基于高光谱数据的纸张粘度反演模型构建

doi:10.3964/j.issn.1000-0593(2024)12-3524-10

摘要
参考文献
相关文章 (15)

全文: PDF (16576 KB)
输出: BibTeX | EndNote (RIS)

摘要：纸质文物对我国历史文化传承和民族精神延续具有重要意义，而纸张老化严重损害了文物寿命。纸张粘度是反映纸张老化程度的重要指标，传统粘度测量方法是一种有损的破坏性实验，对珍贵文物造成不可避免的二次伤害。针对这一问题，高光谱遥感技术能够实现无损快速分析，是建立纸张粘度反演的有效途径。以干热和湿热纸张老化为研究对象，获取了110组测定粘度含量的模拟纸张老化样本，通过光谱滤波处理，建立了纸张老化的高光谱数据库。在此基础上，对比分析了原始光谱的一阶微分、二阶微分、倒数对数、倒数对数一阶微分以及多元散射校正、多元散射校正光谱一阶微分、多元散射校正光谱二阶微分、多元散射校正光谱倒数对数和连续小波变换9种数据处理方法，将其与基于竞争自适应重加权采样算法（CARS）和相关系数法（R）组合作为模型的输入，通过数据集划分验证优选出最优模型。研究结果表明：（1）原始光谱430 nm处光谱数据与粘度相关性最高，R为0.75，经过原始光谱倒数对数一阶微分变换后，430 nm处的R为0.874，显著增加了光谱中纸张粘度信息；（2）原始光谱二阶微分变换后，578 nm处的相关性最高，R为0.57，远低于原始光谱数据的最高R（0.75），说明二阶微分不适应于纸张粘度估算，同时其他变换方法的最大相关系数均大于原始光谱，表明了光谱变换的有效性；（3）对于原始光谱，随着分解尺度的增加，光谱中包含的纸质粘度信息有所减少，最大相关系数出现在分解尺度2的484 nm处，为0.873；（4）在不同的输入组合中，原始光谱倒数对数一阶微分作为输入时的模型精度最高，基于CARS方法的支持向量机回归（SVR）、随机森林（RF）和AdaBoost模型在验证数据集中的R²分别为0.96、0.93和0.93，RMSE分别为14.80、18.33和19.79 mL·g^-1，建议在纸张粘度反演中，优先选择原始光谱倒数对数一阶微分作为输入，CARS特征选择算法的SVR模型。以上研究结论表明了高光谱技术在纸张粘度无损分析的适用性，能够对纸质文物的修复工作提供科学依据。

关键词：纸张粘度；高光谱；特征选择；深度学习

Abstract：Paper-based cultural relics hold crucial significance to the historical and cultural heritage and the continuation of the national spirit in China. At the same time, the aging of paper seriously affects the longevity of artifacts. Paper viscosity is an important indicator reflecting the degree of paper aging. Traditional paper viscosity measurement methods are destructive experiments that cause inevitable secondary damage to valuable cultural relics. Hyperspectral remote sensing technology can achieve non-destructive and rapid analysis to address this issue, providing an effective approach to establishing a paper viscosity inversion model. In this study, drying and moist-heat paper aging were taken as the research subjects, and110 groups of simulated paper aging samples with measured viscosity content were obtained. Through spectral filtering, a hyperspectral database of paper aging was established. Based on this, nine data processing methods, including original spectrum first-order derivative, original spectrum second-order derivative, original spectrum reciprocal logarithm, original spectrum reciprocal logarithm first-order derivative, multiplicative scatter correction, multiplicative scatter correction spectrum first-order derivative, multiplicative scatter correction spectrum second-order derivative, multiplicative scatter correction spectrum reciprocal logarithm, and continuous wavelet transform were analyzed, as well as two feature selection methods, competitive adaptative reweighted sampling (CARS) and correlation coefficients (R), were analyzed in combination as input for the models to identify the best model through dataset partitioning validation. Research results have shown: (1) In the original spectrum, the correlation at 430 nm is the highest with viscosity, with an R-value of 0.75. After data processing, the R-value at 430 nm increases to 0.874 following the application ofthe reciprocal logarithm first-order derivative of the original spectrum method, which significantly enhances the paper viscosity information in the spectrum; (2) After the original spectrum is transformed into the second-order derivative, the correlation at 578 nm is the highest, with an R-value of 0.57, significantly lower than the highest R-value (0.75) in the original spectral data. This finding indicates that the second-order derivative is unsuitable for paper viscosity estimation. Meanwhile, the maximum correlation coefficients of other transformation methods are all higher than the original spectrum, demonstrating the effectiveness of spectral transformation; (3) For the original spectrum, as the decomposition scale increases, the paper viscosity information contained in the spectrum decreases, and the highest correlation coefficient occurs at 484nm at the decomposition scale of 2, with a value of 0.873; (4) Among different input combinations, the model with the highest accuracy uses the reciprocal logarithm first-order derivative of the original spectrum as input. Based on the CARS method, the support vector regression (SVR),random forest (RF), and AdaBoost models have R² values of 0.96, 0.93, and 0.93, respectively, and RMSE values of 14.80, 18.33, and 19.79 mL·g^-1 for the validation dataset. We recommend prioritizing using the reciprocal logarithm first-order derivative of the original spectrum as inputs and the SVR model with the CARS feature selection algorithm for paper viscosity inversion. The above research findings demonstrate the applicability of hyperspectral technology for non-destructive analysis of paper viscosity, providing a scientific basis for the restoration work of paper-based cultural relics.

