Tree Species Classification on Hyperspectral Imagery Using Fewer Training Samples

Abstract

The distribution of tree species within the forest holds significant importance for forest management. Since field surveys in the forest are time-consuming and cost-expensive, automatically extracting tree species distribution maps from remote sensing imagery becomes a trend. For tree species classification using hyperspectral imagery, many existing classification methods require a large number of training samples to achieve high classification accuracy. However, the classification accuracy will decrease rapidly if only a few hundred training samples are used. Given the challenges and expenses associated with collecting abundant training samples in the forest, there is a need to explore methods that achieve good classification performance with a limited number of training samples. In this paper, a classification scheme combining SuperPCA and Active Learning (AL) is proposed to improve the tree species classification using a limited number of training samples. SuperPCA is employed to reduce feature dimensions and harness spectral-spatial information within hyperspectral imagery. Active Learning is employed to select informative samples for the training, thus reducing the requirement for training samples. Experiments on a tree species classification data set demonstrate the effectiveness of the proposed classification scheme.