CORRECTING TREE COUNT BIAS FOR OBJECTS SEGMENTED FROM LIDAR POINT CLOUDS

Abstract

We introduce a new statistical distribution for modeling the number of trees that fall in segmented LiDAR point clouds. Average tree count is modeled as a linear function of segment ground area, but the same methods can be used to fit more complex non-linear models. Although observed mean occurrences can be continuous, occurrences must be a non-negative integer implying the usual assumption of normal errors may not be adequate. The starting point for a new distribution is the Poisson. The Poisson is based on the premise of rare events from a large population. The probability a particular tree falls in a given point cloud segment is small and the number of segments is large, hence the Poisson distribution should be an appropriate error distribution for modeling the number of trees that fall in a point cloud segment. Deviation from the Poisson occurs as a result of the point cloud segmentation process. The purpose of segmentation is to provide segments that contain a single tree. This implies that a Poisson with deflated probability of zero occurrences and inflated probability of one occurrence is appropriate. LiDAR point cloud data on 20 stem mapped plots are used to show the utility of this approach.