基于深度自编码器多任务联合学习的高光谱影像非线性解混

混合像元分解(解混)是实现高光谱遥感影像信息精细提取的重要技术。近年来，在深度学习发展的推动下，解混研究取得了显著进展。然而，现有方法的解混精度仍然严重受制于散射效应、地物阴影等问题的影响。鉴于此，本项目拟以深度自编码器多任务联合学习为主要手段，构建可以抑制散射效应及地物阴影干扰的非线性解混方法。首先，从多层次角度量化散射效应，构建无监督和有监督的非线性解混方法，在多任务联合学习框架下估算出散射作用对光谱混合的干扰程度，进而抑制散射效应对解混造成影响；然后，在有监督的非线性解混方法模型中加入地物阴影因子和空间加权因子，利用阴影区域的散射特性及空间信息的隐藏特征估算出地物阴影对光谱混合的干扰程度，进而抑制地物阴影对解混造成影响；最后，整合上述方法，提出一套可以满足多种场景需求的高光谱遥感影像解混方法。研究成果将为进一步提升高光谱遥感影像亚像元解译的准确性提供理论基础与方法支撑。

Hyperspectral unmixing is an important technique that deals with mixed pixels to improve the accurate analysis and interpretation of hyperspectral data. Recently, the fast development of deep learning has fostered the learning-based approaches for unmixing. However, for the current situation of unmixing approaches, most of them do not consider physical interactions between the light scattered by multiple materials in the scene, nor do they consider the effects of shadows of ground objects. In order to address these mentioned problems, this project intends to propose some unmixing algorithms that can reduce the effects from scattering interactions and surface shadows, by adopting deep auto-encoders with multi-task joint learning as the main technology. Firstly, considering the spectral mixing mechanism from the multi-layer nonlinearity, the unsupervised and supervised nonlinear unmixing algorithms are established to deal with multiple scattering effectively. Then, a shadow factor and a spatial weighting factor are added to the supervised unmixing algorithm, aiming at reducing these shadows to impact unmixing. Finally, we integrate all proposed algorithms to obtain a hyperspectral remote sensing image unmixing system which can be applied to requirements in several scenarios. The research achievements could provide the theoretical basis and technical support for improving the sub-pixel interpretation accuracy of hyperspectral remote sensing.