Research On Neural Network Transfer Learning Algorithm For Remote Sensing Image Classification

Remote sensing image classification is a critical research subject in machine learning field.However,the lack of labeled sample data is the biggest problem in remote sensing image classification task.Transfer learning technology can effectively reduce the number of the required labeled samples by integrating the labeled samples of other classification tasks into remote sensing image classification tasks.The characteristics of ground image and remote sensing image are different.How to distinguish the similar data with the remote sensing images and apply these data to the transfer learning needs to be studied in depth.In this paper,we study the neural network transfer learning algorithm in remote sensing image classification application.The main contents of this paper are as follows:Firstly,this paper designs a convolutional neural network framework based on the building blocks to extract the features of remote sensing images.The neural network is trained and tested under different numbers of training samples of SAT-4,SAT-6,CIFAR-10 and CIFAR-100 datasets respectively.The effectiveness of convolutional neural network to remote sensing image classification is verified,and it is found that the number of training samples can affect the classification performance of the neural network.Secondly,the similarity measurement evaluation algorithm between source dataset and target dataset is designed to select the source dataset which is most similar to the target dataset for transfer learning.By measuring the similarity between the source dataset and the target dataset,negative transfer can be effectively avoided,so that the classification accuracy of the target dataset can be improved by using the transfer learning technology.Experimental results show that the proposed similarity measurement algorithm can effectively measure the similarity between datasets and the classification accuracy is improved.Selecting the dataset dissimilar to the target dataset to transfer learning will reduce the classification performance of the neural network.Thirdly,we study the role of neurons of different hidden layers in transfer learning process.Aiming at the problem of how to choose the layer of convolution neural network which is suitable for transfer learning,we compare and analyze the classification performance from different layers of transfer learning,so as to select a suitable layer for transfer learning.Finally,BN-Cluster transfer learning algorithm is proposed,which can effectively improve the classification accuracy of the target dataset.