Land-cover Semantic Change Detection Of Remote Sensing Digital Images

With the constant development of earth observation technologies and continuous updating of sensor equipments,the multi-temporal remote sensing digital image data,which covers increasingly vast areas,is becoming progressively abundant.The research on efficient digital image processing to excavate rich information contained in image data is of great significance to make full use of massive remote sensing data.Multitemporal remote sensing images with vast areas covered contain rich land-cover distribution information and change information,which can serve as sufficient data bases for the research on land-cover classification and change detection.Change detection can locate and identify the land-cover changes by simultaneously analyzing multi-temporal images taken on the same area,which can provide technical supports for water resource management,vegetation cover monitoring,climate and environment monitoring,ecosystem research and so on.Meanwhile,change detection can further provide the bases of macroscopic decision-making for the government in cultivated land protection and urban planning.However,traditional change detection algorithms only focus on the location of change areas and overlook the analyses of semantic information for landcover categories,which have been gradually unable to meet the increasingly complicated demands of practical applications such as urban planning and natural resource management.Therefore,experts and scholars are gradually committed to the research on semantic change detection,which analyzes the semantic categories and identifies the change types,so as to improve practical values of the research.Nevertheless,due to the complexity of land-cover distributions over vast areas within multi-temporal images,the research on semantic change detection still faces many challenges.Firstly,there is a lack of high-quality semantic change detection dataset containing vast areas,sufficient data samples and fine semantic annotations,which makes it difficult to support the research on semantic change detection algorithms with high generalization ability.Secondly,there is a lack of algorithms that fully consider in the diversity of land-cover distributions,which makes it difficult to fit on different land-cover distributions over vast areas within multi-temporal images simultaneously.Thirdly,there is a lack of algorithms with suitable outputs for practical demands,which makes it difficult to obtain pixel-wise change types and interpret the change information into comprehensible text conclusions simultaneously.Thus,existing algorithms can not directly serve for the practical applications such as urban planning.Based on the aforementioned key challenges and remote sensing image properties,including the land-cover distribution diversities over different areas among multitemporal images,the data annotation criteria are formulated and a large-scale well annotated semantic change detection dataset is created.Besides,concentrating on the land-cover semantic category identification in the semantic change detection research,theoretical analyses in terms of parameter optimization are given in this dissertation and a dynamic semantic segmentation structure for remote sensing images is designed.Meanwhile,considering in the remote sensing image properties,a dynamic deep architecture for semantic change detection is proposed.Finally,a text interpretation benchmark dataset for remote sensing semantic change detection is created and a multi task framework to implement semantic change detection with text interpretation outputs is designed.The main contents include the following aspects:1)Construction of the large-scale well annotated semantic change detection dataset.The annotation criteria of existing semantic change detection datasets mainly follow the research on land-cover classification.Thus,change types between the same land-cover category that are closely related to urban planning,such as building demolition and reconstruction,are difficult to be annotated and retained.In addition,existing semantic change detection datasets are defective to different extents in terms of the data quantity,annotation quality and geographical coverage.In order to solve these problems,a large-scale well annotated remote sensing semantic change detection benchmark dataset(SEmantic Change detecti ON Dataset,SECOND)is created from the prospect of image sampling,change type definition and annotation method,which distinguishes changed regions together with unchanged regions when annotating the land-cover categories so as to retain change types such as building demolition and reconstruction.Containing sufficient well-annotated data samples with vast areas covered,SECOND can better support the research on semantic change detection algorithms.2)Research on dynamic deep structures focusing on land-cover distribution diversities of remote sensing images.Existing semantic change detection algorithms apply homogeneous calculation procedure on each pixel to obtain land-cover classification results,which fails to fully consider in the land-cover distribution diversities over different areas.Meanwhile,among existing semantic change detection algorithms,homogeneous calculation procedures applied on multi-temporal images also overlook the land-cover distribution diversities over different image acquisition times.In order to solve these problems,starting from the land-cover classification in the semantic change detection task,theoretical analyses on existing dynamic deep structures in terms of parameter optimization are given,based on which a pixel-level dynamic semantic segmentation structure(Hidden Path Selection Network,HPS-Net)is designed to dynamically select adaptive filtering procedures for different pixels and better fit on different land-cover distributions over different areas simultaneously with the help of hidden variables.Furthermore,focusing on the properties of remote sensing images,a dynamic semantic change detection network(Asymmetric Siamese Network,ASN)is proposed to utilize different filtering procedures and better fit on different land-cover distributions over different image acquisition times simultaneously.3)Research on semantic change detection with text interpretation results to better meet practical demands.Interpretation processes of existing semantic change detection algorithms solely generate pixel-wise change types without the transformation to comprehensible text conclusions,which hinders the usage on practical applications such as urban planning.Although some researchers have proposed text interpretation models for the remote sensing image change detection task,pixel-wise change types and comprehensible text conclusions can not be produced simultaneously.Limited by the data annotations,the text conclusions of these algorithms are not intact and comprehensible enough.In order to solve these problems,a text interpretation benchmark dataset for remote sensing semantic change detection(SEmantic Change detecti ON Dataset CAPtion,SECOND-CAP)is created,which contains text descriptions about land-cover changes on the basis of change type annotations in the SECOND.Meanwhile,based on the research on dynamic semantic segmentation structures and asymmetric dynamic siamese networks in this dissertation,a multi-task framework(Semantic Change Detection and Text Interpretation,SCDTI)is designed to adjust filtering procedures according to the diversity between different tasks when considering in the different land-cover distributions over diverse areas in multi-temporal images and generate text descriptions,which can better serve for the practical applications.