Accurate knowledge of plant species seasonal and inter-annual distributions are required for many research and management agendas that track ecosystem health. Airborne imaging spectroscopy data have been successfully used to map species, but often only in a single season due to data availability. During California’s severe drought, NASA’s Hyperspectral Infrared Imager (HyspIRI) preparatory airborne campaign flew a visible near infrared/shortwave infrared (VSWIR) imaging spectrometer and a thermal infrared (TIR) multi-spectral imager providing the opportunity to improve species discrimination over a broader temporal range. Here we evaluate: 1) the capability of VSWIR and VSWIR + TIR spectra to discriminate species; 2) the accuracy for seasonal and yearly discrimination among species; and 3) the potential of a multi-temporal library for species classifications. Imagery was acquired in the spring, summer, and fall of 2013 - 2015 spanning from Santa Barbara to Bakersfield, CA with the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) and the MODIS/ASTER Airborne Simulator (MASTER) instruments. Single-date and multi-date spectral libraries were created from AVIRIS (224 bands from 0.4 - 2.5 μm) and AVIRIS + MASTER (5 bands from 7.5 – 12 μm). We used canonical discriminant analysis (CDA) as a dimension reduction technique and then classified species using linear discriminant analysis (LDA).
Results/Conclusions
Overall, our results show that plant species can be discriminated and classified with high accuracy throughout the California drought. However, the additional five TIR spectral bands do not improve overall image classification. Some species were classified with higher accuracies with AVIRIS + MASTER, but for most species the five bands do not contain spectral information that significantly improves discrimination. Secondly, overall classification was fairly uniform between seasons with accuracies ranging from 84 – 93%. However, individual species classification varied much more between dates with accuracies ranging from 10 – 78%. These results show that while overall image classification across seasons is accurate, classification performance may not be sufficient for applications that focus on a specific species of interest. Lastly, our results show that compared to single date libraries the multi-temporal spectral library decreased overall classification accuracies by 11- 21%. Due to the drought effect on vegetation spectra, the multi-temporal library contained too much spectral variation to yield accurate classification accuracies across all seasons between 2013 and 2015. This research contributes to efforts aimed at monitoring ecosystems across large spatial and temporal scales and ultimately supports many research agendas that are tracking ecosystem health and changes.