Comparing Different Multivariate Calibration Methods Analyses for Measurement of Soil Properties Using Visible and Short Wave-Near
Infrared Spectroscopy Combined With Machine Learning Algorithms
HUANG Zhao-di1, CHEN Zai-liang2, WANG Chen3, TIAN Peng2, ZHANG Hai-liang2, XIE Chao-yong2*, LIU Xue-mei4*
1. Basic Experiment and Engineering Practice Center, East China Jiaotong University, Nanchang 330013, China
2. School of Electrical and Automation Engineering, East China Jiaotong University, Nanchang 330013, China
3. Department of Transportation Engineering, Lunan Technician College, Linyi 276000, China
4. School of Civil Engineering and Architecture, East China Jiaotong University, Nanchang 330013, China
Abstract:Soil fertility is usually determined by the content of organic matter, total nitrogen, available phosphorus, and available potassium. The content of these substances is usually studied by visible/long-wave near-infrared spectroscopy (Visible/near-infrared spectroscopy, Vis/NIRS: 350~2 500 nm), and visible/shortwave near-infrared spectroscopy (Vis/NIRS: 325~1 075 nm) research is scarce, and the combination of visible/short-wave NIR spectroscopy with machine learning algorithms to measure soil nutrients has great potential. In this paper, four villages in Xinjian District, Nanchang City and Anfu County, Ji'an City, were selected as sample acquisition sites, and soil samples with a depth of 10~30 cm in the diagonal area were selected by the 2×2 grid method, including 120 paddy soils (paddy soil 1 and paddy soil 2), 60 parts of brown soil, and 60 parts of red soil. After grinding and air-drying, the samples were evenly divided into two parts by the quartering method, which was used to determine the samples' spectral and physicochemical information. The acquired spectral data were removed from the noise bands of 325~349 and 1 073~1 075 nm and then preprocessed by S-G convolution smoothing combined with the first derivative. Principal component analysis (PCA) was performed on the preprocessed spectral data, and the score map (PC1: 98.44%, PC2: 3.5%, PC3: 0.14%) obtained according to the principal component analysis showed that the samples had obvious clustering and were in two The samples can be separated from each other in the dimensional space, and the samples have obvious clustering phenomenon. PCA can reasonably explain the differences in the spectral characteristics of different soil samples to a certain extent. In addition, full-band principal component regression (PCR) and partial least squares regression (PLSR) models were established on the preprocessed spectral data, and PCA and PLSR dimensionally reduced the spectral data to extract three principal component factors (PCs) and 9 latent variables (LVs), build nonlinear back-propagation neural network (BPNN) and least squares support vector machine (LS-SVM) models. By comparing the prediction accuracy of Vis/SW-NIRS for OM, TN, P, K by PCR, PLSR, BPNN and LS-SVM methods, the following conclusions can be drawn: (1) The LS-SVM-LVs model has good performance in all soil properties All are better than PCR, PLSR, BPNN-PCs, BPNN-LVs and LS-SVM-PCs models; (2) LS-SVM-LVs model has the highest prediction accuracy for OM and N, which is characteristic of spectral response in the NIR region (3) Determination of soil mineral nutrients P and K by Vis/SW-NIRS has different accuracy, which is due to the co-variation of spectral active components. Based on the results obtained in this study, LS-SVM-LVs analysis is recommended as the best model approach for predicting soil properties (OM, TN, P, and K). However, further research is needed to deeply interpret measurements of soil properties that do not have direct spectral responses in the near-infrared region. The research results of this paper can provide theoretical and technical references for the development of local precision agriculture.
Key words:Visible/shortwave near-infrared spectroscopy; Soil; Principal component analysis; Least squares-support vector machine; Back propagation neural network
[1] LUO De-fang, LIU Wei-yang, PENG Jie, et al(罗德芳,柳维扬,彭 杰, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2022, 42(1): 222.
[2] Ahmad J H F, Meder R, Hamid H A, et al. Journal of Near Infrared Spectroscopy, 2021, 29(3): 148.
[3] Guo P, Li T, Gao H, et al. Remote Sensing, 2021, 13(19): 4000.
[4] RodríguezFebereiro M, Dafonte J, Fandiño M, et al. Remote Sensing, 2022, 14(6): 1326.
[5] Said N, Henning B, Joachim H, et al. Soil & Tillage Research, 2016, 155: 510.
[6] Shupei X, Yong H. Molecules, 2019, 24(13): 2486.
[7] Liu X, Liu J. Spectroscopy Letters, 2014, 47(10): 729.
[8] Liu X, Liu J. Measurement, 2013, 46(10): 3808.
[9] Shengxiang X, Yongcun Z, Meiyan W, et al. Geoderma, 2018, 310: 29.
[10] S R A, J W, J A M D, et al. European Journal of Soil Science, 2014, 65(5): 719.
[11] Yongni S, Chunjiang Z, Yidan B, et al. Food and Bioprocess Technology, 2012, 5(1): 100.
[12] Wenjun J, Zhou S, Jingyi H, et al. PLOS ONE, 2014, 9(8): 1.