| Résumé : |
(auteur) Buildings are prominent objects of the constantly changing urban environment. Accurate and up to date Building Polygons (BP) are needed for a variety of applications, e.g. 3D city visualisation, micro climate forecast, and real estate databases. The increasing number of earth observation remote sensing images enables the development of methods for building extraction. For instance, Hyperspectral Images (HSI) are a source of information about the material of the objects in the scene, whereas the Digital Surface Models (DSM) carry information about height of the surface and of objects. Thus, complementary information from multi-modal images, such as HSI and DSM, is needed to provide better understanding of the observed objects. A variation in material and height is represented by an edge in HSI and DSM, respectively. Edges in an image carry large portions of information about the geometry of the objects, because they delineate the boundaries between them. Object extraction and delineation is more reliable if information content from HSI, DSM, and edge information is jointly accounted for. The focus in this thesis is on method development for BP extraction using complementary information from HSI and DSM by accounting for edge information. Furthermore, a new quality measure, which accounts for shape differences and geometric accuracy between extracted and reference polygons, is proposed. Object and edge detection from an image is meaningful only for some range of scales. Edge detection in scale space is motivated by showing that in the same image different edges appear at different scales. Instead of deterministic edge detection, edge probabilities are computed in a linear scale space. Bayesian fusion of edge probabilities is proposed, which employs a Gaussian mixture model. The scale, at which an edge probability is computed, is defined by a confidence probability. The impact of selecting mixing coefficients in the Gaussian mixture model according to a prior knowledge or by a fully automatic data-driven approach is investigated. Main limitations of joining the edge probabilities from different datasets are the coregistration between the datasets and the inaccuracies in the datasets. The rectilinear BP are adjusted by means of weighted least squares, where the weights are defined on the basis of joint edge probabilities. Two mathematical models for rectilinear BP are proposed, one with a strict rectilinearity constraint and the second one, which introduces a relaxed rectilinearity constraint through weighting. The experiments on synthetic images show that the model with strict constraint gives better results, if the BP under consideration are all rectilinear. Otherwise, the relaxed rectilinearity constraint through weighting balances better between the rectilinearity assumption and fitness to the data. The approximate BP are created by a Minimum Bounding Rectangle (MBR) method. A main contribution of the proposed iterative MBR method is the automatic selection of a level of complexity of MBR through analysis of a cost function. A metric for comparison of polygons and line segments, named PoLiS metric, is defined. It compares polygons with different number of vertices, is insensitive to the number of vertices on polygon's edges, is monotonic, and has a nearly linear response to small changes in translation, rotation, and scale. Its characteristics are discussed and compared to the commonly used measures for BP evaluation. In all experiments the BP are evaluated by computing the newly proposed PoLiS metric and quality rate. The feasibility of joining all the proposed methods in one workflow is shown through the experiment, which is carried out on 17 HSI-DSM dataset pairs with four different ground sampling distances. The main finding of the experiment is that joining the information from multi-modal images, i.e. HSI and DSM, results in better quality of the adjusted BP. For instance, even for datasets with 4 m ground sampling distance, the completeness, correctness and quality rate values of extracted BP are better than 0.83, 0.68, and 0.60. Inaccuracies of the images, such as holes in DSM or imperfect DSM for 1151 orthorectification, are influencing the accuracy and localisation of edge probabilities and consequently also the accuracy of adjusted BP. |
Fusion of hyperspectral images and digital surface models for urban object extraction (2016)
