Feature assessment in object-based forest classification using airborne LiDAR data and high spatial resolution satellite imagery

Abstract

The last decade has witnessed an increase in interest in the application of airborne LiDAR data and high spatial resolution satellite imagery for forest structure modelling, tree species identification and classification. The integration of LiDAR data and WorldView-2 satellite imagery produced different combinations of input data layers for image segmentations and a large number of variables derived from these data layers for object-based classifications. Assessment of different features (including the input data layers and subsequently derived variables) for object-based forest classification is important. In this study, five image segmentation schemes were explored to test the effectiveness of the different input data layers, in particular, the new WorldView-2 multispectral bands to object-based forest classification. Object-based variables derived from these data layers were assessed to rank their importance before inputting into decision trees for forest classifications. It demonstrated that, using methods developed in this study, the integration of airborne LiDAR and eight WorldView-2 bands can significantly improve the accuracy of forest classification in our study area. The variable importance was ranked, indicating how important a variable contributes to the classification in a particular decision tree. The results showed that using LiDAR data alone or four conventional bands only, the overall accuracies achieved were 61.39% and 61.42% respectively, but the overall accuracy increased to 82.35% when all eight bands and the LiDAR data were used.