基于多元互信息和快速稀疏多核学习的高光谱遥感影像地物分类

项目名称： 基于多元互信息和快速稀疏多核学习的高光谱遥感影像地物分类

项目编号： No.61501353

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 冯婕

作者单位： 西安电子科技大学

项目金额： 21万元

中文摘要： 目前高光谱遥感影像地物分类存在计算复杂度高、Hughes现象及波段协作性难以利用的难题。其核心在于波段间缺乏有效的测度及脱离了当前的分类任务。而多元互信息测度则为其测量提供了一种有效的途径。为此，本项目拟建立描述波段协作性的多元互信息测度。在此基础上，构造以任务为驱动，同时考虑冗余性和协作性的高光谱波段选择方法，突破了目前普遍采用的以冗余性测量为主的波段选择框架；此外，建立快速稀疏多核学习模型，给出自适应选择分类器最优核参数组合的方法，达到快速精确的地物分类目的。本项目可为高光谱遥感影像的分类与识别提供潜在的应用价值和指导意义。

中文关键词： 高光谱影像分类；波段选择；多元互信息；波段协作性；快速稀疏多核学习

英文摘要： Currently, some difficulties are encountered in land-cover classification of hyperspectral remote sensing images, such as high computation complexity, Hughes phenomenon, and low utilization rate of band synergy. The key problem is lack of effective band measure and independent of current classification task. Multivariate mutual information just provides a solution for band measure. Thus, this project aims to establish a novel multivariate mutual information measure to describe band synergy. Based on this measure, a task-driven hyperspectral band selection method is constructed. Compared with the widely-used framework based on redundancy measure, this method considers band redundancy and band synergy simultaneously. Moreover, a fast sparse multiple kernel learning model is established to select the best combination of kernel parameters in classifiers adaptively. It achieves fast and accurate land-cover classification . This project has potential significance and instruction for hyperspectral remote image classification and recognition.

英文关键词： Hyperspectral image classification;band selection;multivariate mutual information;band synergy;fast sparse multiple kernel learning