| Building change detection is an important research topic in the field of remote sensing ground analysis,which refers to the design of relevant algorithm models to extract building change features in multiple temporal images of the same area,which can be used for urban development process research,unreasonable building investigation and provide reliable data support for decision makers to formulate relevant policies.The development of remote sensing technology has led to the continuous improvement of image quality,the detailed texture characteristics of high-resolution images and the rich spectral characteristics of multispectral images provide the possibility of fusing and extracting feature change characteristics from multiple perspectives.The powerful image processing capability of deep learning technology makes it achieve good results in remote sensing change detection,but there are still the following problems: Firstly,the data used for change detection research are mostly single-source high-resolution remote sensing images,and too few spectral features limit the improvement of building change detection accuracy to a certain extent.Secondly,the complex texture features of high-resolution images pose a great challenge to the network model to accurately extract building changes,and the current deep learning model cannot better eliminate the influence of factors such as roads on the building extraction results,and it is also lacking in extracting building features at different scales simultaneously.To address the above shortcomings,this thesis uses deep learning techniques to conduct research on multi sensors high-resolution remote sensing images and building change detection methods that fuse high-resolution remote sensing images with multispectral remote sensing images,and the main research contents and conclusions are as follows.(1)Make a multi-source building change detection dataset.To address the lack of multi-source building change detection dataset,this thesis used Arc GIS for building change annotation based on the GF-1 satellite image data(spatial resolution of 2m)and Sentinel-2multispectral satellite image data(spatial resolution of 10,20 and 60m)for two periods in some areas of Huangdao District,Qingdao City,combined with Python language and Open CV,GDAL and other image and spatial data processing libraries,produced a multi-source building change detection dataset(MS-High BCD Dataset).The dataset had600 image pairs,of which 540 image pairs are used for training,30 for validation and 30 for testing,and the spatial resolution of the processed images is 2.5m,which provided a data basis for the research of building change detection method based on multi-source remote sensing data.(2)Research on building change detection methods for multi sensors high-resolution remote sensing images.Based on the open source building change detection dataset(LEVIR-CD),this thesis designed a multi-scale supervised fusion network(MSF-Net),introduced the channel attention mechanism,selective convolution kernel mechanism,and designed a dual-context fusion module and a multi-scale supervised fusion module.The model can effectively remove the influence of irrelevant elements on building change detection and can extract building change features at different scales simultaneously.Compared with current advanced deep neural network models for change detection,MSF-Net effectively improved the change detection accuracy of buildings,and achieved F1 of 88.66% and IOU of 81.3% in two important evaluation metrics.(3)Research on building change detection method by fusing high-resolution remote sensing images and multispectral remote sensing images.Based on the self-labeled MS-High BCD Dataset multi-source building change detection dataset,this thesis investigated the improvement of building change detection accuracy by multi-source data.Based on the different spatial resolution of each band of Sentinel-2 remote sensing images,three combinations of RGB bands of the high resolution remote sensing images of GF-1and the bands of the processed L2A-class Sentinel-2 multispectral remote sensing images are used,and six neural network models are selected for building change detection research.The experiments showed that adding multispectral image data to the high-resolution data source of MSF-Net,the F1 and IOU improved 0.67% and 1.09%,respectively,over the high-resolution images,and 7.57% and 6.21%,respectively,over the multispectral images,indicating that fusing multisource remote sensing data can effectively improve the change detection accuracy of buildings. |
