Research On Urban High-resolution Remote Sensing Image Vegetation Information Extraction Based On UNet++ Network

As an important part of the urban ecosystem,urban vegetation is essential for reasonable planning and monitoring.The wide application of high-resolution remote sensing images enables the extraction of urban vegetation information with low cost,short cycle and high accuracy.However,high-resolution remote sensing images not only improve the recognizability of the naked eye,but also introduce more complex background information.Therefore,the limitations of traditional vegetation extraction methods such as visual interpretation and traditional machine learning methods become more and more obvious.The visual interpretation method is time-consuming,expensive,and has a low degree of automation,which can no longer meet the needs of production.Although traditional machine learning methods have achieved certain results in the research of vegetation extraction,relying on underlying-level features still cannot satisfy the extraction of high-quality vegetation information.In recent years,with the great success of deep learning,especially Deep Convolutional Neural Network(DCNN)in the field of natural image processing,it has been extended to the analysis and processing of remote sensing images,which has promoted the development of high-resolution remote sensing images related tasks,including semantic segmentation.In this dissertation,the convolutional neural network method is used to extract roads from high-resolution remote sensing images automatically,quickly and efficiently.The main research elements are as follows:(1)Aiming at the problems of intra-class dissimilarity and inter-class similarity in vegetation extraction from urban remote sensing images,a deep supervision method of multilevel output fusion is designed and introduced into UNet++ network.The multi-level output fusion deep supervision strengthens the control of the detailed feature information,learns the distinguishing and effective features,enhances the extraction performance of the model,and improves the agreement between the extraction results and the ground truth.(2)In order to solve the problem of ambiguous segmentation boundary and noise in the extracted vegetation area in the application of deep semantic segmentation network model to the task of high-resolution remote sensing image information extraction,a deep semantic segmentation method combined with superpixels segmentation was proposed.The method can further improve the accuracy of vegetation extraction.Firstly,the UNet++ network model with pruning is used in the depth semantic segmentation part to reduce the number of parameters and accelerate the computation speed.Secondly,the SLIC superpixels segmentation with fused texture features is used in the superpixels segmentation part,which can play a good segmentation effect on the boundary contour of the object.Finally,the prediction map of the deep semantic segmentation model and the superpixels segmentation result map are organically integrated,and the pixel labels are redistributed using the designed voting algorithm.The final vegetation extraction result map has clear boundaries and complete vegetation areas without noise.(3)The programming implements a user-oriented urban high-resolution remote sensing image vegetation extraction system,and a visual interface is designed to facilitate user operation.This system integrates all the vegetation extraction models and methods used in this paper,and efficiently realizes the automatic extraction of vegetation information from urban highresolution remote sensing images.It can also provide an intuitive comparison of the vegetation extraction effects of various algorithms to meet the needs of users to extract vegetation areas and calculate greening rates.