Application of fractional-order differential and ensemble learning to predict soil organic matter from hyperspectra

• Published in: Journal of soils and sediments 2024, Vol.24 (1), p.361-372
• Main Authors: Liu, Ya, Zhang, Yuewen, Lu, Huizhong, Yang, Yong, Xie, Jingying, Chen, Danyan
• Format: Article
• Language: English
• Subjects: Accuracy; Agricultural production; Algorithms; Correlation coefficient; Correlation coefficients; Earth and Environmental Science; Ensemble learning; Environment; Environmental Physics; Learning algorithms; Least squares method; Machine learning; Organic matter; Organic soils; Predictions; Root-mean-square errors; Sec 5 • Soil and Landscape Ecology • Research Article; Soil; Soil organic matter; Soil Science & Conservation; Soils
• ISSN: 1439-0108; 1614-7480
• DOI: 10.1007/s11368-023-03647-z

Purpose Accurate estimation of soil organic matter (SOM) content is crucial for agricultural production. The integral-order differential hyperspectral processing might easily lead to the loss of trace hyperspectral information of substances, especially in moist condition. The aim of this study was to propose a method which can enhance the characteristics of SOM and improve the prediction accuracy of SOM in moist condition. Materials and methods This study firstly used the fractional-order differential (FOD) technique to refine the hyperspectral information, performing 0 ~ 2 order differential preprocessing on the hyperspectral spectrum with a step of 0.1, and then, the significance level at 0.05 was taken to complete the screening. Afterwards, traditional machine learning algorithms (partial least squares regression) and ensemble learning algorithms (eXtreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and random forest (RF)) were applied to predict SOM content. Based on 21 different differential orders and 4 learning algorithms, a total of 84 models were established. Coefficient of determination ( R 2 ), root mean square error (RMSE), residual prediction deviation (RPD), and Lin’s consistency correlation coefficient (LCCC) were used to evaluate the prediction accuracy. Standard deviation was applied to represent the uncertainty of the models. Results and discussion The results showed that FOD processing of hyperspectra could effectively mine highly efficacious information and improve prediction accuracy of SOM. The best model was believed to be the XGBoost algorithm with 0.8-order differential, and its values of R t 2 , RMSE t , RPD t , and LCCC t were 0.81 ± 0.02, 6.95 ± 0.31 g kg −1 , 2.24 ± 0.02, and 0.90 ± 0.01, respectively. Conclusions Briefly, FOD showed tremendous potential in enhancing spectral characteristics of SOM; its combination with ensemble learning can improve the prediction accuracy of SOM. Furthermore, it also provides a new approach to rapidly and accurately acquire other soil attributes.