Individual Tree Crown Delineation Using Multi-scale Segmentation of Aerial Imagery

With the development of remote sensing techniques, parameters of individual trees for forest inventory can be extracted efficiently from high-resolution remote sensing imagery or LiDAR (light detection and ranging) data rather than using field surveys [1]-[4]. As a prerequisite step, individual tree crown (ITC) delineation from highresolution imagery or LiDAR data is one critical issue in current forest study. Many ITC delineation algorithms were developed, such as the valley-following [5], region growing [6], edge detection [7]-[11], template matching [12], [13], watershed segmentation [14], [15], and 3D model-based methods [16]. In most of ITC delineation algorithms, treetops are first detected and then used as reference points for crown delineation, and it is typically assumed that a treetop is a radiometric maximum and is geometrically adjacent to the center of the corresponding tree crown. However, it is difficult to find an appropriate low-pass filter to extract all of the trees of different sizes simultaneously. Even if the filter size varies with tree heights, the low correlation between tree height and crown width [15] makes it difficult to obtain accurate treetops. Based on the similarity between the tree crown surfaces of forests and terrain surfaces, the watershed segmentation algorithm is used to segment monochromic images for ITC delineation. Since the traditional algorithm typically yields serious commission errors, mark-controlled watershed algorithm is widely used. Objects in a forest, such as branches, small trees, large trees, and stands, are of multi-scale. Although it is difficult to extract the different-size objects of interest simultaneously, the objects can be separated to multi-scale layers, in each of which the objects have a similar size. Then, the extracted objects of different scales are linked to form a hierarchy. Finally, the trees of different sizes are merged into a thematic map. This proposed ITC delineation method is based on Watershed segmentation and Multi-scale analysis techniques and thus called WM.