Higher temperatures and lower rainfall amounts may result in drought, is a common weather phenomenon and costly natural hazard. Drought is a recurrent climate process that occurs in association with temporally and spatially uneven rainfall over broad areas and extended periods [1
]. The regional temperatures of Southern Mongolia have increased by 0.1–3.7 °C over the past 60 years, spring precipitation has decreased by 17%, and summer precipitation has increased by 11% [2
]. These changes in temperature and precipitation are likely to intensify the occurrence of drought, especially during the onset of vegetation green-up. Moreover, the frequency of drought in the spring and summer has been reported to increase as much as four times every five years in the Gobi region [3
]. Drought has a disturbing effect not only on agricultural productivity and hydrological resources but also on natural vegetation; therefore, it may accelerate the desertification processes associated with destructive human activities (that is, overgrazing) in semi-arid grassland areas in Mongolia.
Many drought indices derived from remote sensing (RS-derived) data have been developed and used to effectively detect drought conditions all over the world. Because drought causes stress to vegetation, the extent of a drought can be reflected by changes in the vegetation index. The normalized difference vegetation index (NDVI) is the normalized difference between near-infrared (NIR) and red reflectance and is simple and effective and is now the most widely used index for detecting drought [4
]. Many vegetation indices based on the NDVI have been proposed for diverse regions. To measure the impacts of weather and ecosystem components on vegetation and reduce their deviations, a vegetation condition index (VCI) was developed by Kogan [7
]. Water stress in plants results in reduced chlorophyll absorption in the blue and red bands [8
], so the blue band can be viewed as the band least sensitive to vegetation moisture variation. Zhang et al. [9
] proposed the visible and shortwave infrared drought index (VSDI) by combining the blue, red, and SWIR optical spectral bands; this index is suitable for drought monitoring throughout the America growing season. Shortwave infrared reflectance (SWIR) is sensitive to leaf liquid water content, and a combination of NIR and SWIR data has been used to derive water-sensitive indices for monitoring drought occurrences. Based on NIR and SWIR, the normalized difference water index (NDWI) [10
] is sensitive to changes in soil moisture that are strongly related to vegetation drought conditions in the grass and crop land of the Oklahoma Mesonet. Analysis has revealed that combining information from multiple near infrared and short-wave infrared channels into the normalized multi-band drought index (NMDI) [11
] enhances its sensitivity to drought severity, a method that is suited to estimate both soil and vegetation moisture. Gu et al. [12
] found that the NDWI responds more quickly to drought than the NDVI, and they then developed the drought indicator known as the normalized difference drought index (NDDI), which also has been adopted by Mongolia’s National Remote Sensing Center for monitoring grassland drought [13
]. Land surface temperature (LST) derived from thermal radiance bands is a good indicator of the energy balance of the earth’s surface, because temperatures can rise quickly under water stress. Gutman [14
] showed that thermal data might be useful for detecting inter-annual changes in surface moisture. Kogan [7
] developed the temperature condition index (TCI), which is an initial indicator of water stress and drought.
A combination of different indices representing vegetation stress, water deficit, and soil moisture status can describe the severity of and changes in drought better than each index in isolation. Previously, Carlson [15
] constructed the vegetation supply water index (VSWI) using the NDVI and LST to assess summer drought, because the ratio of LST to NDVI is shown to increase during drought. VSWI describes the soil moisture changes in agricultural land well and is a rapid and cost-effective method for monitoring drought conditions [16
]. Sandholt et al. [17
] found that the scatter of LST and NDVI data occupies a triangular space, the NDVI–LST spectral space, and that an index based on this relationship (the temperature vegetation dryness index, TVDI) can be used to better monitor regional drought. A similar indicator, the vegetation temperature condition index (VTCI), was applied to drought monitoring by Wang et al. [18
]. The VTCI had better performance than NDVI in classifying relative drought occurrence levels and in studying the distribution of drought occurrences. A typical drought indicator, the vegetation health index (VHI), was proposed by Kogan [7
] and was based on the combination of vegetation greenness (VCI) and temperature (TCI) indices. AVHRR-based drought indices (VCI, TCI, and VHI) were also proposed by Kogan [19
] for monitoring grass conditions in Mongolia; that study found that these indices coherently indicated intensive drought in poor grass.
