Research And Application On Deep Learning Based Highway Change Detection With High-Resolution Remote Sensing Images

The remote sensing image change detection algorithm determines the change among features by comparing remote sensing images from different periods.With the continuous development of remote sensing technology and sensors,we are able to acquire more remote sensing images with high spatial and temporal resolution,and these data have complex texture features and rich image details,which make the traditional change detection methods unable to handle efficiently.Deep learning-based change detection methods can effectively overcome the shortcomings of traditional change detection,but still have some challenges.In this dissertation,we conduct research based on the deep learning method,combine the change detection process,propose our own improvement and innovation method for the problems of remote sensing image resolution difference and change detection model robustness in change detection,and design and develop the change detection system along the highway based on the improvement method.Specifically,the research proposed in this dissertation is as follows:(1)In response to the issue of resolution differences caused by different sensors used in remote sensing change detection datasets,we use an improved SRGAN(Super-Resolution Generative Adversarial Network)to convert low-resolution remote sensing images into higher resolution ones through a generative adversarial approach,thereby overcoming the issue of remote sensing data resolution differences.In the generator part,we use composite convolution modules to more effectively extract image features and generate images closer to the real distribution.In the discriminator part,an attention mechanism is added to adaptively enhance image feature information,thereby improving the ability to distinguish between real and fake data.(2)To address the problems of single scale and failure to fully utilize deep multiscale features in change detection leading to low robustness of the model,this paper designs a deeper multiscale encoding and decoding feature fusion network DMEDNet(Deeper Multiscale Encoding and Decoding Feature Fusion Network).The model first performs deeper multiscale feature extraction for diachronic remote sensing images,and then performs absolute difference operation on the extracted features;the absolute difference information obtained in the encoding stage is fused with deeper multiscale features in the decoding stage,so that the lost information can be better complemented in the corresponding decoding process;finally,the multi-level features output from each layer in the decoding stage are fused,and then an improved attention mechanism for feature enhancement,thus improving the accuracy of change detection.The proposed model achieves an F1 score of 97.4% on CDD(Change Detection Dataset).(3)Aiming at the current rapid development of highways in China and the need for efficient change detection of features around roads,a high-definition remote sensing highway perimeter change detection system was designed and implemented.Based on the specific process of change detection by high-definition remote sensing deep learning,and accordingly realized and developed the corresponding high-definition remote sensing highway perimeter change detection system,which greatly improved the efficiency and accuracy of change detection.The development of this system provides feasibility for realizing change monitoring of remote sensing images around highways and provides an important reference for research and application in related fields.