首页|基于多重注意力与混合残差卷积的高光谱地物分类

基于多重注意力与混合残差卷积的高光谱地物分类

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
针对高光谱数据集普遍存在小样本、高维度、波段之间光谱相关性高、无法对图像进行深层次数据挖掘等问题，提出一种基于多重注意力与混合卷积残差网络的高光谱地物分类模型(Residual Convolutional Attention Neural Networks，RCANN-Net)。首先，采用主成分分析对高光谱图像进行降维，并通过多尺度3D卷积运算得到多尺度特征信息。其次，将此特征信息输入改进的3D残差空间-通道注意力模块中，在学习特征的同时传递参数并校正特征层权重，得到高光谱图像光谱-空间联合精细特征。同时，引入并行的深度可分离卷积残差空间注意力模块，使模型偏向于学习高光谱图像的空间特征，最后通过结果预测模块根据其特征信息得到分类结果。在 3 个公开高光谱数据集上的多次对比表明，该方法在总体精度(OA)、平均精度(AA)、KAPPA系数和平均训练时间上均优于其他4 种对比方法。
High-spectral-resolution Hyperspectral Land Cover Classification Based on Multi-head attention and Hybrid Residual Convolution
To address common challenges in hyperspectral datasets such as small sample sizes,high dimensionality,high spectral correlation between bands,and the inability to perform deep-level data mining on images,we propose a high-spectral-resolution hyperspectral land cover classification based on multi-head attention and hybrid residual convolution networks(RCANN-Net).Firstly,principal component analysis(PCA)is employed to reduce the dimensionality of hyperspectral images,and multi-scale 3D convolutional operations are performed to extract multi-scale feature information.Subsequently,this feature information is input into an improved 3D residual spatial-channel attention module,which not only learns features but also transmits parameters and corrects the weights of feature layers,resulting in joint fine-grained spectral-spatial features of hyperspectral images.Simultaneously,parallel deep separable convolutional residual spatial attention modules are introduced to bias the model towards learning spatial features of hyperspectral images.Finally,the classification results are obtained through the result prediction module based on the feature information.Through multiple comparisons on three publicly available hyperspectral datasets,the proposed method outperforms four other comparative methods in terms of overall accuracy(OA),average accuracy(AA),KAPPA coefficient,and average training time.

hyperspectral image classificationdeep learningfeature fusiondeep separable convolutionattention mechanismresidual network

彭逸清、闫晓奇、任小玲

展开 >

西安工程大学 计算机科学学院,陕西 西安 710600

高光谱图像分类 深度学习 特征融合 深度可分离卷积 注意力机制 残差网络

国家自然科学基金面上项目陕西省自然科学基础研究计划重点项目西安工程大学研究生创新基金项目

619713392018JZ6002chx2023022

2024

10.20165/j.cnki.ISSN1673-629X.2024.0189

计算机技术与发展
陕西省计算机学会

计算机技术与发展

CSTPCD
影响因子：0.621
ISSN：1673-629X
年,卷(期)：2024.34(10)