Drought impact in the Bolivian Altiplano agriculture associated with the El Niño–Southern Oscillation using satellite imagery data

Drought is a major natural hazard in the Bolivian Altiplano that causes large agricultural losses. However, the drought effect on agriculture varies largely on a local scale due to diverse factors such as climatological and hydrological conditions, sensitivity of crop yield to water stress, and crop phenological stage among others. To improve the knowledge of drought impact on agriculture, this study aims to classify drought severity using vegetation and land surface temperature data, analyse the relationship between drought and climate anomalies, and examine the spatio-temporal variability of drought using vegetation and climate data. Empirical data for drought assessment purposes in this area are scarce and spatially unevenly distributed. Due to these limitations we used vegetation, land surface temperature (LST), precipitation derived from satellite imagery, and gridded air temperature data products. Initially, we tested the performance of satellite precipitation and gridded air temperature data on a local level. Then, the normalized difference vegetation index (NDVI) and LST were used to classify drought events associated with past El Niño–Southern Oscillation (ENSO) phases. It was found that the most severe drought events generally occur during a positive ENSO phase (El Niño years). In addition, we found that a decrease in vegetation is mainly driven by low precipitation and high temperature, and we identified areas where agricultural losses will be most pronounced under such conditions. The results show that droughts can be monitored using satellite imagery data when ground data are scarce or of poor data quality. The results can be especially beneficial for emergency response operations and for enabling a proactive approach to disaster risk management against droughts.

Abstract. Drought is a major natural hazard in the Bolivian Altiplano that causes large agricultural losses. However, the drought effect on agriculture varies largely on a local scale due to diverse factors such as climatological and hydrological conditions, sensitivity of crop yield to water stress, and crop phenological stage among others. To improve the knowledge of drought impact on agriculture, this study aims to classify drought severity using vegetation and land surface temperature data, analyse the relationship between drought and climate anomalies, and examine the spatio-temporal variability of drought using vegetation and climate data. Empirical data for drought assessment purposes in this area are scarce and spatially unevenly distributed. Due to these limitations we used vegetation, land surface temperature (LST), precipitation derived from satellite imagery, and gridded air temperature data products. Initially, we tested the performance of satellite precipitation and gridded air temperature data on a local level. Then, the normalized difference vegetation index (NDVI) and LST were used to classify drought events associated with past El Niño-Southern Oscillation (ENSO) phases. It was found that the most severe drought events generally occur during a positive ENSO phase (El Niño years). In addition, we found that a decrease in vegetation is mainly driven by low precipitation and high temperature, and we identified areas where agricultural losses will be most pronounced under such conditions. The results show that droughts can be monitored using satellite imagery data when ground data are scarce or of poor data quality. The results can be especially beneficial for emergency response operations and for enabling a proactive approach to disaster risk management against droughts.

Introduction
Agricultural production is highly sensitive to weather extremes, including droughts and heat waves. Losses due to such hazard events pose a significant challenge to farmers as well as governments worldwide (UNISDR, 2009(UNISDR, , 2015. Worryingly, the scientific community predicts an amplification of these negative impacts due to future climate change (IPCC, 2013). Especially in developing countries such as Bolivia, drought is a major natural hazard, and Bolivia has experienced large socio-economic losses in the past due to such events (UNDP, 2011;Garcia and Alavi, 2018). However, the impacts vary on a seasonal and annual timescale, in regards to the hazard intensity, as well as the existing capacity to prevent and respond to droughts (UNISDR, 2009(UNISDR, , 2015. Regarding the former, the El Niño-Southern Oscillation (ENSO) plays an especially important role in several regions of the world, including the Bolivian Altiplano, as it drives drought occurrence that could cause losses of agricultural crops and increase food insecurity (Kogan and Guo, 2017). The most important rainfed crops in the region include quinoa and potato (Garcia et al., 2007). Generally speaking, agricultural productivity in the Bolivian Altiplano is low due to adverse weather and poor soil conditions (Garcia et al., 2003). On the other hand, low agricultural production levels can also be associated with ENSO (Buxton et al., 2013).
