Research On Massive Remote Sensing Image Classification Method Based On Distributed Storage

With the continuous development of Earth observation technology,the classification of remote sensing images plays an important role in military and agricultural fields.Traditional remote sensing image classification methods need to design features and parameters manually,which makes the generalization of the model poor and the classification accuracy low.As an emerging technology,deep learning has made breakthroughs in the field of image classification.It can automatically learn the deep features of images and has high recognition accuracy.Because the deep learning model has many parameters,it is generally trained by the GPU server.However,the amount of remote sensing image data is large,and a single GPU server cannot efficiently manage massive remote sensing images.Therefore,this paper proposes a mass remote sensing image classification method based on distributed storage.This method uses the distributed file system as the underlying storage architecture,and uses the GPU server to train the improved remote sensing image classification model in the upper layer.The main research contents of the paper are as follows:(1)In order to solve the problem that single GPU server can not manage massive remote sensing images efficiently,this paper proposes a distributed storage technology for remote sensing image classification applications.The method utilizes the HDFS distributed file system and the distributed database HBase to manage and publish large-scale remote sensing images.In this paper,the combination of Hilbert curve and grid index is used to ensure that the remote sensing data has high spatial proximity in the physical storage structure of HBase.At the same time,in the process of image pyramid construction and data storage,a parallel construction and storage method based on MapReduce is proposed.In the remote sensing image query,the data query time is reduced by setting the filter column family.The experimental results show that the HBase-based distributed storage remote sensing image method can quickly process large-scale remote sensing images.Compared with Oracle and MongoDB storage methods,this method has high scalability and short processing time,and can provide good data management services for large-scale remote sensing image classification.(2)In order to improve the training time and classification accuracy of remote sensing image classification model,this paper proposes a U-Net network model based on migration learning improvement.Firstly,a parallel sampling algorithm MRSW is designed based on distributed storage structure,which makes full use of remote sensing image pyramid data and shortens the construction time of training data.The method uses the convolution parameters of the VGG16 model to initialize the U-Net model and accelerates the convergence speed of the model.In order to avoid over-fitting of the model,this paper combines the pre-sampling and up-sanpling channels of the U-Net model to perform the Batch Normalization operation.At the same time,in order to solve the classification accuracy caused by the category imbalance in the remote sensing data,this paper uses focal loss as the loss function to increase the loss weight of the misclassified sample.Improve the classification accuracy of the model.The experimental results show that the improved U-Net model proposed in this paper is more stable during the training process,and has higher generalization ability while reducing the training time of the model.According to the results on the test set,the method effectively improves the accuracy of remote sensing image classification.It achieves an accuracy of 94.12%on the test data set,which is 5.88%higher than the original U-Net model.