Drought is a complex, globally occurring phenomenon that affects humans and nature alike [1
]. It can be examined from different perspectives: meteorological (precipitation deficit), agricultural (soil moisture deficit), hydrological (runoff and water storage deficit), and socio-economic drought (consideration of water supply and demand) [2
]. Drought events in Europe have far-reaching impacts and economic costs within different sectors of society [5
]. Within the European Union and United Kingdom, the annual economic losses (1981–2010) from drought are estimated at EUR 9 billion year−1
]. In Central Europe, and thus also in Bavaria, extreme drought events in recent years have repeatedly resulted in massive damage and losses in agriculture and forestry [8
Since the beginning of the 21st century, Central Europe has experienced recurrent periods of exceptional drought, with the years 2003, 2015, and 2018, in particular, being extremely hot and/or dry [12
]. While climate data from the past do not yet allow a consistent statement on drought trends in Central Europe [16
], future projections mainly assume an increase in drought frequency and severity during the growing season [20
]. In addition, there could be a significant increase in the risk of consecutive droughts and their affected areas in the future [15
In this context, the detection and forecast of drought are becoming more and more important, a major challenge being the monitoring of the impacts of agricultural drought on vegetation. In addition to in-situ measurements, remote sensing based observations allow a broad scale detection and can identify and fill relevant knowledge gaps within this thematic area. A variety of approaches based on remote sensing exist to detect drought impacts on vegetation [24
], some of which have already been applied in Central Europe, with different methodologies. The majority of existing studies made use of remotely sensed vegetation indices (VIs).
One possibility is the comparison of VIs with meteorological drought indices. It has been shown that the Normalized Difference Vegetation Index (NDVI) and Vegetation Optical Depth (VOD) could capture agricultural drought events in response to meteorological droughts within Europe [27
]. In addition, it has been found that meteorological drought indices for shorter periods were associated with crop stress (defined via the Vegetation Condition Index (VCI) and Vegetation Health Index (VHI)), while longer accumulation periods correlated better with the vegetation status of forest areas. At the same time, the magnitude of regional differences in drought impacts within Europe was pointed out [28
]. Equally, when the NDVI and Standardized Precipitation Index (SPI) data in spring and summer were compared to each other, high positive correlations were detected, especially in Eastern Europe and on the Iberian Peninsula [29
Another approach within VI use is the analysis and comparison with soil moisture data. Comparing the NDVI with soil moisture indices such as the Palmer Drought Severity Index, the Self-Calibrating Palmer Drought Severity Index and the Normalized total depth Soil Moisture (NSM), clear differences in correlations have been observed across Europe [29
]. Here, the NSM achieved the highest correlations (also in Eastern Europe and on the Iberian Peninsula), whereby this was explained primarily with the additional information included in NSM compared to the other indices about soil moisture in deeper soil layers. Using the NSM as a soil moisture variable, it was also found that in European areas with a warmer summer climate, the NDVI only responded to fluctuations in soil moisture, but not in temperature. In spring, this dependency was reversed. In colder regions, the NDVI was only dependent on temperature in both seasons. Furthermore, during the 2018 agricultural drought in the Netherlands, negative soil moisture anomalies occurred 2–3 weeks before the first VI reduction, using near-infrared reflectance of terrestrial vegetation and VOD as VI [30
]. When comparing NDVI and Climate Change Initiative soil moisture data, an offset between soil moisture drought and vegetation drought in Europe was also found [31
In Central Europe, analyses of VIs have been used to assess the impact of agricultural droughts on corn and winter wheat yields. Corn yields at the NUTS3 level in Germany correlated well with VCI and VHI in August, although clear spatial correlation differences could also be seen [28
]. On the other hand, comparing Temperature Condition Index (TCI), VCI and VHI with winter wheat yields in northern and eastern Germany, VHI, in particular, was able to achieve higher correlations than meteorological indices [32
]. Comparing the VIs NDVI and Enhanced Vegetation Index (EVI) with climatic datasets, it was shown that during the extreme drought in July 2018, 1.5 times more area within Europe was negatively impacted than during the extreme drought in August 2003. Differences within land-use types were also found [33
Within remote sensing of agricultural drought, there are also studies which developed combined drought monitoring in Central Europe. Sepulcre-Canto et al. [34
], for example, applied a combination of SPI, soil moisture anomaly, and the fraction of Absorbed Photosynthetically Active Radiation (fAPAR), while Trnka et al. [35
] used a combination of soil moisture data (remotely sensed, modelled, and reported) and EVI in the Czech Republic and Slovakia. Both approaches were shown to be promising. Finally, approaches also exist within this topic to determine areas and time periods where water is the primary limiting factor for plant growth. In Europe, a correlation analysis between LST and NDVI has revealed that energy was a limiting factor mainly in northern Europe, at high altitudes, and in spring, while water limited plant growth mainly in southern Europe and in summer [36