The ENSO is a climate phenomenon that affects the precipitation variability of the Bolivian Altiplano (Thompson et al., 1984;Aceituno, 1988;Vuille, 1999). The ENSO is defined as a periodical variation of the sea surface temperature over the tropical Pacific Ocean, and it represents neutral, warm (El Niño), and cold (La Niña) phases. The positive phase of ENSO is generally associated with warmer and dryer conditions, while the negative phase is associated with cooling and wetter conditions (Garreaud and Aceituno, 2001;Garreaud et al., 2003;Thibeault et al., 2012). Thus, droughts are generally driven by the positive ENSO phase in the study area (Thompson et al., 1984;Garreaud and Aceituno, 2001;Vicente-Serrano et al., 2015). Previous research has addressed the influence of ENSO on agriculture in South America and the globe (see Iizumi et al., 2014;Ramirez-Rodrigues et al., 2014;Anderson et al., 2017). These studies were calling for a better understanding of the association between ENSO and agriculture to improve crop management practices and food security. However, predicting the ENSO effects is challenging, since the ENSO evolution depends not only on the tropic Pacific Ocean temperature, but also on atmospheric convection, climate variability, and teleconnection with other climate anomalies (Santoso et al., 2019).
The implementation of drought risk management approaches is now seen as fundamental (see e.g. the Sustainable Development Goals or the Sendai Framework for Risk Reduction) for sustainable development in vulnerable regions, including Latin American countries such as Bolivia (Verbist et al., 2016). To lessen the long-term impacts of these extreme events, the national government in Bolivia has taken several steps, e.g. to allocate budgets for emergency operations to compensate for part of the losses incurred. Most of these measures are implemented ex post (i.e. after a disaster event). However, based on ENSO forecasting, an El Niño event can be predicted 1 to 7 months ahead (Tippett et al., 2012), and consequently there is an opportunity to implement additional ex ante policies (i.e. before the event) to reduce societal impacts of droughts, increase preparedness, and generally improve current risk management strategies.
We embed our research within the IPCC framework (IPCC, 2012) and conceptually define disaster risk as a function of the hazard, exposure, and vulnerability. Here, drought risk was defined as the likelihood of severe alterations in the normal functioning due to drought hazard interacting with the vulnerabilities of the exposed socio-environmental system, leading to potential adverse effects. Furthermore, disaster risk usually comprises different types of potential losses, which are sometimes very difficult to quantify (UNISDR, 2009). In the case of drought, usually four different types are distinguished (UNISDR, 2009): meteorological, hydro-logical, agricultural, and socio-economic droughts. In more detail, a meteorological drought manifests if certain climate variables (e.g. precipitation) remain under predefined threshold levels over a certain time period, while hydrological drought is usually determined through reduced water levels in water bodies and groundwater. Agricultural drought occurs when insufficient soil moisture and precipitation negatively affect crop yields, while agricultural drought may turn into socio-economic drought if the supply or demand of agricultural products is negatively affected (see Wilhite and Glantz, 1985).
Using these concepts and definitions our research aims to (1) classify agricultural drought severity by applying the normalized difference vegetation index (NDVI; as a proxy for crop yields), land surface temperature data, and climate data (as the hazard component); (2) analyse the relationship between drought and ENSO; and (3) assess drought through examining the spatio-temporal variability of vegetation and its association with climate data (implicitly including the vulnerability component through the spatio-temporal variability). One major constraint for drought risk management in Bolivia is the scarce and uneven distribution of weather-and agricultural-production-related ground data. To circumvent this problem, we test satellite-based and gridded data products (compared to available empirically gauged data) to provide a full coverage (in respect to land area) for drought assessment and its spatial distribution across the region. Due to the particular importance of ENSO for drought risk management, we additionally assess the impacts associated with ENSO on agriculture for the Bolivian Altiplano. Furthermore, we give indications regarding which climate variables may be most important in which regions to predict drought losses that can further be used for hotspot selection. The paper is organized as follows, Sect. 2 provides a discussion of the data used, Sect. 3 present the methodology applied, Sect. 4 presents the corresponding results found, and Sect. 5 ends with a conclusion and outlook for the future.
2 Data Used

Climate data
Our methodology is very much related to the data-scarce situation for the Bolivian Altiplano, and we therefore start with an introduction of the available datasets that were used for our purpose. In regards to the climate dimension, the Altiplano has a pronounced southwest-northeast precipitation gradient (200-900 mm yr −1 ) during the wet season occurring from November to March (Garreaud et al., 2003). Over 70 % of total precipitation occurs during the summer months (from December to February; see Fig. 1a) in association with the South American Monsoon (see Zhou and Lau, 1998;Garreaud et al., 2003). Time series of monthly precipitation at 12 locations as well as mean, maximum, and minimum tem-perature at 8 locations from September 1981 to August 2015 were available from the National Service of Meteorology and Hydrology (SENAMHI) of Bolivia (see Appendix Table A1). These datasets had less than 10 % of missing data and therefore served well for our analysis.
