Grassland areas cover around 40% of the terrestrial surface [1
]. According to the CORINE 2012 Land Cover dataset, grasslands cover approximately 23% of the European Alps [3
]. Grasslands are an important habitat in mountain areas for many plant and animal species, resulting in an increased biodiversity on a regional scale [4
]. Additionally, grasslands are among the biggest terrestrial carbon sinks and mediums for water storage and purification [2
]. On alpine sites, grassland cover and composition is also important for site or soil stabilization, decreasing the risk of erosion processes. Although the profit from agricultural crops is higher per cultivated unit, alpine grasslands are a cheap source of fodder for animals. Additionally, they are considered to be recreational spaces and of high relevance for the tourism sector, which is especially true for alpine meadows and pastures [2
]. Climatic changes are presumed to particularly impact alpine areas and their vegetation in terms of growing length, nutrient uptake, CO2
concentration, water availability as well as species composition, migration, and richness [9
]. At the same time, alpine grasslands are increasingly moved from traditional to more intensive management while, simultaneously, suffering from land abandonment [5
]. For this reason, multi-scale monitoring of this ecosystem type is advisable in order to better understand the ecosystem responses to climate and, consequently, provide tools to land managers to preserve the ecosystem.
The growth and status of grasslands can be monitored with a wide variety of optical sensors. Detecting the percentage of reflected sunlight of a canopy throughout its spectral signature offers a huge potential to monitor the photosynthetic activity and status of vegetation. Typically, the spectral signal of vegetation is determined by a rapid increase in reflectance from the red to the near infrared (NIR) wavelengths caused by high chlorophyll absorption in the visible light and high reflectance by the leaf tissue in the near infrared spectrum [12
]. Especially, passive optical sensors have the characteristics necessary to analyze the plants’ reflectance in red and NIR spectra as they allow the generation of vegetation indices (VI), such as the commonly used Normalized Difference Vegetation Index (NDVI) [14
]. The number of sensors for deriving spectral indices within suitable band ranges for vegetation studies as well as sensor platforms constantly increased over the past years. This allows the investigation of dynamic changes during a vegetation period on diverse spatial scales ranging from ground-based measurements by spectral reflectance sensors, spectroradiometers, repeated digital imagery as well as unmanned aerial vehicles (UAV) to extensive airborne or remote sensing measurements. The combined use of multiple sensors provides new opportunities and facilitates addressing errors of a sensor or the sensor implementation, characterizes advantages and disadvantages for monitoring a particular event during the growing phase of vegetation, or to jointly use them to bridge gaps in time and/or space [15
A first category of optical sensors is installed at the plot of interest, collecting point-based measurements. Optical sensors measuring in-situ can be installed on fixed environmental stations often organized in geo-sensor networks or on moving vehicles [16
]. Commonly used in-situ optical sensors are spectral reflectance sensors [17
], repeated digital imagery [18
], or field spectroradiometers, integrating the signal of only a few centimeters to several meters. Optical in-situ sensors experienced notable advances in data transmission in combination with low-cost solutions, energy efficient technology, and the reduced sensor size [16
]. At the same time, standards for data storage and production have been formulated to simplify the access and interaction with databases composed of field data collections [23
]. Technological advancements and growing availability of ground truth data provide the possibility to study the dynamic growth of vegetation over the year. This leads to more scientific interest in installing and evaluating different sensors for long-term ecological monitoring on a regional to global scale [24
]. Digital imagery has gained increasing scientific interest over the past years especially when deployed to research vegetation activities, and digital cameras are mainly referred to as “Phenocams” [15
]. They automatically acquire and store images every few minutes to hours in red, green, blue (RGB) and/or NIR bands, thus providing an important data source for vegetation analyses, such as phenology [25
], and recognition of different plant types [19
