Research On Semi-supervised Learning Method Of Remote Sensing Image Based On Deep Neural Network

Remote sensing image classification has become an important method of detecting surface objects and has been widely used in various fields of data analysis.However,the small number of labeled samples and the complexity of feature extraction make remote sensing image classification still facing many challenges.Deep learning method can extract complex features from images through hierarchical learning,but it requires a large number of labeled training samples to support the model's construction.Semi-supervised learning method can learn image features from a small number of labeled samples and a large number of unlabeled samples.Combine the two methods provides a new way to solve the problem of remote sensing image classification.Therefore,in this paper,we monitor the transition from deep learning to semi-supervised deep learning as a whole context,and study the impact of different methods on the classification performance of hyperspectral remote sensing images,and completed the following works:1.Analyze the performance of deep learning method in hyperspectral remote sensing image classification and eliminate the redundant information in the spectrum by principal component analysis.The method of building block combination is used to design the HSI-CNN.In view of the disadvantage of pooling layer,HSI-CNN is used in the form of full convolution network,and the advantages of this method in the classification of hyperspectral remote sensing images of small samples are verified by experiments.2.Apply the unlabeled samples to deep learning.Combined with the network structure of HSI-CNN,this paper proposed a semi-supervised autoencoder network Semi-DCAENet with local feature augmentation method.By adding noise disturbance to the feature of some coding layers,enhance the stability of the extracted features of the coding layer,and then improving the image classification accuracy effectively.3.Based on the analysis of supervised learning method and semi-supervised autoencoder learning method,semi-supervised residual ladder network Semi-RLNet is proposed.Compared the effect of different feature enhancement methods on the classification accuracy of hyperspectral remote sensing images,from this,the noise characteristics of hyperspectral remote sensing images are analyzed,and the optimal feature augmentation method is applied to Semi-RLNet.Experiments shows that the Semi-RLNet can further improve the classification accuracy of hyperspectral remote sensing images training with little labeled sample.This paper studies the feasibility of supervised deep learning and semi-supervised deep learning in the classification of hyperspectral remote sensing images,and paves the way for the further study of deep learning methods in the application of hyperspectral remote sensing image classification.