Xuemeng Tian
Research Subjects
Xuemeng Tian has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Leandro Parente
OpenGeoHub Foundation, Doorwerth, The Netherlands
5 shared publications Martin Herold
Laboratory of Geo‐Information Science and Remote Sensing Wageningen University & Research Wageningen The Netherlands
4 shared publicationsMurat Şahin
Department of Geosciences & Engineering, Delft University of Technology, Delft, Netherlands
4 shared publicationsDavide Consoli
OpenGeoHub, Doorwerth, Netherlands
3 shared publications- Davide Consoli
OpenGeoHub Foundation, Doorwerth, The Netherlands
3 shared publications Sytze de Bruin
Laboratory of Geo-Information Science and Remote Sensing Wageningen University & Research Wageningen The Netherlands
3 shared publications
Latest Publications
- Open Access
- 17 July 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-22
Reply on RC2
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- 17 July 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-22
Open AccessReply on RC2
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- Open Access
- 17 July 2026
- Published by Copernicus GmbH
- p. 1-22
Reply on RC1
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- 17 July 2026
- Published by Copernicus GmbH
- p. 1-22
Open AccessReply on RC1
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- Open Access
- 10 April 2026
- Published by Wiley in Global Change Biology
- Vol. 32 (4) , e70813
Modeling Soil Organic Carbon Changes Using Signal‐To‐Noise Analysis: A Case Study Using European Soil Survey Datasets
Soil organic carbon (SOC) is a key indicator of soil health and a crucial component of climate mitigation, making its reliable monitoring increasingly important. While Digital Soil Mapping (DSM) based on Machine Learning and Earth Observation (EO) data enables the generation of time series of spatially explicit SOC predictions, detecting temporal changes from these model predictions remains challenging due to the relatively large associated uncertainties. Although prediction uncertainties are now commonly reported, few studies have explicitly accounted for them when assessing SOC change. This study introduces a model‐based signal‐to‐noise ratio (SNR) framework to assess the detectability of SOC change using both the state‐first approach—modeling SOC states at each time point and then deriving change—and the change‐first approach—modeling SOC change directly from repeated measurements. SNR is defined as the ratio of predicted SOC (concentration, g/kg) change to its modeled uncertainty, enabling evaluation of change‐model reliability at pixel levels. Applied to repeated SOC observations from the pan‐European Land Use and Coverage Area Frame Survey, this framework assesses the reliability of SOC change modeling across multiple land‐cover types using Random Forest and Quantile Regression Forests. At the site level, prediction accuracy was poor and SNR values were consistently low. An illustrative aggregation analysis showed that spatial averaging improved SNR, supporting SOC change assessments at broader scales. However, further work is needed to incorporate land use and management information and to systematically examine how different aggregation schemes affect the results in various contexts, ensuring that aggregated outcomes remain meaningful and policy‐relevant. As an internal metric based on model predictions and their estimated uncertainty, SNR provides a practical diagnostic of change‐model confidence, especially when repeated ground‐truth SOC measurements are not available. We advocate for routine SNR reporting to enhance the transparency and credibility of DSM‐based SOC change monitoring.
- 10 April 2026
- Published by Wiley in Global Change Biology
- Vol. 32 (4) , e70813
Open AccessModeling Soil Organic Carbon Changes Using Signal‐To‐Noise Analysis: A Case Study Using European Soil Survey Datasets
Soil organic carbon (SOC) is a key indicator of soil health and a crucial component of climate mitigation, making its reliable monitoring increasingly important. While Digital Soil Mapping (DSM) based on Machine Learning and Earth Observation (EO) data enables the generation of time series of spatially explicit SOC predictions, detecting temporal changes from these model predictions remains challenging due to the relatively large associated uncertainties. Although prediction uncertainties are now commonly reported, few studies have explicitly accounted for them when assessing SOC change. This study introduces a model‐based signal‐to‐noise ratio (SNR) framework to assess the detectability of SOC change using both the state‐first approach—modeling SOC states at each time point and then deriving change—and the change‐first approach—modeling SOC change directly from repeated measurements. SNR is defined as the ratio of predicted SOC (concentration, g/kg) change to its modeled uncertainty, enabling evaluation of change‐model reliability at pixel levels. Applied to repeated SOC observations from the pan‐European Land Use and Coverage Area Frame Survey, this framework assesses the reliability of SOC change modeling across multiple land‐cover types using Random Forest and Quantile Regression Forests. At the site level, prediction accuracy was poor and SNR values were consistently low. An illustrative aggregation analysis showed that spatial averaging improved SNR, supporting SOC change assessments at broader scales. However, further work is needed to incorporate land use and management information and to systematically examine how different aggregation schemes affect the results in various contexts, ensuring that aggregated outcomes remain meaningful and policy‐relevant. As an internal metric based on model predictions and their estimated uncertainty, SNR provides a practical diagnostic of change‐model confidence, especially when repeated ground‐truth SOC measurements are not available. We advocate for routine SNR reporting to enhance the transparency and credibility of DSM‐based SOC change monitoring.
