Comparison of individual-tree segmentation and area-based approaches using airborne LiDAR for tree density and volume estimation in New Brunswick forests
University of New Brunswick
Airborne LiDAR (Light Detection and Ranging) is a promising remote sensing technology widely used for forest inventory. In this study we explored three approaches for developing LiDAR derived estimates of volume per ha: Area-based nonlinear regression models; individual-tree segmentation and summation; and a combined approach. On the basis of minimum rMSEs, the combination approach was the best (48.23 m3 /ha rMSE), area-based was second (54.98 m3 /ha rMSE), and segmentation method showed the lowest accuracy (64.62 m3 /ha rMSE). Although the three approaches produced different levels of accuracy, the estimates were statistically equivalent to each other based on the two one-sided t-test of equivalence. While the segmentation approach produced acceptable estimates of volume per ha, density (stems/ha) estimation was very poor with estimates often less than 50% compared to field plot data. The Hamraz algorithm produced more accurate estimates than the other segmentation algorithms explored. In addition to fitting models and accessing goodness-of-fit, we explored stand structural factors that influenced the observed errors. Stand totals (Volume, Basal Area, and Density) were more influential for the area-based approach while mean tree size (quadratic mean diameter and height) and species composition (basal area of hardwood, and softwood) were more influential with the segmentation approaches.