Advances in imaging spectroscopy are enabling us to map ecosystem properties with greater accuracies, over broader scales. Better accounting for the sources of image spectral variance is required for improving these techniques. Unsupervised image clustering can summarize spectral variance, but interpretation is difficult. Spectral variance is mainly driven by 1) species composition, 2) canopy structure, 3) functional traits, and 4) sub-pixel fractional cover/composition. Here, we define spectral clusters and evaluate the individual and combined impact of these drivers across an elevational/climatic gradient.
We collected data from three forested ecosystems designated as NEON sites in central California. Field data include leaf traits and spectra, spectral signatures for background materials, and plot-level measurements of canopy composition, cover, and structure. Using hyperspectral imagery collected by AVIRIS, and lidar and hyperspectral data collected by NEON’s Airborne Observation Platform, we created maps of the aforementioned drivers.
The number of image clusters was determined using the HySime algorithm. Unique spectral clusters were defined using hierarchical agglomerative clustering with spatial filtering. The number and spatial distribution of clusters were compared across sites and to the maps of drivers. We identified key drivers by partitioning variance among the four. Unexplained variance was analyzed by cluster and spatially.
Results/Conclusions
Our three sites included a low-elevation oak-pine savanna/woodland, a mid-elevation mixed conifer/broadleaf forest, and a high-elevation conifer forest. Initial clustering results show differences in the amount of spectral variance by site, with the fewest clusters (10) found for the low elevation savanna site, 20 for the mixed, mid-elevation site, and 22 for the high-elevation conifer site. Canopy structural type maps show a similar pattern, with the fewest types defined in the low elevation site (4) and a greater number in the other two sites (7 and 8, respectively). Leaf-level estimates of functional traits were quite accurate, with % RMSEP values ranging from 5-11% for a suite of 15 traits. Canopy-level trait and species composition maps are currently underway.
We expect the importance of each driver will vary by ecosystem type and spatially within an individual ecosystem. Likewise, unexplained variance will represent other drivers not accounted for in this study. Our results will provide an improved understanding of the ecological controls on canopy spectral variance, and how imaging spectroscopy data can provide continuous spatial information on ecosystem properties.