Multi-Level Change Detection For Remote Sensing Imagery

Change detection is one of the earliest processing technologies for remote sensing data, and has been widely used in many applications. Remote sensing change detection can provide a consistent view of landscape variation at a large scale over a long period of time, and is extremely important for the research of global ecosystem change and society development. With the advance of earth observation technology, it is now feasible to obtain various types of remote sensing data covering the same scene for multi-level observation targets, including moderate/low-resolution multispectral image, hyperspectral image and high-resolution image. Multi-temporal remote sensing data with various spectral/spatial resolutions contain multi-level information for landscape variation, such as change/non-change, change type, and semantic change. However, there exist numerous problems for the extraction of multi-level information. First, the radiometric variance due to the difference of environmental factors makes it difficult to extract the real change information; second, high-dimension features contained in multi-temporal hyperspectral images need better interpretation; finally, scene change information at semantic level is hard to analyze in multi-temporal high-resolution images. In order to solve the problems above, in this thesis, we aim at developing new methods and theories for multi-level change information extraction, with the advanced technologies in machine learning and pattern recognition. In detail, the main contributions of this thesis are presented as follows:(1) We introduced the basic procedure and fundamental theory of multi-temporal remote sensing change detection as a review, containing five parts:pre-processing, change detection, thresholding, accuracy assessment and application. In the part of pre-processing, three main steps including data selection, registration and radiometric correction, are shown in detail. Then according to different types of remote sensing data, change detection methods are categorized into traditional change detection, hyperspectral change detection, and high-resolution change detection. The common methods for thresholding and accuracy assessment are introduced in the next part, and the applications of change detection are summarized finally.(2) We proposed slow feature analysis (SFA) for multi-temporal remote sensing images, which aims at extracting invariant features by reducing the radiometric variance due to different environmental factors, so as to highlight real change information. Then unsupervised, supervised and iterative change detection algorithms are proposed, and the changed pixels are excluded from the SFA learning, with the unchanged training samples or by iteratively reweighting approach. The performance of change detection was improved with better learned features. We also proposed automatic radiometric normalization algorithm based on iterative slow feature analysis (ISFA). The weight image obtain by ISFA was used to calculate the linear function between corresponding bands directly. The radiometric variance was corrected relatively with this linear function, to maintain the radiometric consistency between multi-temporal images. For the case there are too many changes in the area, we proposed supervised ISFA for automatic radiometric normalization, which can find the correct linear relationship robust to the selection of initial seeds. Finally, we proposed hyperspectral anomaly change detection algorithm with SFA. It wants to find the transformation features to minimize the spectral change of background, and detect anomalous changes in the change residuals. The proposed algorithm can get a good performance even though the background is mostly changed.(3) We studied the subspace theory for remote sensing change detection, which represents spectral signature with subspace to detect the changes of spectral features and separate change information of different change types. We proposed subspace based hyperspectral change detection algorithm. For each detected pixel, the background subspace is built with the corresponding pixel, and the subspace distance is calculated to detect spectral feature changes. The spatial and spectral information can be added into the subspace to reduce the effect of mis-registration and suppress the change of specific target in the final result. Then we proposed hyperspectral change analysis algorithm with independent component analysis (ICA). The different image of multi-temporal hyperspectral data is transformed by ICA to separate change information of different classes of landscapes. Finally, we proposed targeted change detection algorithm. The specific change type is regarded as a combination of spectral features before and after change. The target detection algorithm is applied in stacked hyperspectral data to separate specific change type.(4) For the first time, we proposed scene change analysis to detect region changes of land-use types at semantic level with multi-temporal high-resolution remote sensing images. We discussed the definition and significance of scene change analysis, and then analyzed its research points and applications. Then we achieved scene classification algorithm with bad-of-visual-word (BOVW) model, and evaluated the parameter settings in experiments. Finally, we proposed scene change analysis framework with BOVW model. The temporal information is considered in the dictionary learning step with three different approaches: separate dictionary, stacked dictionary and union dictionary. The scene changes are detected with the approaches of post-classification and compound classification. The experiments proved the effectiveness of scene change analysis framework and its potential for applications.