Several researchers are currently monitoring drought conditions in Mongolian grassland using satellite-derived indices [1
], and meaningful results have been obtained. As detailed above, other drought indices have performed well in other countries and climate zones. Which of them can best describe Mongolia’s drought conditions, for prevention and mitigation use by researchers, government departments, or stakeholders, remains unclear. Therefore, this paper’s objectives are as follows: (1) to adopt new reference indicators and comprehensive methods for evaluate Mongolian drought in several scales; (2) to build an integrated and novel adaptability analysis framework to determine the optimal satellite-derived drought indices for the accurate and real-time expression of grassland drought in Mongolia.
4.1. Comparisons with the PED
Eighty-six meteorological stations in Mongolian grassland areas were available from 2000 to 2014, and data from these stations was used for analysis and validation. The land cover of these stations can be divided into three types: forest steppe (18), steppe (41), and desert steppe (27). The effects of meteorological variations on vegetation were observed, and lags in vegetation response (drought or normal growth) were identified. Therefore, we calculated the one-month PED using data from the current 10-day and last two 10-day periods. Long-term analyses of PED meteorological index were performed on the grass-growing periods (June–August) from 2000 to 2014.
The NDDI, NMDI, VSWI, and PED indices are positively correlated, whereas the other six indices are negatively correlated (refer to these equations in Table 1
). The correlations between the satellite-based drought indices and the PED are shown in Table 3
, which shows maximum, minimum, average, and standard deviation values of the correlation coefficients (MAX_R, MIN_R, AVE_R, and STDE_R) for all stations. The results indicate that VHI, TCI, and VSWI were highly correlated with satellite-based drought indices throughout the growing period in the steppe zone, with an average correlation of more than 0.52 at the 99% confidence level (this confidence level also applies to the following). The VHI had the highest average correlation (0.66) with a maximum value of 0.76. VHI and TCI exhibited the highest correlation values for forest steppe and desert steppe areas; the MAX_R of VHI was 0.75 (VHI and TCI averages were 0.61 and 0.56, respectively). The results indicate that VHI, TCI, and VSWI are significantly correlated with PED for the three steppe types, and that correlation between VHI and PED is higher than between TCI and VSWI. Additionally, STDE_R for the three land cover zones varied from 0.07 to 0.24. In forest steppe zones, TCI exhibited the lowest STDE_R, and there was less variability figure across stations. TCI is a stable and reliable index for this zone. Similar results were produced for VHI in the steppe zone and TCI and NDWI in the desert steppe zone. NMDI had the largest STDE_R for all three steppe zones, which demonstrates that this index cannot accurately reflect drought in Mongolia as a whole.
Stations Erdenet, Erdenesant, and Saikhan were selected for the forest steppe, steppe, and desert steppe zones, respectively. Figure 3
a–i are composed of time series plots based on nine RS-derived drought indices and the meteorological index (PED) from these three stations over the first 10 days of July from 2000 to 2014. The curves of the RS-derived drought indices were extracted using a 3 × 3 km window over the 86 meteorological stations. All nine RS-derived drought indices present obvious changes relative to the PED. The TCI, VCI, VHI, NDWI, VTCI, and VSDI are negatively correlated with the PED, whereas the NMDI, NDDI, and VSWI are positively correlated. Like the TCI, VCI, and VHI for Erdenet (forest steppe) exhibit less consistent changes than other indices. Moreover, the VSDI is not sensitive to the PED and shows a small value range. The Erdenet station is known to have experienced severe drought in the serious drought year of 2002. The lowest values in the time series occurred for the indices TCI, VCI, VHI, NDWI, VTCI, VSDI, and NMDI, and they corresponded to the largest PED values. The overall trend for the VHI, TCI, and VSWI curves are more consistent with the PED curve than for other curves for Erdenesant station (Figure 3
d–f). The values of the VHI, VCI, TCI, and NDWI decrease rapidly with increases in the PED, and increase rapidly with decreases in the PED. In the desert steppe station (Saikhan), less vegetation resulted in lower sensitivity of the VCI and NDWI to drought. The VHI is obviously the best index for describing grassland drought, because it combines vegetation changes and temperature anomalies, and the PED also reflects water and thermal conditions. The highest correlation between the RS-derived drought indices and the PED is the VHI.
