Applicability of Sentinel-2 time series for drought monitoring
In this study, we used remote sensing tools to assess drought and its impacts in a Central European landscape comprising mostly forests and grasslands. This was done by analyzing either direct indicators of drought, such as LST, and indirect indicators of droughts that assess drought impacts on vegetation, such as anomalies of vegetation indices, using both medium resolution time series from MODIS and high spatial resolution time series from Sentinel-2 (West et al., 2019).
The common remote sensing analysis of droughts implies computing time series of drought indices from medium spatial resolution satellite time series, such as from MODIS or the Advanced Very High-Resolution Radiometer (AVHRR)(Um, Kim, & Park, 2018). Although medium to coarse drought index time series data is a dependable source of drought information, as shown not only in our study but also in previous research (Khorrami & Gunduz, 2021), there is currently a requirement for detailed, high-resolution data on drought characteristics and their impacts at the local level to support land management decision-making and strategies for mitigating droughts. However, utilizing high-resolution image time series, such as those from Sentinel-2, as input for standard remote sensing drought analysis methodologies is impeded by their relatively brief acquisition period when compared to longer time series like those from MODIS (Ghazaryan, Dubovyk, et al., 2020; Kowalski et al., 2022). The reason for this is that the length of the time series affects the calculation of baseline index values, which are used to detect anomalies in drought index time series.
In our study, we developed an approach for drought monitoring in Germany with high resolution 10m Sentinel-2 image time series. By covering several growing seasons with Sentinel-2 data, which were also temporally interpolated to create regular synthetic time series, we included a wide range of vegetation conditions from non-drought-affected to drought-affected vegetation. This in turn allowed the mapping of drought patterns in the study area during the monitoring period with a high level of spatial detail. The applicability of Sentinel-2 time series for drought monitoring was recently demonstrated by another approach in the study by (Kowalski et al., 2022). In their research, the authors used Sentinel-2 time series to compute fractional cover images for 2017–2020 to monitor drought impacts on grasslands in northern Germany. The findings of our study match closely their results in terms of detecting temporal occurrences of drought conditions in grasslands in Germany between 2017 and 2020, as well as the results from in-situ observations herein presented. Furthermore, the spatially detailed drought data extracted from the Sentinel-2 time series spanning 2017–2020 closely matched the drought patterns identified through long-term analysis of MODIS-based time series spanning 2001–2020. This could be due to the diverse range of conditions captured by Sentinel-2 images during this period, ranging from a non-drought year in 2017 to extreme drought conditions in 2018.
During the 2018 drought, soil moisture conditions in the study area were not affected until August due to the abnormally high levels of soil moisture from December 2017, January 2018, and February 2018. However, the remote sensing drought indices from MODIS and Sentinel-2 were able to identify a substantial number of negative anomalies already in May with LST anomalies and in June with NDVI anomalies, especially in grasslands. These anomalies are more common for grasslands and forests of highly intensive use, suggesting a higher vulnerability of intensively managed grasslands and forests to drought, which is in agreement with other studies (Lüscher et al.). Intensively used grasslands as well as forests tend to be composed of only a few species that maximize the abundant nutrient intake, and these are less resilient to drought conditions compared to natural or semi-natural forests (Abdel-Hamid et al., 2021).
Moreover, our research conducted in Germany contributes to the small pool of comparable studies by evaluating and mapping drought index data using synthetic time series obtained from Sentinel-2 imagery. These studies used high spatial resolution image time series for drought impact monitoring in grasslands in Germany with Sentinel-2 and spectral unmixing analysis (Kowalski et al., 2022), grasslands in South Africa with Sentinel-1 (Abdel-Hamid, 2020; Abdel-Hamid et al., 2021), and croplands in Ukraine with Landsat and Sentinel-1/2 (Ghazaryan, Dubovyk, et al., 2020). These types of studies hold significant value as they explore novel methods and data sources for monitoring and evaluating drought, providing new opportunities for drought assessment (Jiao et al., 2021).
