev-NDVI, September SPI-3; July, September and October SPI-12. The variable also had a moderate negative correlation with annual and June Dev-NDVI; July, September and October SPI-1; October SPI-3 values.
The most relevant predictors of SCY in Endamehoni are September Dev-NDVI and the 2017 Kiremt Dev-NDVI in a two-step computation (Supplementary Table S8). The coefficient of determination (r 2 ) show that only the September Dev-NDVI as a predictor of SCY explain 29.5% of the variation, while the addition of the kiremt season Dev-NDVI for the year explains 46.9% variation in standardized crop yield in the area for the year under review (Table 3). Similarly, the September Dev-NDVI is the sole predictor of YpHa in the area with 45.2% of variations explained in the one-step regression implemented (Table 3). September SPI-3 is the only, most relevant and significant predictor of crop losses in this location for the year 2017. The coefficient of determination (r 2 ) shows that September SPI-3 explains 45.4% of the variation in yield loss ratio in the area (Table 3). A nearly-significant predictor is the June Dev-NDVI with (p=.052).

Entire study area
The results depict a moderate positive significant relationship between annual rainfall, kiremt season rainfall, annual NDVI and YpHa (Supplementary Table S4 Annual NDVI and July SPI-12 are the key predictors of SCY from the two-step multiple regression analysis implemented (Supplementary Table S8). Annual NDVI alone explains 19.6% of the variation in the standardized crop yield while the combination of both predictors -annual NDVI and the July SPI-12 values for the year explain 33.1% of the variation in SCY (Table 3). However, for the YpHa, July SPI-12 as the only and most relevant and significant predictor with the coefficient of determination (r 2 ) obtained shows that 24.4% of the variations in the yield per hectare obtained in the study area is attributed to the July SPI-12 values ( Table 3).
The most significant predictors of yield loss in the study area are kiremt rainfall and September SPI-3. The coefficient of determination (r 2 ) show that kiremt rainfall alone explains 19% of the variation in yield loss ratio; while the combination of kiremt rainfall and September SPI-3 explain 33% variations in yield loss within the entire study area (Table 3).

Discussion
The year 2015 is adjudged as the most recent drought year of high magnitude and one of the worst events over the past 30 years covering a large area and exacerbating food insecurity in Ethiopia WFP, 2016). These literature reports are supported by an evidence of low crop yield in a larger part of the study area ( Figure 2). In addition, the year 2017 is recorded as the most recent drought year in some parts of Tigray/Ethiopia with Gross Domestic Product (GDP) losses due to climate variability estimated to be around 1% to 4% excluding human losses (FDRE, 2017). These events were expected to have impacted on crop yield and yield losses in the study area.
From this study, drought has again been proven to be location-speci c, and not a broad spatio-temporal phenomenon at any given time. In an earlier study, Viste et al. (2013) describe most historic droughts as being more of 'local or regional character' with dissimilar effects at different places and seasons. Each drought year reviewed (2015 and 2017) stands out uniquely in its nature and characteristics. The only common denominator is precipitation-de cit, which in itself may be of different magnitude per location. Overall, the seasonal rainfall benchmark of 500mm as stipulated by Haile et al. (2019) for a rough drought indicator in a season helps to present an outlook of the drought. In 2015 the entire study area fell below the 500mm mark, while in 2017, only one tabia/kabele (village) did. Hence the 2015 drought was more widespread than that of 2017, with majority of farmers in the lowland reporting entire crop losses. The predictors of yield and yield losses indicate that droughts coinciding with planting and maturing stages of crops are critical and inimical to crop production and agricultural sustainability and food security.
In both years, and for the entire study area, both annual and the kiremt seasonal rainfall were related to the crop yield reported by farmers in these years. Crop yield is largely a function of adequate soil moisture at all stages of crops. Hence, the shortage of rainfall (occurrence of drought) brought about reduced crop yield in the area. These assertions are corroborated by the ndings of Gidey, Dikinya, Sebego, Segosebe, and Zenebe (2018), who report that agricultural drought responds positively to seasonal rainfall. Also, part of the ndings of Molla and Fitsume (2017) Legesse and Suryabhagavan (2014); and Legesse (2010) indicate that rainfall signi cantly determines the occurrence of drought, hence manifesting high in uence on the growth and development of vegetation.
Different timescales of SPI, especially the 3-and 12-month scales were recurring predictors of crop yield in the area. In relating meteorological drought index for drought monitoring, several studies have found a signi cant relationship between SPI and crop yield (Elagib and Elhag, 2011;Li et al., 2014;Melese et al., 2018). Also, the vegetation anomaly (Dev-NDVI) is able to detect agricultural drought, and predict crop yield in the area. These rainfall and vegetation based indices used in this study considered vis-à-vis speci c crop timelines are helpful to clarify the times drought were more critical. For example, the SPI-3 for July 2015 predicted sorghum yield for the lowland portion of the study area, because sorghum crops are at their early stages of growth at this time. In the mid and highland areas, the Dev-NDVI for October 2015 predicted yield and yield loss as the late maturity period for barley includes October. Hence, the satellites detected low vegetation vigor, indicating drought at the time. For the Dev-NDVI to detect agricultural drought in October, it also implies that preceding moisture de cit has occurred.
The reduced impact of drought on crop yield in 2017, especially in the lowland area is attributable to the increased rainfall received. Most of the tabias received rainfall higher than 500mm, which supports healthy agriculture. Additionally, the report of

Conclusion
This study critically assessed the recent drought years in northern Ethiopia, with a focus on a small area containing high, mid and lowland agroecologies. Results obtained show droughts as having spatiotemporal variations and impacts, with its primary and common re ection being reduced vegetation amount. This translates to crop failures, reduced income of smallholders, food shortages and food insecurity. SPI, NDVI and Dev-NDVI were found quite useful as indicators of drought, and considered able to be applied in the preparation of early warning for droughts when they fall below a certain established threshold for the area in each cropping season, especially at critical crop growth and maturity stages. Findings of this study can provide information for further discussions in improving drought monitoring at lower locational scales. Spate irrigation should be further popularized in the low-lying areas of Raya Azebo as a means to augmenting for de ciencies in the kiremt rainfall. This will provide adequate water supply to meet the crops' water requirements especially during the growing and maturity periods, during future drought years, thus improving chances of future food security even in the event of future droughts. There are more factors responsible for crop losses not captured in this study, hence providing an aspect for future further research.

Limitations and further research
Page 14/22 The ndings of this study would have been more generalizable if the number of farmers that constituted the sample size were more in the number. The exclusion of temperature as a variable in this study is an obvious limitation, which would have provided improved outcome for drought assessment, especially with the use of SPEI, instead of SPI. Future studies can address these de ciencies, in addition to more detailed analysis of speci c crops' growth stages with climatic variables, and the impact of the use of spate irrigation systems, for more robust outcomes. Spate irrigation was not considered a priori as a factor that could affect the impacts of droughts in the area, hence further inquiry was not considered in this regard.

Declarations
Data availability The datasets used for this study are available and can be provided by the corresponding author on request.

Declaration of competing interests
The authors hereby declare no competing interests.

Funding
No external funding was received for this study.
Author's contributions All authors contributed, read and approved the nal manuscript.