Study On Fuzzy Unsupervised Change Detection Methods Based On Remotely Sensed Image

Unsupervised change detection technique plays an important role in identifying land cover changes using multi-temporal remote sensing images. It also has a promising prospect as no ground truth involved when performing change detection. However, there are many issues such as same spectral from different materials, same material with different spectral, noise, mixed pixels, and fuzzy boundary deteriorating the quality of remote sensing images. A certain degree of uncertainties are inevitably introduced and disseminated during the process of unsupervised change detection, reducing the ultimate change detection accuracy. This study is to provide reliable unsupervised change detection methods. Uncertainties of the existing change detection methods are deeply analyzed, and then reliable change detection methods to reduce the uncertainties in the results so as to improve the final detection accuracy are presented based on the fuzzy set, fuzzy topology, D-S evidence theory, fuzzy integral and indicator Kriging, etc. Main efforts focus on the two key steps, difference image generation and difference image analysis. The specific details are given as follows:(1) Taking into account the sensitivity to threshold and weak ability of dealing with uncertain boundary pixels being as the major factors resulting in the low accuracy of threshold change detection, the dissertation introduces fuzzy topology and presents a novel change detection framework for optimizing the threshold change detection (termed as FTT). First, the difference image is divided into unchanged and changed classes by a traditional threshold technique. Second, the results are fuzzified and a membership function is computed for each class by Bayesian theory. Third, the fuzzy boundary regions of the change detection results from the threshold technique are determined using fuzzy topology according to the difference of pixel membership. Finally, the boundary pixels are reclassified using the supported connectivity in fuzzy topology. By using fuzzy boundary and reclassifying the fuzzy boundary pixels, the proposed method reduces the sensitivity to the selection of threshold and enhances the recognition rate of uncertain pixels, providing improved change detection accuracy.(2) To make up the shortcomings of Euclidean distance and the defuzzification process of the maximum membership rule used in the change detection methods of fuzzy C means (FCM) algorithms, the dissertation presents a unsupervised change-detection framework based on adaptive distance idea and fuzzy topology theory (termed as FATCD framework). FATCD first presents an adaptive method for calculating the distance from samples to cluster centers using adaptive distance function to modify the way of evaluating the membership for each pixel, by which the accuracy of the obtained fuzzy membership function is increased; and then, it uses fuzzy topology theory to improve the maximum membership procedure, which optimizes the defuzzification process and increases the change-detection accuracy on the fuzzy boundary pixels.In virtue of the above two points, FATCD can enhance the change detection performance of FCM-type algorithms.(3) Generally, the difference image pixel values close to 0 represent areas of no change and magnitudes close to 255 depict areas of change. However, the existing change detection methods mainly use grey values and spatial context, considering no the labeling knowledge of the pixels at both sides of the of difference image histogram. In view of this, the dissertation presents a novel RSFCM change detection method based on the standard FCM algorithm. First, the pixels with a high probability of belonging to the changed or unchanged class are identified by selectively thresholding the histogram. Then, via a supervised component and a fuzzy spatial term, RSFCM incorporates labeling knowledge and spatial information into the FCM, respectively, which mainly uses the gray-level intensity. The former is used to supervise the clustering process of the difference image for enhancing change information and achieving more accurate membership, and the latter is used to modify the membership for obtaining spatially smooth membership functions and thus reducing the effect of noise pixels and error labels. Thus, RSFCM can detect more changes and provide noise immunity(4) In order to integrate the merits of different difference images and solve the misclassification problems on conflicting pixels of traditional fusion techniques, novel decision-fusion change detection methods based on indicator Kriging theory are proposed, namely, FMVK, DSK and FIK. The proposed methods are in five steps. First, the algorithms for generating difference images are summed up, and six difference images including complementation change information are created by selecting typical difference operators. Second, each difference image is processed by the standard FCM, and the algorithm provides the change detection map and membership functions of no-change and change classes. Third, on the basis of the obtained results, the expert opinion of majority voting (MV), the basic belief assignment function of D-S evidence theory (DS), and the fuzzy density function of fuzzy integral (FI) are defined using fuzziness measures of fuzzy set, similarity of change maps, and Jaccard distance. Forth, the six difference images are combined by the three fusion methods, MV, DS and FI. Finally, the conflicting pixels are determined, and are further processed by indicator Kriging theory. As a result, FMVK, DSK and FIK reduce the uncertainties of single difference image change detection, and yield better change detection results than those of single difference image change detection, and MV, DS and FI.