基于注意力双通道残差网络的高光谱影像分类

Hyperspectral Image Classification Based on Attention Dual-channel Residual Network

针对现有基于卷积神经网络的高光谱影像分类模型在特征提取过程中缺乏关注重要特征信息和特征提取不充分的问题,提出了一种端到端的注意力双通道残差网络模型。该模型采用波段-空间注意力模块来加强影像中的重要特征并抑制次要特征,并将处理后的数据输入双通道残差网络中完成对空谱特征的提取,使得模型在充分提取原始空谱特征的同时实现对重要特征的关注,从而提高分类精度。在Indian Pines、University of Pavia和Salinas Valley 3种公开数据集上进行测试,并与其他深度学习分类方法相比较。实验结果表明:所提方法具有更好的分类效果,在3种公开数据集上总体分类精度分别达到了97.15%、96.67%和98.42%。

Aiming at the problem that the existing hyperspectral image classification models based on convolution neural network lack attention to important feature information and feature extraction is insufficient in the process of feature extraction,an end-to-end attention-based dual-channel residual network(ADRN)is proposed.The model adopts the band-spatial attention module to strengthen the important features in the image and suppress the secondary features,and enters the processed data into the dual-channel residual network to complete the extraction of spatial spectral features,so that the model can fully extract the original spatial spectral features while paying attention to important features,thereby improving the classification accuracy.Through testing on three open hyperspectral data sets of Indian Pines,University of Pavia and Salinas Valley,and comparing with other deep learning classification methods.The experimental results show that,the proposed method has better classification effect,the overall classification accuracy reached 97.15%,96.67%and 98.42%respectively.