改进Transformer的高光谱图像地物分类方法——以黄河三角洲为例

Improved Transformer-based hyperspectral image classification method for surface features:A case study of the Yellow River Delta

高光谱技术已成为沿海湿地监测的主要手段,但传统高光谱分类方法通常存在特征提取不充分、同物异谱和场景碎片化等问题。针对这些问题,该文将Transformer用于高光谱分类,提出一种新的分类方法。该方法基于视觉自注意力模型(Vision Transformer,ViT),利用Non-local技术学习全局空间特征,扩大感受野解决提取判别特征不足的问题;同时,通过自适应跨层残差连接加强层间信息交换,解决信息损失的问题。选取NC16和NC13黄河三角洲湿地数据集作为实验数据,并将提出的方法与支持向量机(support vector machine,SVM)、一维卷积神经网络(one dimensional convolution neural network,1DCNN)、上下文深度卷积神经网络(contextual deep convolution neural network,CDCNN)、光谱空间残差网络(spectral-spatial residual network,SSRN)、混合光谱网络(hybrid spectral network,HybridSN)和ViT进行比较分析。结果表明,所提方法的总体精度(overall accuracy,OA)、平均精度(average accuracy,AA)和Kappa系数均有显著提高,OA分别达到96.24%和73.84%,AA分别达到83.42%和74.87%,Kappa分别达到94.80%和68.94%。

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.