基于波段贡献率预测的高光谱图像半监督分类方法研究

Hyperspectral images (HSIs) contain rich spectral and spatial information, which make them advantageous to image classification and recognition..However, there exist some challenging problems that hinder the classification and the practical application of HSIs. For example, enormous spectral bands but limited number of training samples tends to cause the curse of dimensionality. In distinguishing different classes, ignoring the potential contribution differences of each band may lead to the overuse or underuse of the spectral information. Additionally, there exist ‘the same spectral from different materials’ and ‘the same material with different spectral’ phenomena..To this end, the project will study a semi-supervised HSIs classification method with effective incorporation of the band contribution (BC), spectral and spatial information. First, around BC, F-measure will be adopted to characterize them, and then a theoretical prediction model of F-measure will be built. Second, based on the prediction model, a semi-supervised F-measure prediction method will be developed, which is capable of identifying the class attributes of F-measures. Third, the predicted F-measures will be introduced into the spectral space for better utilization of the spectral information, and then further research on integrating F-measures, spectral information, spatial information, and label propagation into a semi-supervised classification objective function will be conducted. Finally, to improve the accuracy of the results to the objective function, the project will study an alternate multi-parameter optimization method, which can decrease the impact of parameter update order. In addition, for evaluation of the classification method, a software evaluation system will be designed.

高光谱图像拥有丰富的光谱与空间信息，因此，在地物分类与识别领域具有广阔的应用前景。然而，波段众多而样本有限引起的“维数灾难”现象、各波段对区分不同地物的潜在贡献差异被忽略而导致的光谱利用率低的问题，以及“同物异谱”或“异物同谱”现象，都极大地影响高光谱图像的分类精度与应用水平。.为此，本项目将研究有效融合波段贡献率、光谱与空间信息的高光谱图像半监督分类方法。首先围绕波段贡献率，探索采用F-measure对其进行刻画，并通过数学推导建立F-measure的理论预测模型；其次在预测模型的基础上，研究可确定F-measure对应地物类别的半监督预测方法；再次将F-measure引入光谱空间，进而构建融合F-measure、光谱与空间信息，以及标签传播的半监督分类目标函数；最后为提高目标函数的求解精度，将研究有效降低参数对更新顺序敏感性的交替优化方法。另外，本项目将开发评估方法性能的软件评估系统。

在高光谱图像分类任务中，不仅存在因训练样本有限引发的“维数灾难”问题，而且面临众多冗余、低质甚至含噪波段，造成光谱可靠度低，进而导致分类性能的下降。为此，本项目从波段重要性评价、优质波段选择、融合波段重要性的高光谱图像分类等方面展开研究。其中，联合流行保持-弱冗余的波段重要性评价实现了流行保持波段的强化与冗余波段的抑制，极大地提高了对优质波段的鉴别能力。基于多图整合嵌入和鲁棒自回归模型的波段显著性评价，通过邻近波段的显著性一致性原则，提高了波段显著性评价的鲁棒性与准确性。鉴于各波段对区分不同地物的贡献差异，本项目进一步构建了依类别的波段重要性显式预测模型，并提出了弱监督的依类别波段重要性预测方法，为实现分类任务驱动的波段重要性评价提供了有效途径。针对高光谱图像分类任务，本项目进一步提出了依类别波段贡献率加权的光谱-空间半监督分类方法，极大地提升了高光谱图像的分类性能。考虑到深度特征在提高分类精度方面的巨大优势，本项目提出了基于波段重要性的深度空-谱联合高光谱图像分类方法等，通过优质波段的增强与低可用波段的抑制，进一步提升了分类性能。.项目研究期间，共发表学术论文8篇，其中SCI收录7篇，包括顶级期刊IEEE Transactions on Geoscience and Remote Sensing上的论文2篇，申请国家发明专利2项，开发软件系统2套，本项目的研究丰富了波段性能评价及高光谱图像分类方面的理论成果，推动了高光谱图像在土地利用调查、矿物勘测、海洋监测等领域的应用。.