The time series from Sentinel-2 is still limited only to recent years and is characterized by lower temporal resolution than conventional medium spatial resolution sensors, such as AVHRR and MODIS. In this respect, the analyses performed herein of the historic MODIS time series, spanning from the 2000s, and respective deviations of intra-annual observations are crucial for a better understanding of recent drought impacts in comparison to past years. In this regard, it is crucial to conduct further research on the impact of the duration of image time series, as well as the variety of drought index values within those time series, on the identification of drought characteristics using remote sensing data (Gerdener, Kusche, Schulze, Ghazaryan, & Dubovyk, 2022; Kowalski et al., 2022). Future research endeavors should thus focus on exploring the potential of utilizing high-resolution image time series for drought analysis, which should involve experimenting with diverse techniques for calibrating baseline conditions to identify anomalies within short and noisy time series data, across a range of (agro)ecosystems and geographical regions. Moreover, the option of data fusion between medium resolution satellite time series and medium to high data from, e.g., Landsat and Sentinel-2, to increase the spatial resolution of the drought index time series while retaining temporal resolution and duration of the time series should be further studied.
Comparison of drought indices computed from Sentinel-2 and MODIS time series
The results presented in Section 3 show that there are spatial and temporal differences between the drought indices computed from one sensor (i.e., MODIS or Sentinel-2) and/or from different indices. Several factors may contribute to these differences. In our comparison of drought indices that employed various methods, we discovered that NDVI anomaly and VCI, which are calculated using near-infrared and red reflectance, displayed strong agreement. On the other hand, the investigated LST anomalies exhibited less similarity to the results of the analyzed condition-based drought indices. That finding was also confirmed by the study (L. Zhang et al., 2017), which compared different drought indices across the USA. One possible reason for this is that condition-based drought indices are based on time-series analysis, where the first step of the index calculation is to determine the maximum and minimum values in a particular area from the same date of different years. Then, drought information from other years can be used to determine the drought condition of the year in question. Thus, the values of these indices depend on the maximum and minimum values found in the time series for the area/pixel analyzed. This may suggest that normalizing the Seninel-2-based drought indices with the data coming from other sensors with a longer time, such as MODIS; could potentially be a good way to improve the performance of Sentinel-2 indices and should be further investigated. Ideally, a longer high-resolution time series would provide a more comprehensive understanding of the variability and patterns in the data. It would allow for the calculation of more reliable long-term means and standard deviations, which can be useful for detecting anomalies and assessing changes over time, like in the case of the MODIS time series.
However, it's important to note that the adequacy of the dataset length depends on the specific objectives of the analysis, the temporal and spatial scale of the phenomenon being studied, and the availability of alternative data sources. In some cases, even a limited dataset can still provide valuable insights, especially if combined with other sources of information or analyzed within the appropriate context (Moore & McCabe, 1998). In the case of our analysis, the important focus of our work was to assess drought with high-resolution time series and compare it to the results of MODIS; thus, the use of Sentinel-2 time series is justified. To overcome this limitation, further studies should investigate fusion approaches to generate artificial data sets with both high temporal and spatial resolution and sufficient length (Tewes et al., 2015).
In addition, the range of values of all the indices used in this study varies considerably for the same drought condition, suggesting the need for varying thresholds for categorizing a drought (Ghazaryan, Dubovyk, et al., 2020; L. Zhang et al., 2017). Further, the detection of drought conditions by remote sensing indices may be affected by additional factors, such as land management, land cover and land use, irrigation, precipitation, elevation, and soil moisture (Quiring & Papakryiakou, 2003). Some additional factors could also affect the analysis of remotely sensed drought indices. Residual cloud contamination in satellite images can affect the drought condition of a particular geographic location. Lastly, vegetation management factors, such as moving and grazing, crop rotation, irrigation, and use of fertilizers and chemicals, can alter the agricultural climate and impact the identification of drought conditions from satellite time series (Abdel-Hamid, Dubovyk, Graw, & Greve, 2020; Dubovyk, 2019; Ghazaryan, König, et al., 2020).
Last, while case studies, like those performed herein, do provide valuable insights, they are limited in their ability to generalize findings to broader areas with various conditions. Thus, we recommend a large-scale study for drawing further and more robust conclusions.