As already indicated, precipitation and temperature gauge locations are unevenly distributed and mainly concentrated in the northern Bolivian Altiplano. To improve the spatial coverage of climate-related data, monthly quasi-rainfall time series from satellite data the Climate Hazards Group In-fraRed Precipitation with Station data (CHIRPS) were included in our study. CHIRPS represents a 0.05 • spatial resolution satellite imagery and a quasi-global rainfall dataset from 1981 to the near present . The advantage of using CHIRPS is the high spatial resolution of data, obtained with resampling of TMPA 3B42 (with a 0.25 • grid cell). The spatial resolution represents a better option for agricultural studies as well and therefore is most appropriate for our approach (CHIRPS is described in detail at https://www.chc.ucsb.edu/data/chirps, last access: 22 February 2021).
Additionally, a gridded dataset of monthly mean air temperature was obtained from the Physical Sciences Division (PSD) of the US National Oceanic and Atmospheric Administration (NOAA, https://www.esrl.noaa.gov/psd/, last access: 1 May 2020) defined by Willmott and Matsuura, using a spatial interpolation of composite stations records from the Global Historical Climatology Network (GHCN version 2) and Legates and Willmott (Legates and Willmott, 1990a;Legates and Willmott, 1990b). The gridded air temperature dataset has a resolution of 0.5 • by 0.5 • and was available during the study period from September 1981 to August 2015. This dataset incorporates station-height information through an average air temperature lapse rate (Willmott and Matsuura, 1995). Here, a digital elevation model was used for the interpolation to adjust air temperatures in relation to sea level.

Land surface data: vegetation and temperature
Apart from climate datasets, NDVI was assembled from the Advanced Very High Resolution Radiometer (AVHRR) sensors by the Global Inventory Monitoring and Modelling System (GIMMS) at semi-monthly (15 d) time steps with a spatial resolution of 0.08 • . NDVI 3g.v1 (third generation GIMMS NDVI from AVHRR sensors) was available from September 1981 to August 2015. The NDVI is an index that presents a range of values from 0 to 1, and bare soil values are closer to 0, while dense vegetation is close to 1 (Holben, 1986). NDVI 3g.v1 GIMMS provides information to differentiate valid values from possible errors due to snow, cloud, and interpolation. These errors were removed from the dataset and replaced with the nearest-neighbour value.
Additionally, the monthly land surface temperature (LST) was obtained from the Global Land Data Assimilation Sys-tem (GLDAS) by the Noah Land Surface Model L4 monthly version 2.0 (Beaudoing and Rodell, 2019). The dataset was available for the study period September 1981 to August 2015. The LST estimations from GLDAS were based on remotely sensed observations of AVHRR (Rodell et al., 2004) and include an algorithm that relies on an optimal interpolation routine (Ottlé and Vidal-Madjar, 1992) to assimilate the LST onto a 0.25 • by 0.25 • grid. This data record was selected due to its temporal resolution; however, it is important to mention that a higher spatial resolution could improve the accuracy of agricultural analyses and further reduce the uncertainties of the data noise originating from land heterogeneity.
Finally, the study was conducted for the agricultural land in the Bolivian Altiplano (∼ 200 000 km 2 ). The agricultural land was spatially identified based on the land use map developed by the Autonomous Authority of the Lake Titicaca (for the northern Altiplano) in 1995 at a scale of 1 : 250 000 (UNEP, 1996) and the Ministry of Development Planning in 2002 using Landsat imagery and ground information at a scale 1 : 1 000 000 (http://geo.gob.bo/portal/, last access: 15 May 2019, for the southern Altiplano).

Methodology
The analysis of drought impact on agriculture for the Bolivian Altiplano and its relationship with the ENSO is based on the following three steps. Firstly, an evaluation of satellite precipitation and gridded air temperature against gauged datasets was performed to investigate the accuracy of these estimates compared to empirical on-the-ground data. Secondly, the severity of drought was classified using vegetation and land surface temperature data, and using this information drought events were associated with the ENSO variability. Finally, a stepwise regression approach was used to study the variability of vegetation and its relationship with corresponding climate variables. The overall aim of our study is to investigate drought effects on agriculture through the analysis of land surface and climate variations and their relation to the ENSO anomalies.

