Research On High Resolution Remote Sensing Image Rode Extraction Based On Semantic Segmentation

With the development of satellite technology,the clarity of remote sensing satellite images obtained is also correspondingly improving.High resolution remote sensing images have the advantages of wide coverage area,high resolution,and diverse land information.Road information has significant application value in urban planning,digital map updates,and other fields.For complex environments,road extraction is often disturbed by plants,shadows,and buildings,making road extraction one of the difficulties in high-resolution remote sensing image data applications.Semantic segmentation models are widely used in the field of image segmentation,and their application in high-resolution remote sensing images has important practical significance.This article conducts in-depth research on the existing semantic segmentation models,focusing on the phenomenon of incorrect or missed extraction of obstructed or blocked roads in existing methods,as well as the problems of complex models and high computational complexity.And based on the semantic segmentation model Deep Labv3+,the original network model was improved.Firstly,the lightweight network Mobile Net V2 was used to replace the Xception backbone feature extraction network in the original model;Secondly,using deep separable convolution DSC to replace the ordinary standard convolution in the ASPP module;Once again,CBAM module is introduced;Finally,in order to solve the problem of unbalanced road extraction categories in remote sensing images,Dice Loss and binary cross entropy loss function are assigned different weights,and the two are superimposed as the final loss function.And evaluate the improvement effect using indicators such as accuracy,recall,and F1-score.The optimized Deep Labv3+ model was tested on the Massachusetts Roads dataset,and the accuracy,recall,and F1-score of the original Deep Labv3+ algorithm were 84.05%,82.45%,and 83.70%,respectively.The accuracy,recall,and F1-score of the improved Deep Labv3+ model were 85.46%,85.27%,and 86.42%,respectively.The experiments demonstrate the effectiveness of the proposed method.Compared with other semantic segmentation models,the accuracy,recall,and F1-score of the FCN-8s algorithm are 80.30%,78.72%,and 79.66%,respectively.The accuracy,recall,and F1-score of the U-Net algorithm are 82.26%,82.58%,and 81.13%,respectively.The experiment shows that compared with other methods,the proposed method combines both extraction speed and accuracy,making it more suitable for road extraction tasks in remote sensing images.