Annotation Of High Resolution Remote Sensing Images Based On Active Spectral Clustering

How to interpret the remote sensing image data with high spatial resolution is a practical and challenging problem in current age of remote sensing "big data".The annotation of Remote sensing image,that is,assigning the pixels or specific image areas corresponding to the contents with semantic label,is an important task of remote sensing image understanding.However,the existing solution for annotation of high-resolution remote sensing image has some problems in the face of massive remote sensing data processing:1)The traditional manual visual interpretation method has high reliability,but suffers from high cost of labor and time.At the same time,it can use class label as prior information,which is strongly dependent of expert knowledge.2)For the existing remote sensing image intelligent interpretation algorithm,high accuracy of the results still rely on the quality of labeled data.The active semi-supervised classification method is intended to take the initiative to select effective and less training samples,but it still suffer from a single form of supervision,and dependence on the random initial training set.In view of the above problems,this paper aims to find a means of efficiently annotating remote sensing images with little expert knowledge.In this paper,we mainly use active clustering algorithm to solve the annotation task of high-resolution remote sensing image.The main research content and contribution are as follows:Firstly,we study the active spectral clustering algorithm based on k-NN graph,and use the simple annotation method(pairwise constraint),which is simple to annotate,to complete the annotation work,so as to reduce the working threshold and operation difficulty for the labeler.At the same time,the active selection model based on the neighborhood uncertainty of k-NN graph is designed,and the sample pairs with the largest amount of information are selected for each iteration.The two-step k-NN graph adjustment,"Cut" and "Reallocate",is designed to improve the structure of k-NN graph.Through the above operation,less number of questions is required to complete the k-NN graph and improve the performance of clustering,thereby reducing the workload of annotation.Second,we develop an active spectral clustering algorithm,which can automatic update category number.It can handle high-resolution remote sensing image marking tasks with special cases of the unknown number of categories or unknown categories.The number of categories is updated dynamically during the algorithm.Finally,we consider the use of pairwise constraint propagation,which can propagate the existing pairwise constraints to global.This can highly reduce the need for prior information,thereby reducing labor costs.In order to get a global pairwise constraint score,the information fusion is carried out,which is based on soft pair wise constraint propagation and hard pair constraint propagation.The global adjustment of k-NN is achieved by using the score,and it is used for the two active selection strategy.In order to verify the performance of the proposed algorithm,the experiments are carried out on three high-resolution remote sensing images.Comparison with the existing algorithms for remote sensing image interpretation is also designed.The results verify the effectiveness of the proposed algorithm for high-resolution remote sensing image annotation.