A Multiple Feature Fully Convolutional Network For Road Extraction From High-resolution Remote Sensing Image Under Complex Terrain Condition

With the advent of the era of big data,various remote sensing satellites have been launched,and it has become easier to obtain remote sensing images.The extraction of regions of interest in remote sensing images has become a research hotspot.The road is the main part of the modern transportation system,and it has important geographical,political and economic significance.The extraction of roads from high-resolution remote sensing images has been widely used in map mapping and post-disaster rescue.The development of artificial inteligence makes it possible to extract roads quickly and accurately from remote sensing images.In recent years,convolutional neural networks have performed well in the field of speech recognition and image recognition.On this basis,a ful y convolutional neural network has been generated.This method is a semantic segmentation model,it can extract specific information from images,and so it is chosen and improved as the road extraction method in this paper.This paper choose China-Nepal highway as the research area and use the domestic highresolution remote sensing satelite data.This paper establishes a convolutional network road classification model and a ful y convolutional neural network road extraction model.Due to the complex terrain of the study area,the digital elevation information is innovatively added to enhance the model's extraction ability,and the original FCN model is improved.A multi-feature ful y convolutional neural network is proposed on the basis of the original FCN model.The MFFCN model could extract both spectral and topographic features of the road.The specific research contents are as follows:First,preprocessing the remote sensing data.This paper uses the GF-2 and ASTER GDEM data as the input of the model.Firstly,processing the GF-2 data with ortho-correction and image registration,and second,fusing the panchromatic band(1m)and multi-spectral band(4m)to obtain multi-spectrum image with 1m resolution.Then the slope and aspect of the ASTER GDEM data(30m)are extracted,the result is interpolated into 1m resolution,and merge with the GF-2 image to obtain a RGB,slope,aspect and elevation 6-channels image.Second,in order to adapt to the input of the network,the 6-channels image is cut into a size of 100×100,and is made into a data set.Since the data set contains a large number of nonroad images,a road classification model based on convolutional neural network is proposed to screen out the images containing roads,which saves a lot of extraction time for non-road images.Thirdly,a road extraction model based on multi-features ful y convolutional neural network is proposed.Firstly,the different feature fusion modes of FCN are compared,and the best one among FCN-32 s,FCN-16 s and FCN-8s is chosen.Then the FCN model is improved,some convolutional layers are deleted,and ASTER GDEM data is combined to propose a multifeatured ful y convolutional neural network.This model can process multi-channel data and extract spectral features and terrain features,it has an outstanding performance in mountain road extraction.Experimental comparison results show that the improved model greatly improves the efficiency and accuracy of road extraction.