High-Resolution Remote Sensing Image Classification Based On Deep Transfer Learning And Multi-feature Network Fusion

Nowadays,due to the special acquisition method of high-resolution remote sensing images,the sample size is relatively small.For the classification task of high-resolution remote sensing images with a small sample size,traditional machine learning can be effectively completed.However,the use of artificially extracted features in traditional machine learning has certain limitations and poor robustness.Therefore,the performance of traditional machine learning in high-resolution remote sensing images classification is poor.The convolutional neural network in deep learning can automatically extract the feature information in the image by using the unique convolution operation,so it can complete the classification task well.Deep learning has made a lot of important progress in various image research.Howerer,the deep learning models with excellent performance often require huge training set samples,and the training of small samples can easily lead to overfitting of deep learning models.In addition,for classification tasks,the extraction of image features is particularly critical.The features extracted by a single deep learning model are often insufficient,which limits the improvement of model classification performance.Therefore,this paper uses the convolutional neural network as the basis to design a deep learning model with better classification effect for a small sample of high-resolution remote sensing image datasets.The research work of this paper are as follows:1.Aiming at the overfitting phenomenon of deep convolutional neural network in training caused by small sample size,a deep transfer convolutional neural network model based on deep transfer learning is proposed.The basic deep learning network model used is pre-trained on other similar large data sets,and the training experience obtained after the training is transferred to the convolution part of the deep transfer convolutional neural network model built and frozen,and then Only the fully connected layers of the deep transfer convolutional neural network model need to be trained.This method can alleviate the overfitting caused by the small sample size,thereby improving the classification performance of the network.Finally,adopting this deep transfer convolutional neural network achieves a classification accuracy of 94.7% on a small-sample high-resolution remote sensing image dataset.2.Aiming at the problem of insufficient feature extraction and poor robustness of a single deep learning network,a deep learning network with multi-feature network fusion is further proposed based on the deep transfer convolutional neural network.Through different deep learning models of the same image,different features can be extracted at the output of the convolution layer of different models.After these features are fused,more judgmental features can be obtained.Then,the fused features are classified by the fully connected layer to obtain very good classification performance.Finally,the RMBTLMFNF network built by this method achieves 96.8% accuracy on a small sample high-resolution remote sensing image dataset.3.In view of the excellent classification ability of some traditional machine learning methods,a classification optimization strategy is proposed.This method combines convolutional neural networks with traditional machine learning methods to achieve classification tasks.Since the traditional machine learning algorithm has unique advantages in the training of small samples,the traditional machine learning algorithm is connected after the fully connected layer of the convolutional neural network has a very good effect on the classification of small samples.Experiments show that this scheme can promote the classification of small samples of high-resolution remote sensing images to a certain extent,and the SVM classifier is the best classifier,which can achieve a high resolution of 97.8%.