Methodology

Our approach combines three models to predict soil organic carbon (SOC) dynamics based on field measurements, then transforms these predictions into maps.

  • Roth-C (biogeochemical)

Roth-C is a model recognized in the literature for estimating changes in soil organic carbon. It is based on a simplified description of the major processes that control carbon in the soil: input of organic matter via plant residues, transformation by microbial life, and losses related to climate, soil type, and agricultural practices.

  • Allometric Model

This model is based on a simple principle: increased biomass production by a crop or vegetation (quantified, for example, using the NDVI index from satellites) is directly linked to a higher input of organic matter into the soil. By analyzing these correlations, it is possible to anticipate changes in soil organic carbon (SOC) based on changes in vegetation cover or agricultural practices.

  • XGBoost (AI)

XGBoost is a machine learning method that predicts SOC based on variables such as NDVI, pH, and clay content. The machine learning algorithm combines and corrects these predictions step by step to increase accuracy. Through this process, it is able to detect complex relationships between climate, soil, and vegetation and to project changes in soil carbon over time based on observed data.

Data

The dataset used, LimeSoDa (Precision Liming Soil Datasets), comprises 31 distinct agricultural sites spread across several continents (South America, Europe, Asia, and Oceania), covering a wide variety of soil and climate conditions. 

The areas studied range from 1.4 ha to 420 ha, allowing the robustness of the models to be tested at both the experimental plot and farm levels.

Each site systematically includes three target variables, soil organic carbon (SOC), pH, and clay content, to which additional properties are added depending on the site (e.g., sand, silt, moisture, bulk density).

SOC measurements are derived from standardized laboratory protocols. 

The covariates used for modeling come from several complementary sources:

  • Satellite remote sensing: Sentinel-2 images (level 2A, Copernicus Open Access Hub), used to generate spectral indices (VI, RSS) during the growing season and on bare soil, with a spatial resolution of 10–20 m.
  • Proximal sensors and spectroscopy: vis-NIR and MIR measurements (soil spectroscopy), gamma radiometry, ERa (apparent resistivity), and XRF (X-ray fluorescence spectroscopy), a non-destructive method in which X-ray excitation induces characteristic fluorescence that can be used to quantify the elemental composition of soils.
  • Topography: DEM and topographic derivatives.
  • Soil analyses: CSMoist (measured soil moisture), pH-ISE (ion-selective electrode), particle size fractions (sand, silt, clay).

All data are provided in a harmonized tabular format, directly usable for digital soil mapping.

Reference: Schmidinger et al., LimeSoDa: A Dataset Collection for Benchmarking of Machine Learning Regressors in Digital Soil Mapping, 2025.