Small Sample Classification Of Remote Sensing Images Based On Deep Transfer Learning

When dealing with the classification problem for small sample of remote sensing image,the traditional machine learning classification method can effectively solve the problem.But it relies too much on the prior knowledge of the remote sensing field and the manual experience of efficiently extracting the distinguishable features of the remote sensing image,which has poor versatility and weak scalability.However,the convolutional neural network model in deep learning has made extraordinary achievements in image classification tasks.And the classification accuracy obtained based on deep learning methods is much higher than traditional machine learning classification methods.However,deep learning classification methods can only achieve satisfactory classification results when they are fully trained on a large number of labeled images,and have a high dependence on the number of samples,which is obviously not suitable for small sample classification tasks of remote sensing images.At the same time,in actual experiments,the similarity between different categories of remote sensing images is high,and the same category is extremely different.The deep convolutional neural network model obtained from remote sensing image training has poor classification performance on new remote sensing images.The obtained deep convolutional neural network model does not have good reusability.Based on the deep transfer learning method,this thesis constructs a basic transfer learning framework by using the convolutional neural network model whose weight parameters have been trained in large-scale image datasets.And on this basis,this thesis studies the methods to solve the classification problem of remote sensing image small samples.The main research contents of this thesis are as follows:1.Use the efficient extraction ability of the transfer learning model for image features and the good performance of the support vector machine classifier in the classification of small samples to construct a new combination model.At the feature extraction level,adopt specific learning strategies to improve the feature representation ability of the deep transfer model;At the level of small sample classification,a binary learning method based on error correction output coding is adopted,and Bayesian automatic optimization of hyperparameters is used to improve the generalization performance of support vector machines.2.Based on the idea of siamese network architecture,using triplet loss and cross entropy loss to construct a joint loss function.It further accelerates the convergence speed of the small sample classification task of remote sensing images and improves its classification accuracy.3.Finally,based on the deep transfer learning strategy,the attention mechanism is introduced,and a specific multi-layer perceptron is designed.At the same time,a penalty item is added on the basis of the loss function to increase the degree of attention for difficult samples of remote sensing images.