] or plant stress [27
]. Benefitting from the technological advances, the availability of Phenocams has been increasing globally in the past years. This has led to the creation of regional as well as global phenocam networks [20
]. The most spatially extensive scale of vegetation monitoring is Remote Sensing. Spaceborne optical sensors, such as Landsat-8 Operational Land Imager (OLI), Moderate-resolution Imaging Spectrometer (MODIS) Terra and Aqua. Satellite Pour l’Observation de la Terre (SPOT) Vegetation or the Advanced Very High Resolution Radiometer (AVHRR) are equipped with bands detecting the reflectance in both the red and the NIR spectra and therefore are well-suited for monitoring vegetation. Each of the sensors possesses its own characteristics in spatial resolution and temporal coverage. MODIS revisits plots on a daily basis with a resolution from 250 to 1000 m, similar to AVHRR. The Landsat-8 OLI Sensor offers a better spatial resolution of 30 m, but the temporal revisit occurs every 16 days [29
]. Despite not being openly accessible and irregularly acquired, Very-High-Resolution (VHR) optical sensors, such as Quickbird, Rapid Eye, or Planetscope, are highly adequate for detailed analysis of vegetation with a spatial resolution ranging from 0.5 m to 5 m [30
]. The launch of the Sentinel-2 A and B satellites, designed within the Copernicus Programme, offers a new possibility for vegetation monitoring. The combination of the same sensor on multiple spaceborne platforms considerably lowers the revisit time of the sensor to a few days and simultaneously provides information with a spatial resolution from 10 to 20 m [31
]. With its spatial and spectral characteristics, Sentinel-2 is considered to be well-suited for synergetic applications with other remote sensing platforms, such as Landsat 8 [34
The integration of optical indices has become a research topic of increasing interest in vegetation studies over the past years [2
]. Especially, the use of satellite sensors in combination with ground truth spectral information has been explicitly studied (e.g., as reference data or calibration/validation sites). In recent years, the amount of studies utilizing the VIs in order to evaluate the similarity of its derivation by different sensors has been steadily increasing [19
]. However, the evaluation of scale for monitoring grassland communities often results in a big issue in research and combining VIs calculated by multiple sensors, such as Sentinel-2 and Phenocams or Spectral Reflectance Sensors (SRS), as they are still not fully explored [6
]. Additionally, the synergetic use of diverse sensors detecting in similar spectral ranges has not yet been fully explored. Combining sensors is advantageous when generating consistent time series (gap filling) or analyzing or compensating sensor-specific characteristics. This study is intended to analyze the suitability of one VI calculated from diverse sensors to analyze the dynamic growth of grassland vegetation in alpine meadows for future approaches integrating multiple optical sensors. We selected four meadows spread among the Autonomous Province of South Tyrol and covering different altitudes, expositions, management types, and precipitation conditions. We analyzed the reflectance values measured by four different sensors in different altitudes, angles, and extents and therefore in different spatial and temporal scales: (i) Ground spectroradiometer, (ii) station-based SRS, (iii) Phenocams as well as (iv) spaceborne Sentinel-2 A and B Multispectral Instrument (MSI). Since all the instruments provide the spectral signal in both red and the NIR wavelengths, we decided to use the widely used NDVI as VI for this study and aim to explain the following:
The similarity of NDVI signal among sensors for each meadow site by visually interpreting the NDVI signatures as well as by calculating linear correlation and cross-correlation among sensors;
the suitability of each sensor for detecting events with short-term impacts on the vegetation cover, such as harvests and snow coverage, by analyzing short time spectral changes in each sensor; and
sensor-specific characteristics of the grassland sites by collecting multiple NDVI measurements on each of the four sites. The different plots are compared visually and with linear correlation analysis to assess the stability of the NDVI signal over time.
The presented study shows that the acquired NDVI signals diverge considerably depending on the sensor being utilized. These are likely due to the specific acquisition conditions of the sensors, namely the spatial scale of detection, the temporal resolution, or detection geometry.
