Research On Classification Method Of Hyperspectral Image With Small Samples

Hyperspectral images with rich spectral and spatial information,are widely used in military,agriculture,geological survey and other fields.Hyperspectral images contain a wealth of information that can be used for classification,but there are also information redundancy and the phenomenon of "different objects which have the same spectrum ".In the case of limited labeled samples,it is difficult for the classification model to obtain a good classification result.With the rapid development of remote sensing technology,hyperspectral images can be easily obtained,but precise category labeling requires a lot of manpower.In the case of limited labeled samples,how to make full use of the spatial spectrum information of hyperspectral images to improve the feature representation and generalization capabilities of the model,and to perform hyperspectral image classification tasks is a current hot issue.Therefore,this thesis proposes three more efficient hyperspectral image classification methods based on meta-learning,representation learning and spatialspectral attention feature extraction mechanisms.The main contents are as follows:(1)A hyperspectral classification model based on adaptive subspace and feature-wise transformation with small samples is proposed.First,a three-dimensional local channel attention residual network is designed to extract spatial-spectral features of hyperspectral images,and then a feature-wise transformation strategy is introduced to enhance feature diversity to avoid model over-fitting.Finally,a subspace classifier is used to construct a category subspace based on the feature representation of the limited labeled samples of each category,and the hyperspectral image classification is performed using spatial projection and distance metric.Introduce the meta-learning training method to improve the generalization ability of the model.For newly acquired hyperspectral remote sensing images,only a limited number of labeled samples can effectively achieve the task of full image classification.Experiments have proved that this method can obtain better classification results when the labeled samples are limited.(2)A hyperspectral classification model based on representation learning and feature augmentation with small samples is proposed,which makes full use of the labeled and unlabeled samples of hyperspectral images to perform hyperspectral image classification under limited labeled samples.First,the consistency strategy is used to perform the mixed augmentation of labeled samples and pseudo-label assignment of unlabeled samples.Then design a self-supervised rotation operation,multi-angle rotation of the hyperspectral input data,and use the external rotation classification layer to perform angle prediction to improve the representation learning ability of the feature extraction model.Finally,a feature augmentation strategy for hyperspectral data is proposed to perform feature augmentation at the feature level and apply it to the loss function to fully mine the deep feature of the labeled samples.The weighted superposition of multiple loss functions combined with backpropagation is used to train the model.Experimental results show that this method can effectively perform hyperspectral image classification under the condition of limited labeled samples.(3)A hyperspectral image classification model based on multi-directional rotation attention is proposed.First,for the three-dimensional data block represented by each hyperspectral pixel,multi-directional rotation is performed in the spatial dimension.Then the rotating attention network is used to interact with rich spectral information and spatial information.Finally,the features of the same pixel in multiple directions are adaptively weighted and superimposed to obtain more characteristic features.Due to a large number of parameters,the hyperspectral image model can easily cause over-fitting to a limited training sample.In order to solve this problem,an optimization method for minimum loss sharpness is introduced.During model training,the model parameters with lower loss values in the neighborhood are found to improve the generalization ability of the model.Experiments have proved that this method has a better classification effect when the labeled samples are limited.