项目名称: 基于知识迁移的多时相高分辨率遥感影像分类方法研究
项目编号: No.41301393
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 天文学、地球科学
项目作者: 岳安志
作者单位: 中国科学院遥感与数字地球研究所
项目金额: 25万元
中文摘要: 高分辨率遥感影像是获取土地覆被/利用现状和变化信息的重要数据源。基于不同时期单时相影像的重复分类方法知识共享率低,且没有充分利用多时相影像中蕴含的时间维和空间维信息。本项目以同一地区、不同时间获取的多时相高分辨率遥感影像为研究对象,引入机器学习中的迁移学习方法,重点研究多时相高分辨率遥感影像分类训练样本和特征的知识迁移方法,实现单一时相影像中的训练样本和分类特征向另一时相影像迁移;针对多时相高分辨率遥感影像同一地类存在光谱差异的特点,研究基于条件随机场和高斯混合模型的多时相影像时空特征表达方法,挖掘识别多时相影像典型地类的不变特征;最后研究知识迁移支撑的典型地类分类方法,以及基于知识规则推理控制策略的多特征及分类规则迁移学习方法。通过本项目的研究,确立基于知识迁移的多时相高分辨率影像分类技术体系,提高多时相影像分类效率和精度,促进高分辨率遥感在土地覆被/利用信息提取的深入应用与发展。
中文关键词: 迁移学习;监督分类;训练样本选择;特征提取;隐马尔可夫模型
英文摘要: High spatial resolution satellite (HSRS) images are important source data for extracting information of land-cover/use. The classification knowledge shares rate is low for multi-temporal images by classifying repeatedly single-phase image acquired on the same area, and does not take full advantage of image information in time and space dimension. This project takes HSRS images acquired on the same area at different times as the research object, and introduces the transfer learning methods of machine learning firstly, focusing on the methods of the knowledge transfer of training samples and features for multi-temporal HSRS images supervised classification, achieving a method that can transfer training samples and features of a single-phase image to another phase image; Secondly, for the spectral differences of the same class in multi-temporal HSRS images, based on Conditional Random Fields and Gaussian Mixture Models, mainly researches on multi-temporal images space-time characteristics expression methods and the mining of invariant image features of typically classes; Finally, researches the typically classes classification supported by knowledge transfer, and transfer learning of multi-features and classification rules based on the knowledge rule reasoning control strategy. From our research, establishes multi-
英文关键词: transfer learning;supervised classification;training sample selection;feature extraction;hidden markov model