基于主动深度学习的高光谱影像分类

发布时间：2018-03-16 01:07

本文选题：高光谱遥感影像分类　切入点：空谱特征　出处：《计算机工程与应用》2017年17期 　论文类型：期刊论文

【摘要】：针对当前高光谱遥感影像分类人工标注样本费时费力,大量未标注样本未得到有效利用以及主要利用光谱信息而忽视空间信息等问题,提出了一种空-谱信息与主动深度学习相结合的高光谱影像分类方法。首先利用主成分分析对原始影像进行降维,在此基础上提取像素的一正方形小邻域作为该像素的空间信息并结合其原始光谱信息得到空谱特征。然后,通过稀疏自编码器得到原始数据的稀疏特征表达,并通过逐层无监督学习稀疏自编码器构建深度神经网络,输出原始数据的深度特征,将其连接到softmax分类器,利用少量标记样本以监督学习的方式完成模型的精调。最后,利用主动学习算法选择最不确定性样本对其进行标注,并加入至训练样本以提高分类器的分类效果。分别对Pavia U影像和Pavia C影像进行分类实验的结果表明,该方法在少量标记样本情况下,相对于传统方法能有效地提高分类精度。
[Abstract]:In order to solve the problems such as manual labeling of hyperspectral remote sensing images is time-consuming and laborious, a large number of unlabeled samples are not effectively used and spatial information is ignored by the main use of spectral information. A hyperspectral image classification method based on space-spectral information and active depth learning is proposed. Firstly, the dimension reduction of the original image is performed by principal component analysis (PCA). On this basis, a small square neighborhood of the pixel is extracted as the spatial information of the pixel and the spatial spectral feature is obtained by combining the original spectral information. Then, the sparse feature representation of the original data is obtained by sparse self-encoder. The depth neural network is constructed by layer by layer unsupervised learning sparse self-encoder, and the depth features of the original data are output, which are connected to the softmax classifier, and the fine tuning of the model is accomplished by using a small number of labeled samples to supervise the learning. Active learning algorithm is used to select the most uncertain samples and add them to the training samples to improve the classification effect of the classifier. The results of classification experiments on Pavia U and Pavia C images show that, Compared with the traditional method, this method can effectively improve the classification accuracy in the case of a small number of labeled samples.
【作者单位】：河海大学地球科学与工程学院;
【基金】：国家自然科学基金(No.41271420/D010702)
【分类号】：TP751

【相似文献】