- Open Access
- 23 February 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-22
Comment on egusphere-2026-229
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- 23 February 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-22
Open AccessComment on egusphere-2026-229
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- Top Cited PapersOpen Access
- 6 February 2026
- Published by Copernicus GmbH in Earth System Science Data
- Vol. 18 (2) , 989-1036
OpenLandMap-soildb: global soil information at 30 m spatial resolution for 2000–2022+ based on spatiotemporal Machine Learning and harmonized legacy soil samples and observations
There is increasing interest in global dynamic soil information with changes in soil properties mapped over time and at high spatial resolution. Thanks to long-term, multi-temporal, and fine- and medium-resolution satellite missions such as Landsat, MODIS, Copernicus Sentinel and similar, it is possible to produce globally consistent predictions of key soil variables that match other 10–30 m spatial resolution global data sets. This paper describes data preparation, modeling, and production of OpenLandMap-soildb: global dynamic predictions of soil organic carbon content, soil organic carbon density, bulk density, soil pH in H2O, soil texture fractions (clay, sand and silt) and USDA subgroup soil types (USDA soil taxonomy subgroups) at 30 m spatial resolution based on spatiotemporal Machine Learning (Quantile Regression Random Forest with output predictions showing the mean plus the 68 % probability lower and upper prediction intervals). To train the models, a large compilation of soil samples imported from legacy soil projects was used: 216 000 soil samples with soil carbon density (kg m−3), 408 000 soil samples with soil carbon content (g kg−1), 272 000 soil samples with soil pH in H2O, 363 000 soil samples with clay, silt and sand content (%) and 134 000 samples with bulk density oven dry (t m−3). Soil carbon and soil pH were mapped with 5-year time-intervals; soil texture fractions, bulk density, and soil types were mapped for recent years only. The cross-validation results indicate Root Mean Square Error (RMSE) of 17.7 (kg m−3; 0.486 in log-scale) and Concordance Correlation Coefficient (CCC) of 0.88 for SOC density, RMSE of 51.3 (g kg−1; 0.574 in log-scale) and CCC of 0.87 for SOC content, RMSE of 0.15 (t m−3) and CCC of 0.92 for bulk density of fine-earth, RMSE of 0.51 and CCC of 0.91 for soil pH, RMSE of 8.4 % and CCC of 0.87 for soil clay content, and RMSE of 12.6 % and CCC of 0.84 for soil sand content respectively. The most important variables for predicting soil organic carbon density (kg m−3) were: soil depth, Landsat-based uncalibrated Gross Primary Productivity (GPP), Normalized Difference Vegetation Index (NDVI) and CHELSA bioclimatic indices. The global distribution of soil pH can be primarily explained by the CHELSA Aridity Index (long-term), annual precipitation, and salinity grade. The global stocks for 2020–2022+ period for 0–30 cm depth interval are estimated at 461 Pg (Peta grams); the results further indicate that, in the last 25 years, the world has lost at least 11 Pg of SOC in the top soil. Suggestions are made on how to set up global permanent monitoring stations to accurately track land degradation and enable land restoration projects. The training data set is available at https://doi.org/10.5281/zenodo.4748499 (Hengl and Gupta, 2025), while the resulting data products can be accessed at https://doi.org/10.5281/zenodo.15470431 (Consoli et al., 2025) and https://world.soils.app (OpenGeoHub Foundation, 2026). Both datasets are released under a CC-BY license.
