| Remote sensing is an effective technique for the development,utilization,protection and management of natural resources,with its technical characteristics of objectivity and comprehensiveness.In recent years,the land surface status and change information extracted by satellite remote sensing has played a fundamental role in national major projects and applications such as natural resources monitoring and supervision,land survey,etc.The research and engineering practice of change detection by using remote sensing data are hot and difficult issues.Furthermore,the research on change detection based on artificial intelligence technology with deep learning as the main body is a core development direction in the future.There are some challenges existing in land cover change detection,such as low quality of spatio-temporal spectrum information fusion,and poor compatibility between different deep learning models based on multi-source remote sensing data.This study focuses on spatio-temporal spectrum fusion and feature extraction,remote sensing deep learning network model building,and business-driven multi-source remote sensing image change detection system construction.An intelligent processing flow is designed covering data acquisition,model construction,change information extraction,etc.A change detection system is also developed according to business application requirements.Large-scale applications are carried out,providing technical support for business management,such as natural investigation and monitoring,supervision and law enforcement,and land surface elements information update.The main studies are summarized below:1)An algorithm of spatio-temporal spectrum information enhancement and feature coupling is developed to solve problems,such as single spectrum segment in spectral dimension,shallow time series feature mining in time dimension,and ignoring the spatial correlation between pixels in spatial dimension.Herein,a spatio-temporal spectrum integrated feature extraction and coupling algorithm,STSNet(SpatioTemporal-Spectral Network)is proposed.It realizes the coupling calculation and analysis of spatio-temporal spectrum,and fusion of multi-dimensional information and features in three dimensions of time,space and spectrum,providing fundamental data for land surface information change detection.2)A model of remote sensing change detection is developed based on pyramid self-attention mechanism and multi-application collaborative to reduce the pseudo change resulted by difference of illumination,time and shadow between different remote sensing images.The pyramid pooling and hierarchical global information with different scale and sub zone are calculated by a twin change detection network structure.The two-dimensional vector matrix,attention map and context information is obtained by constructing a hierarchical pyramid and self-attention mechanism.Also,the multi task planning mechanism is constructed,and the parameters of the auxiliary task are used to iteratively optimize the training of the auxiliary main task to obtain the change detection probability information.The datasets of LEVIR-CD and ZY-1 02 D are used to evaluate the performance of the proposed algorithm.It shows that the algorithm proposed in this thesis significantly improves the edge and fine granularity and reduces false rate of change detection results,which is significantly better than the results of CVA,IRMAD,PSPNet and other methods.3)A remote sensing change detection method is proposed based on spatiotemporal transformation and three-dimensional convolution(Conv-3D)model to solve the interference factors in change detection,such as high resolution image registration error,image color difference because of different acquisition time,terrain,building shadow etc.Firstly,a spatiotemporal attention mechanism combined with the optimization processing method of image features in different periods is designed.Secondly,the time classification is used to design the learnable global semantic token parameters,and the semantic token is encoded with spatial information location to achieve the global context information acquisition.The pyramid structure ASPP3 D is then constructed according to different void rates,and the feature aggregation calculation is carried out to obtain the enhanced features with more robustness for simple algebraic operation of bit graph.The sum of Soft Max cross entropy loss and Dice loss is used as the final supervision loss function to enhance the contribution of the enhanced features to the final classification.The WHU-CD dataset is used to test the performance of proposed change detection method.It shows that the proposed method not only ensures the fine granularity of building detection edges and change information but also obtain better values of F1 scores,Io U,recall,accuracy and other indicators,comparing with FC-EF,DASNet,DDCNN and other algorithms.The proposed algorithm is more precise in connectivity and boundary shape of buildings and other changing areas than DDCNN using the dataset of Xiong’an Area.4)A remote sensing change detection method of pre-generating significant map of depth change is proposed based on the analysis of the coupling relationship between model parameters and detection performance,to solve the large number of deep learning model parameters and longtime of training of high resolution and long time series images in complex environment.A new end-to-end lightweight change detection network model is designed with fewer parameters and relatively fast calculation speed.More robust change features are extracted by using the pre-depth separable change map convolution,eliminating the possible semantic errors on the feature map in different periods.The sum of Soft Max cross entropy loss and Dice loss is selected as the final loss function,and the remote sensing change detection network(PDACN)for pregenerating significant depth change map is constructed by generating the change binary map through the Argmax operation.This model obviously has less parameters,which can effectively filter out the noise of unchanged positions,ensuring more robust fusion and extraction of changing features and easily applying in a large scale.The datasets of LEVIR-CD and SYSU-CD are used to test the performance of proposed method.Compared with FC-EF,FC-Siam-Conc,FC-Siam-Diff,BIT and other methods,the proposed method suppresses the noise and achieves good results in all indicators,with the least number of parameters.5)According to large-scale business application requirements using different remote sensing images,an intelligent processing flow and a change detection system are developed and large-scale applications are carried out.A work process including multi-source remote sensing image sample collection and management,multi-source image data feature analysis,change information extraction,statistical analysis and visualization of extraction results is designed to meet the needs of remote sensing image change detection business.The sample collection principle and sample database of land cover change are designed based on the introduction of national vegetation zoning,agricultural zoning,geomorphic zoning,climate zoning,farming system zoning and other fields.The technologies such as model/sample self-evolution based on iterative feedback,model management and dynamic mobilization based on micro-service architecture,training and evaluation based on model feedback optimization in multiuser scenarios,are integrated.A business-driven multi-source remote sensing image change detection system is developed by comprehensively applying algorithms proposed in this thesis.The service capability of "data acquisition-model construction-change information extraction" has then been formed.The overall technical process has been tested and applied in such businesses as rapid extraction of changes detection in ecological environment and surface features,real-time monitoring of land resources in Qingdao,land surface change monitoring of satellite remote sensing in Xiong’an area.The accuracy of remote sensing monitoring data of natural resources and ecological types has been improved by 5-8%,The extracted results can meet various business needs.There are 82 figures,33 tables and 183 references in this thesis. |
PDF Full Text Request