Spatiotemporal controls on soil organic carbon stocks in subtropical grazing lands: An uncertainty-aware digital soil mapping approach
This repository contains the code and reproducible workflows associated with the study:
Zhao, C., Song, J., Dubeux, J., Grunwald, S., Bretas, I. L., Liao, H.-Y., Tziolas, N., Harley, J. B., Zare, A., Babaeian, E., Garcia, L., Queiroz, L., & Mendes, C. T. E. (2026). Spatiotemporal controls on soil organic carbon stocks in subtropical grazing lands: An uncertainty-aware digital soil mapping approach. Geoderma. 473, 118004. https://doi.org/10.1016/j.geoderma.2026.118004
Publication status: Accepted for publication in Geoderma.
Accurate spatial assessment of soil organic carbon (SOC) stocks and their associated uncertainty is important for understanding terrestrial carbon storage and supporting climate-smart land management. However, SOC stocks in grazing systems remain insufficiently characterized, particularly in subtropical regions where soils, hydroclimate, vegetation dynamics, and management practices are highly heterogeneous.
This study developed an uncertainty-aware digital soil mapping framework to estimate topsoil SOC stocks (0–20 cm) across Florida grazing lands at 30 m spatial resolution.
The framework integrates:
- contemporary and legacy soil observations;
- static and temporally dynamic environmental covariates;
- explainable machine learning;
- Quantile Regression Forests (QRF) for probabilistic prediction; and
- data-spiking experiments to evaluate the value and limitations of legacy soil observations.
The analysis operationalizes the STEP–AWBH soil-forming framework using long-term Earth observation and environmental time series representing hydroclimatic variability, soil moisture dynamics, vegetation phenology, grazing intensity, soil properties, and other environmental controls.
- A parsimonious and uncertainty-aware digital soil mapping framework was developed for predicting topsoil SOC stocks across subtropical grazing lands.
- Cross-temporal data spiking using legacy grazing-land observations improved both prediction accuracy and uncertainty estimation.
- Incorporating legacy observations from heterogeneous non-grazing land uses degraded model performance, demonstrating that more legacy data are not necessarily better when environmental and land-use domains are inconsistent.
- SHAP analysis identified pedological properties and temporal signatures of hydroclimate, soil moisture dynamics, and vegetation phenology as important controls of SOC spatial variability.
- The resulting maps provide spatially explicit SOC predictions and prediction intervals that can support carbon monitoring, grazing-land management, and future soil sampling.
The final model estimated approximately 33.95 Tg of SOC stored within the upper 0–20 cm of Florida grazing-land soils, with a mean predicted stock of 19.12 t ha⁻¹.
The spatial dataset associated with the final analysis contains five GeoTIFF layers at 30 m spatial resolution:
- grazing-land mask;
- mean SOC stock prediction;
- 5th percentile prediction (Q5);
- 95th percentile prediction (Q95); and
- 90% prediction interval width.
Recommended dataset: Version 2.0
DOI: https://doi.org/10.5281/zenodo.22000721
Version 2.0 contains the revised SOC prediction and uncertainty products corresponding to the analysis presented in the manuscript accepted for publication in Geoderma.
The earlier Version 1.0, which accompanied the preprint-stage analysis, remains available through Zenodo for provenance and reproducibility:
Version 1.0: https://doi.org/10.5281/zenodo.19192952
Users seeking data associated with the final peer-reviewed study should use Version 2.0.
The SOC prediction and uncertainty products can also be explored through an interactive Web GIS application.
The application allows users to interactively explore the spatial distribution of predicted topsoil SOC stocks and associated prediction uncertainty across Florida grazing lands.
The analytical workflow is implemented primarily as Jupyter notebooks and covers the major steps of the study:
- Environmental covariate acquisition and preprocessing
- Feature engineering
- Soil observation harmonization and legacy-data integration
- Feature selection and model development
- Quantile Regression Forest modeling
- Prediction uncertainty assessment
- SHAP-based model interpretation
- Spatial prediction and map generation
- Visualization of model results and spatial products
The repository is intended to support transparency and reproducibility of the modeling framework presented in the associated study.
Clone the repository:
git clone https://github.com/Ecosystem-Services-GeoAI/florida-grazing-soc-qrf.git
cd florida-grazing-soc-qrfCreate the Conda environment:
conda env create -f environment.ymlThen activate the environment using the environment name specified in environment.yml and launch Jupyter:
jupyter notebookThe individual notebooks can then be executed according to the workflow documented in the repository.
Note: Some source datasets used in the analysis are maintained by external data providers and are therefore not redistributed directly through this repository. Please refer to the notebooks and associated manuscript for data sources and preprocessing procedures.
If you use the methods, code, or scientific results from this study, please cite the associated publication:
Zhao, C., Song, J., Dubeux, J., Grunwald, S., Bretas, I. L., Liao, H.-Y., Tziolas, N., Harley, J. B., Zare, A., Babaeian, E., Garcia, L., Queiroz, L., & Mendes, C. T. E. (2026). Spatiotemporal controls on soil organic carbon stocks in subtropical grazing lands: An uncertainty-aware digital soil mapping approach. Geoderma. 473, 118004. https://doi.org/10.1016/j.geoderma.2026.118004
If you use the spatial SOC dataset, please also cite:
Zhao, C., Song, J., Dubeux, J., Grunwald, S., Bretas, I. L., Liao, H.-Y., Tziolas, N., Harley, J. B., Zare, A., Babaeian, E., Garcia, L., Queiroz, L., & Mendes, C. T. E. (2026). Topsoil Organic Carbon Stocks and Uncertainty in Florida Grazing Lands Derived from Quantile Regression Forest (30 m Resolution) (Version 2.0) [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.22000721
An earlier version of the manuscript is available as a preprint:
Zhao, C., Song, J., Dubeux, J., Grunwald, S., Bretas, I. L., Liao, H.-Y., Tziolas, N., Harley, J. B., Zare, A., Babaeian, E., Garcia, L., Queiroz, L., & Mendes, C. T. E. (2026). Spatiotemporal controls on soil organic carbon stocks in subtropical grazing lands: An uncertainty-aware digital soil mapping approach. SSRN. https://doi.org/10.2139/ssrn.6459841
The peer-reviewed Geoderma article should be considered the definitive version of the study.