Research On Hyperspectral Remote Sensing Image Classification Method Based On Spectral Reconstruction

Hyperspectral image classification technology is one of the most important hyperspectral remote sensing image analysis methods.By analyzing the spectral curve of each pixel of the hyperspectral image,it can identify the subtle differences in the material and texture of the observed object,and then achieve pixel-by-pixel classification.At present,deep learning methods perform well in the feature extraction of hyperspectral images,so they have gradually become the main method of hyperspectral image classification.However,the particularity of hyperspectral remote sensing images with both spatial and spectral information brings great difficulties to the feature extraction of conventional deep learning networks.At the same time,due to the high cost of sample labeling,deep networks usually have to be trained in few-shot scenarios,which leads to the problem of network overfitting.In order to solve the above problems,this paper constructs a hyperspectral remote sensing image classification method based on spectral reconstruction,which effectively solves the above problems by changing the model structure and introducing unlabeled samples to assist model training.The specific research content is as follows:Firstly,a classification method based on spectral Transformer is proposed,which mainly relies on the good feature modeling ability of self-attention mechanism for sequence data,and solves the problem of insufficient ability of mainstream deep learning methods to extract the spectral-spatial joint features of hyperspectral remote sensing images.Specifically,this method mainly includes three aspects: Firstly,principal component analysis(PCA)is used to remove the data redundancy of the original image,and sufficient spectral information is retained to provide sufficient features for subsequent classification.Secondly,a sampling strategy based on circular region is designed to provide more effective spatial context information for the pixels to be classified.Finally,the self-attention mechanism is used to calculate the correlation of the sample spectral bands,and the sequence characteristics of the spectral information are mined,so as to achieve better classification accuracy.Secondly,a two-stage classification method based on band mask reconstruction is proposed.This method is based on the spectral Transformer,and uses the self-supervised learning strategy to introduce unlabeled samples to pre-train the spectral Transformer,so as to improve the generalization of the model and solve the problem that directly training the spectral Transformer in small sample scenarios is easy to lead to model overfitting.Specifically,the method is divided into a pre-training stage and a model fine-tuning stage.In the pre-training stage,the spectral masked autoencoder network is constructed,which is based on the mask and reconstruction tasks of image bands,and introduces a large number of unlabeled pixels naturally exist in hyperspectral remote sensing images to pre-train the feature extraction network of spectral Transformer.In the model fine-tuning stage,the network weights obtained by pretraining are used to improve the overfitting problem,and the spectral Transformer is fine-tuned by Linear Probe,so as to reduce the number of parameters that need to be adjusted,so that the classification accuracy in few-shot scenarios is further improved.