Remote Sensing Image Classification Based On Improved Fully Convolutional Network

In recent years,remote sensing technology has developed rapidly.Classification of remote sensing images is a common method for analyzing and processing remote sensing image information.The core of existing image classification methods is to extract valuable feature information from images,and then further classify them according to needs.However,when encountering remote sensing images with high resolution,rich feature information,and a huge number of images,these methods still have shortcomings in feature extraction and feature fusion,and it is difficult to achieve the desired classification effect.As deep learning algorithms have made phased progress,the full convolutional network is used as the basic model to drive classification,which can stand out among the solutions to many image classification problems,and has been widely used in computers such as semantic segmentation and image classification.In the field of vision.However,in the field of remote sensing imagery,due to the high resolution of remote sensing image itself and rich feature information,the network model still has problems such as slow training speed and accuracy to be improved,which brings certain obstacles to the practical application of remote sensing imagery.Therefore,this paper selects the full convolutional network as the basic model,and optimizes the network to achieve better classification results in view of the above problems.The research in this paper is as follows:（1）Combing the research status of remote sensing image classification,and summarizing the technical development of deep learning related algorithms in remote sensing image classification,introducing basic structures such as convolutional neural network,full convolutional neural network,attention mechanism and so on.The optimization model U-Net of the current full convolutional network with superior classification performance is selected as the basic model,and its shortcomings are analyzed and improved on the basis of it.And proposed a remote sensing image classification method based on improved fully convolutional network.（2）Aiming at the difficulty of capturing multi-scale feature information in the U-Net network,a method that can use multi-scale extraction of feature values to optimize the network is proposed,and the coding layer and the decoding layer are extracted at the same time when each feature map is generated.Multi-scale features.At the same time,in the process of extracting the coding layer,the hole convolution is introduced to replace the traditional convolution to better retain the detailed features,combined with the multi-scale jump connection calculation of the decoding layer,and then sent to the next network layer.Finally,a two-way attention mechanism is introduced before the network outputs the feature map to obtain the global dependency relationship between the features from the two dimensions of the image channel and space and merge them to enhance the expression ability of the features.Finally,the improved multi-scale jump connection network based on U-Net and the original UNet network were trained on the participating data set of the "CCF Satellite Image AI Classification and Recognition Competition" and compared experiments.Compare the experimental results with quantitative indicators and analyze and evaluate them.Comparing indicators such as the overall classification accuracy and the accuracy of each classification,it is found that the multi-scale jump connection in the overall and single classification has improved compared with the original U-Net network.Among them,m Io U increased by 3.51%to 63.85%.The overall Kappa The coefficient increased by 0.0447 to 0.7139,which proved the superiority of the improved network structure proposed in this paper.