Research On Remote Sensing Classification And Application Of Urban Green Space Based On Deep Learning

High-precision urban green space monitoring is essential for optimizing the spatial structure of urban green space,maintaining urban ecological balance and developing green city construction.On the one hand,high-resolution remote sensing images provide important data support for urban green space information extraction and other fields.On the other hand,the rapid development of deep learning semantic segmentation model also provides algorithm support for urban green space classification of high-resolution remote sensing images.In this paper,the urban green space in the region within the Fifth Ring Road of Beijing is taken as the research object,and the GF-2 multi-temporal remote sensing images are taken as the data sources.We propose a deep learning urban green space automatic classification method combined with feature engineering,which is used to predict urban green space classification in Beijing.On this basis,we use remote sensing technology,geographic information system technology and Fragstats software to analyze the landscape pattern and evaluate the carbon sink function of urban green space within the Fifth Ring Road.The major work is as follows:Firstly,to investigate the classification effects of different methods on urban green space,this paper compares three traditional machine learning classification methods and four deep learning semantic segmentation models.The seven classification methods are listed in order of classification accuracy from large to small:HRNet(93.15%),U-Net(93.08%),Seg Net(92.61%),Deep Labv3+(91.57%),NNs(89.34%),SVM(86.15%),and MLE(78.87%).The results show that the deep learning method is more suitable for urban green space classification than the traditional machine learning classification method.The advantage of deep learning urban green space classification methods is that it can fully mine the feature information of data,effectively reduce the"salt-and-pepper noise"generated during the classification process and realize high-precision pixel-level classification of urban green space.Among all the methods,HRNet performs best in urban green space classification task,with an overall classification accuracy of 93.15%and FWIo U of87.90%.Secondly,to improve the classification performance of HRNet in urban green space classification task,we optimize the loss function and training process of the HRNet model.(1)The loss function of Focal Tversky Loss is introduced to reduce the data imbalance in the task of urban green space classification;(2)The urban green space data enhancement is added to ameliorate the generalization performance of the model;(3)The learning rate automatic optimization strategy and training early termination strategy are added to improve training efficiency of the model and realize the optimal parameter training of the model.The experiments demonstrate that the overall classification accuracy using Focal Tversky Loss is improved by 0.22%and FWIo U by 0.24%compared to the Cross Entropy Loss.The classification effect of evergreen trees and grasslands is improved using data augmentation,and the model has better generalization ability with an overall classification accuracy of 93.55%.Thirdly,this paper proposes a deep learning classification method of urban green space combined with feature engineering,which can effectively improve the HRNet model's feature richness learning of GF-2 images.We construct the urban green space feature engineering from three aspects,including selecting the normalized vegetation index(NDVI)as the vegetation feature,constructing the difference of NDVI between summer and winter images as the phenological feature,using the gray level co-occurrence matrix and LBP circular algorithm to obtain the image texture features,and extracting the best texture feature through the Relief F algorithm.Then,we add vegetation,phenology and optimal texture features as auxiliary bands to HRNet model for training and classification.The results show that the introduction of vegetation feature and phenology feature can effectively improve the classification accuracy,while the addition of texture feature does not improve the classification results.The study further uses the model fusion to optimize the final classification results.The fused model better combines the advantages of vegetation feature and phenology feature,compensates the problem of mixed classification of evergreen and deciduous trees,and solves the misclassification and omission of grassland.The overall classification accuracy of the model fusion is improved from 93.55%to94.50%.The F₁-Score of deciduous tree,evergreen tree and grassland increase by1.4%,6.55%and 4.97%respectively,compared with those without any features.Finally,the method proposed in this paper is used to predict the urban green space classification results in the region within the Fifth Ring Road of Beijing.On this basis,we use Fragstats software to analyze the spatial pattern of green space in the entire area within the Fifth Ring Road and four belts from urban center to fringe.In addition,we construct a 100 m×100 m grid in the region within the Fifth Ring Road of Beijing,and use remote sensing area estimation method to estimate carbon sequestration and carbon dioxide absorption per hectare of urban land within the Fifth Ring Road of Beijing.The results show that deciduous trees occupy the dominant position in the urban green space pattern,accounting for 82.51%of the total urban green space area,while evergreen trees and grassland account for 8.05%and 9.44%respectively.The urban green space in Beijing is affected by urban development and human activities,with high fragmentation and moderate landscape diversity and evenness.Moreover,The carbon sequestration function of urban green space should not be underestimated.The annual carbon sequestration of urban green space within the Fifth Ring Road of Beijing is about 2994.75×10⁶kg,and the annual carbon dioxide absorption is about 11091.63×10⁶kg.