A Study Of Deep Feature Learning-driven Change Detection For Very-High-Resolution Remote Sensing Imagery

As one of the earliest and most widely used techniques in the field of remote sensing imagery processing and analysis,remote sensing imagery change detection can help us detect,analyze and understand the change information on the earth's surface.At present,remote sensing imagery change detection technology has played an important role in land-cover and land-use change analysis,urban expansion studies,ecosystem monitoring,and natural disaster assessment.With the development of remote sensing earth observation technology,more and more remote sensing sensors can provide very-high-resolution remote sensing images.Compared to traditional low-and medium-resolution multispectral remote sensing images,very-high-resolution remote sensing imagery has less spectral information and the accompanying high spatial resolution makes the heterogeneity within the same land-cover objects increased.This makes it difficult for traditional change detection methods to achieve performance that meets the needs of practical applications in interpreting land-cover changes in multi-temporal very-high-resolution remote sensing imagery.Deep learning techniques can effectively and automatically extract hierarchical and representative features from the input data and are therefore well suited for processing very-high-resolution remote sensing imagery.At present,deep learning-based change detection in multi-temporal very-high-resolution remote sensing imagery is still in the development stage.Most of these methods are limited to the basic binary change detection,while the abundance of data sources actually provides data support for more in-depth change detection subtasks.This thesis analyzes and summarizes the shortcomings of existing methods based on a thorough investigation,and carries out in-depth research on multi-class change detection,multi-source change detection,and crossdomain change detection on very-high-resolution imagery respectively.The research work in this thesis mainly includes the following aspects:(1)In the study of multi-class change detection on very-high-resolution remote sensing imagery,to address the two problems that most current unsupervised deep learning models break the spatial dependencies within very-high-resolution images in the process of feature extraction and ignore the semantic change information,this thesis proposes an unsupervised feature extraction model,called kernel principal component analysis convolution,based on the subspace learning algorithm kernel principal component analysis.The proposed model can extract spatialspectral features from very-high-resolution imagery in a pattern similar to the convolutional layers in convolutional neural networks.This thesis then further proposes an eigenvalue-weighted change-informed polar coordinate,where the polar diameter contains information about the presence of change and the polar angle represents the type of changes,which can effectively use deep difference features to separate different kinds of changes.Based on these two key modules,this thesis proposes a deep kernel principal component analysis convolutional mapping network for unsupervised multi-class change detection in very-high-resolution imagery.Experimental results show that the proposed method can accurately detect different types of land-cover changes.(2)In the research of multi-source change detection on very-high-resolution remote sensing imagery,considering the fact that most current deep learning-based change detection methods can only process single-source homogeneous remote sensing images,to better and faster extract,interpret and understand the changes on our Earth's surface,this thesis proposes to combine convolutional neural network and recurrent neural network together to design a deep siamese convolutional multiple-layers recurrent neural network for multi-source change detection.The network first extracts spatial-spectral information from multi-source remote sensing data through a deep siamese convolutional neural network.For homogeneous and heterogeneous remote sensing images,the deep siamese convolutional neural network is designed as a pure and a pseudo-siamese network structure,respectively.Then,to efficiently mine the change information,this thesis constructs a multiple-layer recurrent neural network by stacking long short-term memory units.This structure allows temporal modeling of multi-temporal spatialspectral features to obtain spatial-spectral-temporal features with rich change information.Finally,the final probability of change is predicted by a feature mapping network.Experimental results show that the proposed method can effectively detect land-cover changes using both homogeneous and heterogeneous multi-temporal very-high-resolution remote sensing imagery.(3)The concept of cross-domain change detection is proposed for the first time and three different cases are given depending on the type of sensor to which the source and target domains belong in cross-domain change detection.For the first two cases,this thesis proposes a deep siamese feature generalization network.This thesis first introduces a multi-kernel maximum mean discrepancy,which is used to measure the difference in change information distribution between the source and target domains.This thesis then uses a multilayer embedding strategy in the network to embed the multi-kernel maximum mean discrepancy into the final layers of the network,reducing the difference in feature distribution between the source and target domains by minimizing the multi-kernel maximum mean discrepancy between the source and target domains during the training process.Experiments on three representative cross-domain change detection tasks show that the proposed framework requires only very sparse labeled data from the target domain to effectively transfer from the source domain dataset to the target domain dataset,and thus accurately interpret land-cover changes in the target domain datasets.