Study Area
Our study area is a managed grassland site located in the administrative district Havelland (Brandenburg) in northeastern Germany. In this administrative district, which is located in the North German Plain, grasslands represent a third of all agricultural land. This makes grassland an important land use in the area (Landkreis Havelland 2018). The study area has a size of approximately 43 ha (Fig. 1) and is a permanent grassland which is used for animal feed production. The study area is in the Havelländisches Luch, which is characterized by drained shallow fen soils with varying peat layer thickness and groundwater levels, causing moist or moderate moist sites within the fields. In the western half of the field, groundwater levels are somewhat higher compared to eastern parts. The canopy is dominated by grasses, mainly Lolium perenne L., Phalaris arundinacea L., Elymus repens L., Alopecurus pratensis L. and Festuca arundinacea Schreb.. Phalaris arundinacea and Festuca arundinacea which have a moisture index of 8 and 7 (Dierschke and Briemle 2008), and are very well adapted to the moist conditions on fen grassland. Under these conditions, the grasses can produce very high biomass with heights from 0.5 to 1.0 m when used for feed but can reach more than 1.5 m at older morphological stages. Furthermore, Phalaris arundinacea is an endemic grass species of Havelländisches Luch (common name: Havelmilitz), which was observed in similar sites to produce up to 5 t DM/ha dry biomass (BM) per cut at a LAI > 8. The other grasses occur rather on the moderate moist parts of the site (moisture index 5–6). They are high yielding but of medium plant height with less than 0.5 m when young and up to 0.8–1 m at flowering. Herbs also appear in the canopy, though in lower proportion than grasses. The herbal plants present are Rumex ssp., species of Taraxacum and Ranunculus as well as of the legume Trifolium. The grass sward is usually mown three times per year, depending on the meteorological conditions during the growing season. The climate of the study region is characterized by continental conditions with warm summer temperatures and moderately cold winters. Based on the long-term average (1961–1990), the climate of the area is characterized by a temperature of 8.6 °C and a precipitation of 521 mm. Compared to the long-term average temperature of 13.6 °C in the main vegetation season (April–October), the years 2018 and 2019 were very warm, as both had an average temperature of 16.7 °C. Additionally, the region suffered from drought during the years 2018 and 2019 (ZALF 2020). Although grasslands are more resilient to drought conditions than other crops, drought-related yield losses are expected to increase in the next decades with ongoing climatic changes across the region (Schindler et al. 2007).
Field Measurements
Field measurements were taken on the 9th of August 2019 approximately within 2 h before and after solar noon. The weather conditions were ideal with high solar radiation and few clouds. To capture the variability of vegetation within the field, we based our sampling design on a S-2 enhanced vegetation index (EVI) cluster map from the 28th of July 2019 with six classes of similar EVI values, which helped to identify appropriate sampling locations on site. Based on this map, 21 sample plots were selected as central measurement locations (Fig. 1). To capture the coordinates of the selected positions, we took GPS measurements with a Garmin Oregon device, and averaged the received signal over several minutes to minimize the positional error (Gao 2006). At the central plot, we took a white reference measurement using an Analytical Spectral Devices (ASD) FieldSpec 2 spectroradiometer that covers the spectral wavelength region from 450 to 2500 nm and subsequently a reflectance measurement of the vegetation (average of five measurements). We measured from a height of approximately 1.20 m leading, with an opening angle of 25°, to a ground sampling resolution of around 0.2 m2. Following the spectral measurement, we measured the compressed sward height (CSH) using a falling plate meter with a size of 0.46 m × 0.46 m (0.2 m2) following the instructions in Rayburn and Lozier (2003). We then cut the vegetation within the area of the falling plate meter approximately 5 cm above the ground, put the grasses in sampling bags with air holes and stored in cool box until processing in the laboratory. The fresh weight of the collected samples was measured in a laboratory before they were dried at 60 °C for 24 h in a convection oven after which the dry weight was measured. Around the central plot, we established an adapted star sampling design (Muir et al. 2011), with four 20-m transects in cardinal directions to cover a representative area surrounding each central plot, on each of which we took measurements on a transect point every 5 m (Fig. 2).
We obtained spectral reflectance measurements (using an ASD FieldSpec 3 spectroradiometer), CSH, and leaf area index (LAI) using the SS1 SunScan Canopy Analysis System, which measures the incoming light from the bottom of the canopy with 64 sensors within a lance of 1-m length. In parallel, the irradiance was measured using a hemispherical sensor on a tripod. Accordingly, we collected data at 17 locations per plot. Due to changing weather conditions, we could not use all measured spectra/LAI measurements at all transect points. Overall, we obtained 207 LAI measurements, 21 biomass samples, 229 CSH measurements and 229 spectra (Fig. 3). Dry biomass (BM) values for each sampling point were estimated based on the CSH measurements using a linear regression (Fig. 3). The spectral measurements were used to simulate S-2 bands using the respective spectral response function (ESA 2017).
Satellite Data
For the model comparison, the S-2 scene was chosen which was recorded closest in time to the time of the LAI measurements. A cloudless scene was acquired by S-2 on 28th of July 2019, 13 days before the field measurements were taken.