4.1. Sensor Specifications and Geometry
Unlike the other sensors, the spectrometer NDVI is difficult to interpret in terms of vegetation dynamics since the generation of a meaningful time series requires stable illumination, i.e., weather condition, good accessibility of the plot, workforce, and time for preparation and analysis. Especially on higher altitudes, the number of possible spectrometer measurements is limited in time and space during one single year. In 2017, the growth phase of vegetation was affected by bad weather and late snow in April/May. Therefore, we were not able to conduct appropriate field measurements over this period. Additionally, the weather conditions on the Vimef2000 site were often variable during the field campaigns, resulting in fewer acquisitions. The restrictions for collecting spectrometer measurements make it difficult to integrate them into a combined sensor approach other than as reference or calibration for other sensors [35
]. Spectrometer measurements have been found to correlate best with the Sentinel-2 NDVI signal, most probably because they both measure under near-nadir circumstances and tend to saturate quickly. Considering the most reliable time series measured on the Vimes1500 site, we are not able to visually extract a connection to the first harvesting event due to a delay of 10 days after the harvest resulting in a slightly higher NDVI signal than before the harvest, indicating a full recovery of the plant canopy. The signal before and after the second harvest leaves a clearer mark despite the time gap of roughly one month. For spectrometer NDVI on alpine grassland sites, measurements of one week before and after a harvesting event are crucial. Measurements taken afterwards possibly miss the events due to the quick recovery of the plant canopy density.
The SRS sensor provides the most stable NDVI curves among all sensors. The exclusion of extreme values and time ranges at the beginning and the end of the day, as well as the implementation of a threshold filtering for the 90% percentile for the NDVI signal, results in a steady signature. However, the NDVI signal continuously underestimates the NDVI compared to the spectrometer and Sentinel-2 MSI NDVIs by 0.1 to 0.3, which is most likely due to its 30° inclination angle or due to issues with the calibration of the sensor. Additionally, the SRS sensor renders no information about error sources (e.g., changing light conditions, snow, precipitation, water presence, or human interaction) and is not extensively usable, returning only one single NDVI information of a plot near the station. Collecting NDVI with an SRS sensor mounted on an environmental station, as presented in this study, forces it to detect only marginal parts of the grassland. It could be useful to integrate multiple sensors pointed at different locations of the grassland or to manage the same sensors manually, e.g., during field campaigns, to quantify the NDVI differences between the measured plot and the overall site. Especially when compared to sensors measuring more extensively (e.g., Sentinel-2 MSI and Phenocam) on sites with inconstant and multiple management activities, the SRS NDVI signatures can prove to be inconsistent. This explains the difficulties of recording harvest events on the sites. The plots on 1500 m a.s.l show a better signal response than the plots on 2000 m a.s.l., which is probably due to the complete harvesting of the whole meadow within one day. On the other hand, the signal response shortly before and after snow events varies from slightly higher to considerably lower and noisy NDVI signatures. Here, the snowmelt process and the light conditions could influence the sensor since the snow melts at different rates throughout the day. The high temporal resolution within a day could be interesting for further expanding this analysis on the detection of the snow patchiness on grasslands.
The NDVI signal derived from Phenocam imagery is difficult to standardize and filter. The band range and the mechanical IR filter properties are not clearly defined and documented in the literature. Additionally, the viewing/inclination angle of the Phenocams is probably the source for underestimation of the NDVI signal. Despite these limiting factors, the correlations retrieved between the Phenocams, Sentinel-2, and SRS are very promising. The NDVI signal seems to be very robust through time over the sites of Domef1500, Domef2000, and Vimef2000 [19
]. The only site with less correlation is Vimes1500. The weaker correlation could result from the diverging positioning of the SRS sensor with respect to the Phenocam, the south exposure, or the Phenocam failure shortly after the first harvest period. The approach used for calculating Phenocam NDVI is considered to be an estimate and can be scaled either by spectrometer [26
