Optimal structural and spectral features for tree species classification using combined airborne laser scanning and hyperspectral data

In this paper, the performance of feature selection in tree species classification based on multi source earth observation data was studied. We applied a sequential forward floating feature selection on imaging spectroscopy (IS) and airborne laser scanning (ALS) data, as well as their combination. Qualitative comparison of the fused results shows that the selected spectral features are more distributed across the spectrum, in contrast to an accumulation of features in the near infrared region when using IS alone. A support vector machine (SVM) classifier was used for quantitative comparison of the different datasets. Assessing the classification accuracies confirmed the superiority of the selected subset of spectral and structural features compared to using all available features (improvement of > 7% in kappa accuracy).