Evaluation of climate data
The performance of the satellite-based data (compared to empirical ground data; see Fig. 2) in accurately estimating the amount of rainfall (for example rain detection capability purposes) was based on statistical measures for monthly pairwise time series, including categorical analyses, and follows methodologies suggested and applied in previous studies in this region, which were selected for comparison reasons (Blacutt et al., 2015;Satgé et al., 2016). The mean error (ME), bias, and mean absolute error (MAE) were calculated based on Wilks (2006). These measures evaluate the prediction accuracy of the satellite data compared to gauged data. The ME and bias show the degree of over-or underestimation (Duan et al., 2015). In contrast, as with measuring the absolute deviation, MAE shows only non-negative values. The ME, bias, and MAE perfect match corresponds to zero between gauge observation and satellite-based estimate. Furthermore, and similar to Blacutt et al. (2015) and Satgé et al. (2016), the Spearman rank correlation was computed to estimate the goodness of fit to observations. To evaluate results, and in accordance with similar studies, correlation coefficients larger or equal to 0.7 were considered reliable (Condom et al., 2011;Satgé et al., 2016). The ME, bias, and MAE were calculated according to Eqs. (1), (2), and (3) in Table 1, respectively.
Two statistical indicators based on contingency tables were computed for the categorical statistics, namely probability of detection (POD) and false alarm ratio (FAR). The POD indicates what fraction of the observed events was correctly estimated, and FAR indicates the fraction of the predicted events that did not occur (Bartholmes et al., 2009;Ochoa et al., 2014;Satgé et al., 2016). The POD and FAR range from 0 to1, where 1 is a perfect score for POD, and 0 is a perfect score for FAR. These measures were used to evaluate the satellite precipitation estimations. Here, the rainfall amounts are considered as binary values, i.e. rain occurrence or absence. Based on this approach, three counting variables were taken into account: the number of events when the satellite rain estimation and the rain gauge report a rain event (hit or H), when only the satellite reports a rain event but no rain on the ground is observed (false alarm or F), and when only the rain gauge reports a rain event but not the satellite and therefore is a miss (M). The POD and FAR were calculated according to Eqs. (4) and (5) in Table 1, respectively.
Besides the precipitation data, the gridded air temperature data were evaluated using ground data. The gridded air temperature was correlated with the mean gauged temperature at the same spatial location. The mean temperature of the gauged data was calculated using the arithmetic mean between the maximum and minimum temperature. The regression performance was evaluated using the monthly pairwise time series to define the Spearman rank correlation, relative ME, bias, and MAE. For air temperature, the MAE was defined using absolute values.

Drought associated with ENSO
Healthy vegetation usually shows enlarged near infrared and reduced visible bands and emits less absorbed thermal infrared radiation, resulting in lower surface temperature (Kogan and Guo, 2017). Therefore, vegetation indices and land surface temperature (LST) are widely used for water and energy balance approaches (see Moran et al., 1994;Corbari et al., 2010;Sánchez et al., 2012;Helman et al., 2015). Previous findings indicate a negative (positive) relationship between LST and NDVI caused by limited moisture (energytemperature) availability for vegetation growth (Karnieli et al., 2010). Drought spells typically present low NDVI and high LST due to vegetation deterioration and higher contribution of the soil signal (Kogan, 2000). Here, we study the relationship between LST and NDVI using the vegetation health index (VHI, Eq. 8) developed by Kogan (1995) that combines the vegetation condition index (VCI, Eq. 6) and temperature condition index (TCI, Eq. 7). VCI compares the current NDVI with the range of NDVI observations during the study period that allows us to seek the variability of the signal, showing an increased VCI when NDVI increases. (Kogan, 1995;Kogan, 2000;Kogan and Guo, 2017). In contrast, the TCI formulates a reverse ratio compared to the VCI, decreasing when LST increases, assuming that higher land surface temperatures suggest a decreasing soil moisture causing stress of the vegetation canopy.
The VCI, TCI, and VHI was defined for each month during the growing season (from September to April). We as- Table 1. Accuracy measures for the performance evaluation of climate estimated variables. Here, N is the number of samples, S i is the climate estimation for month i, and G i is the gauged dataset for the same month. H is a hit, F is a false alarm, and M is a miss.

