基于滤波后处理的主动学习高光谱遥感图像分类  

Active learning for hyperspectral remote sensing image classification based on filtering postprocessing

作　　者：孙宁[1] 邓承志[1] 汪胜前[1] 

机构地区：[1]南昌工程学院信息工程学院/协同感知与先进计算研究所,江西南昌330099

出　　处：《南昌工程学院学报》2015年第1期7-11,共5页Journal of Nanchang Institute of Technology

基　　金：国家自然科学基金资助项目(61162022;61362036);江西省高等学校科技落地计划项目(KJLD12098);南昌工程学院研究生创新培养基金项目(2014ycx JJ-B1-006)

摘　　要：有限的样本数据和高维的特征向量使得高光谱图像分类面临巨大挑战.提出一种结合主动学习和滤波器的高光谱遥感图像分类方法.该方法首先选取部分训练样本得到分类模型,然后采用主动学习通过迭代从非训练样本中选择信息量大的样本不断扩大有效样本数,减少了初始训练样本数降低"维数灾难"出现的可能,同时提高了分类器的泛化性能及准确率.通过主动学习和多项逻辑回归分类器对高光谱遥感图像进行初始分类,然后运用滤波器对初始分类结果保边去噪.实验结果表明,本文方法的分类精度高,分类稳定性好.The limited labeled number and high dimensional features are the great challenge that hyperspectral image classification must face. In this paper,combined active learning and filtering for hyperspectral image classification approach is proposed. It attempts to efficiently update existing classifiers and improve the generalization capability and classification accuracy by using fewer labeled data,to reduce the number of training samples and curse of dimensionality. The active learning and multinomial logistic regression are firstly used to classify the hyperspectral remote sensing images,and then the filtering is used to deal with noise. The experiment results show that our algorithm significantly reduced the need of labeled sample and improve classification accuracy.

关 键 词：高光谱 分类 主动学习 多项逻辑回归(MLR) 滤波 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]