Explainable machine learning for predicting POXC and organic carbon from mid-infrared spectra under varying soil scanning conditions

Soil is a vital natural resource, and growing demand for its evaluation requires rapid and reliable tools. Permanganate oxidizable carbon (POXC), a labile fraction of soil organic matter sensitive to microbial activity, is an important indicator of soil carbon mineralization and stabilization. Traditional laboratory methods for POXC measurement are slow, expensive, and labor-intensive, limiting their utility for large-scale soil health monitoring. This study evaluated the potential of mid-infrared (MIR) spectroscopy to predict POXC and organic carbon (OC) in Mississippi soils, focusing on spectral pre-processing, spectrometer type, and soil scanning condition. Three spectrometers were tested: ARCoptix ROCKET (AR), and Bruker Alpha II with either DRIFT (DR) or ATR (AT) modules, under fine-ground (FG), non-fine-ground (NFG), and field-like (fresh) soil conditions. Principal component analysis showed that pre-processing reduced noise and improved clustering, particularly under FG and NFG. Predictive accuracy for POXC and OC improved after pre-processing, with AR and DR outperforming AT in most cases. Fresh soil scans consistently exhibited higher variability and lower prediction accuracy, highlighting challenges under fresh conditions. SHAP analysis revealed key spectral regions associated with O-H, C-H, Si-O, C=O vibrations, with stronger signals for OC than POXC. Overall, MIR spectroscopy combined with pre-processing, right spectrometer choice (AR or DR), and suitable soil preparation (FG or NFG) demonstrated strong potential for rapid, non-destructive soil carbon estimation. These findings support the integration of MIR spectroscopy into soil health monitoring while highlighting the need for advanced calibration transfer techniques to address variability under field conditions.