Unsupervised Remote Sensing Image Change Detection Based On Self-paced Learning And Multi-objective Fuzzy Clustering

Change detection technology aims to detect unchanged and changed areas between images taken from the same scene at different time.The development of remote sensing technology provides effective data support for the development of change detection technology,and further facilitates the wide application of change detection technology in military and civil fields.In recent years,the change detection methods for remote sensing images have evolved from the traditional single-objective and unsupervised methods to multi-objective and supervised classifier-based methods.Many proposed change detection methods have achieved good detection performance.However,facing to increasingly complex remote sensing image data,the detection accuracy of existing change detection methods can hardly meet people's needs in practical application scenarios.Therefore,we need further research and exploration to improve the performance of change detection.Based on the analysis of the related achievements at home and abroad,we propose a change detection framework based on group self-learning,and then propose a novel time-varying self-paced regularizer for the framework to further improve the change detection performance.In addition,we also propose an evolutionary multi-objective change detection method based on self-learning and fuzzy clustering.The main contributions of this thesis are as follows:1.This thesis proposes a change detection framework based on group self-paced learning.This proposed method aims at solving the problem existing in unsupervised change detection method based on supervised classifier,i.e.,it is hard to collect reliable training samples in an unsupervised manner.We propose a group self-paced learning framework to mine the reliable training samples for the supervised classifier.The proposed scheme is able to iteratively learn the weighted samples and update the weights in a self-paced manner for identifying the reliable training samples.Specifically,in the phase of updating weights,group information is considered to avoid that the training samples tend to come from the homogeneous region.Experiments on five remote sensing image datasets demonstrate the validity of the proposed change detection framework.2.This thesis proposes a novel time-varying self-paced regularizer for the change detection framework based on group self-paced learning.The proposed self-paced regularizer can automatically determine the learning scheme of self-paced learning.Specifically,in the early stage of self-paced learning,only a few training samples are selected to train a classifier and they are assigned larger sample weights so that the classifier can more fully learn the mapping relationship between training features and classification labels.In the later stage of self-paced learning,more and more complex samples are added to train a classifier with the increase of step parameter in self-paced learning.In order to weaken the influence of unreliable samples on classifiers,only a few more reliable part of them are assigned higher weights,while the others are assigned much smaller weights.Change detection results on five remote sensing image datasets demonstrate that the proposed time-varying self-paced regularizer can further improve the change detection performance.3.This thesis proposes an evolutionary multi-objective change detection method based on self-paced learning and fuzzy clustering.In order to solve the problem existing in multiobjective fuzzy clustering in change detection task,i.e.,the estimation cost of objective values for this multi-objective optimization problem(MOP)is expensive and offspring's selection driven by simple evaluation is time consuming,we integrate regression techniques to determine the superiority of the offspring solutions in the evolution process.Specifically,in the proposed method,the self-paced learning process is implemented to collect reliable training samples for training a robust regression model.The regression model can help to select promising offspring solutions from the candidate solutions for MOP and help to converge to better estimated Pareto optimal solutions,which will help to achieve better change detection results in multi-objective change detection task.