Mitigating impacts of global change on biodiversity is a pressing goal for land managers but understanding these impacts is often limited by the spatial and temporal constraints of traditional in-situ data. Advances in remote sensing address this challenge, in part, by enabling standardized mapping of biodiversity at large spatial scales and through time. In particular, hyperspectral imagery can detect functional and compositional characteristics of vegetation by measuring subtle differences in reflected light. The spectral variance hypothesis (SVH) expects spectral diversity, or variability in reflectance across pixels, to predict vegetation diversity. Especially when assessing herbaceous ecosystems, however, there is inconsistent evidence for the SVH, potentially due to a mismatch between plant size and the traditionally coarse pixels of satellite and airborne imagery or variation in the biological characteristics of the observed ecosystems, such as vegetation structure and composition, which can impact spectral variability. However, the majority of research testing the SVH to date has been conducted in systems with controlled conditions or spatially homogenous assemblages, with little generalizability to heterogeneous real-world systems. Here, we move the field forward by testing the SVH in a species-rich system with high heterogeneity resulting from variable species composition and a recent fire. We use very high spatial resolution (~1 mm) hyperspectral imagery to compare spectrally derived estimates of vegetation diversity with in-situ measures collected in Boulder, CO, USA. We find that spectral diversity and taxonomic diversity are positively correlated only for low to moderate diversity transects, or in transects that were recently burned where vegetation diversity is low and composed primarily of C3 grasses. Additionally, we find that the relationship between spectral and taxonomic diversity depends on spatial resolution, indicating that pixel size should remain a priority for biodiversity monitoring. The context dependency of this relationship, even with high spatial resolution data, confirms previous work that the SVH does not hold across landscapes and demonstrates the necessity for repeated, high-resolution data in order to tease apart the biological conditions underpinning the SVH. With refinement, however, the remote sensing techniques described here will offer land managers a cost-effective approach to monitor biodiversity across space and time.

Data collection for this project built on a long-term grassland monitoring initiative in the city of Boulder, CO, USA, run by their Open Space & Mountain Parks department. The monitoring area covers a 2,000 ha mixed grass prairie in the Front Range of Colorado, USA (39°56 N, 105°12 W). Permanent plots (hereafter, ‘transects’) of 50 m x 2 m were established in 2009 to implement the city of Boulder’s Grassland Ecosystem Management Plan (GEMAP). Transects were surveyed using a panel design for the first 8 years and thereafter, every three years  for vegetation cover and composition. Cover is surveyed with a point intercept method: along the 50 m center line in a transect, species identity (or soil/rock/litter) is recorded at each meter mark at 0.5 m from both sides of the centerline, resulting in a total of 100 points. Additionally, vascular species that are present in the rest of the plot, but not hit during the point intercept method, are noted.

Here, we utilize data from the 17 transects surveyed in 2022, which was collected at peak biomass for this system between July 18 and August 23. Vegetation survey data is summarized as relative vegetation cover, species richness and an abundance-based taxonomic diversity metric, Shannon Index, to protect the location and identity of species according to Boulder OSMP policy. To calculate vegetation cover, we computed the percentage of points along the transect where vegetation was hit by excluding rock, litter, and bare soil. To compute species richness, we calculated the number of species identified in each transect both as those counted with the point intercept method (here, “hits”) and as hits plus those identified as present in the rest of the plot (here, “full species list”). We also computed the Shannon diversity index of both the hits and the full species list. Species in the full species list that were not counted as hits were given an abundance of 0.1 for incorporation in this abundance-based metric.

We also collected proximal imagery of the 17 GEMAP transects using a handheld Specim IQ hyperspectral camera (Spectral Imaging Ltd., Konica Minolta; www.specim.com). Sampling was conducted from August 11th - August 31st, 2022, such that the maximum time between vegetation survey and imagery was 24 days (average time between: 7 days). This hyperspectral camera is equipped with a built-in pushbroom imaging system (a line scan camera system composed of an imaging spectrograph, grayscale camera and objective) that records reflectance in 204 bands spanning the visible to near infrared regions (400 to 1000 nm; 7 nm spectral resolution FWHM) with a 1 mm spatial resolution when mounted on a tripod at 1.5 m from the ground. For each image, we included a white reference material provided by Specim (with a reflectance close to 100%) in the frame of reference. Sampling was conducted on cloud free days and within 3 hours from solar noon. Raw data was transformed to reflectance using the reflectance from the white reference, the dark reference signal, and the imaging integration time.

Images were taken of square meter plots in 2-meter increments along the transect. We ultimately combined pixels from all images within a transect to represent the transect-level data in the manuscript, but here archive the spectra at the "plot" level. For each plot, there is a .dat file representing reflectance in 204 bands across a 512 pixels by 512 pixels. Subsequent processing, including removal of bad bands and filtering for photosynthetically active vegetation, is detailed in the manuscript.