COMPARISON OF INDIVIDUAL TREE CROWN DETECTION AND DELINEATION METHODS

Efficient forest management increases the demand for detailed, timely information. High spatial resolution remotely sensed imagery provides viable sources and opportunities for automated forest interpretation at an individual tree level. Recent research, which aims at providing tree-based forest inventory measurements, has considered automatic individual tree crown detection and delineation. A range of algorithms have been developed for different types of images, tested on different forest areas and evaluated using different methods of accuracy assessment. However, no research exists that compares the performance of these methods using a common dataset and the same evaluation approach. In this paper, we compared the performances of three algorithms representative of current published methods for tree crown detection and delineation. The three algorithms—marker-controlled watershed segmentation, region growing and valley-following—were tested on Emerge natural color vertical aerial image with 60 cm ground sampled distance (GSD) of a softwood study site and a hardwood study site. Overall, producer’s and user’s accuracy were applied in segmentation evaluation. While forest stand density and variation in tree crown size influenced performance, the results demonstrated that all three algorithms effectively delineate the Norway spruce tree crowns in the softwood stand, with the region growing method obtaining the best overall accuracy. However, no algorithm proved accurate for the hardwood stand. This analysis suggested that each algorithm has advantages and limitations based on stand characteristics. Future research is needed to explore adaptive algorithms that are capable of accurately delineating crowns in stands where trees vary in size and density.