Road Extraction Of High Spatial Resolution Remote Sensing Image Based On Deconvolution Network

Road is a significant influencing factor of urban and regional development,which plays a key role in the construction and prosperity of both city and countryside.However,the traditional road detection method is mainly accomplished by professional surveying and mapping personnel in the field.This manual method spends plenty of manpower,material resources and financial resources which will result in low efficiency and the slow update speed of road information.Meanwhile,with the rapid development of remote sensing technology at home and abroad,high-resolution remote sensing images are also widely applied in the extraction of road information.However,high-resolution remote sensing images usually contain complex terrain information,and it is onerous to extract road network information accurately and completely.The research on road information extraction of high-resolution remote sensing images has been in existence for a long time.However,on account of the demerits of these methods that they only take the pixel or part of the message into consideration,and lack the analysis of multi-level and multi-scale semantic information in remote sensing images,the extracted road network information is incomplete and broken.This paper puts forward a plan which is to extract road information by using U-Net and RCF model on high spatial resolution remote sensing images to achieve more continuous and accurate road network extraction aiming at the above questions.The processing and data enhancement of high-resolution remote sensing images can generate the road sample set of high-resolution remote sensing images that conforming to network training and testing.The U-Net model and RCF model used in this paper have the following advantages:(1)The U-Net model is a coding-decoding structure.The down-sampling process is the encoding process of an encoder that extracts image features through the convolutional layer.The process of up-sampling is the decoding process of a decoder that restores the position information of the image.Meanwhile,the hidden layer ofU-Net model has plenty of feature dimensions,which is beneficial for the model to learn more diverse and comprehensive feature information.(2)The RCF model makes utmost of each convolutional layer to learn road information,and performs complete side output on it,and fuses the feature information output from the side.The side output network structure realizes the extraction and fusion of multi-scale and multi-level road semantic features to take full advantage of the information from the lower layer to the upper layer.(3)The RCF model is supervised deeply by manually labeled sample images during the road learning.The images are utilized to give feedback and adjust the road features learned by the convolutional layer at each stage to learn the optimal road feature parameters and improve the accuracy of road information extraction.Comparative analysis with pixel-based convolutional neural networks and super-pixel convolutional neural networks,the results demonstrate that the road extraction model proposed in this paper has a better road extraction effect.At the same time,in terms of quantitative analysis,the F1-score indicators of the U-Net model and the RCF model showed high scores of 88.9% and 91.5%,respectively.