Road Networks Extraction Research Combined Classification Of Object-based Deep Learning With Template Matching

As important man-made objects and basic geographic data,roads have always been the focus of map updating.Accurate extraction of road information is not only a key issue but also a challenging one.The methods to extract road networks form sensing images can be divided into two representative categories:classification-based methods and template matching-based methods.Classification-based methods,such as object-based classification,the accuracy and morphological accuracy are limited by the image segmentation and other technical aspects.While deep learning-based classification has high requirements on the volume and quality of training data.On the other hand,template matching technology uses the template set by the user to identify specific objects from images,which can avoid the occurrence of inaccurate shapes during segmentation.However,creating and editing templates may be of great complexity.Moreover,it is difficult to ensure the completeness and universality of the template in matching remote sensing images with wide range.In this paper,a new method which combined by object-based method,deep learning method and template matching technologies are proposed to realize automatic extraction of road network in high-resolution remote sensing images.The main works of the paper are as follows:(1)Initial road networks are extracted by classification combined object-baed methods with deep learning methods.A sample library is designed for deep learning classification,and a deep learning classification model is obtained through sample library training.In the road network extraction based on classification,first,the region primitives are obtained by hard-boundary-constrained image segmentation,and then the primitive deep learning classification is realized by the way of voting image blocks in each region primitive.Secondly,road segments in the classification results are extracted for skeleton calculation,and the road skeletons are filtered to obtain the initial road skeleton networks.(2)Road networks are refined based on automatic template matching.For the initial road skeleton networks,there are defects such as morphological inaccuracy and positional offset.Focus on these defects,two template matching technologies,rigid template matching and deformable template matching,are introduced to modify shape and position of road skeleton networks,and the double lines of road are formed as the final results.Differ from traditional template matching approaches,our initial road templates are automatically generated based on the initial road skeleton networks.Multiple high spatial resolution remote sensing images with different sensors and different environmental backgrounds are selected as experimental data sources for method testing.The experimental results show that rigid template matching has higher positioning accuracy in rural areas,while deformable template matching technology has higher positioning accuracy for urban areas with complex backgrounds.In practice,it is necessary to comprehensively consider the background environment of the road and select the appropriate template matching method to refine the road network.It is turned out that the combination of object-based,deep learning and template matching techniques can correctly extract the main road information from the image.The proposed method has high extraction precision and better universality.