Key words：Paper viscosity; Hyperspectral; Feature selection; Deep learning

收稿日期: 2023-11-14 修订日期: 2024-03-13

中图分类号:

TP79

基金资助: 国家重点研发计划项目(2020YFE0204600)，北京故宫文物保护基金会和万科公益基金会专项经费资助

通讯作者: 曲亮 E-mail: quliang@dpm.org.cn

作者简介: 王飒, 1990年生, 故宫博物院博士后 e-mail: wangsa_01@yeah.net

引用本文:

王飒，曲亮，张立福，高宇，李广华，常晶晶. 基于高光谱数据的纸张粘度反演模型构建[J]. 光谱学与光谱分析, 2024, 44(12): 3524-3533.
WANG Sa, QU Liang, ZHANG Li-fu, GAO Yu, LI Guang-hua, CHANG Jing-jing. Research on the Inverse Model of Paper Viscosity Based on Hyperspectral Data. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2024, 44(12): 3524-3533.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2024)12-3524-10 或 https://www.gpxygpfx.com/CN/Y2024/V44/I12/3524

[1] WANG Hao(王昊). China Paper Newsletters(造纸信息)，2023，(6): 124.
[2] LI Bin(李彬). Culture Industry (文化产业)，2023，(24): 76.
[3] TONG Qing-xi，ZHANG Bing，ZHANG Li-fu(童庆禧，张兵，张立福). National Remote Sensing Bulletin(遥感学报)，2016，20(5): 689.
[4] ZHANG Li-fu，WANG Sa，ZHANG Yan，et al(张立福，王飒，张燕，等). Acta Geodaetica et Cartographica Sinica (测绘学报)，2023，52(7): 1126.
[5] Balas C, Papadakis V, Papadakis N, et al. Journal of Cultural Heritage, 2003, 4(S1): 330.
[6] Sun M, Zhang D, Wang Z, et al. Scientific Reports, 2015, 5: 14371.
[7] WANG Le-le, LI Zhi-min, MA Qing-lin, et al(王乐乐, 李志敏, 马清林, 等). Dunhuang Research(敦煌研究)，2015，(3): 128.
[8] Vermeulen M, Miranda A S O, Tamburini D, et al. Heritage Science, 2022, 10(1): 44.
[9] WU Wang-ting, ZHANG Chen-feng, GAO Ai-dong, et al(武望婷, 张陈锋, 高爱东, 等). Sciences of Conservation and Archaeology(文物保护与考古科学), 2017, 29(4): 45.
[10] GUO Xin-lei, ZHANG Li-fu, WU Tai-xia, et al(郭新蕾，张立福，吴太夏，等). Journal of Image and Graphics(中国图象图形学报), 2017, 22(10): 1428.
[11] ZHOU Xin-guang, SHEN Hua, WU Lai-ming (周新光, 沈骅, 吴来明). Sciences of Conservation and Archaeology(文物保护与考古科学), 2020, 32(1): 56.
[12] YI Xiao-hui, LONG Kun, REN Shan-shan, et al (易晓辉，龙堃，任珊珊，等). Sciences of Conservation and Archaeology(文物保护与考古科学), 2018, 30(3): 21.
[13] GAO Yu, SUN Xue-jian, LI Guang-hua, et al(高宇，孙雪剑，李光华，等). Spectroscopy and Spectral Analysis (光谱学与光谱分析), 2023, 43(9): 2960.
[14] LI Yang(李扬). Paper Science and Technology(造纸科学与技术), 2021, 40(2): 46.
[15] Guo Guodong, Fu Yun, Dyer Charles R, et al. IEEE Transactions on Image Processing, 2008, 17（7）: 1178.
[16] Breiman L. Machine Learning, 1996, 24: 123.
[17] Chen T, Guestrin C. XGBoot: A Scalable Tree Boosting System, Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining，2016: 785.
[18] Prokhorenkova L, Gusev G, Vorobev A, et al. CatBoost: Unbiased Boosting With Categorical Features, Proceedings of the 32nd International Conference on Neural Information Processing Systems，2017: 6639.