Using hyperspectral data to classify vegetation at the plant functional type-level in mountain terrain at three spatial resolutions

Hyperspectral data covering the visible, near infrared, and shortwave infrared portions of the electromagnetic spectrum have been used to map vegetation at the plant functional type and species levels in a variety of ecosystems. Vegetation maps are basic to the study and analysis of natural resources and are an important component in documenting and understanding the impacts of environmental changes in ecosystems due to human-induced changes (e.g., climate change). Multiple methods for mapping vegetation have been developed to take advantage of the large number of bands (> 200) and spectral contiguity that hyperspectral data provide. The purpose of this study is to determine what the limitations and potential areas of success are for using hyperspectral remote sensing data to classify vegetation cover in steep mountainous terrain at the plant functional type level. This research utilizes multiple endmember spectral mixture analysis (MESMA), a modified version of Spectral Mixture Analysis (SMA), to classify plant functional type within an Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) scene acquired over the Wasatch Mountain Range just east of Salt Lake City, Utah on August 5, 1998. The AVIRIS scene was originally acquired at 20 m spatial resolution and was resampled to 40 m and 60 m spatial resolution to examine the ability of MESMA to classify plant functional types at coarser spatial resolutions. Endmembers were extracted and selected from each spatial resolution AVIRIS scene so that three sets of endmembers were available to attempt to classify the three AVIRIS images. The 20 m extracted endmembers were used to classify the 40 m and 60 m AVIRIS images to examine the effects of coarsening the spatial resolution of the image while using a higher resolution set of endmembers. When used to model the AVIRIS images, the 20 m extracted endmembers classified the three resolutions with good success: overall accuracies of 85.65% at 20 m, 85.56% at 40 m, and 83.67 at 60 m. The spectral libraries extracted from the 40 m and 60 m AVIRIS images produced more confusion between classes, resulting in less accurate classifications at these resolutions. This endmember confusion was also responsible for an increase in the percentage of unclassified pixels at the 40 m and 60 m resolutions. The results provide insight into the ability of future spaceborne hyperspectral sensors, which are likely to operate at coarser spatial resolutions, to map vegetation.