Research On Remote Sensing Image Feature Classification Algorithm Of Island Coastal Zone Based On Deep Learning

Island coastal zone is an important element of geographical resources in China.It is one of the key factors affecting China’s coastal economic development.It has rich land resources and significant social value.High resolution remote sensing images provide accurate and effective data sources for island coastal zone related researches,and Geographic Information Science also provides reliable technical support.In recent years,the extensive application of deep learning in the field of computer vision provides a new idea for remote sensing image feature classification,which can meet the challenges of high-resolution remote sensing image and increasingly complex feature classification,and obtain higher target recognition and interpretation accuracy.However,there are some problems in the current researches on island coastal zone,such as less research contents,lack of standard sample database required for in-depth learning,and poor accuracy of traditional remote sensing image classification methods.Based on this,this thesis proposes a new method,which improves the classical U-Net network model of deep learning,adjusts the hidden layer structure,introduces VGG-16 feature extraction network,adds spatial dropout layer,debugs super parameters and other methods to optimize the algorithm model,so that it can be applied to the target feature recognition and detection of island coastal zone.This thesis systematically studies the theoretical knowledge and framework practice of deep learning,and the proposed method provides a meaningful reference for the classification of surface features in high-resolution remote sensing images of island coastal zone.This thesis mainly carries out the following work:(1)This thesis describes the characteristics,advantages and disadvantages of remote sensing image classification methods based on spectral statistical features and machine learning,especially the application of deep learning in the field of remote sensing image target recognition and detection.This thesis deeply studies the basic theoretical knowledge of convolution layer,activation function layer,deconvolution layer and spatial dropout layer involved in target recognition and detection,and introduces in detail the classical model U-Net network structure of deep learning and typical feature extraction network,which lays a solid foundation for building a target feature recognition and detection algorithm suitable for sea island coastal zone.(2)Produce high-quality standard dataset of island coastal zone,Multi-Object Coastal Supervision Semantic Segmentation Dataset.Collect and preprocess the high-resolution remote sensing image data of the island coastal zone,determine to select mangrove,floating raft aquaculture and coastal aquaculture as the research target features,and conduct vector annotation,vector grid conversion,batch cutting,data enhancement and other operations on the remote sensing image of the island coastal zone with reference to the large open deep learning sample dataset,so as to finally form a multi-object coastal supervision semantic segmentation dataset containing 10934 images,which is a supplement to the high-resolution sample data set of island coastal zone for deep learning.(3)Create a remote sensing image feature classification algorithm for island coastal zone based on the classical U-Net network model.The classification algorithm of remote sensing images of island and coastal zone based on classical U-Net network model is constructed.The experimental model in this thesis adopts the classical deep learning semantic segmentation network of U-Net as the overall architecture,uses the first 13 layers of VGG-16 network as the feature extraction module,introduces the batch normalization method to accelerate the convergence of the model,adds a spatial dropout layer to prevent the model from over fitting,makes up for the loss of detail information caused by down sampling through the feature fusion method,and restores the semantic information and size of the image with the help of up sampling.Finally,the experimental model successfully realizes the task of surface feature classification of remote sensing images of island coastal zone.(4)Build the model running environment,optimize the algorithm model,and analyze the target recognition and detection results.In this thesis,Anaconda platform is used as the algorithm interpretation environment,and Tensorflow,the mainstream deep learning framework,is used to compile the surface feature classification algorithm of remote sensing image of island coastal zone.The batch-size,learning-rate,initialization method and the presence or absence of spatial dropout layer,which are the key factors affecting the experimental classification results and accuracy,are compared and tested respectively.After forming the final algorithm model,it is applied to the multi-object coastal supervision semantic segmentation dataset of island coastal zone,and compared and analyzed with the classification results and accuracy based on Segnet network model.The experimental results show that the experimental model in this thesis can not only obtain better classification effect and higher accuracy value,but also meet the detailed requirements of land resources in complex situations.It has certain practicability and generalization ability,and can provide meaningful reference for the classification of island coastal zone land resources in high-resolution remote sensing images based on U-Net network model.