Despite the large numbers of existing studies on agricultural and vegetation-based drought in Central Europe using remote sensing, uncertainties and larger knowledge gaps still exist. For example, the spatial as well as temporal resolution of results indicating where and when water is the limiting factor for vegetation growth is still rather limited. There are also hardly any evaluations that deal with other explanatory environmental variables for vegetation stress caused by drought. In Bavaria in particular, there are still hardly any approaches for continuous drought monitoring via remote sensing-based VIs that can demonstrate an appropriate spatial and temporal resolution. Therefore, our study aims, for the first time, to provide results with a much better spatial and temporal resolution than currently exists as to where and when in Central Europe water is the limiting factor for vegetation growth. Second, the analysis of this limitation incorporates other environmental conditions to explain it, which has rarely been done before. Finally, an approach not yet practiced in Bavaria is presented to perform continuous drought monitoring via remote sensing-based VIs at an appropriate spatial and temporal resolution.
Addressing the aforementioned gaps in knowledge, the general objective of this study is to detect agricultural drought using remote sensing in Central Europe on water-limited areas for vegetation growth. Specifically, this means, with a spatial resolution of up to 250 m and a temporal resolution of up to 8 days:
To determine, by means of a correlation analysis between Moderate Resolution Imaging Spectroradiometer (MODIS) NDVI and LST, areas and time periods in which water is the limiting factor for vegetation growth. In addition, it will be examined whether and to what extent the factors of land cover and altitude influence these conditions;
To carry out drought monitoring using the TCI, VCI, and VHI, as well as to evaluate their results by soil moisture and agricultural yields. The question to be addressed here is whether and to what extent these indices can be used to detect agricultural drought.
This study addresses three research questions that are not satisfactorily answered in the current literature. First, it provides results with much better spatial and temporal resolution than the current one to determine where and when water is the limiting factor for vegetation growth in Central Europe. Second, the analysis of this constraint incorporates other environmental conditions to explain it, which has rarely been the case before. Finally, an approach not yet practiced in Bavaria is presented to perform continuous drought monitoring using remote sensing-based VIs at appropriate spatial and temporal resolution.
Our study showed that the correlation between NDVI and LST depends on the season, land cover, and altitude, and that the TCI and VHI correlate well with both soil moisture and yield data in Bavaria. Considering that the NDVI-LST correlation is negative when water is the limiting factor for vegetation growth and positive when energy is the limiting factor, several conclusions can be drawn from the results presented. In Bavaria, this relationship is, firstly, dependent on the season: At the beginning and end of the vegetation period the correlation between NDVI and LST is predominantly positive, whereas in the summer months it is predominantly negative. Accordingly, the primary growth-limiting factor of energy is replaced by the factor of water, especially in July and August. This seasonal dependence of the NDVI-LST relationship has already been demonstrated globally, for example in China [67
] and North America [54
] and even both from temporal and spatial perspectives in Europe [36
], although not in such a detailed spatial and temporal resolution.
Classifying the correlations according to land cover, it is noticeable that only agricultural land and grassland show negative NDVI-LST correlations in the summer months, while forest shows none or only positive correlations throughout the year. This indicates that forests have a greater water storage capacity and/or a higher resilience to drought stress than agricultural plants and grasses. Accordingly, the growth of forests in Bavaria in summer seems in this study primarily energy-limited, while water is the primary limiting factor for arable plants and grasses. However, it is important to emphasize that this is an averaging analysis over 20 years: Extreme drought years also affect forest areas in Central Europe and are associated with tree mortality [9
The fact that the relationship between NDVI and LST varies depending on the land cover or vegetation type has also already been demonstrated in several regions of interest, e.g., on the Iberian Peninsula [69
], in the Arctic [70
], Mongolia [71
], and North America [54
]. The differences in forest areas and other vegetation types have only been investigated in Mongolia [71
] and North America [54
]. Similar correlation differences were found in Mongolia as well, although the analysis does not include seasonal differentiation. In North America, it was also shown that forest areas are less prone to negative NDVI-LST correlations than agricultural areas or grasslands. However, the different data, classification, and geographical conditions must be taken into account when making these comparisons.
