Understanding limits of species identification using simulated imaging spectroscopy

Imaging spectroscopy is a powerful tool for mapping and monitoring the spatial distribution of species compositions. Most spectroscopy studies rely on extensive field campaigns to assess classification accuracies against ground-truth samples. Their findings are correlative and not based on first principles and by design they often do not provide for an end-to-end traceability of error budgets. In this study, we present a simulation approach for modeling canopy-level reflectance using three-dimensional plant models arranged on a synthetic landscape. This type of simulation offers an important avenue for understanding the interplay between instrument parameters and landscape characteristics for different ecosystems. The simulation approach was applied using species-level leaf reflectance data from the biodiversity hotspot in the Greater Cape Floristic Region (GCFR) of South Africa. We simulated 140 random realizations, each of 20x20m in size, encompassing 10 to 70 species to understand how changing the diversity and image resolution (pixel size) affected spectral mixing and the ability to discern species within the plot. Overall, across all resolutions and species mixtures, the mean classification accuracy was 0.690±0.184 (mean F1-score±standard deviation across realizations), however, this varied widely from 0.812±0.057 at a pixel size of 60 cm for landscapes with 30 species to 0.196±0.106 for landscapes with 70 species and a pixel size of 240 cm. As expected, classification accuracy decreased with increasing spatial grain from 0.763±0.064 at 15 cm resolution to only 0.414±0.214 at 240 cm. Classification accuracies across different levels of species richness remained fairly constant except in the coarsest grain size (240 cm). This suggests an increasing challenge in species identification as grain sizes become coarser than the constituent organisms. In summary, this approach can facilitate detailed simulation of reflectance while exploring various ecological scenarios such as changing species composition or the impacts of environmental variability such as drought.