Differentiating plant species within and across diverse ecosystems with imaging spectroscopy

- We classified dominant plant species with AVIRIS data in 5 diverse ecosystems.
- LDA & MESMA mapped species to a high degree of accuracy in all but one ecosystem.
- Spectral dimension reduction generally improved results, particularly for MESMA.
- We achieved 70% overall accuracy for 65 classes when all ecosystems were combined.
- Our results demonstrate a sensor like HyspIRI could map species at broad scales.

Imaging spectroscopy has been used successfully to map species across diverse ecosystems, and with several spaceborne imaging spectrometer missions underway (e.g., Hyperspectral Infrared Imager (HyspIRI), Environmental Mapping and Analysis Program (EnMAP)), these data may soon be available globally. Still, most studies have focused only on single ecosystems, and many different classification strategies have been used, making it difficult to assess the potential for mapping dominant species on a broader scale. Here we compare a number of classification approaches across five contrasting ecosystems containing an expansive diversity of species and plant functional types in an effort to find a robust strategy for discriminating among dominant plant species within and across ecosystems. We evaluated the performance of combinations of methods of training data selection (stratified random selection and iterative endmember selection (IES)), spectral dimension reduction methods (canonical discriminant analysis (CDA) and partial least squares regression (PLSR)) and classification algorithms (linear discriminant analysis (LDA) and Multiple Endmember Spectral Mixture Analysis (MESMA)). Accuracy was assessed using an independent validation data set. Mean kappa coefficients for all strategies ranged from 0.48 to 0.85 for each ecosystem. Maximum kappa values and overall accuracies within each ecosystem ranged from 0.56 to 0.90 and 61-92%, respectively. Our findings show that both LDA and MESMA are able to discriminate among species to a high degree of accuracy in most ecosystems, with LDA performing slightly better. Spectral dimension reduction generally improved these results, particularly in conjunction with MESMA. Within each ecosystem, both the number and identities of functional types present, as well as the spatial distribution of dominant species, played a strong role in classification accuracy. In a pooled ecosystem classification, using CDA and LDA, we discriminated among 65 classes with an overall accuracy of 70% for the validation library, using only a 6% training sample. Our results suggest that a spaceborne imaging spectrometer such as HyspIRI will be able to map dominant plant species on a broader scale.