A Comparison of Four Methods for Automatic Delineation of Tree Stands from Grids of LiDAR Metrics

Abstract

Increased use of laser scanning in forest inventories is leading to the adoption and development of automated stand delineation methods. The most common categories of these methods are region merging and region growing. However, recent literature proposes alternative methods that are based on the ideas of cellular automata, self-organizing maps, and combinatorial optimization. The studies where these methods have been described suggest that the new methods are potential options for the automated segmentation of a forest into homogeneous stands. However, no studies are available that compare the new methods to each other and to the traditional region-merging and region-growing algorithms. This study provided a detailed comparison of four methods using LiDAR metrics calculated for grids of 5 m by 5 m raster cells as the data. The tested segmentation methods were region growing (RG), cellular automaton (CA), self-organizing map (SOM), and simulated annealing (SA), which is a heuristic algorithm developed for combinatorial optimization. The case study area was located in the Heilongjiang province of northeast China. The LiDAR data were collected from an unmanned aerial vehicle for three 1500-ha test areas. The proportion of variation in the LiDAR metrics that was explained by the segmentation was mostly the best for the SA method. The RG method produced more heterogeneous segments than the other methods. The CA method resulted in the smallest number of segments and the largest average segment area. The proportion of small segments (smaller than 0.3 ha) was the highest in the RG method while the SA method always produced the fewest small stands. The shapes of the segments were the best (most circular) for the CA and SA methods, but the shape metrics were good for all methods. The results of the study suggest that CA, SOM, and SA may all outperform RG in automated stand delineation.