Research On Hyperspectral Image Classification Based On Unsupervised Regularization And Capsule Network

Hyperspectral image(HSI)has a strong pixel representation ability.Hyperspectral imaging is based on the spectral reflectance of ground objects,so its resolution is not easily affected by color shading.Therefore,it has unique advantages in military reconnaissance,agricultural observation,geological prospecting,and transportation planning.As a prerequisite for the practical application of hyperspectral images,the classification of hyperspectral images is of great significance.Hyperspectral image classification refers to dividing each pixel into a specific feature category according to the spectral curve provided by each pixel.There are many researches on hyperspectral image classification,but there are still some problems: 1)Because the ground annotation of remote sensing images is expensive,hyperspectral image classification lacks enough training samples,and hyperspectral images follow As the spectral resolution increases,the data dimensions also increase,making the classification task easy to reduce accuracy due to model overfitting;2)Due to the natural state the distribution of ground objects is often uneven.In a fixed imaging range,samples will appear uneven.That is,the proportion of pixels representing certain types of objects in the total number of pixels is too small.These categories are called minority The categories make the models trained by the overall classification accuracy as feedback do not perform well on these few categories.In order to improve the classification accuracy of hyperspectral images,this paper starts from using the unsupervised information contained in the entire sample to provide the model with regularity and minority class synthetic oversampling ideas,and proposes a few shot classification algorithm that sharing unsupervised information during supervised classification and a classification algorithm for imbalanced hyperspectral data based on capsule network as follows:(1)A shared unsupervised information classification method based on three-dimensional convolutional neural network for HSI classification is proposed.Considering that there is too little labeled information of features in the remote sensing field and unlabeled samples are easier to obtain,from the perspective of improving the utilization of unsupervised information contained in the sample as a whole,a method for jointly extracting the features of labeled and unlabeled samples is designed.The KL divergence sparse stacked autoencoder unsupervisedly extracts the features of the full set of samples and pre-trains the model.At the same time,the K-means clustering algorithm is used to cluster the complete set of samples and annotate the pseudo labels from the clusters.In the supervised classification process based on 3D＿CNN,unsupervised information from clustering is introduced to provide regularity for the model and effectively alleviate the over-fitting caused by too few training samples.Finally,comparative experiments on multiple data sets verify the effectiveness of the algorithm in the case of limited training samples.(2)A classification method of imbalanced hyperspectral data based on capsule network and synthetic minority oversampling technology is proposed.First,the original capsule network is transformed into a three-dimensional form.At the same time,in order to reduce the parameters,the capsule network is designed as a locally connected form by imitating the convolutional neural network.In addition,the synthetic minority oversampling idea for the problem of class imbalance is introduced into the capsule network,To achieve a threedimensional unbalanced hyperspectral image classification model based on capsule network and synthetic minority oversampling.Finally,the experimental analysis on the UAV-borne data set with higher spatial resolution and category imbalance problem proves the effectiveness of the algorithm proposed in this paper under the category imbalance input.