Please use this identifier to cite or link to this item:
Title: Toward generic object extraction from remote sensing images using spectral and spatial contextual information
Authors: Li, Zhongbin
Degree: Ph.D.
Issue Date: 2016
Abstract: Extracting geographic objects has long been an essential research topic in geographical sciences and remote sensing. Past research has been primarily focused on tackling specific types of objects by using particular features, and little attention has been paid to the following three aspects: 1) developing more operational methods that can be applied to handle multiple types of objects by exploring more generic features; 2) improving the computational efficiency of existing methods; and 3) increasing the degree of automation of object extraction to reduce the load on users. Thus, object extraction from remote sensing images remains challenging. In recent years, level set evolution (LSE) has proven effective at object extraction. It can handle topological changes automatically while achieving high accuracy. However, the application of the state-of-the-art LSE methods is compromised by laborious parameter tuning and expensive computation. For these reasons, two fast LSE methods are proposed to extract anthropogenic objects from high spatial resolution remote sensing images. The significant advantages of the proposed LSE methods are as follows: 1) two novel data terms (one is an edge-based and the other is a region-based) are proposed based on the conventional LSE methods and they are more practical and efficient than the existing approaches; 2) the traditionally used mean curvature-based regularization term is replaced by a Gaussian kernel in the proposed methods, which makes it possible to use a larger time step in the numerical scheme, thus expediting the proposed methods considerably; 3) for computational efficiency, the level set function in the proposed methods is initialized as a binary function rather than the traditionally used signed distance function; and 4) a long straight line finder and a change detection technique are introduced into the proposed LSE methods, automating the proposed methods substantially.
Although LSE methods have shown promising performance in homogeneous object extraction, they generally have difficulty in handling heterogeneous objects as they only take advantage of the intensity (spectral) information. To develop a more general and reliable object extraction system, an enhanced binary Markov random field (MRF) is thus proposed, which takes advantage of both the spectral and spatial contextual information of objects of interest simultaneously. The principal novelties of the proposed method are as follows. First, a new mixture model (MM) is proposed for spectral learning. Compared to the existing Gaussian mixture model (GMM), the proposed MM is more robust to objects that do not follow the trained Gaussian distributions. Then, for computational efficiency, the parameters in MM are estimated by using tree-structured vector quantizer (TSVQ) instead of the expectation-maximization (EM) algorithm widely used in GMM. Next, a morphology-based post-processing mechanism is particularly devised to improve its performance. Finally, the proposed enhanced MRF is automated by using a multi-threshold method and is then applied to automatic burned area mapping from Landsat 8 images.The efficiency and accuracy of the proposed methods (both LSE methods and the enhanced binary MRF) are finally corroborated by a wide range of experiments. Compared with other state-of-the-art approaches, they have following significant advantages: 1) they are capable of dealing with multiple types of objects (including both anthropogenic and natural objects) effectively, 2) they can achieve much better performance, 3) they are computationally much more efficient, and 4) they are operational and reliable in real applications due to less parameter tuning and appropriate manual interaction.
Subjects: Remote sensing.
Hong Kong Polytechnic University -- Dissertations
Pages: xxv, 235 pages : color illustrations
Appears in Collections:Thesis

Show full item record

Page views

Last Week
Last month
Citations as of Jun 4, 2023

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.