Uncertainty in soil elemental prediction using machine learning and hyperspectral remote sensing

Potentially toxic elements (PTE) in soils pose significant environmental risks due to their persistence and bioaccumulation. Integrating hyperspectral remote sensing with machine learning models is a promising approach for quantifying and mapping soil PTE distributions, enabling advanced pollution monitoring. However, comprehensive evaluations of model accuracy remain limited. Here, we conducted a meta-analysis using data from 87 studies across 97 locations, 7 soil elements, 42 spectral transformation methods, 16 band optimization methods, and 34 ML techniques. Our findings indicate that among transformation methods, first derivative (FD), second derivative (SD), wavelet transform (WT), and continuum removal (CR) achieve superior model accuracy. For band optimization methods, principal component correlation (PCC), principal component analysis (PCA), expert knowledge (EK), and the combination (C_2) method effectively enhance predictive performance. In terms of model algorithms, random forest (RF), support vector machine (SVM), artificial neural networks (ANN), extreme learning machine (ELM), and partial least squares regression (PLSR) achieve high accuracy. Furthermore, the selection of FD transformation, PCC method, and RF algorithm yields R² of 79.55 % ± 13.26 %, 82.55 % ± 9.81 %, and 79.55 % ± 13.26 %, respectively. While environmental conditions, sampling design, and covariates influence model accuracy, optimizing preprocessing is key to accurate predictions. Applying scientifically optimized preprocessing methods to field sampling data can significantly enhance model performance by maximizing the utility of available samples. Therefore, we highly recommend that researchers prioritize the FD-PCC-RF strategy combination for exploratory modeling after acquiring soil samples. This study highlights the importance of advanced preprocessing and model integration for soil PTE predictions, enhancing environmental risk assessment and management. Future research should focus on adaptive preprocessing approaches based on element spectral characteristics, hybrid optimization frameworks, advances in predictive algorithms, and multimodal environmental data fusion modeling to enhance model robustness and accuracy.