Remote Sensing Image Classification Application Research Based On Deep Learning

Remote sensing images record rich spectral information and spatial structure information of ground objects.It is the most intuitive representation of the overall appearance of a ground object.Using these pieces of information to identify and classify targets in remote sensing images is the main way to get the target information.The classification results of these ground objects are of great significance to agricultural production,environmental protection,military reconnaissance,and geographic mapping.However,due to the complex types of ground objects in remote sensing images and the lack of spectral information in high-resolution remote sensing images,the difficulty of classification of remote sensing images is increased.The traditional machine learning methods that used in remote sensing image classification is not only need to overcome the difficulties that brought by the amount of remote sensing image data,but also need to carry out heavy feature analysis and feature extraction work manually.At the same time,the feature analysis and extraction of remote sensing images also requires more professional expert knowledge,and the accuracy of the classification results depends on the knowledge of prior experts.Therefore,the traditional machine learning method has obvious limitations in the classification of remote sensing images.Deep learning techniques that have emerged in recent years can automatically learn the deep features of a large number of image data,and accurately identify and classify the targets in the image according to the extracted features,and improve classification accuracy of image data to some extent.In view of the close relationship between remote sensing image classification and image classification technology,it is very feasible to apply deep learning technology to remote sensing image classification.At the same time,this technology can overcome the limitations of traditional methods in remote sensing image classification,this study is of great value.At present,although the deep learning technology applied in the field of remote sensing image classification has achieved certain results,it still has problems such as low classification accuracy and no effective solution for high-resolution remote sensing image classification under large scenes.In view of the above problems,this paper applies the full convolutional neural network model based on deep learning technology to remote sensing image classification.The main research contents are as follows:(1)For the problem that the target is of large difference in size and the target cannot be effectively identified appears in large-resolution remote sensing images in large scenes,this paper proposes an improved full convolutional remote sensing image classification model(FCN-16s)based on the tailoring strategy.The model is based on the Full Convolutional Neural Network(FCN).Then based on the FCN-16 s model,combined with the improved Skip structure,a new FCN-16s+Skip model is obtained.The experimental results show that compared with the traditional FCN model,the two improved models can effectively extract the target and have higher accuracy in the classification of high-resolution remote sensing images in large scenes.(2)An ensemble remote sensing image classification method is proposed to further improve the classification accuracy of high-resolution remote sensing images in large scenes.This method integrates the predicted classification results of FCN-16 s and FCN-16s+Skip models.It enables the integrated classification result to have the advantages of two models,and solves the problem that the accuracy is not high.The experimental results show that the ensemble remote sensing image classification method has higher classification accuracy than the single model.(3)In view of the good scalability of the U-Net model,a fusion structure which consist of residual structure and recurrent structure is added to the U-Net model.It forms a remote sensing image classification model based on recurrent residual convolutional structure which can further improve the efficiency and accuracy of remote sensing image classification.The experimental results show that the model can automatically extract and utilize the features of remote sensing images to obtain high-precision remote sensing image classification results.