- 6 February 2026
- Published by Copernicus GmbH in Earth System Science Data
- Vol. 18 (2) , 989-1036
Top Cited PapersOpen AccessOpenLandMap-soildb: global soil information at 30 m spatial resolution for 2000–2022+ based on spatiotemporal Machine Learning and harmonized legacy soil samples and observations
There is increasing interest in global dynamic soil information with changes in soil properties mapped over time and at high spatial resolution. Thanks to long-term, multi-temporal, and fine- and medium-resolution satellite missions such as Landsat, MODIS, Copernicus Sentinel and similar, it is possible to produce globally consistent predictions of key soil variables that match other 10–30 m spatial resolution global data sets. This paper describes data preparation, modeling, and production of OpenLandMap-soildb: global dynamic predictions of soil organic carbon content, soil organic carbon density, bulk density, soil pH in H2O, soil texture fractions (clay, sand and silt) and USDA subgroup soil types (USDA soil taxonomy subgroups) at 30 m spatial resolution based on spatiotemporal Machine Learning (Quantile Regression Random Forest with output predictions showing the mean plus the 68 % probability lower and upper prediction intervals). To train the models, a large compilation of soil samples imported from legacy soil projects was used: 216 000 soil samples with soil carbon density (kg m−3), 408 000 soil samples with soil carbon content (g kg−1), 272 000 soil samples with soil pH in H2O, 363 000 soil samples with clay, silt and sand content (%) and 134 000 samples with bulk density oven dry (t m−3). Soil carbon and soil pH were mapped with 5-year time-intervals; soil texture fractions, bulk density, and soil types were mapped for recent years only. The cross-validation results indicate Root Mean Square Error (RMSE) of 17.7 (kg m−3; 0.486 in log-scale) and Concordance Correlation Coefficient (CCC) of 0.88 for SOC density, RMSE of 51.3 (g kg−1; 0.574 in log-scale) and CCC of 0.87 for SOC content, RMSE of 0.15 (t m−3) and CCC of 0.92 for bulk density of fine-earth, RMSE of 0.51 and CCC of 0.91 for soil pH, RMSE of 8.4 % and CCC of 0.87 for soil clay content, and RMSE of 12.6 % and CCC of 0.84 for soil sand content respectively. The most important variables for predicting soil organic carbon density (kg m−3) were: soil depth, Landsat-based uncalibrated Gross Primary Productivity (GPP), Normalized Difference Vegetation Index (NDVI) and CHELSA bioclimatic indices. The global distribution of soil pH can be primarily explained by the CHELSA Aridity Index (long-term), annual precipitation, and salinity grade. The global stocks for 2020–2022+ period for 0–30 cm depth interval are estimated at 461 Pg (Peta grams); the results further indicate that, in the last 25 years, the world has lost at least 11 Pg of SOC in the top soil. Suggestions are made on how to set up global permanent monitoring stations to accurately track land degradation and enable land restoration projects. The training data set is available at https://doi.org/10.5281/zenodo.4748499 (Hengl and Gupta, 2025), while the resulting data products can be accessed at https://doi.org/10.5281/zenodo.15470431 (Consoli et al., 2025) and https://world.soils.app (OpenGeoHub Foundation, 2026). Both datasets are released under a CC-BY license.