4.2. Comparisons with Soil Moisture
For this study, soil moisture observation data were obtained at depths of 5 cm and 10 cm for each 10 day period from 2000 to 2011 from 39 soil moisture sites (9 forest steppe sites, 23 steppe, and 7 desert steppe). We found that soil moisture at 10 cm was better associated with grass growth and drought conditions than that at 5 cm; therefore, the 10 cm depth soil moisture values were adopted for the analysis in this paper. To account for spatial consistency between soil moisture measurements and RS-derived drought indices, we selected the corresponding nearest pixel for each of the 39 soil moisture site locations and extracted each pixel’s value. Thus we created a dataset that included 39 pairs of site-based soil moisture and RS-derived drought index values for each drought index in every 10-day period.
The Rs were determined from satellite-derived drought indices and ground soil moisture data; the results indicate that the TCI, NDWI, and VHI have higher Rs. For each steppe type, the MAX_R, MIN_R, AVE_R, and STDE_R were produced by many stations. In the forest zone, TCI presented the largest maximum and average R, as can be seen in Table 4
, though larger STDE_R values occurred elsewhere. The correlation between ground-observed soil moisture at 10 cm and the NDWI in the steppe zone was the highest of the indices (0.69, which was significant at p
< 0.01), followed by the VHI (0.67). However, in the desert steppe zone, VHI featured the highest correlation coefficient (0.61) among the indices. Lower STDE_R and higher VHI averages in the desert steppe zone show that these values are stable across stations.
Point-derived data (such as soil moisture) only reflect information from a small region, while the drought extent determined by satellite data over large areas and long time periods are pixel values. Errors and uncertainties may occur in both data sets because of the difference in spatial scale.
4.3. Comparisons with the NorBio
To test the regional effectiveness of RS drought indices and determine the best index during the grass growing period for different land cover types (across 18 forest steppe stations, 41 steppe, and 27 desert steppe), we compared the RS drought indices with the NorBio index.
Generally, the grass-growing period in Mongolia is from May to September. In May, the grass is short. July and August are critical grass-growing periods, and grass biomass data is available from field observations. Therefore, we selected observed biomass data from late July for correlation analysis with the RS-derived drought indices. Using Formula (3), we calculated the NorBio values and constructed a correlation graph for July using the nine RS-derived drought indices and station-based biomass data observed over 15 years (2000–2014). For each steppe type, the maximum, average, minimum, and STDE_R correlation values were produced by many stations (seen in Table 5
). The results show high R values for some indices, with maximum values up to and above 0.90. For the forest steppe areas, the NDWI, VHI, VCI, and VSWI have better correlations than the other indices, with AVE_R values of 0.70, 0.62, 0.64, and 0.59, respectively, and MAX_R values of 0.92, 0.87, 0.83, and 0.78, respectively. For the steppe regions, the VHI, NDWI, VCI, and VSWI have higher correlations with the ground-based NorBio values than the other indices, with averages of 0.60, 0.61, 0.57, and 0.59, respectively, and MAX_R values of 0.94, 0.95, 0.94, and 0.88, respectively. VCI, VSWI, VHI, and NDWI have AVE_R values of 0.67, 0.60, 0.59, and 0.48, respectively, and MAX_R values of 0.92, 0.83, 0.82, and 0.86, respectively, in the desert steppe areas. The NDWI coefficient varies greatly among the stations with high STDE_R values, with a MAX_R value of 0.86 and MIN_R value of 0.08. The signs of the VSDI and NMDI correlation coefficients vary: some stations have positive correlations, whereas others have negative correlations.
