Tree species discrimination in temperate woodland using high spatial resolution Formosat-2 time series

Assessment and mapping of the tree species distribution is an important technical task for forest ecosystem services and habitat monitoring. Since traditional methods (e.g. field surveys) used for the mapping of the tree species tend to be time consuming, date lagged and too expensive, a technology of remote sensing might potentially offer a practical solution for the problem of tree species mapping, especially over large areas. The main purpose of this study was to investigate the potential of Formosat-2 multi-spectral image time series for classification of the tree species in temperate woodlands. Since phenological variations might increase spectral separability of the trees species, additional aim of the study was to assess the possibility of using multispectral-image time series as an alternative to hyper-spectral data for forest type mapping. Noise from the Formosat-2 images was removed with the Whittaker smoother algorithm, which performed quite well although some additional work might be needed during the selection of the optimal regularization parameter. Several supervised classification methods, Support Vector Machines (SVM), Random Forest (RF) and Gaussian Mixture Model (GMM), were used to discriminate tree species from the image time series. All of the classifiers performed reasonably well, with classification accuracies from 88.5 % to 99.2 % (Kappa statistic), although SVM model was the most accurate, while GMM was the most efficient in terms of computing time. High classification accuracy also indicated that the multi-spectral image time series and remote sensing might be a useful method for the mapping of tree species.