Conifer species recognition: An exploratory analysis of in situ hyperspectral data

Abstract

In situ hyperspectral data measured above sunlit and shaded sides of canopies using a high spectral resolution radiometer were analyzed for identification of six conifer tree species. An artificial neural network algorithm was assessed for the identification purpose. Linear discriminant analysis was compared with the neural network algorithm. The hyperspectral data were further processed to smoothed reflectances and first derivative spectra and were separately used in tree species identification. Tree species recognition with data collected from six study sites was tested in seven experiments. The average accuracy of species recognition was obtained at every site. The overall performance of the neural network algorithm was better than that of linear discriminant analysis for species recognition when the same number of training samples and test samples were used. The discriminant analysis produced better accuracy than neural network at one site where many samples (10) were taken from six individual trees. Use of the average spectra of all samples for a particular tree species in training may not result in higher accuracy than use of individual spectral samples in training. Use of sunlit samples alone resulted in an overall accuracy of greater than 91%. The effects of site background including illuminating conditions on tree species spectra were large. Neural networks are sensitive to subtle spectral details and can be trained to separate samples from the same species at different sites. Our experiments indicate that the discriminating power of visible bands is stronger than that of near-infrared bands. Higher recognition accuracies can be obtained in the blue to green or the red-edge spectral region as compared with four other spectral regions. A smaller set of selected bands can generate more accurate identification than all spectral bands.