Unsupervised selection of training samples for tree species classification using hyperspectral data

In this study, we introduced a novel unsupervised selection method for collecting training samples for tree species classification at individual tree crown (ITC) level using hyperspectral data. The selection process is based on a distance metric defined among the spectral signatures of the pixels inside the ITCs, and a search strategy based on the Sequential Forward Floating Selection algorithm. The method was developed using two kinds of samples: plots and ITCs. The experimental results indicated that the method allows reducing the amount of training samples needed for the classification process, without significantly decreasing the classification accuracy. Applying the proposed method, the kappa accuracies obtained using about half of the total number of plots ($hbox{kappa accuracy}=0.84$) and approximately one-third of the total number of ITCs ($hbox{kappa accuracy} = 0.83$) were not statistically different from the results obtained using the full set of training samples ($hbox{kappa accuracy} = {0.86}$). The proposed method demonstrates that using a priori information derived from the hyperspectral data can substantially reduce the amount of field work and, consequently, the forest inventory costs