Research On Change Detection Approaches For High Resolution Remotely Sensed Images Based On Superpixel And Active Learning Sampling Strategies

Change detection is an important research topic in the field of remote sensing application.The spatial resolution of images becomes higher and higher with the continuous progress of remote sensing technology,thus more and more attention has been paid to change detection research based on High Spatial Resolution Remote Sensing Images(HSRRSI),a new research hotspot.The outcome is far from perfect in terms of directly applying change detection methods derived from low or medium resolution images to high resolution ones,while other existing methods particularly aiming at high resolution images have their own shortcomings and limitations respectively.At the basis of previous related research achievements,against the defects of presently available change detection approaches for HSRRSI in accuracy and efficiency,this thesis proposed a new solution by integrating superpixel segmentation and Active Learning(AL)sample selection strategies.Its main contentsand achievements are as follows.(1)Several representatives of prominent works in the field of remote sensing change detection are enumerated and then the research status of high resolution image change detection is summarized;after analyzing the main challenges,our proposed method that combines the superpixel segmentation and the active learning sample selection strategy and its supporting technological route to overcome those challenges is detailed.(2)Based on the lucubration of properties of the existing mainstream superpixel segmentation algorithms,a comparison of the performance of 4 commonly used superpixel segmentation algorithms is conducted and this comes to a conclusion that the simple linear iterative clustering(SLIC)is of comparative preponderance.(3)Margin Sampling(SLIC-MS)and Gaussian Process active learning sampling strategy(SLIC-GP)are respectively combined with the superpixel segmentation,therefore 2 high-resolution image change detection methods are proposed.The former focuses on the problem of sample information redundancy in SLIC-MS and employs angle-based diversity(ABD)to measure the similarity between samples to reduce similar samples,simultaneously ensuring the diversity of samples and lower manual labeling costs;SLIC-GP implementation is a sampling strategy that considers the new sample's impact(SLIC-GPimpact)on whole model the most to distinguish determined samples.Experiments were both conducted in WorldVew II and QuickBird images.Results showed that two proposed methods both achieved better detection quality than other existing methods,and SLIC-GP outperformed SLIC-MS.