Road Extraction From High-Resolution Remote Sensing Images Based On Convolution Neural Network-Transformer

Extracting road information from high-resolution remote sensing imagery is of great significance for digital mapping,intelligent transportation,flood response,and disaster relief.Traditional methods have become inadequate to handle the massive growth of remote sensing data,which has prompted the development of deep learning techniques in the field of road extraction from remote sensing imagery.In particular,the application of encoder-decoder networks,such as the UNet model,has enabled efficient and intelligent road extraction methods in the remote sensing domain.However,convolutional neural networks(CNNs)have limitations in extracting long-range information due to the constraints of their convolutional kernels,which poses challenges for roads with special elongated structures.To fully utilize global information,this paper proposes a framework based on the encoder-decoder architecture and designs a parallel convolution and Transformer hybrid module.This module aims to extract both local and global road features,establish interactions between short and long-range information,and improve the accuracy and completeness of road extraction.The specific research content and innovations of this paper are as follows:A comprehensive overview of road extraction methods from high-resolution remote sensing imagery is provided,along with an analysis of the current research status in domestic and international contexts.The fundamental theoretical knowledge of deep learning CNNs and Transformer network models,as well as the encoder-decoder basic network structure,are discussed to provide a theoretical foundation for the proposed model.Furthermore,three classic remote sensing road datasets and their preprocessing methods are detailed to support the experimental analysis conducted in the paper.To enhance both the inference speed and accuracy of the model,a lightweight asymmetric convolution is designed to strengthen the backbone information.Additionally,an adaptive dense connection ASPP(Atrous Spatial Pyramid Pooling)is devised to improve the receptive field and effectively utilize multi-scale information from the imagery.The proposed methods are evaluated for their effectiveness and applicability on three remote sensing road datasets.A parallel CNN-Transformer hybrid module is introduced,and based on this module,a novel encoder-decoder network model named CTMUNet is proposed.Through ablation experiments,it is demonstrated that this module effectively leverages the advantages of asymmetric depthwise separable convolution and lightweight cross-strip attention to capture both local and global road features,resulting in improved completeness of road extraction.Comparative experimental results with existing methods on the three remote sensing road datasets demonstrate the effectiveness of the proposed approach.In summary,this paper presents an efficient and intelligent road extraction method by incorporating deep learning techniques and a parallel CNN-Transformer hybrid module.The proposed approach achieves promising results on high-resolution remote sensing imagery,demonstrating its significance in digital mapping,intelligent transportation,and flood response..