Hyperspectral Image Classification Method Based On Label Propagation And Few Shot Learning

With the rapid development of sensors and remote sensing platform,hyperspectral remote sensing technology has gradually become an important way to detect ground objects.As an important branch of hyperspectral remote sensing technology attach unique category labels to the target ground objects according to their different abilities reflecting electromagnetic waves.At present,the application of this technology in geological exploration,precision agriculture,food safety and other fields is increasing,and has become a research hotspot.Hyperspectral image has the characteristics of containing a large number of spectral bands with strong correlation and large amount of information carried by the data itself,which lays a foundation for the classification of hyperspectral image.However,the lack of labeled samples and spectral variability that are inherent in hyperspectral data pose challenges for obtaining a better classification model.To solve the above problems,this paper proposes three space-spectral classification methods for hyperspectral images based on semisupervised learning,which is achieved by combining spatial-spectral information and making full use of the similarity between samples.The main research results are as follows:1.In order to solve the problem of “different objects with the same spectrum,different spectrum with the same object” in hyperspectral data,a classification method of hyperspectral images based on label propagation was proposed.By introducing spatial information using a variety of spatial filtering methods,and making full use of the partial smoothness of hyperspectral data,the hyperspectral images were classified by applying label propagation algorithm.The experimental results show that the similarity relationship between spatial information and samples is fully used in the proposed method,which can significantly reduce the probability of "misclassification" of sample labels caused by spectral variability,and can effectively improve the classification accuracy.2.To solve the problem of inadequate characteristics extracted through the hyperspectral image classification process of classification model,an hyperspectral image classification based on a deep learning and label propagation self-learning method was proposed.By using a closed loop network classification based on self-learning,this method can pass the deeper features of extracted by the convolutional neural network to label propagation model to generate a posteriori information without tag samples,then feedback is delivered to the training set,and finally after several iteration training the classification completes.Experimental results show that the proposed method can improve the feature extraction ability of the model,playing the role of a posteriori information in unlabeled samples fully,and effectively improve the classification accuracy.3.Aiming at solving the problem that hyperspectral image classification model is easy to overfit in small sample scenes,a semi-supervised classification method based on small sample learning was proposed.Based on the existing hyperspectral image classification model of prototype network,labeled samples were combined with unlabeled samples to correct the prototype positions of each category.Experimental results show that the proposed method can better present the category distribution of data,alleviate the overfitting problem of classification model,and effectively improve the classification accuracy.