Research On Remote Sensing Image Road Segmentation Based On Deep Learning

Since the successful launch of the first remote sensing satellite,countries around the world have been actively engaged in research on remote sensing technology.Nowadays,with abundant sources of remote sensing data,the quality and resolution of images obtained are constantly improving.These massive highresolution remote sensing images provide powerful data support for the extraction of ground object information.Road is the infrastructure for pedestrians and vehicles,and plays the role of "node" in the whole social traffic network.Road extraction based on remote sensing images has a variety of application scenarios,such as:urban planning,automatic driving,road navigation,emergency command and other fields.The intricate background and abundant ground details in high-resolution remote sensing images make it challenging for traditional road extraction techniques to achieve accurate results.However,the emergence of novel approaches utilizing deep learning convolutional neural networks and their impressive performance in image segmentation have opened up new possibilities for road extraction from high-resolution remote sensing images.However,there are still some issues that need to be addressed,such as the lack of domestic road extraction datasets from high-resolution remote sensing images,the loss of texture details,omission,and misclassification when using convolutional neural networks for road extraction from remote sensing images,and the significant differences in sample categories between road and non-road areas in high-resolution remote sensing images..In view of this,the work content of this paper is as follows:(1)Aiming at the problem that the traditional method has a low degree of automation and is difficult to extract road information comprehensively.The road extraction algorithms U-Net,U-net++,LinkNet,D-LinkNet and CE-Net using convolutional neural network in image segmentation are compared.According to the characteristics of remote sensing image road extraction,U-Net with simple structure and stable performance is selected as the reference model of this paper.(2)Aiming at the lack of high-resolution remote sensing image road information extraction data set in China.The road information of remote sensing image obtained by Gaofen-2 satellite was used for preprocessing by ENVI software.The size of the preprocessing remote sensing image was cut to 512×512 small map,and 1509 remote sensing images containing roads were screened out.The road topology was marked by Labelme software to generate road label map.After summing up with CHN6-CUG data set,a total of 6200 pieces were constructed to construct domestic high-resolution remote sensing satellite road extraction data set.The contrast data set was DeepGlobe data set,which was jointly used for improving algorithm training.(3)To solve the imbalance between the sample road and non-road area categories and the problems of missing judgment and misjudgment when dividing roads.An improved road extraction model based on U-Net baseline was proposed.ResNet-34 network of ImageNet data set and SMU activation function were used as feature extraction module,which avoided gradient explosion and enhanced road detail feature extraction.The improved pyramid pool module is used as the context extractor to enlarge the receptive field and obtain rich road information integrating local and global features.Double attention mechanism is used in the upper sampling cascade to increase the network’s attention to frequency domain channel and spatial feature information.During network training,the joint optimization of Dice loss function and BCE loss function was employed to enhance the model’s performance and accelerate its convergence rate in the early stages.(4)After 100 training sessions on the domestic high-resolution remote sensing satellite road extraction dataset and DeepGlobe dataset,the F1_score and MIoU of the improved RSD-UNet reached 83.69%,83.66%,74.68%,78.56%in the two datasets,respectively.Compared with U-Net,U-net++,LinkNet,D-LinkNet and CE-Net,the indexes of RSD-UNet are higher than those of other classical remote sensing image segmentation algorithms.The ablation results of high-resolution remote sensing data sets in China show that compared with the benchmark network UNet,the scores of index IoU and F1 are increased by 4.4%and 3.07%,which verifies that RSD-UNet has excellent segmentation performance and can better realize road extraction of high-resolution remote sensing images.