The spectral variability hypothesis does not hold across landscapes

Abstract

One of the biodiversity metrics to track from space is the spatial variability in reflectance that has previously been proposed as a proxy of species counts per unit area. The corresponding hypothesis is known as the spectral variability hypothesis (SVH). Little attention has been paid so far to the questions whether the SVH holds over broader regions and across time. Here, we addressed these questions by using a spatially contiguous dataset of vascular plant species occurrences in Southern Germany along with MODIS data at 14 time steps. The floristic dataset consists of species occurrence data for 815 areas of 10 longitudinal by 6 latitudinal minutes (approximately 12km by 11km, referred to as mapping units). The spectral variability in space (within these units) was determined using MODIS pixels of 0.5km by 0.5km. We used two different measures of spectral variability in combination with a moving window approach to derive statistical links between spectral variability and species counts through space and time. The moving windows consisting of several mapping units were shifted in space and meanwhile used as target areas for correlation analyses.The performance of the spectral variability to predict species counts was influenced by the location and the extent of the reference windows. In some regions, high spectral variability was connected to high species counts. In other regions, comparably low spectral variability was linked to high species counts and vice versa. Furthermore, the relation between spectral variability and species varied with season. Certain areas changed from almost no correlation to very high correlation depending on the applied scene. Also, the applied spectral variability measure had a notable influence on the observed results.Based on these results, we conclude that the spectral variability hypothesis does not hold across landscapes at this spatial grain. Using spectral variability alone as a proxy for species counts in a monitoring approach for larger extents and grains is therefore unlikely to work in many parts of the world. This does not mean that it cannot help as a covariate in analyses with more predictors.