a–i show the time series plots of nine RS-derived drought indices and the NorBio values at three stations (representing the three types of steppes) for the last 10-day period of July from 2000 to 2014. The dynamic trend of the NorBio is similar to that of the RS-derived drought indices throughout the studied period. The TCI, VCI, and VHI also showed fluctuation trends consistent with the NorBio. Generally, the TCI, VCI, VHI, NDWI, VSDI, and VTCI had positive relationships with the NorBio, and these values were also substantially lower in the severe drought year of 2002 than in a weak drought year (2003) for the three representative stations, except for the VCI at the Erdenet station (forest steppe, Figure 4
a). The TCI, VCI, and VHI values increased more than the NDWI, VSDI, and VTCI values during the severe summer drought of 2003, which suggests that the former are more sensitive than the latter to drought conditions. The relationship between NMDI and NorBio was negative, although opposite results were observed in certain years, as shown in Figure 4
b,e,h, indicating that this index is not suitable for detecting drought in Mongolia. When the NDDI was compared with the NorBio, the change trends in certain years (such as 2001, 2004, and 2012–2014 at the Erdenet station) were inconsistent; the VSWI and NorBio curves showed greater consistency.
4.4. Comparisons with the RDA in County
The RDA have a specific characterization area due to their nature as field observations, and are used to evaluate vegetation conditions at the county level. The DSC based on RS-derived drought indices represents drought status at the county scale. We adopted CPs (explained in Section 3.3.2
) to describe the abilities of the RS-derived indices to monitor summer drought using RDA as field-derived reference data. We calculated the DSC, RDA, and CP for 37 counties and analyzed the results.
The CP values for the three land cover types are shown in Figure 5
. The CPs of the VHI, VCI, TCI, and NDWI relative to the RDA were higher than those of the other indices. In the forest steppe regions, the highest average CP values were from the VHI, VCI, TCI, and NDWI at 74, 73, 73, and 71%, respectively. The maximum values for these indices were 84, 74, 81, and 87%, respectively. VHI, VCI, TCI, and NDWI predicted the same level of drought severity as RDA for approximately two-thirds of all 10-day intervals during summers from 2000 to 2014. For the steppe areas, the drought CPs between the VHI, VCI, and RDA were higher than those of the other indices, at 74%, although the average CP between the NDWI and the RDA was 73%. The maximum CP values for VHI, VCI, and NDWI were 89, 89, and 88%, respectively. The highest average drought CP (67%) was between VCI and RDA in the desert steppe zone (maximum value of 79%), and an average value of 66% was observed between VHI and RDA across all stations (maximum value of 81%).
These results show good consistencies between RS-derived drought indices and the field-derived RDA. The VCI, TCI, VHI, and NDWI are more sensitive to grassland drought and had stronger relationships with the RDA. These indices can accurately describe changes in drought in Mongolia at different temporal and spatial scales. In addition, the CP value for steppe is generally higher than for the other two land cover types, and the lowest CP is found in the desert steppe zone.
4.5. Spatial Consistence Comparsions
Mongolia experienced heavy drought conditions in 2002 and slight drought conditions in 2003. Two typical years (the first 10 days of July in 2002 and 2003) of drought monitoring results (Figure 6
and Figure 7
) were selected for the visual comparison of the spatial monitoring characteristics of the nine RS-derived drought indices with the reference RDA data. In 2002, drought occurred throughout most of Mongolia except for eastern Mongolia and a few other areas. The VHI and VCI maps showed similar distributions as the reference RDA map in central and southern Mongolia (mainly steppe and desert steppe areas), whereas the NDWI and NDDI maps exhibit good correlations in the northern forest steppe region. The VTCI and VSWI maps showed strong drought conditions, and the NMDI map showed drought in only the southern and western regions. Similar maps were generated for the slight drought year of 2003. The VHI and TCI maps were consistent with the distributions in reference to the RDA map in the central and southern Mongolia. The NDDI and VSWI maps showed heavier drought in the south but slight or no drought in the north or central Mongolia, which was completely inconsistent with the results in the RDA map. In comparison to the RDA and the VHI, TCI, and VCI map, the VTCI showed drought distribution over a larger area, whereas the VSDI showed less drought distribution. In the north forest steppe region, the NDWI and NDDI maps showed only the approximations of the drought distribution. Thus, for the weak drought year of 2003, we find that the VHI, TCI, and NDWI maps were more consistent with the reference RDA map than those of the other indices.