- Open Access
- 28 January 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-22
Soil science-informed neural networks for soil organic carbon density modelling under scarce bulk density data
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- 28 January 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-22
Open AccessSoil science-informed neural networks for soil organic carbon density modelling under scarce bulk density data
Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- Open Access
- 7 August 2025
- Published by Springer Nature in Research Square
Modeling soil organic carbon changes using signal-to-noise analysis: a case study using European soil survey datasets
Soil organic carbon (SOC) is a key indicator of soil health and a crucial component for assessing climate mitigation, making its reliable monitoring increasingly relevant. While digital soil mapping (DSM) using machine learning (ML) and Earth observation (EO) data enables time series of spatially explicit SOC predictions, detecting temporal soil carbon stock changes remains challenging and raises the question whether multi-temporal SOC measurements are sufficient to quantify changes with confidence. This study introduces a model-based signal-to-noise ratio (SNR) framework to assess the detectability of SOC change using both model-then-derive and derive-then-model approaches. We define SNR as the ratio of predicted SOC change to its modeled uncertainty. This enables the evaluation of change modeling reliability at both pixel and aggregated spatial levels. Applied to repeated SOC observations from the LUCAS soil survey at the European scale, the framework assesses the reliability of SOC change modeling across multiple land cover types using Random Forest (RF) and Quantile Regression Forests (QRF). At the site level prediction accuracy was poor, and SNR values were consistently low (<1). This highlights the limitations of current data for modeling SOC dynamics. However, spatial aggregation improved the SNR, supporting SOC change assessments at broader scales. SNR offers a practical diagnostic of model confidence, with a weak correlation to observation-based error, reflecting its role as an internal metric rather than a direct measure of accuracy. We advocate for routine SNR reporting to enhance the transparency and credibility of DSM-based SOC change monitoring while EO datasets enhance and ground-measurement surveys expand in time and quality.
- 7 August 2025
- Published by Springer Nature in Research Square
Open AccessModeling soil organic carbon changes using signal-to-noise analysis: a case study using European soil survey datasets
Soil organic carbon (SOC) is a key indicator of soil health and a crucial component for assessing climate mitigation, making its reliable monitoring increasingly relevant. While digital soil mapping (DSM) using machine learning (ML) and Earth observation (EO) data enables time series of spatially explicit SOC predictions, detecting temporal soil carbon stock changes remains challenging and raises the question whether multi-temporal SOC measurements are sufficient to quantify changes with confidence. This study introduces a model-based signal-to-noise ratio (SNR) framework to assess the detectability of SOC change using both model-then-derive and derive-then-model approaches. We define SNR as the ratio of predicted SOC change to its modeled uncertainty. This enables the evaluation of change modeling reliability at both pixel and aggregated spatial levels. Applied to repeated SOC observations from the LUCAS soil survey at the European scale, the framework assesses the reliability of SOC change modeling across multiple land cover types using Random Forest (RF) and Quantile Regression Forests (QRF). At the site level prediction accuracy was poor, and SNR values were consistently low (<1). This highlights the limitations of current data for modeling SOC dynamics. However, spatial aggregation improved the SNR, supporting SOC change assessments at broader scales. SNR offers a practical diagnostic of model confidence, with a weak correlation to observation-based error, reflecting its role as an internal metric rather than a direct measure of accuracy. We advocate for routine SNR reporting to enhance the transparency and credibility of DSM-based SOC change monitoring while EO datasets enhance and ground-measurement surveys expand in time and quality.
- Open Access
Spatiotemporal prediction of soil organic carbon density in Europe (2000–2022) using earth observation and machine learning
This article describes a comprehensive framework for soil organic carbon density (SOCD, kg/m3) modeling and mapping, based on spatiotemporal random forest (RF) and quantile regression forests (QRF). A total of 45,616 SOCD observations and various Earth observation (EO) feature layers were used to produce 30 m SOCD maps for the EU at four-year intervals (2000–2022) and four soil depth intervals (0–20 cm, 20–50 cm, 50–100 cm, and 100–200 cm). Per-pixel 95% probability prediction intervals (PIs) and extrapolation risk probabilities are also provided. Model evaluation indicates good overall accuracy (R2 = 0.63 and CCC = 0.76 for hold-out independent tests). Prediction accuracy varies by land cover, depth interval and year of prediction with the worst accuracy for shrubland and deeper soils 100–200 cm. The PI validation confirmed effective uncertainty estimation, though with reduced accuracy for higher SOCD values. Shapley analysis identified soil depth as the most influential feature, followed by vegetation, long-term bioclimate, and topographic features. While pixel-level uncertainty is substantial, spatial aggregation reduces uncertainty by approximately 66%. Detecting SOCD changes remains challenging but offers a baseline for future improvements. Maps, based primarily on topsoil data from cropland, grassland, and woodland, are best suited for applications related to these land covers and depths. We recommend that users interpret the maps in conjunction with local knowledge and consider the accompanying uncertainty and extrapolation risk layers. All data and code are available under an open license at https://doi.org/10.5281/zenodo.13754343 and https://github.com/AI4SoilHealth/SoilHealthDataCube/.
