Research On Green Space Information Extraction And Application Based On Deep Learning Algorithms

Remote sensing images contain a wealth of information and how to use them to extract information of land types quickly and accurately is one of the key issues in the application of remote sensing technology.Green space information is of great importance for urban planning as well as ecological environment monitoring and management.When it comes to using remote sensing images to extract green space information,traditional extraction algorithms often fail to meet the requirement for accuracy and are prone to make mistakes in classification and omit classification.With the rapid development of remote sensing and computer technologies,artificial intelligence has been increasingly used in extracting information of remote sensing images and how to do this quickly and accurately is of great significance for urban planning and ecological environment monitoring.Based on deep learning algorithms and semantic segmentation methods of convolutional neural networks,this paper studies the use of remote sensing images in extracting green space information.This research and its conclusions are mainly as follows:(1)In addition to summarize deep learning algorithms and study semantic segmentation algorithms commonly used in convolutional neural networks,this thesis discusses the FCN,VGG,SegNet and U-Net models and comparatively analyzes their advantages and disadvantages for future research.(2)This paper uses the structure of VGG network and U-Net model to construct a deep learning algorithm based on the U-Net model of VGG structure.The analysis in this thesis of the advantages and disadvantages of the commonly used FCN,VGG,SegNet and U-Net models finds that U-Net model is of higher extraction accuracy and less parameter calculation while VGG16 model is better for feature extraction despite of its comparatively greater parameter calculation.Based on this,this thesis improves the semantic segmentation algorithms of deep learning and constructs a deep learning algorithm based on the U-Net model of VGG structure.The comparative analysis of the model constructed in this paper and the commonly used FCN,SegNet and U-Net models shows that the U-Net model based on VGG structure is better than other models extraction accuracy of green space information.(3)Based on the analysis of commonly used activation functions,this thesis constructs a combined activation function of the U-Net model of VGG structure by combining the advantages of TReLU and ReLU activation functions.The combined activation function constructed enhances the nonlinearity of the model mainly by alternately using the two activation functions of TReLU and ReLU and controls the negative half-axis unsaturated interval by adjusting parameters to obtain the required activation value.The experimental results show that the combined activation function is superior to the commonly used activation functions for its soft saturation,mitigation of gradient disappearance and greater robustness against noise.(4)This paper uses the combined activation function of TReLU and ReLU and the deep learning algorithm based on the U-Net model of VGG structure in researching green space information extraction,and constructs an extraction method of green space information based on the U-Net model of VGG structure.The comparative analysis of the constructed method and the traditional extraction methods of green space information show that the extraction method of green space information based on the U-Net model of VGG structure is better in accuracy.