Forests are an essential part of both the natural environment and human existence. Locally and regionally, they provide a variety of ecosystem services, including wood related products, the filtering of water, climate regulatory functions and the potential for recreation and tourism. Forestry is among the most important economic sectors in Europe, especially in Germany. Forests are biodiversity hosts, a key part of global land-atmosphere interactions and a major net carbon sink. The amount of carbon stored in temperate forests increased by 17% in 2000-2007 compared to 1990-1999, contrasting tropical and boreal forests and partly outweighing anthropogenic emissions. Hence, temperate forests play a key role in mitigating climate change.
Natural disturbances, such as windthrow, fire and insect outbreaks, are vital for ecosystem function and dynamics of forests. They change forest's composition, structure and function and alter the resources available to the organisms in it. Consequently, disturbances can support landscape heterogeneity, increase biodiversity and benefit succession and renewal throughout the ecosystem. Disturbance regimes in forests, however, are changing globally, with global warming being a key driver. With increasing climatic variability and ongoing climate change, the frequency and intensity of droughts has increased recently. This has profound implications also on the forest ecosystems in Germany and results in increased tree mortality. Droughts affect forests directly by increasing evaporation and heat stress during summer and decreasing the share of snow in total precipitation that otherwise would act as a moisture reservoir. Indirectly, droughts increase temperate forest's vulnerability to other disturbance agents, especially insect outbreaks, pathogens and wildfire while at the same time facilitating their activity.
Consequently, it is crucial to study drought's effects on temperate forests and monitor their response operationally. Here, we present a novel remote sensing-based approach that compares the phenological characteristics of forests in drought affected years to their typical pattern of vegetation development of a reference period. We used the severe drought that affected major parts of Central Europe in 2018 as an example for this study. This drought was accompanied by both exceedingly low precipitation as well as high temperatures for the months of April to August. While this drought was not uniform over all of Europe, Germany was largely affected, with temperatures being exceeded only by conditions of the major heatwave of 2003 and a strong rainfall decrease especially in the country's North and central East. Consequently, our approach was tested for two areas in Germany - one in the south of North Rhine Westphalia in the country's western part and one in Brandenburg in its east, both being characterized by mixed forest, but experiencing somewhat different precipitation patterns.
For both areas, all available Landsat and Sentinel-2 images from January 2000 until September 2019 were acquired. Data was atmospherically corrected and stacked into one consistent data cube of Normalized Difference Vegetation Index (NDVI) values with a resolution of 30 m. Processing was performed using the software Framework for Operational Radiometric Correction for Environmental monitoring (FORCE) and included a removal of cloud- or snow-covered as well as otherwise unsuitable pixels. Using Google Earth Engine, we created a mask of all pixels that did not experience any stand-replacing disturbance event between 2000 and 2017 by utilizing the Global Forest Change Dataset. After masking all areas that did exhibit such a disturbance event, each remaining pixel's time series was divided into two parts. All observations between 2000 and 2017 were used as the reference of vegetation development and the years 2018 and 2019 constituted the monitoring period. 2019 was included to also observe ongoing drought effects during the vegetation period after the event.
For reference period of every pixel, a 2-dimensional kernel density estimation was made, with day of the year (DOY) as the and NDVI value as the variable. By standardization to sum 1, this density estimation is transferred into a probability distribution. The most likely NDVI value per DOY is then used as the baseline phenological cycle. This procedure has several advantages. First, this is a non-parametric approach that does not rely on a predefined function to describe vegetation development throughout the year and that doesn’t require any calibration. Second, it enables the estimation of an observation in the monitoring period being an anomaly. This is achieved by comparing its NDVI value to the probability distribution.
For both areas, NDVI values of 2018 and 2019 were compared to the estimated benchmark of vegetation development. While the difference in NDVI between the modelled baseline and the reference period shows some variability and faulty values, there is a detectable pattern. NDVI values where lower in 2018 compared to the phenological baseline in most areas, but the probability of these of being an anomaly was often low (less than 0.5). This suggests that values where not low beyond historical record and it is likely that this NDVI decrease can be attributed to natural variability, especially because observations were typically in line with or higher than past values in 2019. For some smaller areas however, NDVI values were exceptionally low in 2018, followed by consistently very low rates in 2019. This indicates that trees in this area were possibly impacted by drought or another disturbance agent that may interact with it such as bark beetles, causing die-off of trees.
While the approach presented here appears useful to detect the impact of droughts on forest areas and this study serves as a proof of concept, there are some aspects that should be further considered. First, the procedure was originally developed for MODIS Enhanced Vegetation Index (EVI) data, and transferring it to a combined Landsat-Sentinel-2 NDVI time series may require some additional adjustments such as time series smoothing to remove some of the variability mentioned above. Second, the use of other remotely sensed indices such as the Normalized Difference Moisture Index (NDMI) or the vegetation Condition Index (VCI) that may capture the impact of droughts on trees better may result in more accurate results. Third, while disturbance impacts can be quantified, we cannot yet infer the actual disturbance agent from the data using this approach. Since this study considered a drought year, drought is possibly a major influence, yet it is not clear whether the NDVI response, especially in case of detected mortality, is caused by water stress directly or rather by a subtler drought impact or a different driver. Still, it exhibits some promising results that should further be elaborated and tested in different locations to gain more insight into drought impacts.