Object-oriented Change Detection With Multi-feature In Urban High-Resolution Remote Sensing Imagery

The mass information consisted in the remote sensing images with increasing spatial resolution make it possible to detect changes with a finer scale in urban areas. It makes the change detection to be a hot topic in researches on the high-resolution remote sensing images, whereas taking challenges to the existing change detection techniques as follows:1) the spectral variation of a single pixel can not reflect the transformation of the observed object covering it since the object is consisted of a cluster of pixels spatially connected;2) the increasing spatial resolution and limited spectral resolution lead to the increasing spectral deviation inter-classes and spectral overlaping intra-classes in the high-resolution images, which restrains the partibility of image spectrum;3) the diversity on receiving status of mutlitemporal images results in the spatial shifting and stretching of observed objects, especially the ones with elevation; and4) the increasing data amount claims higher automaticity of techniques.In order to detect the changes in the high-resolution remote sensing images, several change detection models are proposed in this paper. They are respectively devoted on ameliorating the problem of "smoothed object" in the object-oriented methods, improving the ability of automatically and globally searching the optimal solution, resolving the diversity of spectral resolution between multi-source images and the errors from compound segmentation of multitemporal images, measuring the similarity of multitemporal images by novel means, and enhancing the fault-tolerance to the "building pseudo changes" resulting from different receiving angles of multitemporal images. The effectiveness of these models was confirmed in the experiments with real high-resolution remote sensing images.As the background of our works, we introduce the four crucial procedures in the change analysis of remote sensing images, including the image preprocessing, change extraction, threshold selection and accuracy assessment. In the reviewing of change information extraction, the existing methods are categorized into two kinds, the methods considering only the pixel spectrum or together with the spatial information. The methods only considering the pixels spectrum includes the methods based on the algebraic operations and image transformations. And the methods combining the spectral and spatial information are based on the object-oriented theary and neural networks. The experiments with a dataset of multitemporal QuickBird images proved that the methods considering only the pixel spectrum could not meet the need of high-resolution images as neglecting the spatial context information, whereas the methods considering spectrum together with the space could solve it. However, the existing techniques considering the spectrum and space are comfronted with some problems, such as the "smoothed object" and the segmentation errors of compound image.First of all, two innovative models are proposed to respectively solve the two critical problems of the object-oriented change detection in the remote sensing images. The searching of the optimal resolution resolved by taking using of the mechanism of automatically and globally searching the optimal solution in the genetic algorithm (GA), and the K-S test is employed to measure the statistic characteristic of multitemporal objects to solve the "smoothed object" problem. The experiments with two sets of QuickBird images proved that the method based on GA could avoid the effect of threshold selection, and the method based on the K-S test could effectively reserve the spectral variance in the image objects. Both of the methods improve the object-oriented change detection.Secondly, comparing the change detection in multi-source images to the one in single-source images, we introduce the difficulty of the former, namely the different spectral resolutions between the multi-source images. Our solution is defining a similarity measure of the multitemporal objects based on the spatial relations between the changed areas and image objects, and judging the objects with lower similarity as the changed areas. Meanwhile, the errors of compound segmentation is overcome by segmentation mapping between the multitemporal images. The experiments with two sets of multi-source images, which was respectively acquired by the QuickBird and IKONOS satellites, proved the effect of this method to detect changes in the multi-source remote sensing images.In our works on the building change detection, we conclude the two main problems:the change dominance measurement and the effect of different multitemporal viewing angles. In order to define a new change dominance measure, we propose to use the pulse-coupled neural network (PCNN) to detect building changes, and exploit several correlative measures to investigate the change probability of each building object. The experiments with two sets of QuickBird datasets proved the effectiveness of this method. On the other hand, we define a fault-tolerant building change detection method to recognize and delete the "pseudo" changing buildings by local registration of the multitemporal building feature points. This method was proved to be effective to suppress the commissions from different multitemporal viewing anlgles, and improve the accuracy of building change detection.