基于多任务学习的高光谱图像目标探测方法研究

The spectral variability is a bottleneck problem for hyperspectral target detection, which limits its application potential in various fields. The use of optimal priori target spectrum and spatially similar spectra can improve the spectral signature expression and alleviate the spectral variability effect. However, due to the quality and quantity difference of the priori target spectra and the existence of the observation noise and error, there is still a lack of robust prior target spectra learning method and effective spectral signature modeling method, as a result, reducing the accuracy of target detection. Therefore, respectively focusing on the correlation and complementary of multiple priori target spectra, adjacent single band images, and spatially similar spectra, this project will develop a systematic, automatic and robust target detection method for hyperspectral imagery (HSI) based on multi-task learning technology. Firstly, it proposes a joint learning method for multiple priori target spectra; then generates a multi-task learning framework for target detection based on multi-level dimensional reduction method; and formulates a collaborative representation model for the pixel spectrum by simultaneously exploring the spatial and spectral correlation feature; finally applying the total reconstruction error difference accumulated over all the tasks to detect the targets. The proposed method of this project will improve the ability of spectral signature expression, and reduce the impact of spectral variability on target detection, which has important applications in military reconnaissance, environmental monitoring, and disaster assessment.

高光谱图像的光谱可变性是实现高精度目标探测所面临的瓶颈问题，限制了目标探测在各领域的应用潜力。利用优质先验目标光谱和空间相似光谱可以提高像元光谱特征表达效果、缓解光谱可变效应。然而，由于先验目标光谱质量和数量上的差异以及观测噪声和误差的存在，现有目标探测技术缺乏稳健的先验目标光谱学习方法和有效的光谱特征建模方法，大大降低了目标探测精度。为此，本项目拟深入研究多先验目标光谱、相邻单波段图像和空间相似光谱等角度的相关性和互补性，围绕多任务学习理论，发展系统、自动、稳健的高光谱图像光谱目标探测方法。主要研究内容包括：多先验目标光谱联合学习、多层次降维的多任务学习目标探测框架、融合空-谱相关特征的多任务学习目标探测。该研究成果有望增强地物光谱特征表达能力，降低光谱可变问题对目标探测的影响，在军事侦察、环境监测、灾害评估等领域具有重要的应用价值。

高光谱图像的光谱可变性是实现高精度目标探测所面临的瓶颈问题，限制了目标探测在各领域的应用潜力。利用优质先验目标光谱和空间相似光谱可以提高像元光谱特征表达效果、缓解光谱可变效应。然而，由于先验目标光谱质量和数量上的差异以及观测噪声和误差的存在，现有目标探测技术缺乏稳健的先验目标光谱学习方法和有效的光谱特征建模方法，大大降低了目标探测精度。为此，本项目深入研究多先验目标光谱、相邻单波段图像和空间相似光谱等角度的相关性和互补性，围绕多任务学习理论，发展系统、自动、稳健的高光谱图像光谱目标探测方法。主要研究内容包括：多先验目标光谱联合学习、多任务学习目标探测框架、融合空-谱相关特征的多任务学习目标探测。相关研究成果为光谱可变问题的解决提供了新的思路和处理框架，通过先验目标光谱优化学习和和空谱相关特征协同建模，提高了目标与背景的光谱可分性，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9篇，其中7篇SCI检索；出版《智能化的遥感影像亚像元定位技术》专著1部；获得2019湖北省自然科学奖一等奖；申请专利1项。.