Intelligent Extraction Method Of Urban Green Space Based On High-resolution Remote Sensing Images

As an important part of the urban environment,urban green space plays an important role in many aspects such as human settlements,urban ecology,cultural entertainment,green economy,etc.Therefore,high-precision acquisition of spatial distribution information of urban green spaces has strong practical significance for optimizing urban green space,integrating green space resources,and building a green city.Therefore,highprecision acquisition of urban green space information has strong practical significance for optimizing urban green space,integrating green space resources,and building a green city.With the development of remote sensing technology,more and more high-quality domestic high-resolution remote sensing images have emerged,which can provide an excellent data foundation for obtaining spatial distribution information of urban green spaces.However,traditional methods for obtaining spatial distribution information of urban green spaces are costly,inefficient,and low in automation,making it difficult to meet the requirements of modern development for informatized and automated,high speed and high quality for urban green space extraction.Therefore,exploring an efficient,accurate and intelligent method for extracting urban green space is necessary.The rapid development of deep learning technology provides sufficient algorithmic support for intelligent extraction of urban green space,and also opens a new door for efficient and intelligent research on urban green space extraction methods.The remote sensing image of GF-1D satellite is used as the basic data for research,and the U-Net network is used as the basic algorithm,combining the high reflectance characteristics of vegetation in the near-infrared region,an intelligent extraction methods for extracting urban green space based on SD-UNet and Standard false color composite sample sets is proposed.The main work is as follows:(1)A series of methods have been adopted to optimize the network to improve the urban green space extraction performance of the U-Net network: increasing the convolution depth of the network encoder to enhance the network’s ability to learn features of urban green space;dense connection modules are used to strengthen the network’s information connection of different levels of green space feature maps;the problem of poor model fitting caused by a small number of sample sets is solved by adding batch standardization layer and tanh activation function;finally,separable convolutions are inserted in the middle of the encoder to reduce the number of network parameters and reduce the operational burden of the network.The experimental results show that the green space extraction accuracy index of the SD-UNet model(ACC(Accurancy):0.9447;IOU(Intersection over union):0.8740;Recall:0.9412)is improved compared to the U-Net model(ACC:0.9182;IOU:0.8305;Recall:0.9163).(2)Aiming at the problem of unclear recognition of some buildings and green spaces in the model caused by the relatively lack of spectral information in the R,G,and B band synthetic true color sample sets,selects standard false color synthetic image and the Normalized Difference Vegetation Index(NDVI),Red band,Green band synthetic false color images as spectral sample sets to participate in deep learning model training.The experimental results show that the spectral sample set can effectively improve the extraction accuracy of urban green space,and the SD-UNet combined with the standard false color sample set has the best accuracy,which compared with the true color sample set model,the ACC has increased by 0.0134,and the IOU has increased by 0.0237;Recall increased by 0.0165.Moreover,the application results of extracting urban green spaces from models of buildings,clouds,parks,suburbs scenes show that the model trained by SD-UNet and the standard false color sample set has better generalization performance.(3)In order to make the comparison between deep learning methods and traditional methods in extracting urban green space more intuitive,the best deep learning model has been used to compare with two traditional algorithms,object-oriented based classification and random forest,the ACC rates of various methods are: the best model of deep learning is 0.9581,and the object-oriented based classification is 0.8266,the random forest is0.8903.The results show that the deep learning method can fully mine the deep information of green space in the sample set,effectively improving the classification accuracy of the model,which proves that the deep learning method is more suitable for extracting urban green space than traditional methods.