基于测度学习的高光谱遥感图像目标探测方法研究

The limitations of existing hyperspectral target detection methods include: the classical target detection methods depend on specific statistical hypothesis test; it needs to be addressed in target detection is the training samples are very limited in number for constructing the detection model; most methods fail to mine the correlation of spectral data. These shortcomings restrict the information extraction ability of hyperspectral remote sensing image. Metric learning theory has the advantages of not relying on specific assumptions, needing fewer adjusted parameters, and can effectively using fewer training samples to obtain better target detection results. Therefore, this project will study metric learning theory under the framework of “global feature learning–local feature learning–multitask detector construction” in depth. Using this framework, we can find the distance metric of original feature space with different metric learning methods, and maximally separate the target samples from the background ones. In this project, we will mainly work on global metric learning detector, local metric learning detector and multi–task metric learning detector for maximizing the distance. The resulted algorithms of this project are expected to reduce the dependence on the number of statistical hypothesis test and increase the ability of detection, which have important applications in disaster evaluation, military reconnaissance, mining surveying, etc.

高光谱遥感图像目标探测方法往往需要依赖特定的统计假设、缺乏足够的训练样本、无法挖掘光谱相似性，制约了高光谱遥感图像信息提取能力。测度学习理论具有不依赖特定假设条件、不需要过多的参数调节及能有效地利用有限的先验样本等优势。为此，本项目拟深入研究测度学习的高光谱遥感图像目标探测理论，提出“全局特征学习—局域特征学习—多任务探测器构建”的研究主线，构造不同的测度学习框架，建立特征空间的距离测度描述目标和背景之间的距离，以距离最大化为目标建立全局测度学习探测器、局部测度学习探测器和多任务测度学习探测器等，实现目标和背景最大程度地分离。本研究成果有望降低传统探测算法对先验假设模型的依赖，增强目标的探测能力，在灾害评估、军事侦察和矿区勘测等领域具有重要的实际应用价值。

高光谱遥感图像目标探测方法往往需要依赖特定的统计假设、缺乏足够的训练样本、无法挖掘光谱相似性，制约了高光谱遥感图像信息提取能力。测度学习理论具有不依赖特定假设条件、不需要过多的参数调节及能有效地利用有限的先验样本等优势。为此，本项目深入研究测度学习的高光谱遥感图像目标探测理论，构造了不同的测度学习框架，建立特征空间的距离测度描述目标和背景之间的距离，以距离最大化为目标建立全局测度学习探测器、局部测度学习探测器和多任务测度学习探测器等，实现目标和背景最大程度地分离。相关研究成果降低传统探测算法对先验假设模型的依赖，有望满足高精度目标探测在军事侦察、环境监测、城市规划、灾害评估等应用需求。目前，在该项目的支持下，项目组发表了科研论文11篇，其中10篇SCI论文，申请专利3项。