Study On Change Detection Based Algorithm For Multitemporal Remote Sensing Image Classification

With the rapid development of remote sensing satellite technology,a large number of multitemporal remote sensing images have been formed.The classification of multitemporal images has inevitably attracted extensive attention owing to its importance in monitoring the land cover changes.The classification of image time series aims to classify the unlabeled target image taking advantage of the labels in another image(the source image)which is acquired on the same scene but different time.Traditional semi supervised methods need to collect enough reference labels for the target image,while the domain adaptive methods assume that the data distribution of the source and the target image is correlated.These are difficult to be satisfied in real scenes.Therefore,in recent years,change detection technologies have been introduced to solve these problems.The change detection based classification methods employ change detection techniques to detect the unchanged pixels and transfer their labels from the source to the target image in the unchanged regions.Finally,a classifier for target image is trained with these transferred labels.However,this kind of methods also has some problems.First,a potential downside is the introduction of error from change detection.Secondly,most of the existing methods ignore the wrongly propagated labels after propagation.To solve these two problems,this thesis proposes two algorithms.The main contents of this thesis can be summarized as follows:First,this thesis proposed a cost-sensitive change detection based multitemporal images classification algorithm.The main idea is to improve the accuracy of unchanged class and provide reliable propagated labels for classification of the target image.In the proposed method,self paced learning is employed to assign a weight to each sample and update it dynamically.Based on the acquired weights,self paced learning is able to select reliable training samples.Then the sample weight and cost sensitive strategy are incorporated into the objective function of the model,so that the model will tend to reduce the errors in the detected unchanged classes.Specifically,this is achieved by setting a greater cost value to the error in the unchanged class and minimizes the total cost.When the category of a sample is difficult to determine,it is more inclined to be considered as changed.Experimental results show that the proposed algorithm provides a set of reliable samples for the training of classifier and achieves a promising improvement on classification accuracy.Second,the difference information of change detection can reflect the reliability of the propagated labels.In order to further exploit the difference information and learn a robust classifier,this thesis proposes a curriculum learning based classification method incorporating difference information to classify multitemporal remote sensing images.In the proposed method,a superpixel-based label transfer technique is designed to reduce the errors of the unchanged class and transferred the labels of source domain to the target domain in the unchanged region.Then curriculum learning is utilized to formulate a learning objective by assigning a weight to each sample in the training set for measuring the reliability of each sample.In the process of curriculum learning,a tunable curriculum learning function incorporating difference information is designed to constrain the curriculum region in an implicit way.Experimental results on six data sets(three areas)show that the proposed method achieves remarkable classification performances,thus confirming its suitability when applied to multitemporal remote sensing image classification problems.