摘要
在有监督高光谱图像分类问题中,深度学习方法已成为主流,但是这类方法依赖于大量且准确的标签信息。而且,在高光谱图像分类的实际应用中,由于现场调查困难以及人工标注出现错误等,易出现噪声标签,导致模型性能下降。针对上述问题,文中提出了一种端到端的高光谱图像分类方法,基于分组洗牌缓解噪声标签的影响,并引入跳连接操作与鲁棒的损失函数,实现含噪声标签的高光谱图像分类。首先,通过捕捉高光谱图像初步的光谱信息,利用错误标签样本中的有用信息,跳连接操作实现高光谱图像特征重利用,对特征进行分组和洗牌,可以减缓高光谱图像特征之间的紧耦合性,增强模型的噪声鲁棒性;然后,利用鲁棒性损失函数训练模型。在带有人工生成的噪声标签的Salinas Valley、UP和KSC三个公开高光谱图像数据集上进行实验,证明了本网络优秀的噪声鲁棒性,提高了总体分类精度、平均分类精度和Kappa系数这三个指标。
In supervised hyperspectral image classification problems,deep learning methods have become mainstream,but such methods rely on large and accurate label information.Moreover,in the practical application of hyperspectral image classification,due to the difficulty of field surveys and errors in manual labeling,noise labels are prone to appear,resulting in model performance degradation.In view of the above problems,this paper proposes an end-to-end hyperspectral image classification method that alleviates the influence of noisy labels based on grouping and shuffling,introduces skip connection operations,and uses robust loss functions to achieve hyperspectral image classification with noisy labels.This paper captures the preliminary spectral information of the hyperspectral images,uses the useful information in wrongly labeled samples,and skips connection operations to realize the reuse of hyperspectral image features.Grouping and shuffling features can slow down the tight coupling between hyperspectral image features,enhance the model's robustness to noise,and finally train the model with a robust loss function.Experiments were carried out on three public hyperspectral image datasets of Salinas Valley,UP,and KSC with artificially generated noise labels,which proved the excellent noise robustness of this network and improved the overall classification accuracy,average classification accuracy,and Kappa coefficient for these three indicators.
作者
王雷全
朱同川
秦智超
刘金云
张悦
WANG Lei-quan;ZHU Tong-chuan;QIN Zhi-chao;LIU Jin-yun;ZHANG Yue(China University of Petroleum(East China),Qingdao 266580,China;China Academy of Electronics and Information Technology,Beijing 100041,China;SINO-pipeline International Company Limited,Beijing 100028,China)
出处
《中国电子科学研究院学报》
北大核心
2022年第12期1180-1189,共10页
Journal of China Academy of Electronics and Information Technology
基金
国家自然科学基金(62071491)
关键词
高光谱图像
图像分类
噪声鲁棒深度学习
噪声标签
hyperspectral images
image classification
noise-robust deep learning
noise labels
