Research On Change Detection In High Resolution Remote Sensing Images Based On High Level Feature Analysis

With the rapid development of remote sensing,the availability of remote sensing images is increasing.A large amount of available remote sensing images and the related progressing techniques are providing more potentials for remote sensing-based applications in business and science research communities,as well as provide strong supports for researches in land use and land cover change detection.Remote sensing change detection aims to extract information from change areas by comparing and analyzing differences between images that observing the same area but are taken at different times,which is vital for earth surface cover change researches.With the increased spatial resolution of remote sensing images,traditional image change detection methods are difficult to obtain satisfactory performances when dealing with the change detection task in high resolution remote sensing images,and the deep learning-based method can achieve superior change detection performances by highlevel feature extraction and analysis.At present,deep learning-based change detection in high resolution remote sensing images is still in the development stage.Though some deep learning-based methods have been proposed for the change detection tasks of high resolution images,change detection performances still has room for improvement in supervised,unsupervised change detection tasks and season-varying remote sensing image change detection tasks.This paper analyzes and summarizes the general ideas and problems of existing deep learning-based change detection methods,and conducts in-depth researches on the task of change detection of supervised,unsupervised and cross-season images,respectively.The work in this paper mainly includes the following aspects:(1)In the research of supervised change detection in high-resolution remote sensing images,this paper classifies and compares the existing deep learning-based change detection methods.A framework with two independently trained sub-networks for informative raw image deep feature extraction is proposed to overcome the less representative raw image feature problem and the vanishing gradient problem.For the fusion of heterogeneous original image features and image difference features,a fusion method based on attention mechanism is proposed.Multi-level deep supervision is proposed to improve the difference discrimination network.The superiority of this method is verified by the performance comparison experiment on the public data set and the personal collection data set.(2)In the research of unsupervised change detection in high-resolution remote sensing images,the general process method of change detection method based on image deep feature extraction is summarized.For the defect of poor change results based on redundant high level deep image difference features,a method for extracting significant high level image difference features is proposed.A sample collection method on change samples,significant non-change samples and non-significant non-change samples is proposed.Based on the collected samples,a similarity evaluation network is proposed for discriminating change samples and unchanged samples by considering not only the spectral change characteristics of a single pixel,but also the neighboring space texture features around the pixel.Experiments that comparing the proposed method with the benchmark method verify the effectiveness of the method.(3)For the very challenging task of cross-seasonal image change detection,the key challenge of the task is illustrated.This paper proposes to use bi-lateral domain transferring network to simplify the cross-domain change detection task to an easier unique-domain change detection task.A fusion method is proposed to fuse multiple change detection results produced by different image pairs.The classic object-based unsupervised change detection method is used to verify the reliability of the domain transferring network for unsupervised change detection.