Authors: Lee Ann Nolan*, West Virginia University, Tim Warner, West Virginia University, Aaron Maxwell, West Virginia University
Topics: Remote Sensing, Land Use
Keywords: classification, shrub, shrubland, random forest, SVM, NLCD
Session Type: Paper
Presentation File: No File Uploaded
Accurate classification of shrubland using satellite imagery is becoming increasingly important to conservation management due to continued declines in both total shrubland area and shrub-dependent species’ populations. The 2011 National Land Cover Dataset (NLCD) legend defines the shrub class as dominated by shrubs less than 5 meters tall with shrub canopy typically greater than 20% of total vegetation. This definition indicates that unlike other classes, shrub cover can be the minority of the pixel. The research objective is to examine the impact of shrub cover proportion on the classification accuracy of shrubland.
June 2014 GeoEye-1 imagery for a 577 km2 area in Pocahontas County, WV, was resampled to 8, 16, and 32 m resolutions. Vegetation data was acquired through 323 random and non-random plots with four 11.3 m transects radiating outward with 5 equidistant sample points. The percentage of each of three vegetation classes (shrub, forest, grass) found within a pixel was calculated. Four vegetation classifications were created with shrub proportion thresholds of 20%, 40%, 50%, 60%, and 80%. Each vegetation classification was merged with randomly sampled points of water and impervious surfaces creating an “other” class to create the reference data set. The image was classified using random forest and support vector machine classifiers.
A clear trend emerged where shrub producer’s and user’s accuracies decreased with increasing proportion of shrubland regardless of spatial resolution and classifier, although overall accuracy remained relatively steady. These results indicate that shrub proportion plays a strong role in the classification accuracy of shrubland.