The third factor that significantly influences the relationship between NDVI and LST is altitude. An increase in the correlation coefficient between NDVI and LST with altitude has similarly been observed in other parts of the world before, for example in North America [54
], Mongolia [71
], and Europe [36
Additionally, in Bavaria the correlation coefficient increases with altitude in all months, with negative correlations only occurring below 800 m a.s.l. on average. Thus, especially in summer at low altitudes in Bavaria (<800 m), water is the primary limiting factor for vegetation growth, while above this altitude energy continues to be the primary factor.
Looking at the calculated drought index maps from TCI, VCI, and VHI (Figure 6
), it can be seen that MODIS data can primarily capture the temporal and spatial dynamics of NDVI and LST. Comparing this with drought index calculations from other satellites (e.g., [72
]), this is an advantage in drought monitoring, especially in the temporal dimension. However, some compromises have to be made within the spatial resolution. In general, it can thus be argued that MODIS data are particularly useful for near-real-time drought monitoring, while general classifications at selected time periods with high spatial resolution are more useful with other platforms.
When evaluating the drought indices TCI, VCI, and VHI, with the soil moisture index, two aspects are particularly relevant. Firstly, TCI and VHI show higher correlations with soil moisture than VCI, with TCI resulting in the highest correlation overall. It can be concluded from this that surface temperature generally reacts faster to changes in soil moisture than vegetation state. Further studies with similar methodology exist mainly for Asia and North America. Positive correlations between soil moisture and VHI have been observed also in [74
]. In Mongolia, the correlation between TCI and soil moisture is higher than that for VCI, firstly, at different soil depths (but only with lagged measurements) [77
] and secondly, in areas with predominant vegetation cover [78
]. In China, soil moisture measurements and TCI also correlate higher than VCI in different months [79
]. A similar picture emerges for rice fields in Vietnam, and only on forest land do both indices show a similar soil moisture correlation [80
]. The results for Bavaria are thus largely supported in other regions of the world.
The second relevant aspect of the correlation analysis between soil moisture and the drought indices is the seasonal course. In summer, all indices correlate more strongly with soil moisture than at the beginning and end of the vegetation period. This suggests that soil moisture in summer has a stronger influence on both the surface temperature and the status of the vegetation. This can be explained by the fact that water is the primary limiting factor for plant growth in the warmer or hot summer months when the biomass to be sustained is high: If soil moisture shows negative or positive anomalies in summer, this has a stronger influence on the water balance of the vegetation and thus also on the temperature [81
] than in the other seasons, in which energy is the primary limiting factor.
When comparing the agricultural yield data with the drought indices, three aspects are particularly central. Firstly, all three indices (TCI, VCI, and VHI) show similarly high positive correlations in general, i.e., as indicated in the weighted yields. Also, there is no general tendency for the superiority of any index for the individual crops. All three indices thus reflect the annual anomaly of agricultural yields in Bavaria well. Two aspects are noticeable here: VHI shows only slightly superior correlations here than its two components separately. On the other hand, VCI and TCI correlate with yield at the same level, although VCI is assumed to have a much more direct link to agricultural yield via NDVI than the TCI via LST. This suggests that drought, which is reflected in higher than normal vegetation thermal conditions, was the main cause of yield losses in the study area. Other correlation studies in Europe involving yield data have also shown positive correlations with the three drought indices almost without exception [61
]. However, a direct comparison is usually difficult due to the different methodologies, data situations, selected crops, and geographical locations. When comparing with the most methodologically similar study by Bachmair et al. [28
], the results are confirmed in terms of both correlation strength and spatial distribution, although in the other study the VCI shows slightly higher correlations than the VHI. The added value in the present paper compared to this study is mainly the analysis of several crops and the inclusion of the TCI.
The second aspect worth mentioning is the difference in correlation strength between the individual crops. Especially sugar beet and silage corn show higher correlations than the others, which can be explained primarily in the monitoring period: While most crops are harvested at the end of July/beginning of August, sugar beet and silage corn are harvested at the end of September/beginning of October (Table 3
). Thus, for these two crops, more growing time and area (especially in August) is included in the correlation analysis, which is obviously reflected in the drought indices. In comparison with other studies, it is noticeable that winter wheat yields in north-eastern Germany [32
] show higher correlations with the VIs used than in Bavaria. The decisive difference here, however, is in the methodological choice to include only relevant areas in north-eastern Germany in the monitoring or the correlation analysis, whereas in our study all vegetation areas with negative NDVI-LST correlation were included. Our results of the yields for wheat and barley in connection with the drought indices are also confirmed by a Europe-wide temporal correlation analysis between NDVI and wheat and barley yields [84
]. Here, both correlations in Germany decrease significantly from June onwards (Day 153) and reach similar value ranges as in our analysis.