] or by satellite-based NDVI values [19
]. Results suggest that Phenocam NDVI is underestimated continuously, but differently depending on the site and growing phase. Other than the NDVI, the Green Chromatic Composite (GCC, [19
]) or the Excess Green Index (EXG, [66
]) discount the RGB+IR image have been widely researched and are considered to relate to phenological stages in grasslands [59
Monitoring grasslands with remote sensing methods mostly lacks a standardized approach for integrating multiple sensors [6
]. Sentinel-2 MSI imagery, with the advantages in temporal availability and spatial resolution, showed that both the annual [68
] growing phases, as well as dynamic short-term changes within grasslands, are well representable. Sentinel-2 images on Vimes1500 are less constant than for the other three sites, which is most probably due to the influence of the slope and aspect of the site. An additional topographic correction could be very helpful to correct the images in a proper manner although it is difficult to be adequately geolocated in heterogeneous mountainous areas. As weather, or cloudiness, is the crucial constraint of optical Sentinel-2 data, the combination with other remote sensing data, such as Landsat 8, may lead to a slightly better time series [32
]. However, the combination of remote sensing imagery with other optical sensors, such as Phenocams and SRS, provides reliable information on the slightest changes of the spectral signal each day and makes it easier to monitor the dynamics in grassland communities. At the same time, the NDVI acquired with Phenocam and especially SRS tend to be stable under inconsistent weather conditions, such as clouds or shadows, which is most probably due to the number of measurements and the possibility to reduce noise by filtering and the availability of images during cloudy periods. Therefore, both sensors are equally useful when creating a combined sensor workflow. By analyzing the meadows on a subplot level, we noticed considerable differences within the NDVI signals of one plot (up to 0.2). This may result in the time lag hampering a simultaneous acquisition on each scale, the linear filtering of the NDVI values, or heterogeneity within vegetation of the study sites. We found that sensors detecting the ground in a near-nadir position (Sentinel-2, Spectroradiometer) have higher overall NDVI values than SRS installed at a 30° inclination as well as the very variable inclination of the Phenocams recording the lowest NDVI values. This indicates that the viewing geometry of the optical sensors constantly influences the NDVI signal over time, resulting in an underestimation of the NDVI [70
4.2. Temporal and Spectral Resolution
The temporal scale of acquisition influences the detection of short-term events. In the case of both meadows on 1500 m a.s.l., the NDVI signal experiences the highest change within the 2 weeks after the event, but is not always applicable to each of the events (e.g., the second harvest on Vimef1500) for the Sentinel-2 MSI image. Within the sites on 2000 m a.s.l., the temporal gaps in Sentinel-2 imagery made it difficult to determine the exact harvest event when multiple harvests took place during the year. For the detection of snow events, the weather conditions during the measurement must be stable. Especially for Sentinel-2 imagery, a snow event lowers the availability of images augmenting the usefulness of alternative sensors, such as SRS or Phenocam. Later snow events are particularly difficult to characterize by remote-sensing images due to a rapid snowmelt process caused by the elevated temperatures later in the year. Especially Phenocams in combination with SRS sensors have the appropriate temporal characteristics to assess snowmelt and snow patches on grassland sites [49
The spectral resolution of the instruments limited the research, arguably due to the lack of computable and comparable VIs from each sensor as well as the diverse or poorly documented length of the Phenocam spectral bands. The SRS is able to reproduce only NDVI and PRI indices. Phenocams can be used for generating more VIs, such as GCC or EXG, but are less accurate when it comes to combining both images (RGB and RGB+IR) in one single index (e.g., NDVI). On the other hand, Sentinel-2 is able to create a wide variety of indices and the spectral resolution of the spectrometer allows the creations of most of the vegetation indices aside the ones requiring thermal information.
