Research And Application Of Deep Learning Model For Road Extraction From High-resolution Remote Sensing Images

With the rapid development of artificial intelligence and economy,the population is increasing,and traffic problems are becoming more and more serious.Congestion has become the focus of people’s daily travel.The extraction and analysis of road information in remote sensing images can be updated in time,which has far-reaching significance for vehicle navigation and automatic driving technology.With the rapid development of science and technology,the image resolution is constantly improving,and the information is more abundant and clear,High resolution remote sensing has become one of the important means to accurately obtain land surface information.But how to extract effective information quickly and accurately is also an urgent problem.Nowadays,deep learning algorithm is gradually mature,and image semantic segmentation technology has become more accurate.As one of the most basic geographic information data,road information has always been paid attention to its extraction method and accuracy.The traditional semi-automatic remote sensing image road The extraction method mainly relies on pattern recognition,which is not only inefficient and not strong in generalization ability,so the use of deep learning algorithms to extract road information from high-resolution remote sensing images has become a development trend today.Over the years,a variety of different deep learning road extraction frameworks have emerged in endlessly.In this paper,through the comparison experiment of full convolutional neural network(FCN)and Unet model on the same sample data set,it is found that these two different extraction methods have some spatial information.The problem of loss and the degree of refinement of road semantic segmentation is not high,which can not solve the problem of interference of non-road information such as vegetation shadows,highrise buildings and rivers on target extraction.Because the ResNet residual structure has a very rich shortcut connection,which can greatly enhance the characterization ability of the network itself,this paper proposes a feasible road segmentation method based on the improved Unet network model,by using the ResNet50 residual module to replace the original Unet framework.The two convolution modules select the binary cross-entropy loss function in the loss function and combine the Dice coefficient to perform back propagation training,update and optimize the model,so that the road can be better refined semantic segmentation.Experiments show that the improved model based on this article has great advantages in highresolution image road extraction.It can not only effectively solve the interference problems of vegetation shadows,high-rise buildings occlusion,rivers and other non-road information,but also improve the segmentation accuracy of the model.It makes the connectivity and detail processing of roads more accurate,and has the highest accuracy rate compared with other road extraction models.Its research methods and technical routes have certain theoretical significance and practical value.