主动学习联合聚类分组网络的高光谱遥感图像分类  

Active learning-clustering-group convolutions network for hyperspectral images classification

作　　者：刘敬 李银桥 刘逸[2] LIU Jing;LI Yinqiao;LIU Yi(School of Electronic Engineering,Xi’an University of Posts and Telecommunications,Xi'an 710121,China;School of Electronic Engineering,Xidian University,Xi'an 710071,China)

机构地区：[1]西安邮电大学电子工程学院,陕西西安710121 [2]西安电子科技大学电子工程学院,陕西西安710071

出　　处：《光学精密工程》2024年第9期1395-1407,共13页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.62077038,No.61672405)。

摘　　要：基于卷积神经网络的高光谱图像(Hyperspectral Image,HSI)分类面临网络参数量大,带类标样本少的现状,针对这些问题,提出了基于主动学习和聚类分组网络的高光谱图像分类方法(AL-CGNet)。AL-CGNet采用主动学习和聚类联合卷积神经网络进行HSI的特征提取与分类,设计了基于分组卷积的轻量化网络模型以降低网络参数量。对线性判别分析(Linear Discriminant Analysis,LDA)降维后的高光谱图像采用小批量K均值聚类算法划分成不同的簇,并用簇中心的光谱特征代表不同的簇,以利用无类标样本的信息。在分组网络中将生成的特征图沿光谱维划分成一系列小组,每组通过多个残差块依次提取空间-光谱特征,这种分组策略可以充分利用波段的冗余性和差异性,降低网络参数,并实现轻量化。最后,采用主动学习选取信息量大的样本作为训练样本集,以解决带类标样本少的问题。实验结果表明,AL-CGNet在使用相同比例的6%训练样本的情况下,在Indian Pines,Botswana,Houston 3个数据集下的分类结果明显高于ClusterCNN,SSRN和HybridSN等方法,其OA分别为99.57%,99.23%,98.82%,甚至在训练样本更少5%的小样本情况下也是有效的。该方法不仅大大提高了HSI的分类效率,在获得高精度的同时还能高效率地提取特征。Hyperspectral image(HSI)classification using convolutional neural networks often grapples with a large number of network parameters and a scarcity of class-labeled samples.To tackle these issues,we propose a method called AL-CGNet,which integrates active learning and clustering with group convo⁃lutions network for efficient HSI classification.AL-CGNet combines a convolutional neural network with active learning and clustering to enhance feature extraction and classification,while a group convolutionsbased lightweight network model significantly reduces parameter count.Initially,HSI reduced in dimensionality through linear discriminant analysis is segmented into clusters via the mini-batch K-means algo⁃rithm. The central feature of each cluster substitutes the samples within, leveraging information from unla⁃beled samples. Subsequently, feature maps are segmented into groups along the spectral dimension in thegroup convolutions network, where each group sequentially extracts spatial-spectral features through multi⁃ple residual blocks. This grouping strategy optimizes band redundancy and diversity, cuts down networkparameters, and achieves lightweighting. Active learning then selects informative samples for the trainingset, mitigating the issue of limited labeled samples. Experimental results demonstrate that AL-CGNet,with only 6% training samples, significantly outperforms ClusterCNN, SSRN, and HybridSN on the Indi⁃an Pines, Botswana, and Houston datasets, achieving overall accuracies of 99.57%, 99.23%, and98.82%, respectively. Remarkably, AL-CGNet remains effective even with a smaller training sample sizeof 5%. This method not only boosts HSI classification efficiency but also ensures robust feature extractionand high accuracy.

关 键 词：高光谱图像分类 卷积神经网络 分组卷积 聚类 轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]