Study On Remote Sensing Image Classification Based On Deep Learning

Remote sensing image data acquisition and processing has entered the era of big data application.Accurate acquisition of useful information from different types of data is the core content of remote sensing big data processing and application.Among them,classification and recognition of remote sensing images are important application fields.The traditional remote sensing image classification technology can no longer meet the new requirements and tasks of remote sensing big data processing.Deep learning technology developed in recent years has been successfully applied in image pattern recognition especially in man-machine competition.In order to solve the problem of classification and application of large data of remote sensing images,this paper studies the classification of remote sensing images by using the full convolution dense convolution network(FCDenseNet)in depth learning,the DeepLab system based on dilated convolution(AC),atrous space pyramid pooling(ASPP)and conditional random field(CRF)technology and the improved DeepLabV3 model and proposes two models based on gradient optimization theory,FC-DenseNet and DeepLabV3.The main work and contents of this paper are as follows:(1)Established a ship remote sensing image data set suitable for deep convolution neural network classification training to provide data for deep learning model training,testing and prediction.At the same time,for the problem of small amount of data for some ground objects or regions,a method of data enhancement is proposed to expand and process a small amount of data so that the data set can meet the requirements of deep convolution network model training.(2)Three depth volume integration models of FC-DenseNet,DeepLabV3 and DeepLabV3+ are deeply discussed and analyzed,classification experiments of ship remote sensing images,INRIA aerial remote sensing images and urban remote sensing images are designed.The research results show that the deep convolution learning method can effectively classify remote sensing images and obtain good experimental results.(3)Two improved optimization models are proposed through gradient descent optimization algorithm,namely,the optimized FC-DenseNet model and DeepLabV3 model,and test experiments are carried out with the previously established ship remote sensing data set and INRIA aerial remote sensing image dataset.The experimental results show that among the different gradient descent optimization algorithms,the RMSProp optimization algorithm is better than the Adaptive Moment Estimation(Adam),Adaptive Learning Rate(AdaDelta),Adaptive Gradient(AdaGrad)algorithms and has better robustness and convergence.