Hyperspectral inversion of soil organic matter based on improved ensemble learning method

Abstract

Soil organic matter (SOM) is a vital component of soil, and its rapid and accurate detection is crucial for ensuring land health and stabilizing atmospheric carbon dioxide levels. Soil hyperspectroscopy has demonstrated its efficiency and cost-effectiveness as a method for detecting SOM. In the field of soil spectroscopy, the Ensemble Model (EM) holds substantial promise due to its robust nature and strong generalization capabilities. However, the efficacy of EM is largely contingent upon the judicious selection of the base learner count and the strategic allocation of weights. Traditional practices is mainly relying on empirical weight distribution or a singular index, R², of the base learners, with scant clarity on the optimal base learner count for varying ensemble techniques. To address this gap, our study utilizes Vis-NIR spectroscopy to quantitatively assess SOM across 704 samples from the Tarim River Basin in Xinjiang, China. Our objective is to innovate base learner weight assignment methods and identify the differing optimal counts of EM base learners, thereby refining the ensemble approach and augmenting EM performance. Subsequently, we examined the impact of various weight coefficient assignment methods and base learner counts on EM performance within Weighted Averaging (WA), Blending, and Stacking frameworks. Our findings reveal that a weight coefficient assignment method incorporating R², RMSE, and MAE significantly enhances EM performance. This improvement surpasses traditional methods relying solely on base learner R², yielding an increased EM R² of 0.006–0.024, with reductions in RMSE and MAE by 0.014–0.085 g kg⁻¹ and 0.03–0.085 g kg⁻¹, respectively. Though the number of base learners is crucial, it does not establish a linear relationship; an increase does not invariably translate to enhanced performance. Notably, when the base learner count is 12, Blending and Stacking exhibit peak performance, whereas WA’s precision continues to ascend with 15 base learners. Among the ensemble methods, Stacking demonstrates the highest precision, achieving a validation R² of 0.889, RMSE of 0.957 g kg⁻¹, and MAE of 0.803 g kg⁻¹. In summary, configuring the base learner count to 12 and employing a multi-index comprehensive evaluation for weight assignment within the Stacking method emerges as the optimal integration strategy for SOM hyperspectral inversion.