The third and last aspect to be noted is the spatial difference within the correlation analysis. The general impression is that for all indices and crops the correlation between yield and drought index is higher in northern/eastern Bavaria than in southern Bavaria. The reason for this can be traced in the different precipitation conditions (Figure S3
): In drier northern Bavaria, the sensitivity of both the drought indices and agricultural yields to a change in the water balance is higher than that in wetter southern Bavaria. The higher sensitivity causes a higher correlation and thus a stronger relationship between the two variables.
Nevertheless, certain limitations of this study should be noted. The spatial resolution of the underlying data should be thoroughly considered. Due to the small-scale features of the landscape in Bavaria, mixed pixels often occur, especially in the remote sensing data, which can lead to inaccuracies both in the correlation analysis between NDVI and LST and in the index calculation. Within the evaluation data, this indication is equally true for both the soil moisture data (4 km) and the agricultural yield data (county level). In addition, it must be noted that the data in this paper show significant differences within temporal and spatial resolution. In order to match the data, they are adjusted on both levels (e.g., via resampling or interpolation). This causes additional inaccuracies and could be an explanation for partially low correlations within the results. In addition, the assumption of a 19-year identical distribution of crop shares within the counties introduces—to some extent—further uncertainties.
Another aspect worth discussing is the general use of VIs for estimating vegetation growth and the application of NDVI in particular. Thus, VIs only ever provide an indirect vegetation proxy, and several limitations must be considered here: On the one hand, detection on heterogeneous surfaces and at different canopy heights is problematic, and on the other hand, factors such as sensor calibration, sensor viewing conditions, solar illumination geometry and soil moisture, also influence the data quality of the indices [85
]. NDVI in particular is one of the most widely used vegetation indices and thus offers a wide range of comparisons. It allows a large number of estimations of different vegetation properties and, in addition, it has a sensitivity to green vegetation even on areas with low vegetation cover. Nevertheless, the NDVI also brings uncertainties in the detection due to the influence of soil brightness, soil color, clouds and cloud shadows as well as leaf canopy shadows and a saturation problem in the presence of high vegetation diversity [86
From a methodological point of view, two aspects, in particular, should always be kept in mind: Firstly, the application of the drought indices is in part severely limited in terms of area due to the threshold value set for the correlation coefficient between NDVI and LST (which refers to the entire period and partially masks individual drier years). The evaluation of the indices is thus only of limited significance, especially in the marginal months of the vegetation period and in counties with little area included. On the other hand, the area included in the analysis is always the same when evaluating the drought indices with the individual crop yields and no differentiation is made between the individual crops.
In this study, we aimed at identifying by means of a correlation analysis between NDVI and LST when and where in Bavaria vegetation growth is primarily water-limited and how this is related to the factors of land cover and altitude. In addition, we investigated whether and to what extent the remote sensing-based drought indices TCI, VCI, and VHI can capture agricultural drought and yield losses within these areas. The indices were calculated from 2001 to 2020 within the growing season and evaluated with both soil moisture and yield anomalies of agricultural crops.
We found that in Bavaria, especially in the summer months of July and August, on agricultural land and grassland and below 800 m, water is the primary limiting factor for plant growth. Within these areas and periods, the remote sensing-based drought indices TCI and VHI correlate strongly with soil moisture and agricultural yield anomalies. From both a soil- and a vegetation-based perspective, both indices have the potential to detect agricultural and vegetation-based drought, respectively.
However, there are also further research potentials within this thematic focus. With regard to the LST-NDVI relationship, it would be particularly interesting to include in the correlation analysis other relevant variables, such as evapotranspiration, precipitation, and global radiation and to specifically analyze individual drought years with regard to this relationship. Within the evaluation of the drought indices, further soil- and vegetation-related variables could be included in the analysis. In addition, a temporally differentiated (e.g., monthly) correlation analysis between the drought indices and the field crop yield anomalies would be desirable to determine possible seasonal focal points in the context.