Research On Hyperspectral Image Classification Algorithm Based On BERT Network Spatial-Spectral Combined

Hyperspectral Image(HSI)is a three-dimensional hyperspectral image cube composed of two two-dimensional spaces and a one-dimensional spectrum.Although the features of high spectral resolution and many bands of hyperspectral image have obvious advantages for hyperspectral image feature extraction,the complex spectral-spatial distribution of hyperspectral image contains hundreds of spectral bands and data redundancy,which causes great difficulties for classification of hyperspectral image.In order to solve the above problems,this thesis introduces the Bidirectional Encoder Representations from the Transformers(BERT)model in the field of natural language processing into the field of hyperspectral image classification.The following is the main work of this thesis.1)This thesis proposes a Spatially Augmented guided Sequential BERT(SAS-BERT)network to improve the classification performance of BERT models in hyperspectral images.The method explores the effectiveness of HSI feature descriptions in an image elementto-sequence manner.The introduction of affine transformation in BERT effectively enables spatial feature representation and solves the local,restricted geometric architecture convolution kernel problem of Convolution Neural Network(CNN).In addition,spectral bands are used to extract discriminative spectral features while capturing effective,adaptive non-local spatial features.Experimental results show that SAS-BERT outperforms the BERT-only network for classification.2)In this thesis,a Probabilistic Sparse Weighting based Sequential BERT(PSWS-BERT)network is proposed to improve the model’s ability to focus on the main contributing features.The method addresses the sparsity of the self-attention mechanism by introducing a probabilistic sparse self-attention mechanism to extract word vectors Q and K.This part of the vectors with significant attention weights is subjected to self-attention feature extraction and spatial correlation is established with the spectra to extract the spectralspatial features of the images and improve the decoding capability of hyperspectral images.3)This thesis designs and implements a hyperspectral image classification system for userfriendly operation and visualisation of hyperspectral image classification algorithms.The system is aimed at users who do not understand computer programming.After importing data and clicking on the algorithm that needs to be run,the algorithm classification results and the classification graph can be obtained and compared with other comparative algorithms to visualise the differences between algorithms,making it more operable.This thesis uses the Indian Pines dataset,the Pavia University dataset and the Houston dataset for algorithm experiments.It uses the overall classification accuracy,the average classification accuracy and the Kappa coefficient as qualitative evaluation metrics.Compared with the current state-of-the-art BERT-based hyperspectral image classification algorithm,the algorithm proposed in this thesis achieves better results.