Statistical indicator Abbreviation Units Equation
Mean error ME mm, sumed the occurrence of a drought event when the indices were lower than 40 %. The classification of drought was established based on the severity of the event in which five classes were defined: extreme (≤ 10), severe, (≤ 20), moderate (≤ 30), mild (≤ 40), and no (> 40) drought (Bhuiyan and Kogan, 2010). The drought events were further classified based on the occurrence of El Niño and La Niña events (Table 3). The classification ENSO was obtained from Null (2018). El Niño and La Niña events were identified from five consecutive overlapping 3-month mean sea surface temperatures for the Niño 3.4 region (in the tropical Pacific Ocean). A moderate El Niño (La Niña) was defined as five consecutive overlapping 3-month periods at or above the +1.0 to +1.4 • C anomaly (−1.0 to −1.4 • C), a strong El Niño (La Niña) event for a threshold between the +1.5 and +1.9 • C anomaly (−1.5 to −1.9 • C anomaly), and a very strong El Niño event for a threshold equal to or greater than the +2 • C anomaly (https: //ggweather.com/enso/oni.htm, last access: 15 May 2020).
For this study, a neutral or weak phase was defined as a threshold between the −0.9 and +0.9 • C anomaly.

Regression of vegetation and climate variables
A stepwise regression approach was used to quantify the dependency between vegetation and climate variables (satellitebased precipitation and gridded air temperature; Eq. 10) to be used for the hotspot selection process. In more detail, the results presented here are a combination of forward and backward selection techniques to increase the robustness of the results (in terms of explanatory power, i.e. variability explained, as well as variable selection, i.e. same variable selected across a range of possible models). The independent variable considered was NDVI, and the dependent variables were selected to include precipitation and air temperature (for the same spatial location across the study region). We assumed that NDVI represents the crop phenological stages of the growing season that is from September to April (Fig. 1). Precipitation was selected as a predictor due to its relevance Here, NDVI i (LST i ) is the monthly NDVI (LST) and NDVI max and NDVI min (LST max and LST min ) are its multi-year absolute maximum and minimum , respectively. We took a mean of VCI and TCI assuming that they equally contribute to the VHI.  1986-1987 1987-1988 1982-1983 1995-1996 1988-1989 1994-1995 1991-1992 1997-1998 2011-2012 1998-1999 2002-2003 2015-2016 1999-2000 2009-2010 2007-2008 2010-2011 for water availability for vegetation growth. Precipitation is the main source of water in the Altiplano because only 9 % of the Bolivian cropped surface area is irrigated (INE, 2015). Air temperature is a relevant variable due to photosynthetic and respiration processes (Karnieli et al., 2010). Firstly, the NDVI was related to CHIRPS rainfall datasets. Secondly, air temperature was included in the analysis. Here, only the NDVI grids for agricultural land were selected. Since, agricultural production data are scarce in the region, we suggest that crop yield data can be improved using the NDVI. Besides improving the crop yield resolution, the NDVI also allows us to analyse the variability of vegetation at a monthly timescale. This makes it possible to analyse the phenology of the studied crops through to the growth phases. NDVI estimates the vegetation vigour (Ji and Peters, 2003) and crop phenology (Beck et al., 2006). The final regression model for each spatial unit was defined as NDVI = β 0 + β 1 precipitation + β 2 air temperature.
For the forward selection, the variables were entered into the model one at a time in an order determined by the strength of their correlation with the criterion variable (only including variables if they present a confidence level of 95 %). The effect of adding each variable was assessed during its entering stage, and variables that did not significantly add to the fit of the model were excluded (Kutner et al., 2004). For backward selection, all predictor variables were entered into the model first. The weakest predictor variable was then removed and the regression fit re-calculated. If this significantly weakened the model, then the predictor variable was re-entered; otherwise it was deleted. This procedure was repeated until only useful predictor variables (in a statistical sense, e.g. significant as well as model fit) remained in the model (Rencher, 1995). The results were compared with results from the literature regarding phenology and weather-related characteristics of crops.
It should be noted that the precipitation in the Altiplano shows a marked rainy season from November to March. The peak of precipitation is in December and January (Fig. 1a). Additionally, NDVI displays a peak in March and April (Fig. 1b). The lag between the precipitation and NDVI is reasonable since vegetation requires time to grow (e.g. Shinoda, 1995;Cui and Shi, 2010;Chuai et al., 2013). Considering this lag time, the 3-month time series of NDVI was regressed with the 3-month time series of the climate variables (satellite-based data product of precipitation and gridded data of air temperature) during the growing period for the agricultural land. First, the NDVI and the climate variables were related considering the overlapping 3-month time series, and afterwards a relation was developed considering a lag from 1 to 4 months between NDVI and climate variables, resulting in 22 regressions per NDVI grid. The regressions were developed for each NDVI grid separately, associated with the nearest precipitation and air temperature dataset. Prior to the stepwise regression analysis, the 3-month time series of NDVI, satellite precipitation, and gridded air temperature data were standardized.

