Detection of harvested trees in forests from repeated high density airborne laser scanning

Abstract

Identification of harvested and fallen trees is a prerequisite for the detection and measurement of changes in forests. This paper presents a three step approach to monitor harvested and fallen trees based on direct comparison of repeated high density airborne LIDAR data. In a first step differences between data sets are obtained from a point to point comparison, such that the data can be reduced to the deviating points only. Secondly, the resulting points are clustered into spatially connected regions using region growing. Finally, individual trees are extracted from the clusters by analysing their relative proximity and by analysing geometric properties of points in the clusters. Two data sets, acquired at a four year interval and covering a forest with mainly deciduous trees, are compared. First results show that most points relating to a change can be extracted and that clustering of these with region growing enables us to efficiently separate harvested and fallen trees from the remaining trees. Grouped harvested trees could not be separated using the region growing approach due to touching crowns. Segmentation of these using spectral clustering however identified individual regions well, but the results depend mainly on the pre-defined number of clusters. Crowns of grouped trees can be therefore separated if the number of trees is known.