The Research On Road Extraction In High-resolution Remote Sensing Images Based On Deep Learning

Road is a typical ground object in remote sensing images,and accurate road information is extremely important in applications such as autonomous driving,urban planning,and map drawing.With the perpetual development of the remote sensing observation technology,rich data resources for road extraction are provided by enormous high-resolution remote sensing images.Compared with on-site data acquisition and manual labeling by experts,extracting road information from high-resolution remote sensing images automatically is more efficient,which has become the current research hot-spot.The method based on deep learning has become the mainstream of road extraction research in highresolution remote sensing images by reason of its excellent feature extraction competence.Nevertheless,due to the complexity of the road itself and the background environment in high-resolution remote sensing images,the performance of most road extraction methods is not ideal.Therefore,based on deep learning,the thesis carried out the following work on the road extraction methods of high-resolution remote sensing images.Firstly,the road extraction method was studied based on encoder-decoder network structure.In order to effectively extract and reserve road features,the thesis first executed road extraction analysis based on Convolutional Neural Network and Transformer respectively,and then selected Link Net with excellent capability.Aiming at the loss of road information in the encoder structure of the network,the thesis also designed a Dense Feature Enhancement Module,which has been verified that it can effectively enhance the network’s ability to represent road features from comparative experimental results.Finally,the basic skeleton networks for subsequent road extraction experiments were built.Secondly,the road extraction method was studied combining the multi-scale information fusion approaches.For the purpose of improving the ability to extract road features of the basic skeleton network,the thesis designed a Dense Multi-scale Information Fusion Module,which used densely cascaded multi-scale atrous convolution operations and parallel global pooling operation to extract crucial multi-scale road feature information.According to the results of contrastive experiments,the Dense Multi-scale Information Fusion Module can more effectively improve the road extraction ability of the basic skeleton networks,and through the combination with the Dense Feature Enhancement Module,the accuracy and completeness of the road extraction results were superior to the comparison methods.Thirdly,the road extraction method was studied combining the attention mechanism.In response to the problem that the multi-scale information fusion method based on local calculations cannot effectively perceive the relationship between long-distance road pixels,the thesis designed a Double Criss-cross Strip Attention Module,which simultaneously paid close attention to different directions,such as horizontal,vertical,left diagonal and right diagonal,and computed long-range dependencies between road pixels within global range.On the basis of the results of comparative experiments,the Double Criss-cross Strip Attention Module can help the model to focus on more effective and pivotal road features,and finally obtain the best road extraction performance in terms of evaluation indicators and visualization results.