Quantitative Method to Near-Infrared Spectroscopy With Multi-Feature Fusion Convolutional Neural Network Based on Wavelength Attention
ZHU Yu-kang1, LU Chang-hua1, ZHANG Yu-jun2, JIANG Wei-wei1*
1. School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China
2. Hefei Institute of Physical Science, Anhui Institute of Optics Fine Mechanics, Chinese Academy of Sciences, Hefei 230031, China
Abstract：In recent years, deep learning technology has been applied more and more in the quantitative analysis of near-infrared spectroscopy. However, the traditional convolutional neural network is applied to the spectral analysis due to the problems of a small amount of spectral data and insufficient data quality in near-infrared spectral data. Overfitting problems will occur in quantitative analysis. To improve the ability of convolutional neural networks to extract spectral information and enhance the ge-neralization of the network, this paper proposes a multi-feature fusion convolutional neural network model (MWA-CNN) based on wavelength attention to quantitative analyze the dry matter content in mango by near-infrared spectroscopy. MWA-CNN adds an attention mechanism and a multi-feature fusion mechanism based on the traditional convolutional neural network. The network can learn different spectral feature maps and weight information of different wave bands during the training process, thereby extracting high-quality spectral information to alleviate the overfitting problem in traditional convolutional neural networks and improve the accuracy of regression analysis.In the study, the near-infrared spectrum data of 11 691 mango samples were used, 80% of the samples were used as the training set, 20% of the samples were used as the test set by random method, and the test set root mean square error (RMSEP) and the training set root mean square error were passed. (RMSEC), coefficient of determination (R2), and mean absolute error (MAE) for model evaluation. In this paper, we first standardize the spectral data for pre-processing and then compare the prediction results with four traditional models of partial least squares regression (PLS), extreme learning machine regression (ELM), support vector machine regression (SVR), and traditional convolutional neural net-work (CNN) under the original spectral conditions.The prediction results show that the MWA-CNN network performs the best among the five methods, and the RMSE of MWA-CNN in the test set is 0.669 9. The traditional CNN effect is second only to MWA-CNN with an RMSE of 0.740 8, and the degree of over fitting of MWA-CNN decreases significantly compared to the traditional CNN. The RMSE of the test set in MWA-CNN compared to the training set increased by 15.69%, while the RMSE of the test set in the CNN compared to the training set increased by 151.45%. By adding noise with different signal-to-noise ratios to the spectra and then predicting the spectra with five models respectively after adding noise, the experimental results show that the MWA-CNN model can achieve the best results among the five models under various signal-to-noise conditions. It can be seen from the experimental results that the MWA-CNN has high prediction accuracy and generalization ability in NIR spectral quantile regression and a certain noise immunity capability.
[1] WANG Wei-yan, FENG Wen-qiang, CHANG Nai-jie, et al(王韦燕, 冯文强, 常乃杰，等). Soil and Fertilizer Sciences(中国土壤与肥料), 2023，(3): 194.
[2] Hair J, Alamer A. Research Methods in Applied Linguistics, 2022, 1(3): 100027.
[3] Wang J, Lu S, Wang S H, et al. Multimedia Tools and Applications, 2022, 81(29): 41611.
[4] Fu P, Wen Y, Zhang Y, et al. Journal of Innovative Optical Health Sciences, 2022, 15(3): 2250021.
[5] CHEN Guo-xi, ZHOU Song-bin, CHEN Xin, et al(陈国喜, 周松斌, 陈 欣，等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2022, 42(12): 3811.
[6] Guo T, Xu F, Ma J, et al. Journal of Spectroscopy, 2022, 2022: 6875022.
[7] Jiao Q, Guo X, Liu M, et al. Chemometrics and Intelligent Laboratory Systems, 2023, 235: 104779.
[8] Qin X, Wang Z, Bai Y, et al. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(7): 11908.
[9] Soares S F C, Gomes A A, Araujo M C U, et al. TrAC Trends in Analytical Chemistry, 2013, 42: 84.
[10] Jiang W, Lu C, Zhang Y, et al. Journal of Spectroscopy, 2020, 2020: 3590301.
[11] Krizhevsky A, Sutskever I, Hinton G E. Communications of the ACM, 2017, 60(6): 84.
