Infrared Light and Explainable AI Reveal How Much Carbon Soil Can Still Store

Soil is one of the planet’s largest carbon reservoirs, and unlocking its full capacity to stash away atmospheric carbon has long been a slow, expensive guessing game. Now researchers at Curtin University in Australia have shown that a beam of infrared light, paired with transparent machine learning, can rapidly predict both how much stable carbon a soil currently holds and how much more it could store — with an accuracy that rivals laborious laboratory fractionation. The study, published in the journal SOIL, analysed 482 Australian topsoil samples and demonstrated that mid-infrared spectroscopy can estimate mineral-associated organic carbon with a coefficient of determination of 0.86 and the carbon storage deficit with 0.89, while also revealing which soil constituents drive the predictions.

The carbon in question is not the visible, chunky debris of decaying leaves and roots. Plants capture atmospheric carbon dioxide through photosynthesis, and this carbon enters the soil first as particulate organic carbon. Soil microorganisms then consume this material, breaking some of it into smaller molecules. A portion of those molecules becomes protected from further decomposition by adsorbing onto the surfaces of mineral particles in the soil’s fine fraction — particles of clay and silt no larger than 20 micrometres. This protected pool is known as mineral-associated organic carbon, or MAOC, and it is the long-term vault of the soil carbon economy. Soils richer in silt and clay offer more mineral surface area, and therefore a greater capacity to adsorb and stabilise carbon.

Crucially, that capacity is finite. Scientists refer to a soil’s maximum ability to stabilise organic carbon as its carbon saturation capacity, which depends on the proportion of reactive minerals present. The gap between the carbon a soil currently holds and that saturation ceiling is the carbon saturation deficit — the sequestration potential that land managers and policymakers urgently need to quantify. The United Nations Framework Convention on Climate Change has identified soil carbon sequestration as a critical nature-based process for withdrawing carbon dioxide from the atmosphere, and better estimates of the deficit underpin climate adaptation strategies, soil health assessments and emerging carbon credit schemes.

Estimating the deficit has traditionally required two things: many soil samples, and painstaking measurements. Measuring MAOC involves physically fractionating soil to isolate the carbon in the fine fraction and then quantifying its organic carbon content with an elemental analyser. Earlier approaches to defining the saturation ceiling relied on linear relationships between fine-fraction carbon and clay-plus-silt content, but studies found these methods underestimated capacity because they fitted lines through the middle of the data rather than capturing its maximum values. Quantile regression at the 95th percentile improved matters, yet still cut through the data cloud. The Curtin team instead used a bootstrapped frontier lines analysis, which fits a smooth envelope to the upper boundary of the relationship between MAOC and clay-plus-silt content, preventing underestimation and providing uncertainty estimates.

The 488 topsoil samples came from 275 sites spanning Australia’s main Köppen-Geiger climate zones, from arid hot deserts to tropical savannahs, and covered 11 of the 14 Australian soil classification orders. Most were collected from areas of minimal human impact, such as nature conservation sites and native vegetation grazing lands, with eucalyptus woodlands the most common vegetation type. The researchers physically separated each soil using ultrasonic dispersion and automated wet sieving into macroaggregates, microaggregates and the fine fraction, then measured the organic carbon of each fraction. Three hydrosol samples were excluded because waterlogged, anoxic soils store carbon through fundamentally different mechanisms, leaving 482 samples for analysis.

The frontier line revealed striking numbers. The maximum attainable carbon storage ranged from 5.29 to 45.79 grams of carbon per kilogram of soil, with a mean of 32.76 grams per kilogram. The carbon saturation deficit ranged from essentially none to 45.17 grams per kilogram, averaging 26.31 grams per kilogram — a substantial untapped reservoir. The frontier line rose steeply with increasing clay and silt content up to roughly 20 to 45 percent, after which the rate of increase slowed, reflecting the asymptotic approach to a maximum attainable storage under each soil’s environmental conditions. Uncertainty was quantified through 100 bootstrap resamples, with samples from the same site kept together to prevent data leakage.

The real innovation lay in replacing the expensive laboratory workflow with light. The team recorded mid-infrared spectra of finely ground whole soils using diffuse reflectance Fourier-transform spectroscopy, capturing how molecules vibrate at wavelengths from 4000 to 450 wavenumbers. These spectra act as an integrative molecular fingerprint of the soil, encoding its organic matter chemistry, clay and iron-oxide mineralogy, and particle size simultaneously. The spectra were interpolated to 32 wavenumber intervals to reduce collinearity, preprocessed with baseline corrections and a standard normal variate transformation, and regions dominated by noise from water and carbon dioxide were removed before modelling.

To turn spectra into predictions, the researchers used CUBIST, a rule-based regression tree algorithm that balances accuracy with interpretability. Each CUBIST rule corresponds to a subset of the data satisfying a set of if-then conditions, with a linear regression model fitted to each subset. The team deliberately used a single committee rather than an ensemble to preserve transparency, and validated the models with 10-fold cross-validation grouped by site. The MAOC model achieved a root mean squared error of 2.77 grams per kilogram with Lin’s concordance of 0.91, while the deficit model achieved 3.72 grams per kilogram with a concordance of 0.94 — both unbiased, and notably stronger than earlier partial least squares regression approaches applied to similar problems in New Zealand and Australian soils.

The interpretability analysis proved the scientific heart of the work. By examining the regression coefficients within each rule and computing SHAP values — SHapley Additive exPlanations, a game-theory-based method that assigns each spectral feature an instance-level contribution to every prediction — the researchers could see exactly what the model was reading. All rules relied on the region between 2946 and 2850 wavenumbers, associated with carbon-hydrogen vibrations of alkyl groups in organic carbon. Rules covering carbon-poor soils leaned on absorptions from quartz, a chemically inert mineral with negligible reactive surface area, and from carbonates typical of arid regions with low plant productivity. The rule covering the most carbon-rich soils uniquely drew on hydroxyl stretching vibrations from clay minerals, the reactive matrix that facilitates organo-mineral bonding.

For the carbon deficit model, the SHAP analysis revealed an elegant logic. Absorptions from organic matter contributed negatively — existing mineral-associated carbon already occupying reactive mineral surfaces means less remaining capacity — while absorptions from clay minerals and silicates contributed positively, signalling abundant reactive surface area that is available but not yet occupied. The model shifted progressively from organic-carbon-dominated interpretation in carbon-rich soils to mineral-dominated interpretation in carbon-poor, fine-textured soils with the largest deficits. The researchers note limitations, including overlapping absorptions in the fingerprint region and a mismatch between the 50-micrometre fractionation cutoff and the Australian 20-micrometre texture definition, but argue the principles apply across land uses, soil types and climates. With laboratory models potentially underpinning future remote-sensing calibration, the approach could scale soil carbon monitoring for climate mitigation targets under the Paris Agreement and carbon credit schemes such as Australia’s ACCU programme.

Subject of Research: Estimating soil carbon sequestration potential using mid-infrared spectroscopy and explainable machine learning

Article Title: Estimating soil carbon sequestration potential with mid-IR spectroscopy and explainable machine learning

Article References: Estimating soil carbon sequestration potential with mid-IR spectroscopy and explainable machine learning. (n.d.). https://doi.org/10.5194/soil-12-619-2026

Image Credits: AI Generated

DOI: 10.5194/soil-12-619-2026

Keywords: soil carbon, carbon sequestration, mid-infrared spectroscopy, machine learning, mineral-associated organic carbon, carbon saturation deficit, SHAP, CUBIST, frontier line analysis, soil health, climate mitigation, Australian soils