Assessing Floristic Composition with Multispectral Sensors
Thonfeld, Frank1; Feilhauer, Hannes2
1Center for Remote Sensing of Land Surfaces (ZFL), University of Bonn, GERMANY; 2University of Erlangen-Nuremberg, GERMANY

Remote sensing has long been used for land cover mapping. Current requirements often go, however, beyond the mere discrimination of land cover types. Highly variable classes such as natural vegetation types are characterized by spatial and temporal patterns. Their assessment leads to subclass information which improves the understanding and characterization of complex ecosystems. Existing maps of high nature value ecosystems, protected areas or areas with limited access often do not provide detailed and contemporary information and thus do not reflect local differences. However, local conditions within one land cover type may be influenced by human activities (e.g., favoring invasive species or degradation indicators), structural attributes such as shrub and grass encroachment, or varieties of a vegetation type caused by slightly different environmental conditions. Spatial information about intra-class variation is important for rangeland management and nature conservation.
Mapping of subtle floristic varieties of vegetation types has been successfully carried out with hyperspectral data. Current research projects explore multitemporal optical and SAR data to assess the dynamics of protected ecosystems. Since studies like these are still rare we address in this study (Feilhauer et al. 2013) the impact of the spectral coverage and resolution of various multispectral sensors on the estimation of floristic varieties of vegetation types.
An empirical study was conducted for three different vegetation types which occur frequently in Central Europe and are thus seen as representative. These three types are 1) a nutrient-poor and dry grassland, 2) a wet heath in transition to mire, and 3) a nutrient rich and wet floodplain meadow.
For each vegetation type the vegetation was sampled in 57, 35 and 37 plots for the nutrient-poor grassland, the wet heath and the floodplain meadow, respectively, using a random sampling design. The cover fractions of all occurring vascular plants and mosses were estimated with a frame. In order to quantify the within-type floristic variation, the records of each type were subjected to an ordination analysis. This ordination aimed to identify the main floristic gradient within each type. The scores of the records on this gradient were subsequently taken as continuous description of subclass variation. Indicator species were identified in numerical analyses to illustrate the floristic variation along the gradients.
Canopy reflectance of the plots was measured multiple times during the vegetation period with a full-range field spectrometer. After pre-processing, the collected spectra were used to simulate pseudo-radiances for 11 multispectral sensors (ASTER, HRSC-X, IKONOS, Landsat-5 TM, Landsat-7 ETM+, LDCM OLI, Quickbird-2, RapidEye, Sentinel-2, SPOT-5, Worldview-2). The simulated data were regressed against the ordination scores with Partial Least Squares regression.
The inter-correlation between neighboring wavelengths of the field spectrometer data revealed five statistically independent spectral regions. Depending on its spectral characteristics, each sensor covers two to five of these spectral regions with one or more bands.
In a first analysis step we tested via the Jeffries-Matusita-Distance (JMD) if the spectral performances of the different sensors allow for the discrimination of the three vegetation types. In a second step their ability to detect intra-class variation was analyzed.
All analyses were based on simplified assumptions since neither spatial resolution nor signal to noise ratio were considered in the simulation process. The main comparison was based on reflectance data taken during the vegetation optimum since this date showed in a previous study the closest relation between reflectance and floristic patterns. The date of vegetation optimum differs between vegetation types and many of the sensors under consideration have short revisit times which allows them to acquire multiple images within the vegetation period. We thus also quantified the additional explanatory power gained by multiseasonal data considering all available time steps of the collected spectra.
The JMD analysis revealed that the spectral separability of the three vegetation types varied between sensors and in time. Sensors with well distributed bands and covering all spectral regions showed best performance (Sentinel-2, LDCM OLI) with Worldview performing well, too. Jeffries-Matusita-Distances were up to 1,91, indicating good separability.
The ordination analysis revealed one prominent floristic gradient per site. Cross-validated model fits for these gradients were high for most sensor simulations. Floristic variation in the nutrient-poor grassland was mainly related to reflectance in the VIS region. In the floodplain meadow, all regions but the SWIR region showed relations to the floristic patterns. Here, too, the closest relations were observed for the VIS and NIR regions. The wet heath model was based on contributions of all spectral regions. Consideration of multiseasonal data slightly improved the fit for the wet heath. For the floodplain meadow, incorporation of multiseasonal data resulted in a decreasing PLSR-model fit due to spatio-temporal heterogeneities of the local vegetation conditions. Taken together, the potential of multispectral sensors to discriminate vegetation types and to assess their intra-class floristic varieties could be shown. From a spectral perspective, coverage of the full solar electromagnetic spectrum is mandatory to reliably assess floristic gradients.

Feilhauer, H., Thonfeld, F., Faude, U., He, K.S., Rocchini, D., Schmidtlein, S. (2013) Assessing floristic composition with multispectral sensors - a comparison based on monotemporal and multiseasonal field spectra. International Journal of Applied Earth Observation and Geoinformation 21, 218-229.