Results
Validation of the satellite rain data using empirical precipitation data from the weather stations was done for the 12 locations where gauge precipitation data were available (see Fig. 2 and Table A1). The qualitative methods discussed in Sect. 3.1 for the CHIRPS rainfall estimates show differences between summer (from December to March) and winter sea-son (from June to August). CHIRPS data show better accuracy during summer. The precipitation during the austral summer is highly relevant because it concentrates the 70 % of the annual rainfall (Garreaud et al., 2003), and it occurs during the growing season. During May, CHIRPS data show lower accuracy compared to the other months. The precipitation from May to August is almost null in the study area (Fig. 1), and it will be further described as the dry season. This season presents stable atmospheric conditions with few precipitation events (Garreaud et al., 2003).
Interestingly, the Spearman rank correlation between monthly gauged precipitation and satellite rain product datasets was significant (p value < 0.05) for all locations. The correlation coefficients (r) vary from 0.5 to 0.8 (mean = 0.7). The ME and bias disclose an underestimation of precipitation estimation during October, November, and April and an overestimation during the summer season (mean = 5 mm and 7 %, respectively) with a peak in February. For the MAE coefficient, CHIRPS estimations are more accurate during the rainy season (mean = 31 %). In contrast, CHIRPS data indicate poor accuracy during the dry season (mean MAE = 92 %). From June to August, CHIRPS data present an underestimation of the gauged precipitation (mean bias = −39 %). Summarizing these observations, we conclude that the CHIRPS rainfall dataset is more accurate during the rainy season, and it represents an adequate alternative in case of lack of gauged data or in case of poor data quality. However, it should be noted that such data still must be used with caution considering the uncertainties due to the under-or overestimation of precipitation along the heterogeneous topography of the Altiplano (see Paredes-Trejo et al., 2016;Paredes-Trejo et al., 2017;Rivera et al., 2018).
Moving from rainfall to temperature, the inter-annual temperature at the eight locations varied considerably between summer (from December to March) and winter (from June to August), including a larger variance for the minimum temperature (Fig. 1a). The mean monthly air temperature from gridded data was compared with mean temperature of gauged data. The gridded air temperature underestimated the mean gauged temperature, and this error could be due to the heterogeneous topography and high elevation. The Spearman correlation at the eight stations displayed coefficients from 0.1 to 0.7. From November to April, the gridded air temperature data show significant correlations (p value < 0.05). Large correlations are shown during summer season (mean = 0.7), while the other months show rather weak correlations. ME and bias show a slight underestimation from October to April (mean = −0.5 • C and −4 % respectively) and an overestimation from May to August (mean = 0.3 % and 12 % respectively). Finally, MAE is about 1.2 • C from September to April, and higher values develop during winter season (mean = 1.6 • C). In conclusion, the gridded air temperature data product performs better from November to April. Similar to the precipitation data, the application of gridded air temperature data must take into account the potential errors due to the estimation uncertainties, mainly during winter season.
