Research On Hyperspectral Image Classification Method Based On Transformer

A hyperspectral image is an image captured by a hyperspectral sensor that images an object or scene in tens or even hundreds of bands.Traditional RGB images have only three bands and each pixel contains little spectral information; whereas hyperspectral images contain tens or even hundreds of bands of spectral information per pixel point,which makes it possible to classify hyperspectral images at the pixel level.Thus,it can be used for precise analysis of environment or materials,and is widely used in atmospheric environment monitoring,vegetation cover analysis,oil spill detection,etc.Hyperspectral image classification is a hot research topic in the field of hyperspectral remote sensing.In this paper,we study hyperspectral image classification problem based on Transformer model.Combining metric learning,consistency regularization and contrastive learning,we design a high-performance hyperspectral image classifier for supervised,semi-supervised and unsupervised cases.The main research work and innovations in this paper are as follows.1.For the supervised classification of hyperspectral images,a supervised hyperspec- tral image classification technique based on Transformer model and metric learning is designed in this paper.The Vision Transformer model is partially improved and the metric learning method is introduced to improve the accuracy in the supervised case.The overall accuracy is up to 4.87% higher than the optimal convolutional neural network in the experiments on three datasets.2.For the semi-supervised classification of hyperspectral images,this paper designs a semi-supervised hyperspectral image classification technique based on Transformer model and consistency regularization.Consistency regularization relies on data augmentation,and the application of data augmentation in semi-supervised classi- fication of hyperspectral images is explored.To make consistency regularization available for hyperspectral image classification,we use a data enhancement method that differs from traditional image enhancement.The overall accuracy in the exper- iments on three datasets is at most 4.38% higher than the optimal other methods.3.For the unsupervised classification of hyperspectral images,this paper designs an unsupervised hyperspectral image classification technique based on Transformer model and contrastive learning.Unlike previous unsupervised classification methods for hyperspectral images based on deep learning,contrastive learning does not require reconstructed data.We used the simpler Transformer model to achieve better experimental results while ensuring the unsupervised classification effect.The overall accuracy in the experiments on three datasets is at most 4.43% higher than the optimal other methods.