Research On Remote Sensing Image Fusion And Classification Based On Deep Learning

With the development of remote sensing technology,single sensor image data is often constrained by data sources and cannot meet the requirements of multi-layer information data.Due to signal transmission bandwidth and imaging sensor limitations,most remote sensing satellites can only provide low-space,hyperspectral multi-spectral images and high-space,low-spectral panchromatic images.In order to obtain the fusion image and its classification with rich spectral information and clear spatial details,and to reduce the uncertainty of image analysis and interpretation,this paper studies two remote sensing image fusion algorithms based on deep learning and a remote sensing image classification algorithm.The research contents are as follows:1.Under the premise of analyzing the characteristics of remote sensing images,a 6-layer single image enhanced neural network is studied.Multi-scale expansion convolution,local residual unit,global residual compensation,and P-ReLU activation function are introduced in the neural network.The CNN model studied has better approximation effect than the previous model,and the better performance can make full use of the features extracted by the convolution layer.The algorithm of the set enhancement network has better performance for the restoration of the detail after the fusion of the remote sensing image and the maintenance of the spectral information.2.For the advantages and disadvantages of deep convolutional neural networks and shallow convolutional neural networks,a multi-depth and multi-scale convolutional neural network is studied.The research network has two kinds of deep branches,which respectively fit high frequency information and low frequency information from the input remote sensing image.For the inconsistency of the number of input and output channels,the problem of global residuals cannot be used,the data set is separately produced,and the global residual is introduced,which greatly improves the fitting effect of the network.3.A remote sensing classification algorithm is studied for RGB channel remote sensing images.Aiming at the shortcomings of slow convergence and low recognition accuracy of existing neural networks,a method of classification and recognition of remote sensing image farmland based on convolutional neural network is studied.The algorithm uses a large convolution kernel to extract gradient information effectively.The convolutional neural network with a depth of 6 layers is designed to improve the classification effect of the network and greatly reduce the number of training sessions.The larger sample library is used.Train to avoid over-fitting.The experimental results show that the recognition performance of farmland,architecture,desert and vegetation is higher.