Research And Application Of Road Recognition Technology Based On Remote Sensing Images

Remote sensing images are used to detect the electromagnetic radiation information on the earth’s surface through sensors located far from the ground,and then the information is transmitted and processed to form a digital image that can be processed by computers.Due to the optical characteristics of remote sensing images,after image transformation and image enhancement of remote sensing images,based on the extraction of imaging features,can be widely used in land cover monitoring,agricultural crop monitoring,environmental monitoring and even the military and other fields.As an important recognition target in remote sensing images,roads can be widely used in many fields such as urban traffic and earthquake rescue detection,which is of great practical significance.However,roads still face some difficulties in the accurate recognition of remote sensing images.Firstly,remote sensing images have high-resolution features,which provide more detailed features while also generating more complex background interference,affecting the extraction effect;secondly,different spatial resolutions lead to different morphologies of roads in images,such as rural and urban roads of different sizes and different road materials,so the presented optical features are not the same.These cause great difficulties for the detection of road targets.Most of the methods on road detection at this stage are classification methods based on spectral features,while ignoring other high-dimensional features such as morphological features of the road.Therefore,traditional manual feature extraction methods can easily ignore many high-dimensional features and require a lot of labor and time to adjust key parameters.And with the development of artificial intelligence in recent years,deep learning has played its unique and excellent role in many fields,and image target detection has become a hot area with the emergence of many excellent techniques that allow people to find solutions without having to go by empirical and interpretable ways.In particular,the deep neural network structure represented by U-Net has relatively excellent performance in the field of image recognition.However,the traditional U-Net structure has problems such as cumbersome structure,low correct rate and low efficiency.In this paper,we propose Res-U-Net based on U-Net with the addition of residual network for improvement,which makes the improved network more efficient,saves training time and improves the correct recognition rate.This paper further innovatively incorporates the attention mechanism and proposes A-Res-U-Net,which is finally trained on the dataset（Massachusetts Roads Dataset）and performs road detection.The experimental results show that the models proposed in this paper,after training,are all able to identify road targets in remote sensing images efficiently and accurately.Therefore,the model structure is feasible and has reference significance for remote sensing image road recognition.In addition,this paper further innovatively proposes a model for calculating road width on the deep learning model of road recognition,which is proposed based on the deep learning model A-Res-U-Net proposed in this paper,and completes the calculation of road width through breakpoint connection and three iterations,and proposes a thematic map of remote sensing image roads containing richer information.Finally,it is applied to the remote sensing image feature annotation platform,which greatly facilitates the remote sensing annotation efficiency of geography-related unit staff.