基于光谱库优化学习的光谱超分与分类精度提升

基于光谱库优化学习的光谱超分与分类精度提升 Spectral super-resolution using optimized dictionary learning via spectral library and its effects on classification 韩晓琳 1 first-author 张欢 1 孙卫东 1 清华大学 清华大学 光谱库优化学习是指依据日臻完善的光谱库可以涵盖高光谱图像中所有地物类别光谱信息的这一基本特性，将光谱库中的光谱数据作为训练样本，在严格理论推导下构建字典优化学习过程，使高光谱图像中的光谱能够在该字典下进行稀疏表示。本文提出了一种基于光谱库优化学习的光谱超分辨率重建方法，该方法在稀疏表示框架下，通过波段匹配，将光谱库映射为与待重建高光谱图像波段相对应的特定光谱库；利用映射后的特定光谱库与高分多光谱图像，从理论上推导、并构建了基于ADMM算法的光谱字典与稀疏系数优化学习过程；进而，实现了仅由一幅高分多光谱图像到高分高光谱图像的高质量光谱超分辨率重建。多种数据集上的对比分析表明，即使仅使用一幅高分多光谱图像，本文方法仍能恢复重建出高质量的高分高光谱图像，同时光谱超分辨率重建后的高分高光谱图像可显著提升地物分类精度。 Based on the basic characteristic that the spectral library can contain the spectral information about the whole types of ground-surface objects in the observation area of hyperspectral images, the optimized dictionary learning via spectral library refers to the process of constructing optimized spectral dictionary under strict theoretical derivation, in which the spectra in the spectral library are used as training samples. The above process enables the spectra in the hyperspectral image to be sparsely represented under the learned spectral dictionary. To this end, a new spectral super-resolution method using optimized dictionary learning via spectral library is proposed in this paper, which uses only one high-spatial multispectral image to reconstruct high-spatial hyperspectral image. The above problem is formulated in the framework of sparse representation, as an estimation of the band matching matrix, the optimized spectral dictionary and the corresponding sparse coefficients. Specifically, a band matching method is proposed for mapping the common spectral library to a specific spectral library corresponding to the reconstructed high-spatial hyperspectral image. Then, by using both this specific spectral library and the high-spatial multispectral image, an optimization of spectral dictionary and its corresponding sparse coefficients is derived theoretically using the alternating direction method of multipliers (ADMM) algorithm. Comparison results with the relative methods demonstrated that, even only using one high-spatial multispectral image, our method cannot only achieve a high quality reconstruction of the high-spatial hyperspectral image, but also significantly improved the classification accuracy of multispectral images. 光谱超分辨率，光谱库，稀疏表示，优化学习，地物分类 spectral super-resolution, spectral library, sparse representation, optimized dictionary learning, landcover classification 10.11834/jrs.20210591 2021-10-21