An individual tree crown delineation method based on multi-scale segmentation of imagery

Abstract

A forest consists of multi-scale branches, tree crowns, and tree clusters. Similar to small tree crowns in shape and scale, branches normally cause over-segmentation of imagery when a watershed segmentation approach is used to segment imagery for tree crown delineation. In order to eliminate such over-segmentation, a new method for individual tree crown delineation from optical imagery was proposed based on multi-scale filtering and segmentation in this study. In this method, the dominant sizes of tree crowns are first determined; Gaussian filters are designed to fit the three-dimensional radiometric shapes of multi-scale tree crowns; the grayscale image is smoothed using the Gaussian filters and segmented using the watershed segmentation approach; and finally, the resulting multiple segmentation maps are integrated together to generate a tree crown map. In an experiment on aerial imagery of forests consisting of multi-scale tree crowns, the proposed method yielded high-quality tree crown maps.