To clarify the annual variations, we focused on the temporal fluctuations from 2002 to 2003. The NDWI, NMDI, NDDI, VSWI, VSDI, and VTCI maps had features of the same drought pattern in 2002 and 2003, with less drought extent and severity in 2003. The VHI, VCI, and TCI results indicated decreasing drought extent and severity, consistent with the RDA map and local drought characteristics. Furthermore, the NDWI and NDDI maps showed distribution features in the northern forest steppe region that were consistent with significantly decreasing trends as on the RDA map. Figure 6
and Figure 7
illustrate the advantages of the VHI, VCI, TCI, and NDWI for drought monitoring large areas of Mongolia.
4.6. Comprehensive Results
A suitability analysis was conducted by comparing the RS-derived drought indices to the ground-derived results of the PED, soil moisture, NorBio, and RDA for the forest steppe, steppe, and desert steppe areas during the study period at the pixel, county, and regional scales. The output is shown in Table 6
. For each area, two levels (best and second best) were used to express sensitivity to reference indicators. The primary results were VHI/NDWI/TCI for forest steppe, VHI/VCI/NDWI for steppe, and VHI/VCI/VSWI for desert steppe. As the VHI index consists of both the VCI and TCI, the VHI can represent those two indices.
In terms of statistical distribution, there is an obvious trend point that represents the general level of data, known as the mode. The mode is the value that is repeated most often in a data set. Here, we propose using the mode to select the best indices for grassland drought monitoring. Based on the mode, the best indices are VHI and NDWI. Therefore, we find that these two RS-derived drought indices (VHI and NDWI) accurately describe Mongolian drought in the growing stage.
This paper proposes a comprehensive suitability analysis method for monitoring drought in Mongolia. The selected RS-derived drought indices (TCI, VCI, VHI, NDDI, VSWI, VTCI, VSDI, and NMDI) and were derived from multi-band data (visible, NIR, SWIR, and thermal infrared). The new method does not depend on only one reference data for evaluation; it takes full advantage of multiple sources of field data (environmental conditions, grass condition, and drought-affected information) to describe drought status at different scales for different aspects.
Drought had a substantial influence on maximum measured aboveground biomass production [34
]. However, biomass data reflect vegetation conditions in different areas at different times, and are limited in terms of spatial and temporal comparability with drought indices. To reduce these shortcomings, the NorBio was implemented to capture regional climate differences and the effects of short-term weather-related fluctuations on vegetation. Additionally, an AVHRR-based VHI was successfully used as a proxy for biomass in Mongolia [28
], because the VHI has high correlations with biomass anomalies and estimates of crop yield and grassland biomass in other parts of the world [39
]. The biomass production of a barley crop changes in response to drought depending on the timing and duration of the drought [35
]. Drought stress may influence the water supply to vegetation and reduce accumulated biomass and production of crops or grasses. Hence biomass data can reflect the extent and severity of drought; we successfully used biomass index to express drought status in different regions over many years.
VHI and NDWI are the best indices for Mongolian drought, based on comparisons with field RDA. Because RDA is regional data, we analyzed the temporal and spatial distributions of RDA and satellite-derived drought indices for the severe drought in 2002 and the slight drought in 2003. The VHI and NDWI had more consistent spatial–temporal drought variability than the other indices. The spatial distribution data was able to express the general drought status, changes, and development better than the pixel- or point-based data.
Considering all the multi-aspect and multi-reference comparison results, the VHI and NDWI are the optimal selection. Kogan [28
] previously applied the VHI for drought detection and derivation of pastoral biomass in Mongolia and found that VHI could reflect grassland health conditions and water- and temperature-related vegetation stress during drought. Hence the results of this paper are consistent with the findings of pioneering researchers. This index was also used for monitoring drought in other areas. Initially, VHI was used to evaluate the impact of drought on regional agricultural production in South America, Africa, Asia, North America, and Europe [7
], and a very strong correlation was observed between VHI and crop yield, particularly during critical periods of crop growth. In China, VHI has been evaluated for drought monitoring by several researchers [42
], who found that it was a stable and reasonable RS index for monitoring agricultural drought in different agro-meteorological zones of China.
