基于特征子空间学习的跨场景异构高光谱图像分类

Pixel classification is an essential application on hyperspectral image (HSI). One big challenge of HSI classification is the lack of labeled training samples. According to the trend that colossal amount of HSIs are available nowadays, this research project proposes the classification problem across different HSI scenes (datasets), i.e., we are trying to solve the small-sample-size problem by mining the underlying relationship between HSI datasets that share common contents. In classical HSI classification problem, training and test samples are drawn from an identical HSI scene, i.e., the classification is based on a single HSI dataset. However, in cross-scene classification, we are confronted with two challenges, namely heterogeneous feature distribution and heterogeneous feature space, which make the training and test sample significantly different. Thus the classical algorithms for single HSI classification are no longer valid in the cross-scene classification case. This research project is based on heterogeneous HSI datasets. We would model the type of heterogeneity as well as the potential correlation between different hyperspectral scenes. Then several machine learning techniques are utilized, including transfer learning, domain adaptation, etc., combing with some tools such as multi-task dictionary learning and manifold learning, to build a common feature sub-space shared by different datasets. Within this sub-space, the common features are extracted while the differences between datasets are eliminated. Finally, several valid algorithms will be proposed for cross-scene HSI classification problem.

像元分类是高光谱图像应用中的一个重要问题。带标记样本数量有限一直是高光谱图像分类中的一个难题。根据高光谱图像海量化的趋势，本项目提出跨场景（数据集）进行高光谱图像分类的研究，通过借助相似内容的高光谱图像间隐含的关联性，有效解决无训练样本或训练样本不足的问题。传统的高光谱图像分类问题中，训练样本与测试样本选自同一幅高光谱图像，即单数据集分类。而在本课题研究的跨场景分类中，我们面临特征分布异构和特征空间异构两大难题，导致待预测的像元与训练样本像元具有较大差异，从而造成传统的单数据集分类算法不再适用于跨场景分类问题。本课题以异构高光谱图像为研究对象，对异构类型及内在关联进行建模，利用迁移学习、领域适应等机器学习技术，结合多任务字典学习、流形学习等工具，构建不同高光谱数据集之间的共享特征子空间，从而提取出不同高光谱图像场景间的共有内在特征，消减差异性，为跨场景高光谱图像分类问题提供若干可行的算法。

带标记样本不足是高光谱图像分类的一个重大挑战。针对此问题，本项目将类别一致或相似的不同高光谱图像场景进行联合学习，通过含带标记样本较多的源场景信息帮助带标记样本较少的目标场景取得较好的分类结果。跨场景高光谱图像分类主要使用迁移学习、多任务学习等方案。本项目主要探究基于特征的迁移学习方法，列举如下：（1）跨场景特征选择算法。本项目兼顾类别区分性和场景间一致性，提出了基于跨领域信息增益、跨领域ReliefF、跨领域鲸鱼优化与模拟退火等多种跨场景特征选择算法，将传统的单场景特征选择算法拓展到了跨场景应用。（2）低秩表示与流形学习相结合的特征提取算法。本项目从低秩子空间的流形结构出发，使用降维减小差异性，并使用图对同场景、不同场景的样本间相似性进行建模，提出了流形约束多任务非负矩阵分解，解决同构迁移学习问题，并提出了双字典非负矩阵分解、半监督双字典非负矩阵分解、多层跨领域非负矩阵分解等算法，解决异构迁移学习问题。（3）多尺度深度特征提取。本项目通过矩阵分解算法的多层扩展，结合残差传递方式，构建了多尺度深度特征提取架构，提出了深度残差PCA、深度残差非负矩阵分解模型，强化了特征的可分性能。项目成果已覆盖申请书中包含的研究内容，项目成果产出达到了项目申请书中的预期目标。