基于三维经验模态分解与结构化稀疏表示的高光谱影像精细分类研究及其在红树林中的应用

Mangrove is a kind of precious but endangered wetland resource, it is important to obtain the fine category information of mangrove wetland in the conservation. The project focuses on the classification of mangrove hyperspectral images based on 3D empirical mode decomposition and structured sparse representation, whose contents are threefold: the joint spatial-spectral feature extraction, the structured dictionary construction and the block structured sparse representation and classification. (1) Following the fact that hyperspectral images are complicated non-linear and non-stationary “data cubes”, the joint spatial-spectral adaptive feature extraction approach based on 3D empirical mode decomposition is specifically proposed. (2) To fully dig out the structured sparse information hidden within hyperspectral images, group-block sparsity-based dictionary learning method is proposed, in which both the computational speed and the learning effect are taken into consideration. (3) A block-structured sparse reconstruction method is proposed, which utilizes the constructed dictionary and conduct sparse representation on test samples. Class labels of the test samples are determined by sparse coefficients. The project aims to break through the limitations of high dimension, inadequate use of information and difficulty to improve classification accuracy in hyperspectral images, and fine classification results of mangrove can be finally obtained. The research results can promote the application of hyperspectral remote sensing to the mangrove monitoring and conservation.

红树林是一种处于濒危状态的珍贵湿地资源，精细地获取红树林树种类别信息对红树林的保护至关重要。本项目拟采用三维经验模态分解与结构化稀疏表示方法对红树林高光谱影像进行分类，包括空间-光谱联合特征提取、结构化字典构造和块结构稀疏重构与分类三方面的内容。（1）遵循高光谱影像复杂非线性非平稳“数据立方体”的本质，有针对性地提出基于三维经验模态分解的空间-光谱联合自适应特征提取方法。（2）充分挖掘高光谱影像隐藏的结构稀疏信息，提出基于群-块稀疏的字典学习方法，兼顾计算效率与字典学习效果。（3）利用已构造的字典，提出基于块结构的稀疏重构方法对测试样本进行稀疏表示并利用稀疏系数来判定类别。本项目旨在突破高光谱影像维度高、信息利用不充分和分类精度难以提高等限制，实现对红树林树种的精细分类识别。研究成果能促进高光谱遥感在红树林监测与保护方面发挥更大作用。

红树林是《拉姆萨尔公约》中一类重要的湿地类型，红树林生态系统蕴藏极多的生物多样性和极高的生产力，为当地政府和国家提供广泛的服务。如何精细地获取红树林树种类别信息是对红树林进行保护的基础问题同时也是难点问题之一。本项目以红树林高光谱影像分类为出发点，探究了空间-光谱联合特征提取、结构化字典构造和块结构稀疏重构与分类三方面的内容。首先，利用经验模态分解（Empirical Mode Decomposition，EMD）能自适应地提取复杂非线性非平稳信号的本质信息的特点，提取能反映地物类别特性的典型特征。遵循高光谱影像“数据立方体”的本质，用扩展到三维的EMD（3 Dimensional EMD，3D-EMD）来提取空间-光谱联合特征。其次，利用稀疏表示分类器（Sparse Representation Classifier，SRC）进行精细分类。为了充分利用立方体影像的结构特征，借助张量表达将SRC推广到三阶张量情形，并进行结构化字典学习和块结构稀疏重构。总之，利用3D-EMD和结构化SRC能克服高光谱影像带来的维度高、信息利用不充分以及分类精度难以提高等潜在困难。最后，实例研究方面以珠海市淇澳岛红树林为研究对象，该区域是全国少有的紧靠大城市的红树林区之一，对岛上红树林树种进行了识别，并估算红树林叶面积指数。本研究结果能为红树林监测和保护提供科学依据，对推进高光谱技术在红树林中的应用具有重要作用。