Our research has found that the NDWI can also accurately express Mongolian grassland drought over long periods. The NDWI was derived from the NIR and SWIR channels, and it responded to changes in both water content (absorption of SWIR radiation) and spongy mesophyll (reflectance of NIR radiation) in vegetated areas [10
]. The NDWI has further been used to monitor the moisture conditions of vegetation canopies over large areas in several investigations [45
]. Soil moisture is a critical component in interactions between the land surface and the atmosphere, and prolonged soil moisture deficits often lead to drought-induced vegetation stress [47
]. We evaluated the performances of the NDWI for drought monitoring, and found that it is more sensitive to Mongolian grassland drought than other possible indices, especially for forest steppe (seen in Table 4
). The spatial distribution of NDWI values can best express RDA spatial distribution in forest steppe areas. Additionally, high correlation coefficients between NDWI and NorBio exist for forest steppe and steppe regions. Therefore, the NDWI is a very effective and simple index for monitoring grassland drought.
However, certain weaknesses in the proposed RS-derived drought products (VHI and NDWI) and ground data must be mentioned. These differences were caused by several error sources, including the selection of field stations, the distance between repeated points, differences in soil structures, and inconsistencies between the footprint of RS data and the point-based nature of the measured parameter indicators (soil moisture, NorBio, and PED). Soil moisture is a good indicator for vegetative drought; however, it is derived from in situ data, and it is difficult to obtain high correlations between in situ data and the spatial data of drought indices. Future experiments should focus on measurements of regional soil moisture to determine changes in drought. Normally, the PED index requires data over long time periods; however, the currently available dataset only covers approximately 15 years. The RDA is treated as regional field observation data, the data is actually collected by different observers without standardized instruments or equipment, possibly resulting in somewhat subjective results.
The VHI may feature additional errors for high-latitude regions, according to A. Karnieli [48
]. The relationship between NDVI and LST is positive, but the VHI-based drought index hypothesizes that increasing temperatures act negatively on vegetation vigor and consequently cause water stress and drought. However, in high-latitude or equatorial humid regions, higher temperatures accelerate vegetation growth, and vegetation development is mainly limited by the available energy. Consequently, the VHI may not be the best index for drought monitoring in high-latitude regions. Indeed, this paper has shown that, in forest steppe areas, VHI is not always the best indicator for describing drought (Table 4
). Our comparison analysis of NorBio and RDA spatial distribution showed that NDWI was better than VHI. This result is consistent with previous findings [49
]. In this paper, we adopted the VHI equation used by Kogan’s papers [7
]. In this equation, the coefficients of TCI and VCI are the same, that is, to 0.5, which is not strictly accurate because the NDVI and the LST exert varying influences on drought in different regions or ecosystems. In the future, focus should be on these shortcomings to find the optimal coefficients for various regions. This will enable the construction of a comprehensive drought monitoring model for the entirety of Mongolia based on the VHI (with various coefficient of TCI and VCI), and NDWI.
To identify the optimal index or indices for monitoring pasture drought in Mongolia, a new adaptability analysis framework was adopted for evaluating the performances of satellite-derived drought indices. Methods based on comparison to a meteorological index (PED), normalized biomass (NorBio) reference indicator, and RDA-based drought consistent percentage (CP) were proposed. Due to valuations at diverse scales (pixel, county, and region) for three land cover types (forest steppe, steppe, and desert steppe), Pearson’s correlation, CPs, and spatial consistency analysis methods were adopted, and an integrated assessment was developed to fully describe drought status. The mode method of statistical significance was used to identify comprehensive results among the comparisons of satellite-derived drought indices and five different reference indicators (PED, soil moisture, NorBio, RDA, and RDA spatial distribution).
The VHI and NDWI were found to be appropriate for the assessment of drought characteristics and for monitoring drought conditions in Mongolian grassland. These indices were able to detect the timing of drought onset and processes, and provided realistic quantification of drought severity in the study areas. These two indices can therefore be used to develop a combination drought model for accurately monitoring drought in the future. A comprehensive and novel adaptability analysis framework was built to identify the most appropriate satellite-derived drought indices for the accurate and near real-time detection of droughts in other countries or regions.