Research On Scene Change Understanding For High-Resolution Remote Sensing Images

With the rapid development of high-resolution remote sensing technology,the quality and data volume of modern remote sensing images have greatly improved.Nowadays,remote sensing images have sub-meter spatial resolution,clear geometric structure of ground objects,and rich spatial detail information,which exert high value in the applications of geographical space and time domain.Scene changes understanding based on multi-temporal high-resolution remote sensing images is the process of repeatedly observing,identifying and analyzing the dynamic changes of the same scene area or target at different times.In practical applications,repeated observation of long-term multi-temporal phases in the area of a specific scene(e.g.,an airport)is necessary to analyze the law of change of the target(aircraft)or phenomenon(motion trajectory)over time.However,it is a systematic project,and there are still many key technologies that have not been solved or are not perfect.The dissertation utilizes computer vision technology and studies the objectconstrained change detection,remote sensing target analysis,and remote sensing scene semantic change classification involved in the task of high-resolution remote sensing image scene change understanding from the perspective of local areas(or individual objects)to global areas(or entire scenes)at different levels of scene understanding.By enhancing these key technologies,the understanding of high-resolution remote sensing image change scenes can be improved.The main research includes:1.In the aspect of extracting changes in local areas of a scene,object-constrained remote sensing image change detection is proposed,which can be used to analyze changes in land cover in the same area over different periods of time.The visibility complexity of the scene is presented by multiple dimensions,including quantity,scale,shape and location of objects in urban areas.Also,remote sensing systems are affected by lighting and geomorphological structure,and targets with the same semantic concept exhibit different spectral behavior on different spatial and temporal scales.Therefore,the dissertation proposes a remote sensing image change detection based on multi-scale Swin Transformer and depth supervision network(MSTDSNet),which includes three main modules: wider and deeper aggregation module,measurement module based on multi-scale Swin Transformer,and deep supervision module.On the SYSU-CD and LEVIR-CD datasets,the work evaluates the performance of the proposed method,MSTDSNet,and compares it with five state-of-the-art remote sensing image change detection methods: FC-EF,FC-SIAM-DI,FC-Siam-conc,STANet,and DSAMNet.MSTDSNet outperformed all methods on the SYSU-CD and LEVIR-CD datasets,reaching the highest F1.2.To support the understanding of small-scale scene changes,the dissertation focuses on the application scenario of airports in remote sensing images,proposing an aircraft detection technology to assist in airport change analysis and aircraft motion estimation,namely the Rotated Aircraft and Aircraft Head Detector(RAAH-Det)for end-to-end aircraft and its head detection.However,accurate detection of aircraft and its head in high-resolution remote sensing images is a challenging task due to the difficulty of effectively modeling the characteristics of aircraft targets,such as appearance variations,large-scale differences,complex composition,and cluttered backgrounds.Therefore,RAAH-Det performs two main tasks: the extraction of interested region features based on U-Conv Ne Xt and the prediction module with six vector regressions(rotation box prediction and aircraft head key point prediction).The evaluation results of the RAAH-Det method on the DOTA-Plane dataset provide superior performance compared to the seven state-of-the-art target and keypoint detection methods.3.To support the understanding of large-scale scene changes,a framework for remote sensing scene semantic change classification was proposed to determine changes in semantic categories between two or more remote sensing scenes(such as converting vacant land into a residential area).In this process,remote sensing scene classification plays a crucial role as it is an important technique for realizing scene change analysis.Recently,many methods based on data-driven and machine learning methods have been proposed for this task.However,the existing algorithms only focus on image feature representation,and it is still difficult to deal with problems such as intra-class diversity,inter-class similarity,large difference in scene/target scale,and coexistence of multiple types of feature targets.In this dissertation,a new remote sensing scene classification network based on local and gloal semantic relationship(LGSRNet)is proposed,which includes three important modules: feature extraction based on Att Conv Ne Xt,semantic relationship learning module based on graph convolutional network,and joint expression learning module based on cosine similarity.Numerous experiments on two scenario classification datasets(AID and NWPU-RESISC45)have shown that LGSRNet outperforms night other state-of-the-art methods.4.The dissertation establishes an integrated prototype system for understanding scene changes based on multi temporal high-resolution remote sensing images,and conducts application validation experiments.According to the speed and process of analyzing the dynamic change of scene area or target object,the dissertation divides the understanding of scene change in high-resolution remote sensing images into long-time interval(such as expansion or new construction)and short-time interval(such as realtime scheduling management).The experimental results demonstrate that the system is suitable for both long-term and short-term applications in airport scenes,laying a foundation for future practical applications.