4.3. Potential for Combined Sensor Appoach
The results of the analyses demonstrate that the NDVI signals have a high correlation among each other, favoring their combination in a multisensor approach. In general, the signal acquired from diverse sensors are comparable and offer the opportunity to fill data gaps or as a source of calibration among each other. Despite the elevated Sentinel-2 MSI, Phenocam, and SRS correlation among time series during the year, the detection of crucial phases or short-term changes in vegetation growth, offset in the NDVI signal or differences in their signal saturation, are strongly sensor and site dependent [46
]. The analysis shows that the short-term dynamic changes of the vegetation are not equally represented by the optical sensors of choice, differing by management practice, extent and date, revisit frequency, or time gaps. Grassland sites with clear cutting events, such as Vimes1500 and Domef1500, are represented equally well by Sentinel-2 MSI, SRS, and Phenocam (aside the first cut on Vimes1500). The sites on 2000 m a.s.l.—being diverse in their vegetation growth, grassland composition, and management activities—show, however, that one single sensor lacks the capability to evidence all changes. Examples are the detectability of the cutting event on Domef2000 as well as the unstable detectability of the multiple harvesting events on the Vimef2000 site. The detection of snow has proven to be difficult for Sentinel-2 MSI, SRS, and Phenocam sensors. This result was from a lack of images or the inconsistent signal measured by the sensors during phases of snowfall and snow acquisition. Despite clear signal changes being visible, they do not tend to minor (as we would have expected) or elevated NDVI values, but signal an incoherent NDVI value. A common property is that each NDVI signal during and after a snow event suffers a period of less NDVI signal growth (e.g., snow period at the beginning of May on the Domef1500 site). Nevertheless, this is not equally representable throughout the stations since the grassland vegetation onset is diverse on 1500 and 2000 m a.s.l. as well as snow amount and frequency divergences depending on the site. This indicates that a multisensor approach for monitoring mountainous grassland vegetation must address the specific regions of interest as well as their topographic properties and management activities. It should also include the possibility to compute different indices, such as the NDSI, better suited when analyzing or masking the presence of snow periods.
The calibration of the sensor specific signal has to be addressed as one of the most pressing issues when considering a multisensor approach. The comparison of the spectral signatures requires a dedicated a-priori calibration, especially for Phenocam imagery, whose spectral properties make a computation of the NDVI index less comparable to the other sensors. A simple correction of one sensor integrating the spectral signal of another is considered insufficient. Especially, short-term changes are not represented equally throughout the sensors. The combination of NDVIs from the sensors would result in either a lack of knowledge of short-term changes or an erroneous correction of one signal. Examples are the first snow period in Domef1500, the first harvest in Vimes1500, or the single cutting events in Vimef2000. Additionally, the inconsistent increase of the NDVI becomes evident during the growth period on Domef2000 (SRS, Phenocam) or the period between the harvests on Domef2000 (S2 vs. Phenocam or SRS).
This study renders a detailed overview of the behavior of the NDVI acquired by different sensors over diverse spatiotemporal scales. Even though SRS, Phenocams, and Sentinel-2 MSI were able to trace the vegetation growth in a similar manner, we conclude that the same index can have notable differences both within a grassland site and among sensors depending on spatial and temporal resolution and heterogeneity of the targeted grassland. We demonstrated sensor-specific NDVI signal offset as well as different increases in NDVI signals during the growing period in spring. Simultaneously, short-term dynamic changes were represented differently by the sensors in terms of NDVI intensity changes, the management process, and extensiveness as well as the site location (especially concerning the altitude). We demonstrated that the combination of the NDVI by multiple sensors enhanced the possibility for detecting short-term dynamic changes throughout the year for each of the stations.
In the future, it seems promising to unify the spectral response acquired on multiple scales within a standardized workflow, enhancing the characterization of grassland dynamics within mountainous regions. Standardization efforts as formulated by the OGC are useful to collect, sort, store, access, and process data from different sensors in a unified way. Furthermore, the inclusion of other biophysical parameters (e.g., LAI or fAPAR) derived from optical sensors and biophysical models could lead to an augmented detectability of dynamic changes of grassland and enable the assessment of its status, stress levels, phenological stages, or biomass increase throughout the year. At the same time, the synergetic use of other optical sensors or microwave remote sensing indicates great potential in future vegetation analyses, reducing spatial uncertainties or temporal gaps that occur individually among sensors. Especially for mountainous vegetation, factors, such as elevation, topography, sun geometry, or the presence of clouds as well as cloud shadows, impact the optical response of grasslands. Therefore, it is advisable to include corrections of the inclination angle or the correction of topographic effects, such as BRDF, for the analysis of alpine vegetation as well as detailed detection of cloud, cloud shadow, and haze.