Research On Remote Sensing Image Classification Based On Deep Transfer Learning

With the development of remote sensing technology,the number of available remote sensing images has increased significantly,but the annotation of images cannot follow the pace of image acquisition.The contradiction between such big data and less tags is a major difficulty in the classification of remote sensing images.This is particularly prominent in the classification of deep learning remote sensing images that require sufficient training samples.Transfer learning can transfer the old knowledge in the labeled source domain dataset to the new unlabeled target domain dataset to improve the classification performance.Therefore,the combination of deep learning and transfer learning is one of the directions to solve this difficulty.Based on the existing deep transfer learning methods,this paper studies remote sensing images.The main research results are as follows:(1)Exploring the transferability of deep convolutional neural networks for remote sensing images,and implementing remote sensing image classification tasks based on several deep convolutional neural networks trained on ImageNet.Experimental results on UC Merced datasets show the transferability of Resnet50 is better than VGG16 and InceptionV3,and fine-tuning the convolutional layer of the network can improve classification performance.(2)Combining the idea of ensemble learning,a remote sensing image classification algorithm based on multi-model transfer feature fusion is proposed.Based on the transfer model,the transfer features are extracted based on these trained pre-training classification models.According to the complementarity of these transfer features,multiple transfer features are fused,and finally classifier classification is used.And the influence of multi-model transfer feature fusion algorithms is explored when different transfer feature are given different weights.Experiments on three remote sensing image datasets and comparison with different feature extraction algorithms verify the effectiveness of the proposed algorithm.(3)For the case that the remote sensing image training data is unlabeled,an improved unsupervised domain adaption method based on Deep Domain Confusion(DDC)is proposed.It uses high-order statistical criteria as the similarity measure method,and uses the dual-flow structure with weakly parameter sharing as the feature mapping network to realize the remote sensing image classification task which target domain is unlabeled.The experimental results show that compared with the DDC method,the improved method can learn the invariant features and domain-specific features of the domain,make the model more suitable for the target domain task,and improve the generalization ability of the target domain data.