首页 > , Vol. , Issue () : -
在光谱解混过程中，在端元集合中选择有效的端元子集进行解混是至关重要的，但端元子集选择会受到端元光谱变异性的影响，导致选择的结果以及解混精度具有一定的不确定性。本文提出了一种知识图谱嵌入的光谱解混算法KGESU(Knowledge Graph Embedding Spectral Unmixing)，在利用光谱特征进行解混的同时，引入一定的先验知识来进一步提高端元选择的可靠性，从而提升解混的精度。主要涉及两个核心问题：一是地学知识图谱的嵌入，二是引入先验的光谱解混。前者借助了TransE模型来实现对地学知识图谱的图嵌入训练；后者则在知识图谱嵌入的基础上进行知识推理，并把推理结果融入到光谱解混的过程，从而实现了一个引入先验知识的光谱解混算法。为了验证算法的有效性，通过高分五号高光谱相机所采集的真实图像进行实验，与其他经典算法作对比，结果表明本文的算法具有更好的解混效果，证明了在光谱解混过程中使用一定的地学先验知识有助于提升解混精度，KGESU算法具有一定的应用前景。
In the process of spectral unmixing, it is very important to select effective endmembers from a set of endmembers. However, the selection of endmembers will be affected by the spectral variability of endmembers, resulting in a certain uncertainty in the results of selection and the accuracy of unmixing. In order to solve this problem, this study combines geoscience prior knowledge with sparse unmixing, and proposes a Knowledge Graph Embedding Spectral Unmixing(KGESU) algorithm. While utilizing spectral features, certain prior knowledge is introduced to further improve the reliability of endmember selection. The implementation steps of the KGESU algorithm involves two issues, i.e. the embedding training of geoscience knowledge graph and spectral unmixing with priori knowledge. The embedding training of geoscience knowledge graph is to transform geoscience knowledge into a structured expression form through knowledge graph. Then TransE model is used for graph embedding. Considering the second issue, we perform knowledge reasoning according to the knowledge graph embedding. Then develop a reasoning-weighting sparse unmixing algorithm to integrate the process of the reasoning and unmixing. Experiments are conducted to validated the effectiveness of the proposed method. The prior knowledge was instantiated with the aid of auxiliary data such as Landsat8 and GDEMV2. The spectral unmixing data were GF-5 satellite data. The GF-2 data with a resolution of 1 meter after graphics fusion were used as the verification data. Compared with the traditional pixel-by-pixel evaluation, this paper expands the evaluation window. By increasing the overlap area between pixels and allocating the residuals, the sensitivity of different resolution images to registration errors is reduced. The root mean square error of each endmember, the mean of the root mean square error of each endmember and the overall root mean square error of the image are used as evaluation indexes to evaluate the unmixing results.The results demonstrate that the KGESU algorithm outperforms the state-of-the-art algorithms. By the guidance of geo-prior knowledge in the unmixing process, the uncertainty caused by factors such as data itself and external noise can be reduced. And the ability to discriminate endmembers can be improved to a certain extent. At the same time, the method proposed in this paper combines the advantages of knowledge reasoning and numerical computation. Furthermore, we use both geoscience knowledge and spectral characteristics to select the endmembers. The unmixing result can be more reliable. In the future, the research has the following issues that need further consideration. (1)In this paper, a knowledge graph is constructed only from the perspective of land use classification, and prior knowledge is introduced. In the follow-up work, secondary and even more precise classification can be considered to highlight the advantages of hyperspectral data.(2) In the future work, we will consider more complex relationships between ground objects, introduce more abundant geoscience knowledge, and further build a more perfect geoscience knowledge graph. (3) Knowledge reasoning based on graph embedding is a relatively good method to integrate reasoning results into spectral unmixing at present. With the continuous development of technology, we further try to introduce knowledge through other knowledge reasoning mechanisms.