基于深度学习的精细化高光谱特征表达与匹配方法

Spectral feature matching algorithm can take full advantages of the spectral whole and detailed shape features to distinguish refined classes for hyperspectral data. However, the present spectral feature matching algorithms are lack of hierarchical spectral feature expression mechanism, the organization structure of spectral feature is disordered; Spectral feature expression processing is without learning mechanism, and it’s difficult to meet the demands of feature extraction for the complex hyperspectral remote sensing scene; Due to the separation of feature expression and feature matching process, although a huge number of spectral features can be obtained, the effective information, which are extracted from these features, for refined spectral matching is limited. Because of the drawbacks of the present spectral feature matching algorithms, it’s hard to realize automatic and fine spectral feature matching and recognition. Focusing on these problems, the research thoughts and structure is established from "spectral feature expression based on deep learning" to "the selection for hierarchical spectral features" and to "the fine spectral matching and recognition". Based on the research of deep learning theory for hyperspectral data, the hierarchical spectral features expression system is established. The sparse learning and feedback learning algorithms are introduced into the processing to build the adaptive hyperspectral feature expression and matching model. Based on the above processing, the degree of refinement and efficiency of spectral signature recognition processing for hyperspectral data will be enhanced significantly.

充分地利用地物光谱曲线所蕴含的整体形态特征与细节形态特征进行光谱特征匹配，能够精细地区分地物类别。然而当前的光谱特征匹配方法，缺少层次化光谱特征表达机制，光谱特征组织结构混乱；特征表达过程缺乏学习机制，难以满足复杂场景光谱特征提取的要求；特征表达与特征匹配过程割裂，特征匹配陷入特征数量多，有效信息少的窘境，难以实现精细的地物光谱特征匹配识别。本项目针对传统高光谱遥感数据光谱特征表达与匹配识别过程中缺乏层次化表达机制，以及算法自适应能力与自主学习能力不足等问题，以“深度学习光谱特征表达—多层次光谱特征选择—精细化光谱匹配识别”为研究主线，研究高光谱数据光谱特征深度学习理论，构造多层次光谱特征表达体系，利用稀疏学习、反馈学习等研究方法，建立基于深度学习的自适应高光谱特征表达与匹配模型，提升高光谱数据匹配识别的精细化程度与效率。

本课题拟以“深度学习光谱特征表达—多层次光谱特征选择—精细化光谱匹配识别”为研究主线，根据高光谱数据成像机理，利用深度学习理论表达光谱特征，将光谱曲线表达为多层次光谱吸收反射特征与光谱细节特征的组合；基于反馈学习的多层次光谱特征选择机制，建立光谱特征表达与光谱匹配过程的有机联系，获取不同类别的高光谱数据的多层次本质特征组合，实现自动、精细的光谱特征匹配识别。.主要研究思想包括：1）引入高光谱数据特征深度学习表达模型，根据光谱特征结构特征，建立基于深度学习的高光谱数据多层次光谱特征表达机制；2）基于深度学习反馈学习方法，提取不同类别的高光谱数据的多层次本质特征组合，建立光谱特征表达过程与特征匹配过程的有机关联；3）根据后向传播学习机制，发展自适应的多层次光谱特征匹配距离评价策略，实现自动、精细的光谱特征匹配。