基于多视图协同训练的高光谱遥感影像分类

项目名称： 基于多视图协同训练的高光谱遥感影像分类

项目编号： No.41471356

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 天文学、地球科学

项目作者： 谭琨

作者单位： 中国矿业大学

项目金额： 80万元

中文摘要： 高光谱遥感影像分类是目前遥感信息处理领域的研究热点，但在分类器选择、精度稳定性、处理效率等方面仍存在许多问题需要解决。针对高光谱遥感影像分类的难点和机器学习的研究进展，将多视图协同训练引入高光谱遥感影像分类，构建基于多视图协同训练的高光谱遥感影像分类的方法体系，选择不同的高光谱多特征集作为分类视图，研究多视图单分类器、多视图多分类器、多源数据多视图分类等组合方式，探索多视图协同训练的差异性测度、学习机优化算法、多分类器集成学习策略等关键问题，引入CUDA并行遥感技术解决运算效率的问题。通过城市扩展、土地覆盖分类、矿区环境监测等应用实例进行算法评价，归纳适用于高光谱遥感影像分类的多视图协同训练策略。研究成果能促进遥感科学、机器学习、模式识别等学科的交叉融合，能有效提高高光谱遥感影像分类和信息处理的精度和可靠性，推进高光谱遥感信息处理的应用。

中文关键词： 高光谱遥感；分类；协同训练

英文摘要： Hyperspectral remote sensing image classification is currently a hot research field of remote sensing information processing，though there are still many challenges in the selection, accuracy stability, efficiency of classifier algorithms for hyperspectral remote sensing image processing. Faced with those difficulties in hyperspectral remote sensing image processing and the development of advanced machine learning, multiview co-training is introduced into hyperspectral remote sensing image classification. Based on multiview co-training，we intend to design the framework of hyperspectral remote sensing classification. Chosen the different feature sets as multiview classification, it is studied the multiview single classifier, multiview multi-classifier, multiview multi-source data classification etc., and then investigated the differences in multiview co-training diversity measurement, optimized selection of member learners and ensemble strategies in image classification. To solve the problem of computing efficiency, we introduce CUDA (Compute Unified Device Architecture) parallel computing technology. Finally, case studies on urban growth, land cover classification, environment monitoring using hyperspectral remote sensing images will be experimented to validate the proposed methods, and the effective multiview co-training strategies for hyperspectral remote sensing image processing will be concluded. Advance in cross-integration of remote sensing science, machine learning, pattern recognition and other disciplines, this study can effectively improve the accuracy and reliability of hyperspectral remote sensing image classification, and enhance the application of hyperspectral remote sensing information processing.

英文关键词： Hyspectral remote sensing;classification;co-training