Research On Extraction Of Urban Green Space Based On GF-2 High Resolution Image

With the development of China’s economy and the rapid improvement of people’s quality of life,creating a green and low-carbon tourism city is an irresistible development trend in China’s sustainable green economy development path.The area of green space is one of the important indicators for measuring the degree of urban greening.How to efficiently and accurately measure the area of urban green space for monitoring the degree of urban greening has been a problem that scholars in the field of remote sensing image interpretation research need to study carefully.At this stage,algorithms for the classification of high-resolution remote sensing images are emerging,mainly divided into traditional methods and deep learning methods.In the traditional method,its accuracy is subjectively affected by human subjectively,and the workload is large.In reality,the remote sensing image is affected by factors such as mountain shading,chromatic aberration,aerosol,etc.,which easily leads to a decrease in the accuracy of ground object extraction.In the deep learning method,the model generalization ability is poor due to factors such as small training samples,small feature size and insufficient detail information,and the segmentation accuracy is low.In response to the above problems,this article will combine traditional methods and deep learning methods to focus on the method of extracting green space.The main research contents include:(1)In the traditional method,for the extraction of green space in remote sensing images,the problem of low classification accuracy due to the misdivision of green land caused by light,mountain shadows and other factors,this paper proposes a combinatio n of normalized water Exponential method and maximum likelihood method are used to extract green space.(2)In terms of deep learning methods,an improved U-Net algorithm is proposed to solve the problem that the small target segmentation performance is po or and the segmentation edges easily overlap in the U-Net network in multi-segmentation scenarios.Convolutions of different scales are introduced to replace the original single-size convolutions,and BN layers are added to solve the problem of easy over-fitting during network model training.At the same time,a method of combining jump connection and upper pooling operation is proposed to capture detailed information to further improve the accuracy of model recognition.At the same time,an extended training set of MGAN network model is designed.(3)Considering the classification accuracy of the further overall model,this paper proposes the use of an integrated learning strategy to optimize the segmentation results and reduce the problem of misclassification.Increased the overall classification accuracy of the model to 92%.From the perspective of multi-model fusion,the traditional method is used to extract the multi-spectral features and the deep learning method is used to extract the three-channel spectral features of the digital image,which further improves the accuracy of the feature classification.Experiments show that the method of combining normalized water index method and maximum likelihood method in this paper can achieve 85% accuracy in extra cting green space,which is greatly improved compared with the original method.The improved U-Net method has a segmentation accuracy rate of 91%,and the integrated learning method optimizes the result segmentation accuracy rate to 92%.