The scene was pre-processed using the Vista Imaging Analysis algorithm (VIA, Niggemann et al. 2014; Niggemann et al. 2015). Pre-processing within this processing chain includes an atmospheric correction, a land-use classification, masking of clouds and cloud shadows as well as detection and correction of cirrus clouds. Since the chosen scene was not covered by clouds or cirrus on the pixels within the area of interest no cloud masking or cirrus correction was performed on the pixels within the test site. No Bidirectional Reflectance Distribution Function (BRDF) correction was performed within the VIA, since a BRDF correction is part of the radiative transfer modeling process within SLC. For the analysis, the 20-m and 60-m bands of the S-2 scene were resampled to 10-m resolution. To reduce the effect that neighbouring pixels influence each other, an adjacency correction was applied. In this way, the influence of brightness differences of neighbouring pixels is reduced (Verhoef and Bach 2007; Bach 1995).
Modeling Approaches
We used a random forest regression to relate the simulated S-2 spectra with the LAI and biomass field measurements in an empirical model (EMP). Random Forests are an ensemble of individually trained decision trees aiming to average out modeling errors (Breiman 2001). To randomize the training of the decision trees they are grown with a subset of the input training data and only a pre-defined number of input predictors (mtry) is used at each decision tree node to find the optimal split. The empirical modeling was done in R (R Core Team 2018) using the RandomForest package (Liaw and Wiener 2002) in which we set the number of trees to 500 and tuned the model parameters mtry using the tuneRF function. To get statistically robust results and an estimation of uncertainty, we iterated the modeling approaches for each of the variables (LAI and BM) 100 times with random splits of the input data using 70% of the data for model training and 30% of the data for model validation. Within each iteration model, performances were assed using the coefficient of determination (R2) and the root-mean-square error (RMSE). To make the performance metrics comparable between the two response variables, we normalized the RMSE (NRMSE) to the mean value of the respective set of validation data. The individual variable importance was assessed by calculating the increase in mean squared error (MSE) between the initial model and a model in which the variable to be assessed was permutated (Liaw and Wiener 2002).
As a second method to derive LAI estimates from the S-2 scene, the radiative transfer model SLC was used (Verhoef and Bach 2003, 2007). SLC is a physically based surface reflectance model that evolved from the GeoSAIL model (Verhoef and Bach 2003). Direct and diffuse fluxes of incident and reflected radiation are taken into account while modeling the radiation transfer in a so-called four-stream approach. Input variables for SLC are grouped into four different groups of variables including information about the satellite sensor, the observation geometry at the moment of satellite data acquisition, biophysical and biochemical properties of the vegetation canopy and the properties of the soil below the canopy. A two-layer modernized version of the model SAILH (Verhoef 1985) is used for canopy modeling in SLC; whereas, spectral reflectance and transmittance of green and brown leaves are calculated using the PROSPECT sub-model (Jacquemoud and Baret 1990). To account for soil reflectance characteristics a non-Lambertian soil BRDF sub-model for soil reflectance and its variation with moisture is incorporated in SLC (Verhoef and Bach 2007). SLC models potential reflectance spectra by varying the input parameters. For this study, a look-up-table approach was used to calculate all possible solutions resulting of the variation of the inverted parameters. The parameter combination that best describes the conditions of earth’s surface is chosen by comparing modeled and measured reflectance spectra and choosing the result with the lowest RMSE. The soil input dataset was adapted to resemble the soil characteristics of the study site by measuring reference spectra on adjacent arable fields during a moment when the bare soil was visible and converting the derived S-2-spectra into single scattering albedo values to use as soil-background layer for the model calculations in SLC (Verhoef and Bach 2007). By doing this, the influence of the reflectance characteristics of the soil, that can influence the reflectance spectra of a site when the soil is not entirely covered by vegetation, can be taken into account. Green leaf area, leaf chlorophyll content and the leaf angle distribution within the canopy were inverted. Dry biomass values for each pixel were calculated by multiplying the inverted green leaf area with the leaf mass per area (LMA). Leaf mass per area is a species-specific parameter that also varies with the phenological stage of the plants and with LAI. For this study, the LMA value was calculated from the field data by calculating the ratio between dry biomass values and the measured leaf area for each sample site and averaging the values. The site-specific LMA was then applied to calculate the dry biomass values from the inverted green leaf area. Model performance of SLC is evaluated by comparing modeled to input spectra, where input spectra refer to the surface reflectance measured by S-2.
Validation
To assess the model accuracy, the calculated values were compared to the reference data collected in the field. This was done separately for both modeled parameters LAI and BM. Once the LAI values from sample locations with varying lighting conditions had been excluded, a total of 207 LAI and 228 BM (calibrated using CSH) measurements could be used for validation. Based on the transect sampling design, we matched the point measurements of LAI and BM to the estimated pixel values from the SLC and the EMP. CSH and LAI measurements were taken on the 20-m transect lines with a distance of 5 m between each sample point (Fig. 2). This means that often more than one sample point lies within a pixel of the S-2 scene. Therefore, we averaged all values per plot and compared the mean value for each plot with the averaged modeled pixel values located within a 20-m buffer around the central coordinate of the plot. For this validation, only the 13 plots were used, at which all point measurements were valid (Fig. 4). In this way, we accounted for geometric inaccuracies which might be due to GPS positioning errors and the positional accuracy of the S-2 image.