As discussed above, the VCI, TCI, and VHI were calculated during the growing season. The sowing period depends on the initial soil moisture content; therefore the beginning of the growing season oscillates from September to November (Garcia et al., 2015). For this reason, the drought severity was classified considering the mean of VCI, TCI, and VHI for the agricultural land during November-April. Figure A1 shows mean monthly VCI from November 1981 to April 2015. The major drought events (severe or extreme) are visible in 1982-1983, 1983-1984, and 2009-2010, followed by moderate drought events during 1987-1988 and 1993-1994, as well as several mild events. Figure A2 shows the mean monthly TCI, where the major drought events (severe or extreme) occurred in 1982-1983, 1987-1988, 1997-1998, 2004-2005, and 2009-2010, followed by moderate drought events during 1981-1982, 1983-1984, 1994-1995, 2006-2007, and 2008-2009, and several mild events as well. Finally, Fig. A3 shows the VHI results, in which the major drought events occurred during 1982-1983, 2004-2005, and 2009-2010. In a next step we related drought indices with the ENSO phases (Table 4). Extreme and severe droughts were generally found during the El Niño phase. The extreme drought of 1982-1983 coincided with a very strong El Niño phase. For this event, the largest economic losses caused by droughts during the study period are reported (Table 5), followed by the very strong El Niño phase of 1997-1998, which reported the second largest economic losses. Besides these two main drought events, the strong El Niño during 1987-1988 coincided with an extreme/moderate drought (TCI ≤ 10 %, VCI ≤ 30 %) classification. During this period, large economic losses were reported as well (Table 5). In contrast, the strong El Niño during 1991-1992 showed low severity (mild drought VCI ≤ 40 %), and no economic losses were reported. This indicates that despite the fact that the El Niño phenomenon is generally associated with drought in the Altiplano, there are several other mechanisms that drive a drought occurrence and determine its severity. For instance, dry (wet) and warm (cool) conditions during El Niño (La Niña) phases are generally shown in the Altiplano (Garreaud et al., 2003). However, an anomalous location and intensity of zonal wind anomalies could cause disturbances of the warming and cooling air patterns causing rainfall anomalies in the region (Garreaud and Aceituno, 2001). This is the case of the dry La Niña during 1988-1989 that showed a mild drought classification (TCI ≤ 40 %).
One severe (1983)(1984)    in November and December. From January to April of 2004-2005 the VCI and VHI were above 40 %, and there were no claims of drought losses in the Altiplano for this particular year (Table 5). Besides these two events, moderate and mild droughts also occurred during non-El Niño phases. Table 5 shows that five drought events were reported during a neutral ENSO phase. In 2012-2013, the largest impact occurred, affecting about 80 000 people in the Altiplano (Desinventar Sendai, 2020). Despite the fact that the mean of the drought indices indicates no drought during this period (VCI, TCI, and VHI > 40 %), some spatial locations in the study region indicated the occurrence of a drought event in November and December (21 % and 29 % of the total studied grids showed mild and moderate droughts for the TCI and VCI respectively).
The results of the stepwise regression show a larger coefficient of determination (R 2 ) in the northern and central Bolivian Altiplano, starting from the southern Lake Titicaca and moving southwards to Lake Poopó, and close to the rivers' paths. Lower R 2 values are shown along the southwestern Bolivian Altiplano that could be explained through the large variance of the NDVI, which may depend to on other factors besides precipitation and temperature, including crop management. Figure 5 shows the R 2 of the best fit regression in the Bolivian Altiplano for the 3-month period of NDVI and the climate variables (precipitation and temperature) during the beginning and end of the growing season. It can be seen that the NDVI depends largely on the studied climate variables. This may be due to the crop's sensitivity for water stress during specific stages of the growing season. For instance the most sensitive stages of the quinoa crop are the emergence, flowering, and grain development (see Geerts et al., 2008;Geerts et al., 2009), as well as the near absence of irrigation practices in most of these regions.
In more detail, the stepwise regression results for the overlapping 3-month time series of NDVI and climate variables for SON (September, October, and November) show statistically significant coefficients for precipitation and air temperature at 45 % and 98 % of the agricultural area in the Bolivian Altiplano with a median of 0.2 and 0.7, respectively (Fig. 5a). This indicates that the NDVI increases with more rain and higher air temperature. Interestingly, the significant regression coefficients of NDVI for OND (October, November, and December) associated with precipitation and air temperature for SON cover 64 % and 91 % of the agricultural area and have a positive median of 0.3 and 0.4, respectively (Fig. 5b). A time lag of 1 month shows larger spatial coverage of the response of vegetation to precipitation anomalies. Here, the largest coefficient of determination are shown in areas surrounding the Lake Titicaca. Moreover, the response of the NDVI for MAM (March, April, and May) to the studied climate anomalies for FMA (February, March, and April) covers 95 % and 96 % of the agricultural land for precipitation and air temperature, respectively (Fig. 5c). This mostly shows coefficients of determination ranging from 0.4 to 0.8, and positive regression coefficients for precipitation and air temperature have a median of 0.5 and 0.4, respectively. The hours of sun required for crop development could be an explanation for the time lag between vegetation and the climate variables. In addition, the lag differences between vegetation and precipitation can be partly explained by the topography, land cover, groundwater, and soil properties (Quiroz et al., 2011;Yarleque et al., 2016). Finally, the regression for NDVI and climate variables for the overlapped 3-month time series of MAM shows significant coefficients at 23 % and 98 % of the agricultural land, with a median of 0.4 and 0.6 for precipitation and air temperature, respectively (Fig. 5d). Hence, the vegetation response to precipitation is limited for the last overlapped 3-month time series of the growing season. However, it should be noted that air temperature remains an important variable.
