基于多注意力机制与编译图神经网络的高光谱图像分类
Hyperspectral Image Classification Based on Multi-attention Mechanism and Compiled Graph Neural Networks
孙杰 1杨静 2丁书杰 3李少波 2胡建军4
作者信息
- 1. 贵州大学机械工程学院,贵阳 550025
- 2. 贵州大学机械工程学院,贵阳 550025;贵州大学省部共建公共大数据国家重点实验室,贵阳 550025
- 3. 贵州大学省部共建公共大数据国家重点实验室,贵阳 550025
- 4. 南卡罗莱纳州大学计算机科学与工程系,哥伦比亚29208
- 折叠
摘要
针对高光谱图像(Hyperspectral image,HSI)分类研究中小样本学习时,无法达到理想分类效果的问题,以多注意力机制融合、编译图神经网络与卷积神经网络有机结合提出了一种新的高光谱图像分类方法.设计了一种基于混合注意力机制的网络(Multiple mixed attention convolutional neural network,MCNN)与编译图神经网络(Compiled graph neural network,CGNN),在学习样本有限的情况下,其能有效保留HSI的光谱与空间信息.引入的图编码器与图解码器可以有效地映射不规则的HSI地物类别特征信息.设计的多注意力机制可以重点关注一些重要的空间像素特征.研究了不同训练样本下对不同算法学习示例分类的影响,在公共数据集Botswana(BS)的实验表明,本文方法比 CEGCN(CNN-enhanced graph convolutional network)、WFCG(Weighted feature fusion of convolutional neural network)算法总体分类精度(Overall classification accuracy,OA)分别高 2.72、3.86 个百分点.同样在 IndianPines(IP)数据集上仅用3％训练样本数据的实验结果显示,本研究方法比CEGCN与WFCG算法的OA分别高0.44、1.42个百分点.说明本研究提出的方法不仅对HSI具有良好的空间与光谱信息感知能力,而且在微小学习数据下仍然表现出强有力的分类准确性.
Abstract
In recent years,although some scholars have achieved satisfactory research results on hyperspectral image(HSI)classification,they often fail to achieve ideal classification results when facing small sample learning.Aiming at this problem,a hyperspectral image classification method was proposed by the organic combination of multi-attention mechanism fusion,compiled graph neural network and convolutional neural network.Firstly,a type of multiple mixed attention convolutional neural network(MCNN)and compiled graph neural network(CGNN)was designed,which can effectively retain the spectral and spatial information of HSI with limited learning samples;secondly,the introduced graph encoder and graph decoder can effectively map irregular HSI feature information;finally,the designed multi-attention mechanism can focus on some important HSI feature categories.In addition,the effect of different training samples on different algorithms for learning example classification was also investigated.Experiments on the public dataset Botswana(BS)showed that the proposed method improved the overall classification accuracy(OA)by 2.72 percentage points and 3.86 percentage points compared with the current state-of-the-art algorithms(CNN-enhanced graph convolutional network,CEGCN;weighted feature fusion of convolutional neural network,WFCG).Similarly,the experimental results on the IndianPines(IP)dataset with only 3％of the training sample data showed that the method also improved the OA of the current state-of-the-art algorithms(CEGCN and WFCG)by 0.44 percentage points and 1.42 percentage points,respectively.This demonstrated that the proposed method not only had good spatial and spectral information perception for HSI,but also showed strong classification accuracy with small learning data.
关键词
高光谱图像分类/图神经网络/注意力机制/超像素分割Key words
hyperspectral image classification/graph neural network/attention mechanism/superpixel segmentation引用本文复制引用
基金项目
国家自然科学基金(62166005)
国家重点研发计划(2018AAA0101800)
贵州省科技支撑计划(QKH[2022]130)
贵州省科技支撑计划(QKH[2022]003)
贵州省科技支撑计划(QKH[2021]335)
贵阳市科技人才培养对象及培养项目(ZKHT[2023]48-8)
出版年
2024
国家科技期刊平台
NSTL
万方数据