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

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning Applications

Background:

• Potentially toxic elements (PTE) in soils present persistent environmental risks due to bioaccumulation.
• Hyperspectral remote sensing integrated with machine learning offers a promising avenue for soil PTE quantification and mapping.
• Comprehensive evaluations of model accuracy for these integrated approaches are limited.

Purpose of the Study:

• To conduct a meta-analysis evaluating the accuracy of various spectral transformation, band optimization, and machine learning (ML) techniques for soil PTE prediction.
• To identify optimal preprocessing and modeling strategies for enhancing the accuracy of soil PTE mapping.
• To provide recommendations for future research and practical applications in soil pollution monitoring.

Main Methods:

• Meta-analysis of 87 studies encompassing 97 locations and 7 soil elements.
• Evaluation of 42 spectral transformation methods, 16 band optimization methods, and 34 ML techniques.
• Statistical analysis to determine the predictive performance (R²) of different method combinations.

Main Results:

• Superior accuracy was achieved with first derivative (FD), second derivative (SD), wavelet transform (WT), and continuum removal (CR) spectral transformations.
• Principal component correlation (PCC), principal component analysis (PCA), expert knowledge (EK), and combined (C_2) band optimization methods enhanced predictive performance.
• Random forest (RF), support vector machine (SVM), artificial neural networks (ANN), extreme learning machine (ELM), and partial least squares regression (PLSR) demonstrated high accuracy.
• The FD-PCC-RF combination yielded high R² values (e.g., 79.55% ± 13.26% for FD-RF).
• Environmental conditions, sampling design, and covariates influence model accuracy, but preprocessing optimization is critical.

Conclusions:

• Optimized preprocessing methods, particularly the FD-PCC-RF strategy, significantly enhance the accuracy of soil PTE predictions.
• Prioritizing scientifically optimized preprocessing is crucial for maximizing the utility of field sampling data.
• This study underscores the importance of advanced preprocessing and model integration for effective soil PTE assessment and management.