Machine learning prediction of soil carbon fractions using bulk-soil and fraction-specific MIR spectra: Is physical fractionation necessary?

Soil organic carbon (SOC) persistence reflects the balance between relatively labile particulate organic carbon (POC) and more stable mineral-associated organic carbon (MAOC), yet quantifying these fractions at scale remains analytically demanding. Here, we investigate whether physical fractionation is necessary for a scalable soil condition (soil health) monitoring program seeking to estimate MAOC and POC. We compared mid-infrared (MIR) spectra from bulk soils with those from physically isolated fine (MAOC-enriched) and coarse (POC-enriched) fractions to assess how fraction-specific information contributes to predicting and interpreting SOC pools. We analysed 405 soils across Australia spanning contrasting ecosystems using multiple preprocessing pipelines, resampling strategies, and machine-learning models. Unsupervised classification revealed strong site-level spectral structure, motivating the use of site-wise validation. Savitzky–Golay first-derivative preprocessing consistently enhanced the diagnostic relevance of absorptions, and Cubist models delivered the most accurate and robust predictions. Fine-fraction spectra achieved the highest accuracy for MAOC (R 2 = 0.77), whereas bulk soil spectra retained strong predictive power for both MAOC (R 2 = 0.71) and POC (R 2 = 0.68). Spectral attribution revealed distinct, chemically meaningful MIR drivers for each pool. MAOC predictions were dominated by bands in the ∼1520–1750 cm −1 region associated with oxidised, microbially derived functional groups stabilised on mineral surfaces, whereas POC predictions relied on features near ∼1740–1800 cm −1 characteristic of plant-derived and partially decomposed residues. Although fraction-specific spectra enhanced mechanistic interpretation, their predictive advantages over bulk soil spectra were modest once site-level dependence was accounted for. Together, these results show that while fraction-specific spectra support mechanistic interpretation, bulk soil MIR retains sufficient chemical information to enable accurate and scalable prediction of SOC fractions without physical separation, advancing MIR-based soil carbon assessment toward landscape-scale monitoring.