Tree crown delineation and tree species classification in boreal forests using hyperspectral and ALS data

Tree species classification accuracy at the individual tree crown (ITC) level depends on many factors, among which in this paper we analysed: i) the remote sensing data used for the ITC delineation process carried out prior to the classification, and ii) the pixels considered inside each ITC during the classification process. These two factors were analysed on the ITC level classification accuracy of boreal tree species (Pine, Spruce and Broadleaves), considering two remote sensing data types: hyperspectral and airborne laser scanning (ALS). ITCs were delineated automatically on ALS and on hyperspectral data. A manual ITCs delineation was used as reference in the analysis. The pixel level classification was performed on the hyperspectral bands using a non-linear Support Vector Machine. The classification at ITC level was obtained by applying a majority voting rule to the classified pixels confined by each ITC. The results showed that ITCs automatically delineated from hyperspectral data were usually smaller than those from ALS, and the tree detection rate for hyperspectral data was much lower compared to ALS data (28.4 versus 48.5%). Regarding the classification results, using only manually delineated ITCs a kappa accuracy of 0.89 was obtained, while using only automatically delineated ITCs from hyperspectral or ALS data reduced the kappa values to 0.79 and 0.76, respectively. Slightly different results were achieved using semi-automatic approaches based on both manual and automatically delineated ITC (0.81 and 0.74, respectively). A selection of only certain pixels inside each ITC improved the classification accuracy from 1 to 7 percentage points. A selection based on the spectral values of the pixels was found more influential than one based on the ALS-derived canopy height model. The best results were obtained after a selection based on the spectral values in the bands in the blue region of the spectrum using either the Otsu method or an ad-hoc percentile-based thresholding method