Predicting soil organic matter using corrected field spectra and stacking ensemble learning

Abstract

Soil organic matter (SOM), as a key indicator of soil fertility and the carbon cycle, its field rapid and precise quantification is of great scientific significance for precise agricultural management. Visible near-infrared (Vis-NIR) spectroscopy technology is a rapid and highly accurate SOM quantification method. While the laboratory spectra measurement requires a series of processing procedures. Compared with laboratory spectra, field spectra measurement has the advantages of being faster and more convenient. However, achieving high-precision estimation of SOM based on field spectra poses significant challenges, primarily in mitigating the effects of interfering factors, such as soil moisture and developing a highly robust spectra prediction model. In the practical application of field spectroscopy, eliminating interference through spectra correction methods is an effective strategy. The field prediction of SOM using spectra correction algorithms in conjunction with ensemble learning remains a significant and unresolved challenge. In this study, we gathered 180 soil samples from Hancheng City, Shaanxi Province, China, and built Stacking models using field spectra, field corrected spectra, and lab spectra, respectively, and comprehensively compared their predictive abilities. The study aims to assess the ability of spectra correction methods, including non-negative matrix decomposition (NMF), and generalized least squares weight (GLSW), when combined with the Stacking model, to predict SOM. The results showed that it was difficult challenging to accurately predict SOM using models calibrated with field spectra. However, spectra data corrected by NMF and GLSW could effectively mitigate the influence of environmental interference factors and significantly enhance the model’s predictive performance. The GLSW (R² = 0.85, RMSE = 3.74 g kg⁻¹) outperformed the NMF method (R² = 0.69, RMSE = 5.14 g kg⁻¹) and was close to the laboratory spectra model (R² = 0.89, RMSE = 3.81 g kg⁻¹). Combining spectra correction and stacking improves field SOM prediction accuracy, increasing R² value by 0.1 and 0.26, and decreasing RMSE by 1.16 and 2.56 g kg⁻¹. The performance of all Stacking models was superior to that of the best single model. The stacking model could effectively improve the accuracy of SOM model.