A bayesian theory for forestry image understanding

This thesis aims at developing a statistical framework for forestry inventory based on airborne digital imagery. The goal is to infer the plausible 3D interpretation of forestry scene geometry given 2D image data. The cycle of image understanding is considered in terms of a common Bayesian inference approach where components are estimated at each stage by solving optimization sub-problems. Traditional forestry inventory systems have developed many heuristic-based methods that can be roughly classified into two categories: (a) data-driven heuristics, such as Gougeon's valley-following method based on the assumption that tree boundaries are coincident with specific intensity changes along neighboring pixels, and (b) model-driven heuristics, such as Pollock's synthetic tree template which synthesize the 2D tree template by assuming prior knowledge of the 3D geometry of given species and camera/lighting conditions. Obviously these methods fail when the corresponding assumptions are not satisfied. In other words, due to the complexity and variability of the visual appearance of trees, the true 3D forestry scene may lie in a large and complicated solution space. Heuristics, like those used in the above cases, do apply to specific cases, but are too fragile to apply in general. In particular they cannot apply to native boreal forests consisting of mixed species with overlapping canopies—the type of forest of main interest in this thesis. Fortunately, by taking advantage of recent developments in Bayesian statistics and Machine Learning, we are able to propose a generic framework that allows reliable and effective inference of the optimal solution given the adopted statistical models. The proposed framework consists of three main modules: the segmenter, the stereo, and the 3D-fitting modules. Each module is formulated as an optimization problem. The first two modules serve to produce the independent “sketches” of the input image data while the third module integrates evidence to produce 3D models as consistent as possible with what is being sensed. In the following, each of the modules is discussed in details. The segmenter provides annotations of the input images after offline supervised learning. A new image model was proposed, then labelled regions (aspen, spruce, shadow and ground regions) are formed. The stereo module infers the depth/elevation map from an aerial input stereo image pair. A Bayesian stereo algorithm is proposed for approximate inference of the depth map using loopy belief propagation, and for parameter estimation using sampling techniques. The 3D model fitting module aims at finding the optimal solution that is most plausible and consistent with the stereo and annotated region extraction modules in terms of minimizing the fitting errors. Sampling techniques are used to traverse the solution spaces. As much as can be concluded from the literature, this seems to be the first time a fully Bayesian image understanding system has been developed that extracts both annotated image and 3D models based on the fundamental principles of sub-process optimization, feedback and update of hypotheses based on observed data and the integration of predictions from the system's modules. Very encouraging results have been obtained for forestry image data, both synthesized and natural and beyond the typically single species plantation images examined in previous studies.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.