基于高光谱多特征协同稀疏表达的滨海湿地精细分类研究

Our proposal analyzes both the sparsity feature and “data cube combing image with spectrum” feature of hyperspectral imagery (HSI), and tries to introduce sparse representation to solve the uncertain classification problem that results from the phenomena of “spectral mixing” and “different ground objects with same spectrum”. The purpose of our study in the proposal is to satisfy fine classification demand of coastal wetland. . First, our proposal will investigate theory and methods of multi-feature collaborative sparse representation and reconstruction. The method could simultaneously extract multiple collaborative features to represent the “spectrum-space-texture” information from HSI data. The multi-feature combination can deal with the problem in fine classification caused by the “different ground objects with same spectrum”. Second, with the help of multi-feature collaborative sparse representation theory, our proposal will explore a new data fusion method for integrating HSI data with high resolution image. The method combines the corresponding sparse coefficients of multi-features from both data to finish the fusion procedure. It can greatly promote the spatial quality of original HSI data and help to alleviate nonnegative effects from “spectral mixing” on the fine classification. Third, our proposal will study intelligent optimal combinations of different features and different classifiers, and wishes to propose a proper classifier using multiple collaborative features of HSI data to achieve the accurate and fine classification results of coastal wetland. Finally, using the coastal wetland in Hangzhou South Bay as our experimental area, our proposal will carry on numerous classification experiments on coastal wetland HSI data to completely testify and verify the performance of our proposed theory and methods. The study in our proposal could not only promote the theory of sparse representation in HSI data processing, but also provide advanced technique for fine classification of coastal wetland.

本课题以满足滨海湿地的精细分类应用为目标，立足高光谱影像的稀疏和“图谱合一”的数据立方体特性，引入稀疏表达理论来解决地物的“光谱混合”和“异物同谱”现象导致光谱信息的分类结果存在较大不确定性的问题。首先，研究多特征协同稀疏表达与重构的理论和方法，实现“光谱-空间-纹理”多特征的协同挖掘和提取，利用多特征信息来克服“异物同谱”对精细分类造成的困难。其次，提出基于多特征协同稀疏表达的高光谱与高分辨率影像的数据融合方法，利用特征稀疏系数的融合来提升高光谱影像的空间数据质量，减弱“光谱混合”对精细分类的影响。再次，研究多特征协同稀疏解译的分类方法，构建不同特征系数和分类器的智能优化组合，实现滨海湿地的精细分类的应用目的。最后，以杭州湾南岸的滨海湿地为实验区，利用湿地地物的分类实验应用来验证提出的理论与方法。本课题的研究能够拓展高光谱影像的稀疏表达理论，同时为滨海湿地的精细分类提供先进技术支持。

滨海湿地具有独特的水文特性和地物分布格局，在维持生态系统平衡、保持生物多样性、降解污染和调节气候等方面具有重要作用。高光谱遥感具有独特的高光谱分辨率优势，然而其自身存在“异物同谱”和“光谱混合”问题，导致无法发挥其用于湿地地物精细分类的巨大应用潜力。本课题立足湿地高光谱影像的稀疏特性和“图谱合一”的数据立方体特性，引入稀疏表达理论来构建研究构建多特征协同稀疏表达与重构的理论与方法，提出基于多特征稀疏系数的影像数据融合方法，研究多特征协同稀疏解译的分类方法，实现滨海湿地的高精度制图应用。在多特征稀疏表达理论研究方面，引入非负矩阵分解和稀疏自表达模型来挖掘不同波段光谱之间的内在重构关系，提出了一系列的稀疏表达重构模型，包括广义损失非负矩阵分解模型、自步非负矩阵分解模型，地球移动距离的概率加权原型分析模型等，丰富和完善了现有高光谱遥感的协同稀疏表达理论和方法。在高光谱与高分辨率影像的数据融合方面，依托国产高分五号卫星高光谱影像数据，从定量指标、分类精度和计算效率三个方面来评价了现有数据融合方法，研究了空间分辨率比率较大尺度差异下的高分五号（GF-5）与高分一号（GF-1）卫星影像的一体化融合策略框架，提出了稀疏低秩近等距线性嵌入和快速稳健的拉普拉斯图主成分分析，实现了融合影像的快速准确降维计算。在多特征稀疏系数的分类方法研究方面，基于稀疏系数的地物分类重构能力，提出了一系列的稀疏分类模型，从噪声或异常值抑制、样本标签扩展和地物的不同空间尺度表征等方面来提升现有分类模型的稳健性和有效性，并结合异构数据机载激光雷达点云来研究特征级和决策级融合模型，实现多源数据的空间、光谱和纹理特征的有效融合，显著提升了杭州湾、黄河口和盐城等多个滨海湿地的地物分类精度与制图效果。