Change Detection In Remote Sensing Imagery Based On Fisher Classifier And Computational Intelligence

Most change detection algorithms generate the change maps based on the analysisof the difference images, which are obtained by a comparison between themultitemporal images. Statistical models play an important role in the development ofthe change detection techniques. They detect the changes in the multitemporal imageswith different distribution models. However, the single function models used in thestatistical models are too restrictive to fit all the data. Although the improved algorithmsintroduce the spatial information during the detection process, the initial segmentationsstill need the distribution assumption. Therefore, the detection accuracy of the changemaps based on the statistical models is dependent on the parameter model chosen andthe distribution of the difference images.To improve the ability of the change detection algorithms, we focus the attentionon the distribution of the data itself to obtain the changed regions by using Fisherdiscriminant analysis and computational intelligence techniques. Furthermore, to avoidthe instability problem caused by the comparison operator, the processing elements ofthe classifiers are changed to the joint intensity histogram and the joint feature vector.The Wavelet transform and the non-local mean weighted method is used here tointegrate the spatial information. The main contribution of this thesis can besummarized as follows:1. To overcome the restrictive of the single function models, a new approach isproposed by virtue of the double Gaussian mixture model and the Wavelet transform.After the decomposed low pass images are fitted by the double Gaussian mixturemodels, the change maps in different scales are fused using HMT model based onsequential maximum a posteriori estimation. The experiments of the real remote sensingimages confirm the effectiveness of the proposed algorithm.2. The improved dynamic Fisher classifier is used here to detect the changes byanalyzing the joint intensity histogram of the multitemporal remote sensing images. Thealgorithm uses adaptive edge detection to get training data. Considering the spatialinformation, local mean dynamic Fisher discriminant analysis (LMDFDA) is proposedhere, which uses local mean instead of global mean to increase the correlation betweenthe unlabeled data and the training data. The local mean is calculated with the closestimage blocks segmented by the mean shift algorithm. The local mean and theparameters of the Fisher classifier are adjusted according to the current detection resultto avoid the influence of initial condition. The experiments indicate that the proposed algorithm is effective and feasible for real multitemporal remote sensing images.3. A novel change detection approach based on the dynamic fuzzy Fisher classifierfor multitemporal remote sensing images is proposed in this thesis. To increase theseparability of the unlabeled pixels, a non-local mean weighted method is used tointroduce the spatial information. The unlabeled pixels are labeled with a predefinedprobability based on their predictive values. The weights of the unlabeled pixels and theparameters of the dynamic classifier are adjusted according to the updated samples untilall the pixels are classified. The proposed method is distribution free, context-sensitiveand not affected by the comparison operators.4. An unsupervised technique for change detection area between two SAR imagesis proposed in this thesis. The algorithm uses adaptive edge detection to get trainingdata. The joint intensity histograms in different levels are used to decide themembership degree of unlabeled points through Fisher classifier. The fusion modelwhich considers the context relationship and inter-scale information improves theinsensitivity. The simulation results of two real SAR images show that the algorithm iseffective and has better detection results.4. To avoid the instability caused by comparison operators, the joint feature vectorextracted from multitemporal images directly are introduced, which are input to a backpropagation neural network to detect the changes. Furthermore, a non-local meanweighted method is used to integrate the spatial information to increase the separabilityof the unlabeled pixels. The proposed method is distribution free, context sensitive andadjusted dynamically. Experimental results on real SAR images confirm theeffectiveness of the joint feature vector.5. The clustering method is used here to find the change map by minimizing meansquare error (MMSE) with evolution algorithm. After introducing the image character, anew search strategy in Memetic algorithm was given here, which adjusted the localsearch algorithm according to the current detection result. The approach wasdistribution free and did not need priori knowledge. The experimental results obtainedon the real SAR images showed that the proposed method had a higher convergencespeed than GA、 ICSA and original MA, the detection results demonstrated theeffectiveness of the proposed algorithm.