Estimating the spatial distribution of foliar pigments and canopy structural components with remote sensing can serve as an important approach for monitoring plant community characteristics, as spatially explicit measurements of vegetation biochemical and biophysical variables can provide insight into ecosystem composition, processes, and/or disturbance caused by changing environmental conditions. Vegetation monitoring efforts in Arctic regions have been mostly accomplished with nadir-looking broadband instruments, thus leaving multi-angle, spectroscopic retrievals of vegetation biochemical and biophysical variables largely unexplored.
Using field and spaceborne (CHRIS/PROBA) multi-angle spectroscopy, the performance of various modelling techniques was compared for retrieving biochemical and biophysical variables from tundra vegetation situated across a bioclimatic gradient in the Western Canadian Arctic. Specifically, empirically-based multi-band and predefined narrowband vegetation indices (VIs), a machine learning regression algorithm (Gaussian processes regression, GPR), and a physically-based radiative transfer model (PROSAIL) were compared for their capability of retrieving leaf chlorophyll content (LCC), plant area index (PAI), and canopy chlorophyll content (CCC) from multi-angle, multi-scale, high-resolution canopy reflectance data. Reference data for these variables were acquired through laboratory and field-scale leaf and canopy measurements.
Iterative empirical models were the most effective for retrieving LCC, PAI, and CCC irrespective of view angle and spatial scale (p<0.05). GPR produced the best correlation-based modelling results (cross validated r2cv=0.59), however, a multi-band vegetation index (i.e. simple ratio, SR) was shown to provide statistically comparable results while providing a more simplistic methodological approach (r2cv=0.55). Furthermore, SR produced statistically superior (p<0.05) normalized prediction accuracies over GPR (NRMSE=0.13 vs. NRMSE=0.16). Empirically modelled band selections showed that variable covariation is an important consideration when constructing reflectance models used for vegetation variable retrievals in the Arctic, and thus it was concluded that spectroscopic remote sensing provides benefits for such tasks. The overall conclusion drawn from the compiled empirical and physical modelling results, when examined across the field and remote sensing scales, was that a multi-angle approach does not provide a statistically significant advantage over a nadir approach for retrieving LCC, PAI, or CCC in Arctic environments (p>0.05).