To summarize, while acknowledging some important limitations, we found that the CHIRPS dataset is adequate to be used for drought risk assessment in case of severe data scarcity for the Bolivian Altiplano. Furthermore, we found that the vegetation variance can be significantly explained by precipitation and air temperature. More specifically, we point out the relevance of precipitation as the main water source for vegetation development and air temperature as a driver of photosynthetic processes. Precipitation is particularly important at the early and late phenological stages, in which crops are more sensitive to water shortage. This is the case for the main crops in the region, i.e. quinoa and potato. For the quinoa crop, the most sensitive phases to water stress are the emergence, flowering, and grain development (see Geerts et al., 2008;Geerts et al., 2009). The most sensitive phases of the potato crop to water stress is the tuber initiation and bulking (van Loon, 1981;Alva et al., 2012). On the other hand, air temperature is relevant for vegetation productivity, and overall we found a positive relation between vegetation and air temperature. However, in prolonged dry periods, high air temperature could increase the evapotranspiration rates and consequently decrease the soil moisture (Huang et al., 2019). This scenario could impact negatively the vegetation, as this is the case of the drought events of 1982-1983 and 1997-1998, where large production losses were reported (Santos, 2006).

Discussion and conclusion
We employed satellite-based and gridded dataset products and tested the dataset's empirical accuracy as well as performance to similar (but with coarser resolution) datasets available for the Bolivian Altiplano region. Spatio-temporal patterns of satellite precipitation and gridded air temperature anomalies were explored based on monthly time series during the period September 1981 to August 2015. Drought severity was evaluated based on a drought classification scheme using NDVI and LST; this classification was related with the ENSO anomalies. Finally, association between the spatial distribution of NDVI with precipitation and air temperature was examined. Using these datasets, it was shown that drought severity (measured through various drought indices) increases substantially during El Niño years, and as a consequence the socio-economic drought risk of farmers will likely increase during such periods. ENSO forecasts as well as drought severity (through drought indices) can help to determine possible hotspots of crop deficits during the growing season. The empirical relationships of land surface and climate data on the local scale of our approach can support a proactive approach to disaster risk management against droughts, through an evaluation of the evolution of climate anomalies (in this case the ENSO) and their potential adverse effects in the region. As it was shown here, the ENSO-warmphase-related characteristics are especially important in the context of extreme drought events and could therefore be incorporated within early warning systems as standard practice. Despite these challenges for the development of drought early warning systems (see FAO, 2016FAO, , 2017, applications have been successful in the past (e.g. Global Information and Early Warning System (GIEWS) of FAO and Famine Early Warning System (FEWS) of USAID). Monitoring and predicting ENSO can therefore significantly contribute to reduce the risk of disasters. This study is a first attempt to provide an assessment of drought impact on agriculture in relation to the ENSO phenomenon for the Bolivian Altiplano. We focused on where vegetation is more affected by droughts over agricultural land and how this can be clarified using satellite imagery. It is important to note that the variance of drought indices (as well as NDVI) to a large extent is explained by precipitation and air temperature anomalies in the studied region. The agriculture in this semi-arid region is ecologically fragile, and the main water source is precipitation, and thus crop production is considerably affected by precipitation anomalies. However, while an overall response of vegetation variance to precipitation and air temperature is evident, it is important to consider other variables, such as evapotranspiration and soil moisture to improve risk-based models. Another important issue is the time lag of the response of vegetation to precipitation and air temperature anomalies, which shows a hysteresis of 1-2 months. These findings provide information for future drought risk management and early warning system applications. In addition, with such information agricultural models can be set up along with risk management plans with improved accuracy.
Appendix A Table A1. Spatial location of the studied weather stations where gauged precipitation data are available; the stations that also present temperature maximum and minimum data are indicated by T on the temperature column.    Code availability. For analysis we have used the MATLAB programming language. A script file is available in the Supplement.
Author contributions. CCR conceived the study, collected the data, performed the analysis, interpreted the results, and wrote the article. SHS, GP, BC, and RB contributed to the writing and the interpretation of the results.
Competing interests. The authors declare that they have no conflict of interest.