高光谱遥感影像稀疏深度学习与分类研究

Hyperspectral remote sensing image exhibits complex and nonlinear data structure due to the high-dimension, redundant information, homogeneity and heterogeneity in spectral and spatial domain, leading to big challenges for accurate hyperspectral image classification. Therefore, how to represent the complex and nonlinear data, with the aim of achieving accurate classification is a critical scientific problem in this community. Sparse representation can enhance the discriminative power for high-dimensional signals, whereas deep learning is able to effectively model the high-level features of image, leading to great potentials by integrating the two methods. This project will exploit hyperspectral image classification methods via sparse deep learning though theory analysis and method complements of sparse representation and deep learning, which consists of three perspectives as follows. Firstly, explore the link between sparse representation and similarity measure to propose a collaborative sparse graph embedding method for dimensionality reduction. Secondly, introduce spatial correlation and sparse constraint into deep convolutional networks, and propose a sparse deep convolutional networks for feature learning. Finally, propagate the class label information by using sparse graph, and propose a sparse regularization method for classification. The project studies sparse deep feature learning and classification methods, which has significant scientific-research and practical-application values in hyperspectral image feature learning with complex and nonlinear data space, and in image classification with large spatial domain but small training set.

高光谱遥感影像的高维、信息冗余、光谱维与空间维的同质与异质性等特性使得高光谱数据空间呈现出复杂的非线性结构，对准确的影像分类带来巨大挑战。因此，如何更有效的表达复杂、非线性的高光谱数据并实现高精度分类是当前亟需解决的重要科学问题。稀疏表达能够增强高维信号的判别性，而深度学习能够有效刻画影像的高层特征，二者结合的潜力巨大。本项目深入探索高光谱遥感影像稀疏深度学习与分类方法，从三个层面对稀疏表达和深度学习进行理论分析和方法完善：发掘稀疏表达与相似性度量之间的密切关联，研究基于协同稀疏表达的维数约减方法；探索引入空间邻近和稀疏性约束的深度卷积网络，研究基于稀疏深度卷积网络的特征学习方法；借助稀疏图进行类别标记信息的传播，研究基于稀疏图正则化的分类方法。本项目研究稀疏深度学习与分类方法，对复杂、非线性数据空间条件下的高光谱遥感影像特征学习以及大场景影像的小样本分类具有重要的科学研究和实际应用价值。

针对高维特性和现实分类场景中样本的有限性，构建稳定、可靠的特征学习与小样本分类方法是高光谱遥感影像处理领域亟待解决的重要科学问题。本项目系统研究了高光谱遥感影像基于稀疏图嵌入的特征提取和分类方法，重点提出了基于协同稀疏图嵌入和稀疏多流形学习的特征提取方法，以及基于稀疏图正则化的主动半监督分类方法。采用国际通用测试数据集 ROSIS 和 AVIRIS 检验了本文方法的性能，同时针对黑河流域张掖绿洲农业区精细农作物的区分，采用 CASI/SASI 航空高光谱遥感影像对本文方法进行了应用验证。研究揭示了高光谱遥感影像稀疏表达与图构建之间的有机联系，提出了两种改进的稀疏图构建方法；根据构建的稀疏图，应用图嵌入思想提取高光谱遥感影像光谱维特征以提高对地物的区分能力，提出了基于协同稀疏图嵌入和稀疏多流形学习的特征提取方法；借助稀疏多项式逻辑回归模型（SMLR），实现了基于稀疏图嵌入特征提取的高光谱遥感影像分类；采用全约束稀疏表达构建稀疏图，提出了稀疏图正则化的主动半监督分类方法。本项目研究稀疏深度学习与分类方法，对复杂、非线性数据空间条件下的高光谱遥感影像特征学习以及大场景影像的小样本分类具有重要的科学研究和实际应用价值。