Subtropical tree recognition with hyperspectral data analysis in Hong Kong

Abstract

Subtropical environment is characterized by a great diversity of flora that is increasingly vulnerable to human impact. The availability of hyperspectral data is going to insert new insight for environmental studies in this environment. In this study, subtropical tree species were studied based on a controlled experiment. Spectra of 25 tree species were collected from which the original spectra together with their first and second derivatives were used for tree species recognition. Spectra were collected in four seasons using a high resolution spectrometer taking data in 689 bands from 400-900 nm. The data were filtered to 138 bands for tree recognition using both linear discriminant analysis and a feed forward neural network with back propagation training. Neural network performed better but was relatively inefficient due to long computation time. Original spectra produced high overall accuracy than the derivative data. Results from stepwise discriminant analysis showed that hyperspectral data collected in winter and autumn were able to attain an overall accuracy of over 85% and were significantly better than those taken in spring or summer.