Research On Low-grade Road Extraction Algorithms From Remote Sensing Images Based On Deep Learning

Up to now,there are still millions of kilometers or even more roads in the world that are not recorded in the system in real time,especially the low-grade roads in many remote areas.Low-grade roads are a more complex class of objects in the road area,because the diversity of their shapes,the complexity of the environment and the many interference factors make it difficult to obtain higher-quality road extraction results.While high-resolution images provide rich discriminative information for road extraction,they are also mixed with many noise interference factors.In response to the above problems,thesis takes low-grade roads as the research object and high-resolution remote sensing images as the data source to carry out research on road extraction based on deep learning.Considering the serious lack of low-grade road datasets at present,a low-grade road dataset was constructed with the Gaofen-2 satellite image as the data source.The details of the research contents are as follows:（1）A low-grade road extraction algorithm（DPIF-Net）with dual-branch information fusion is proposed,which has a dual-encoding branch structure with local and global information encoding.The former uses the advantages of convolutional neural network inductive bias in the spatial domain to characterize the local details of the road,and combines the multi-path dilated convolution to extract the multi-scale and topological information of the road.The latter builds a Trans Block module with a Transformer with a self-attention mechanism to model the global features of the road.The two are fully fused in the decoder part with contextual information fusion.Finally,a variety of advanced road extraction methods are compared in the Gaofen-2 low-grade road dataset,and the results show that the Io U and F₁ score of DPIF-Net both reach the highest,which are 61.4%and 76.08%,respectively.The low-grade road extracts details and completeness are better than all comparison methods.The generalization performance test was carried out on the Deep Globe dataset and the Massachusetts dataset.The results showed that the Io U was 57%and 53.82%,and the F₁ score was 72.61%and 70%,respectively.The indicators of the model were the best,with strong generalization performance.（2）In view of the problem of misjudgment and missed judgment in the single model in the extremely complex low-grade road environment,inspired by the ensemble learning algorithm,three low-grade road improvement segmentation models are constructed.Finally,it is verified on the Gaofen-2 low-grade road dataset and the Massachusetts dataset.The results show that the three improved segmentation algorithms have better low-grade road extraction performance,and all show strong generalization ability.Among them,Bagging Segmentation has the best extraction effect,increasing the Io U to 67.19%and 63.24%,and the F₁ score to 80.37%and 77.48%,respectively.Boosting Segmentation has the highest accuracy,reaching 90.88%and 84.18%.（3）A road automatic vectorization strategy is proposed,which can effectively generate the corresponding road vector results.Using Py Qt5 and GDAL to develop a low-grade road extraction prototype system with an interface.Finally,the system was tested using existing remote sensing data to confirm the convenience and efficiency of the system.