Forestry Scene Geometry Estimation Via Statistical Learning

In the context of a forest inventory application, given preprocessing of the 2D airborne images of a forest plot, we focus on estimating the parameters which control the 3D geometry of trees, in order to generate a virtual forest. The major contribution of this paper lies in the proposed probabilistic graphical model and the novel sampling scheme for solving this data fusion problem. To deal with the variability introduced from both the image data and the preprocessing procedures, we adopt a Jump-Diffusion Markov Chain Monte Carlo sampling paradigm to traverse the possible state spaces. Within each state space, a stochastic version of the Expectation Maximization algorithm is employed to explore the plausible parameters and latent scene geometry by finding the local maxima. Therefore, the propose algorithm estimates the number of trees and the associated parameters, and also infer the 3D scene geometry that is consistent with the preprocessed data and the expert prior knowledge. Experiments on both synthetic and real forestry data show promising results.