Tree species classification using hyperspectral imagery: a comparison of two classifiers

The use of hyperspectral and LiDAR data for tree species classification can be a useful and efficient tool of forest managers for planning and monitoring purposes. The first objective of this research was to create a high accuracy species map to be used in a tree inventory of Muir Woods National Monument and Kent Creek Canyon in Marin Countv. California. Secondly, a comparative classification approach was used to test the performance of two non-parametric classifiers, support vector machines (SVM) and random forest (RF), in classifying eight tree species, including old growth redwood, found in the steep terrain of the study area. A minimum noise fraction transform was applied to the hyperspectral imagery to reduce data dimensionality and a LiDAR derived canopy height model provided the basis for the object-oriented classification. The influence of training sample size and segmentation size on the classifications was also explored further. Both classifiers were compared together and individually. SVM outperformed RF in overall accuracy (OA) in all comparisons, however, the statistical significance of the improved accuracy was varied. All classifications resulted in high overall accuracies above 90%.