Change Detection In Remotely Sensed Imagery Using Multivariate Statistical Analysis

Change detection in remotely sensed imagery is defined as the procedure of quantitatively analyzing and identifying changes occurred on the earth's surface from remotely sensed imageries acquired at different times. As a key element for many applications of earth observation such as resource inventory, environment monitoring, update of fundamental geographical database, etc., change detection technique is of urgent demands and has great potential in scientific applications. Currently change detection, especially change detection based on multi-temporal multi-channel (multispectral, multi-polarization, etc.) remotely sensed imageries has become a hot topic in research field related to remote sensing applications.Significant efforts have been made in the development of change detection techniques, and quite a lot of methods have been devised. However, there are still some problems that could not be solved properly by traditional methods in change detection, such as concentration of change information on all channels to produce temporal difference images, extraction of changed areas, identification of change types, etc. Under such circumstances, our investigations are carried out around the issues related to how to automatically extract change information rapidly and effectively from multi-temporal spaceborne remotely sensed multispectral imageries with mid-resolution, as well as Synthetic Aperture Radar (SAR) imageries in this dissertation. Most efforts are focused on two key problems, including production of temporal difference images and extraction of changed areas.The problem of producing difference images from multi-temporal multi-channel remotely sensed imageries is investigated in the first part of this dissertation. Compared with change detection based on single-channel imageries, it is more difficult to perform change detection on multi-channel imageries due to impact of inter-channel correlations. And it is necessary to effectively concentrate change information from all channels to produce a temporal difference image to facilitate detection and analysis of changes. From the point of view of multivariate statistical analysis, thorough researches are conducted in following aspects:(1) To eliminate impact of inter-channel correlations, canonical correlation analysis (CCA) and the so-called multivariate alteration detection (MAD) method based on CCA are introduced into bi-temporal multi-channel change detection. According to MAD method, two multi-channel imageries covering the same geographic location and acquired at different times are taken as two sets of random variables, then MAD transformation is performed on these random variable sets to produce a set of result variates that are uncorrelated with each other. In this way correlations between channels can theoretically be removed as much as possible, so that the actual changes in all channels can be simultaneously detected in the resultant difference image.(2) To improve effectiveness of the MAD result, it is proposed to use signal-to-noise ratio (SNR) instead of variance as a measurement for change information distribution, and another multivariate statistical transformation called minimum noise fraction (MNF) is introduced as a post-processing step for MAD transformation. In this way, change information can be separated from noise to the greatest extent, so that the technical problem of effectively concentrating change information and producing difference image could be solved properly.(3) The feasibility of change detection based on direct comparison usng multi-temporal remotely sensed imageries acquired by multi-sensors is explored. The scheme of employing MNF/MAD to produce difference image is proposed for multi-sensor change detection. An experiment on Landsat7 ETM+ and SPOT5 HRG imageries is carried out to demonstrate the effectiveness of the proposed scheme.Experimental results in a few test sites indicate that MNF/MAD method based on CCA is able to extract change information effectively from multi-temporal multi-channel remotely sensed imageries and pool them into a few resultant components of the temporal difference image. Generally these components could manifest some clear physical meanings. A distinguished advantage of the MNF/MAD scheme is its invariant to linear scaling, which means it is insensitive to disagreement in measurement scale, gain settings in measuring devices, and linear radiometric distortions, as a result the requirement on image preprocessing could be reduced.In the second part of this dissertation, the problem of extracting changed areas from difference image produced by change detection is studied. In fact changed area extraction is a typical problem of two-category classification, and can be solved by employing thresholding strategy. However, thresholds are difficult to establish in traditional schemes. In virtue of theories and methods in statistical pattern recognition, thorough researches are conducted in following aspects:(1) A method based on Bayes Rule for Minimum Error is proposed to establish change thresholds in an automatic way. Upon analyzing statistical characteristics of difference image, we firstly assumed that both the pixels of change and that of no change were subject to simple Gaussian density distribution model, and employed the Expectation-Maximization algorithm to estimate distribution parameters and change thresholds, so as to extract changed areas in an automatic way. Then, to account for the difficulty of applying simple Gaussian density distribution model in describing complicated distributions containing multiple classes, the mixed Gaussian density distribution model is used instead to describe distributions of the two pixel classes. And accordingly genetic algorithm is employed to estimate distribution parameters, so as to improve estimation of change thresholds.(2) A deficiency of Bayes scheme is found to be the adoption of pixel independency assumption as well as ignoring contextual information. Contextual Bayes decision method is devised for this problem. In this method, Markov random field (MRF) model is introduced into Bayes decision to depict and utilize contextual information to estimate local prior probability, so as to improve accuracy and reliability of the changed area extraction results.(3) The problem of SAR change detection is studied. The scheme of ratioinp with logarithmic stretching is employed to produce a temporal difference image. According to the approximate Gaussian distribution characteristics of pixels in difference image, a scheme is proposed to apply the contextual Bayes decision method to extract changed areas from the difference image.Experimental results demonstrate that for both multi-temporal optical and SAR imageries acquired by spaceborne sensors, the contextual Bayes decision method could establish change threshold in an automatic and unsupervised way, thus could identify and extract change areas effectively from the difference image.