Improved Transformer-based hyperspectral image classification method for surface features:A case study of the Yellow River Delta
Hyperspectral technology has become the major means of coastal wetland monitoring.However,traditional hyperspectral classification methods usually face challenges such as insufficient feature extraction,the same surface features corresponding to different spectra,and fragmented scenes.To solve these problems,this study proposed a new classification method by applying Transformer to hyperspectral classification.This vision Transformer(ViT)-based method expanded the receptive field by learning global spatial features using non-local technology,thus overcoming the insufficient extraction of discriminant features.Meanwhile,this method enhanced the cross-layer information interchange through cross-layer adaptive residual connection,thus eliminating information loss.This study,taking NC16 and NC13 wetland datasets of the Yellow River Delta as experimental data,compared the classification method proposed in this study to support vector machine(SVM),one-dimensional convolution neural network(1DCNN),contextual deep convolution neural network(CDCNN),spectral-spatial residual network(SSRN),hybrid spectral network(HybridSN),and ViT.The comparison results show that the new method yielded significantly elevated overall accuracy(OA)of up to 96.24％and 73.84％,average accuracy(AA)reaching 83.42％and 74.87％,and Kappa coefficients of up to 94.80％and 68.94％,respectively for the two datasets.
国家科技期刊平台
NSTL
万方数据
