Precise Image Segmentation for Forest Inventory

This paper presents a portable vision system for forest surveying that can estimate the diameter of a target tree with a single " point and click ". Of specific interest is the texture segmentation algorithm used for accurately determining the boundaries of the target tree in a forest. This segmentation is challenging because the target may be occluding other, almost identical looking trees, creating an obscure edge. To compensate for this possibility, we have created an algorithm sensitive to small texture changes which uses both the co-occurrence matrix and the Mahalanobis metric to first find a window that must contain the target texture edge. Then, the edge is precisely localized in this window using a novel method based on co-occurrence matrix differences. Using this algorithm, given no structured shadows, we have been able to achieve consistent accuracies of +/-1 inch, or less than 5% error, in both simulated tests and a field trial.