Research On Intelligent Road Network Detection Technology Of Remote Sensing Images

The advancement of aerospace technology and high-resolution payloads makes remote sensing images broader in breadth and higher in spatial resolution,and the geomatics industry’s application need for remote sensing image analysis and processing is growing.Road distribution has been a research hotspot in the field of remote sensing intelligent processing as a key factor for recognizing characteristics in the geomatics business.Many factors in the imaging process of satellite images,such as sensor noise,light changes,and the complexity of feature distribution,have a negative impact on downstream tasks,while road network detection is highly susceptible to interference from trees,buildings,and vehicles in the images,resulting in missed detection and false detection.Traditional road network detection analyzes and processes geometric and semantic elements of roads,which has significant limits and makes meeting the need for autonomous extraction challenging.Furthermore,most deep learning-based algorithms have low timeliness and connectedness,and their outputs are usually raster binary pictures with no topological links rather than road network vectors.In light of this,this study performs research on the convolutional structure,network design,and post-processing based on real application needs.This paper’s research is summarized below:(1)A road detection network based on attention gate is developed to solve the shortcomings of the existing remote sensing picture road network detection model’s high number of parameters,restricted accuracy,and low operation efficiency.In the coding section,the fully pre-activated convolutional module is designed to improve the model’s detection accuracy,and the encoder structure is designed using small convolutional kernel network multi-layer stacking to reduce the number of model parameters and improve operation efficiency.To improve attention to the road region,the attention module is introduced in the jump link between the respective codec feature layers.The experimental findings reveal that the model developed in this chapter has high detection accuracy and efficiency of operation.(2)A road detection network based on dilated convolution and multi-scale fusion is proposed for further improving the accuracy of road network detection results.Multiple dilated convolution structures are designed to build a multi-dimensional contextual feature extraction module for feature extraction;in addition,a multi-stage road feature fusion module is built to fuse different levels of features to further improve the accuracy of road extraction results.The experimental results show that the designed model can obtain a more accurate road network structure with higher accuracy.(3)To incorporate practical applications,in addition to improving the network model,post-processing operations are designed in this paper to optimize the road network results.The broken road sections and the internal missing sections are connected to maintain the connectivity and integrity of the extracted roads,and the sliced vectorized road network extraction results are vector merged and smoothed.Experiments show that the road detection model proposed in this paper performs well in the detection of road networks with remote sensing images in local areas of Wuhan City and Shenzhen City.