Hyperspectral Image Classification Method Based On Context Feature Extraction

In recent years,deep learning methods have made great progress in hyperspectral image classification.However,the current model generally obtains deep global characteristics by increasing the number of network layers,ignoring the context characteristics of global-local neighborhoods.To solve these problems,two effective hyperspectral image classification models are proposed,which can extract the context information of hyperspectral images effectively without increasing the complexity of the model.The details are as follows:1.This paper first visualized a two-branch network model,which extracts the spatial and spectral local context features,adaptively fuses and refines the features at different levels,and has an efficient ability to aggregate context information.The model uses self-calibrated convolution to extract spatial information.Spectral information is extracted using multiscale dense convolution.The local feature extraction module proposed in this paper makes full use of the background information and the mutual information between locations of each cell to obtain local context feature information.In addition,this paper uses multi-category focal loss to solve the problem of different classification difficulty for each sample.The experimental results show that,with limited samples,the model proposed in this paper not only achieves high classification accuracy,but also greatly reduces the amount of computation and parameters.2.A new feature fusion network model for dual attention mechanism is proposed,which is mainly used to capture more precise global-local context attention features.The model uses self-attention mechanism to extract global context attention features and cross-attention mechanism to extract local context attention features.Considering that attention mechanism is easy to lose location information during the conversion process,this paper presents a location self-calibrated module which can be flexibly embedded in two attention modules.In addition,in order to better fuse global-local features,a multiscale global-local feature fusion module is designed to retain more representative features with less communication costs by aggregating attention features from both global and local sources.In this paper,three commonly used hyperspectral datasets are experimented.The classification results show that this model can achieve high classification accuracy even when the number of samples is limited.