Research On Automatic Road Extraction From High-resolution Remote Sensing Images Based On Deep Learnin

High-resolution remote sensing images contain complex and diverse ground objects,among which roads,as one of the important basic ground objects,have a wide distribution.The automated and accurate extraction of road information from highresolution remote sensing images has significant application potential in fields such as unmanned vehicle driving,electronic map navigation,and smart city construction,and has become a hot research spot for many scholars in recent years.Because unrelated objects such as buildings and sidewalk shielding have greater interference in road information,high-resolution remote sensing image road information extraction is full of challenges.Currently,information extraction technology based on deep learning algorithms has shown excellent performance in the field of image segmentation and has become a research focus for many scholars.How to use deep learning algorithms for extracting road information from high-resolution remote sensing images,improve the accuracy of road category identification,and construct efficient road extraction models to effectively express deep level information features is crucial in research on road information extraction.Therefore,this article takes road information from highresolution remote sensing images as the research object.In response to the problems of information loss and errors caused by fuzzy information features and the influence of "same spectral foreign objects" in road extraction,deep learning algorithms are used for research and analysis,and an improved road extraction model is constructed.The main research work of this paper is summarized as follows:(1)Summarized the current difficulties in the process of extracting road information,classified and summarized representative research methods at home and abroad,and analyzed their advantages and disadvantages.Introduce the principle,composition,and training optimization methods of convolutional neural networks,and outline several classic convolutional neural networks.(2)Preprocess the road datasets of Massachusetts and Deep Globe,and construct road extraction models based on FCN network,PSPNet network,and U-Net network respectively.Analyze and compare the visual results and accuracy evaluation indicators of road extraction in different scenarios of these three models on the two datasets,and select the best road extraction model U-Net.(3)In response to the problems of missing and incorrect extraction caused by the ambiguity of road features and the "same spectrum different substances" phenomenon,a U-Net network-based road extraction model was constructed,which integrates residual and convolution attention mechanisms.Firstly,the Re LU activation function in the residual unit was replaced with a Mish activation function with self-regularization and non-monotonicity to alleviate the problem of gradient disappearance.Secondly,the improved residual unit was added to the U-Net network to deepen the network hierarchy,enhance feature learning ability,and maintain network stability.Finally,to improve the model’s representation ability for road information and refine the model’s segmentation ability,the convolution attention mechanism was embedded in the skip connection part to enhance the model’s performance and obtain more road detail features.(4)The improved model was tested through ablation and comparative experiments on both the Massachusetts and Deep Globe road datasets.In order to demonstrate its universality,four representative scenarios were selected for comparative analysis.The experimental results show that on both datasets,the overall accuracy,precision,recall,F1 value and other evaluation indicators of the improved model are better than those of the original U-Net model,and the extraction effect is better.When compared with the Deeplabv3+ and CE-Net models,the visual results of road extraction and various accuracy evaluation indicators have been improved to a certain extent.In summary,the performance of the improved network model in this paper is excellent,with obvious advantages and good semantic segmentation ability.