Research On Remote Sensing Image Classification Based On Deep Learning

Remote sensing image classification is the basic research of remote sensing technology,and improving the classification accuracy of remote sensing images is of great significance to the development of application field such as precision agriculture,military identification,environmental monitoring et al.However,remote sensing image classification currently faces issues such as spectral data redundancy,insufficient utilization of spatial information,and limited training samples et al.This thesis starts with the comprehensive and effective use of spectral-spatial information,the theory and model of deep learning are studied and applied to improve the extraction accuracy of spectral features and spatial features in remote sensing image.And then the spectral-spatial features are combined to achieve higher classification accuracy.The main work in this thesis focuses on the following aspects:(1)Based on the decision fusion method,a spectral-spatial feature fusion method is designed to solve the problem of spectral-spatial information in remote sensing image classification.In this method,the spectral features and their prediction results,the spatial features and their prediction results are fused and classified.This method can not only independently extract spectral features and spatial features,but also reduce the number of parameters and calculations in the classification process.What's more,it is important to improve the classification accuracy by effectively integrating the features of different structures.(2)Combined with the spectral-spatial feature fusion method,a remote sensing image classification algorithm based on the convolutional neural network is designed.The algorithm includes a one-dimensional convolutional neural network to extract spectral features and a two-dimensional convolutional neural network to extract spatial features.Finally,the extracted spectral features and spatial features are fused and classified.Experiments show that this algorithm implements the spectral-spatial feature classification and obtains better classification performance than other traditional classification methods,which demonstrates the feasibility of this algorithm.(3)In order to adequately excavate more spatial features of remote sensing images and reduce the waste of spatial information,the remote sensing image classification algorithm based on convolution neural network is improved,and then a remote sensing image classification algorithm based on the cross-connect convolutional neural network is proposed.On the basis of two-dimensional convolutional neural network which is based on spatial information,this algorithm joins the concept of the cross-layer connection to construct a new two-dimensional convolutional neural network which is used to extract spatial features.Experiments show that the algorithm adds detail features to high-level features,which can extract more effective features,improve the utilization of spatial information,and obtain better classification results.As well as,experiments demonstrate the spatial information has significant meaning for remote sensing classification.