Bayesian approach to single-tree detection in airborne laser scanning - Use of training data for prior and likelihood modeling

Abstract

This paper introduces a novel computational approach to handling remote sensing data from forests. More specifically, we consider the problem of detecting an unknown number of trees based on airborne laser scanning (ALS) data. In addition to detecting the locations of individual trees, their heights and crown shapes are estimated. This detection-estimation problem is treated in the Bayesian inversion framework. We use simplified, rotationally symmetric models for the tree canopies to model the echoes of laser pulses from the canopies. To account for the associated modeling errors, we use training data consisting of ALS data and field measurements to build a likelihood function which models statistically the propagation of a laser beam in the presence of a canopy. The training data is utilized also for constructing empirical prior models for the crown height/shape parameters. As a Bayesian point estimate, we consider the maximum a posteriori estimate. The proposed approach is tested with ALS measurement data from boreal forest, and validated with field measurements.