• Vermillion, Stephanie C.
  University of Maine Graduate School
  

Recent developments in remote sensing technology have provided foresters with a new generation of sensors. The Airborne Visible/Infared Imaging Spectrometer (AVIRIS) provides higher spectral resolution, compared to broad band sensors, resulting in mo information for each pixel in a digital image. Discreet differences in spectral reflectance among the narrow wavebands may be exploited to increase the accuracy of forest type mapping, possibly on a species level.

The research reported in this thesis was performed in two parts. The first chapter examines the ability of transformed AVIRIS data to distinguish spectral differences between species types for forest classification. Three types of transforms were applied to AVIRIS data: principal component analysis, minimum noise fraction, and wavelets. Supervised classification using maximum likelihood decision rule was performed on each data set and the accuracy for identifying forest types was assessed. Due to confusion between morphologically similar forest types, traditional accuracy assessment methods were determined to have shortcomings in evaluating the forest type classifications. A fuzzy accuracy assessment, in addition to the traditional accuracy assessment, was performed on all three data sets. Results of the accuracy assessment showed three important findings. First, there was no significant difference between any of the classifications (using transformed and non-transformed data) for the identification of northern forest types. Second, the fuzzy accuracy assessment was a useful technique for evaluating forest type classifications and errors in classification, providing an improvement of the overall accuracy for each of the data sets examined. Finally, AVIRIS data produced slightly higher forest classification accuracy compared to Landsat TM for the Orono, Maine study site.