Comparison of the Different Classifiers in Vegetation Species Discrimination Using Hyperspectral Reflectance Data

Feature selection methods play an important role in Hyperspectral Remote Sensing applications, especially in classification. This paper proposed a new Feature selection strategy for Hyperspectral dataset. This strategy was designed to help refine vegetation classification of 4 categories with 13 species vegetation which are the most common species in central China. An ASD field spectrometer (Analytical Spectral Device) was used to collect spectrum information of plant leaves from each species through 400 nm to 900 nm with 1 nm spectral resolution. Firstly, correlation between the physical/chemical characteristics of the leaves and the separability of each vegetation species was tested. Then, two feature selection methods, spectral angle and spectral distance, and the feature parameters extracted from spectral curves (FPESC) were used to build the feature space which would be the input space for the classifiers. At last, two linear classifiers, mahalanobis distance (MDC), and fisher linear discriminate analysis (FLDA), and a quadratic classifier, maximum likelihood (MLC), were used for vegetation species refine classification. The results showed that (1) there were no significant differences among 13 species on the leaf dry weight (physical parameter) and leaf chlorophyll content (chemical parameter); (2) FPESC of 13 species have distinctive differences and could be ideal features to discriminate these species; (3) The linear classifiers, MDC and FLDA, have better classification results in the experiments compared to the quadratic classifier MLC, where MDC has the highest classification accuracy which is above 96.2 %.