Rapid characterization of heavy metals in soil using a novel integrated strategy for near-infrared spectroscopy...

As one of the critical natural resources sustaining human survival and sustainable development, soil quality directly underpins ecological security and the sustainability of agricultural production. Given the escalating severity of soil pollution, particularly heavy metals (HMs) contamination, the development of a standardized, efficient, and high-precision technology for acquiring soil HMs information has become an urgent necessity. Traditional research approaches have typically focused on using diverse models to independently predict and analyze different physical and chemical properties of soil. The research focus has largely centered on comparing the performance of various models, while overlooking the fundamental challenges of single models, such as limitations in universality and insufficient reproducibility. In this study, linear and nonlinear modeling approaches were employed, combined with multi-step spectral data processing procedures, to construct near-infrared spectral prediction models for HMs (taking Pb, Cd, and Cr in soil as examples). Furthermore, the inversion methods of soil HM contents and model selection were analyzed. On this basis, by integrating the model averaging strategy and weighted fusion strategy respectively, the limitations and instability of single-model estimation were addressed by leveraging the operational characteristics of six sub-models. The results indicated that the complex architecture and parameter settings of nonlinear algorithms exhibit greater potential for inverting soil biochemical properties. Assigning different weights based on the predictive performance of sub-models can correct outliers occurring in single-model estimation. After applying the weighted fusion strategy, the average prediction errors for Pb, Cd, and Cr in soil were reduced to 2.12 %, 5.35 %, and 3.51 %, respectively. Compared with single-model prediction, the residual prediction deviations of the test set were improved by 0.43, 0.15, and 0.20, respectively. The integration strategy elevates the reference-level single models to a practically applicable level, thereby providing new insights for the subsequent development and optimization of soil regression models.