Canadian Forest Service Publications
As assessment of both visual and automated tree counting and species identification with high spatial resolution multispectral imagery. 1999. Leckie, D.G.; Gougeon, F.A. Pages 141-152 in D.A. Hill and D.G. Leckie, Editors. International forum: automated interpretation of high spatial resolution digital imagery for forestry, Proceedings: Symposium. February 10-12, 1998, Victoria, British Columbia. Natural Resources Canada, Canadian Forest Service, Pacific Forestry Centre, Victoria, BC.
Issued by: Pacific Forestry Centre
Catalog ID: 5179
Availability: PDF (download)
Thirty six centimeter MEIS multispectral imagery was acquired over a test site containing boreal and temperate forest softwood and hardwood species in eastern Ontario, Canada. Twenty four 20x20 meter field plots were established identifying the location, species, crown diameters, dominance and openness of each tree within the plot. Trees were manually outlined on the MEIS imagery and used to classify tree species with a maximum likelihood classifier. A visual enhancement was produced to highlight the different species. Trees were automatically isolated using a valley following approach and tree counts and delineations compared to those of the ground truth trees. Operational photo-interpreters from the provinces of Quebec and Ontario were trained in interpretation of the enhancement and an interpretation test conducted on an individual tree basis. Species interpretation accuracy was in the order of 70-90% for softwood species and down to 50-65% or lower for hardwoods. Accuracy was assessed against the dominance, crown size, crown openness, and species relative to adjacent trees. Preliminary analysis of manually delineated trees produced classifications comparable for some species, but generally in the order of 15% less accurate than the visual interpretation. Visual detection of trees for imagery resampled to different resolutions indicated that optimum detection occurred with resolutions of 10-45 pixels per tree (10 for hardwoods and 25-45 for softwood species). Omission errors dropped rapidly to near 10% at 10-15 pixels/tree then decreased slowly. Commission errors (counting one crown as more than one crown) increased gradually with number of pixels per crown. Accuracies again were assessed against dominance, crown size, crown openness and species. Preliminary tree detection with the valley following approach on the 36 cm data resulted in close estimates of overall average stems/ha, but only 40% direct 1 to 1 correspondence with ground truth trees. The visual counting acc