Open AccessSpatiotemporal prediction of soil organic carbon density in Europe (2000–2022) using earth observation and machine learning
This article describes a comprehensive framework for soil organic carbon density (SOCD, kg/m3) modeling and mapping, based on spatiotemporal random forest (RF) and quantile regression forests (QRF). A total of 45,616 SOCD observations and various Earth observation (EO) feature layers were used to produce 30 m SOCD maps for the EU at four-year intervals (2000–2022) and four soil depth intervals (0–20 cm, 20–50 cm, 50–100 cm, and 100–200 cm). Per-pixel 95% probability prediction intervals (PIs) and extrapolation risk probabilities are also provided. Model evaluation indicates good overall accuracy (R2 = 0.63 and CCC = 0.76 for hold-out independent tests). Prediction accuracy varies by land cover, depth interval and year of prediction with the worst accuracy for shrubland and deeper soils 100–200 cm. The PI validation confirmed effective uncertainty estimation, though with reduced accuracy for higher SOCD values. Shapley analysis identified soil depth as the most influential feature, followed by vegetation, long-term bioclimate, and topographic features. While pixel-level uncertainty is substantial, spatial aggregation reduces uncertainty by approximately 66%. Detecting SOCD changes remains challenging but offers a baseline for future improvements. Maps, based primarily on topsoil data from cropland, grassland, and woodland, are best suited for applications related to these land covers and depths. We recommend that users interpret the maps in conjunction with local knowledge and consider the accompanying uncertainty and extrapolation risk layers. All data and code are available under an open license at https://doi.org/10.5281/zenodo.13754343 and https://github.com/AI4SoilHealth/SoilHealthDataCube/.
- Open Access
- 24 June 2025
- Published by Copernicus GmbH in EGUsphere
- Vol. 2025, 1-66
OpenLandMap-soildb: global soil information at 30 m spatial resolution for 2000–2022+ based on spatiotemporal Machine Learning and harmonized legacy soil samples and observations
There is increasing interest in global dynamic soil information with changes in soil properties mapped over time and at high spatial resolution. Thanks to long-term, multi-temporal, and fine- and medium-resolution satellite missions such as Landsat, MODIS, Copernicus Sentinel and similar, it is possible to produce globally consistent predictions of key soil variables that match other 10–30 m spatial resolution global data sets. This paper describes data preparation, modeling, and production of OpenLandMap-soildb: global dynamic predictions of soil organic carbon content, soil organic carbon density, bulk density, soil pH in H2O, soil texture fractions (clay, sand and slit) and USDA subgroup soil types (USDA soil taxonomy subgroups) at 30 m spatial resolution based on spatiotemporal Machine Learning (Quantile Regression Random Forest with output predictions showing the mean plus the lower and upper prediction intervals of 68 % probability). To train the models, a large compilation of soil samples imported from legacy soil projects was used: 216,000 soil samples with soil carbon density (kg m-3), 408,000 soil samples with soil carbon content (g kg-1), 272,000 samples with soil pH in H2O, 363,000 samples with clay, silt, and sand (%), and 134,000 samples with bulk density oven dry (t m-3). Soil carbon and soil pH were mapped with 5 year time-intervals; soil texture fractions, bulk density, and soil types were mapped for recent years only. The cross-validation results indicate RMSE of 17.7 (kg m-3; 0.486 in log-scale) and CCC of 0.88 for SOC density, RMSE of 51.3 (g kg-1; 0.574 in log-scale) and CCC of 0.87 for SOC content, RMSE of 0.15 (t m-3) and CCC of 0.92 for bulk density of fine-earth, RMSE of 0.51 and CCC of 0.91 for soil pH, RMSE of 8.4 % and CCC of 0.87 for soil clay content, and RMSE of 12.6 % and CCC of 0.84 for soil sand content respectively. The most important variables for predicting soil organic carbon density (kg m-3) were: soil depth, Landsat-based uncalibrated Gross Primary Productivity (GPP), Normalized Difference Vegetation Index (NDVI) and CHELSA bioclimatic indices. The global distribution of soil pH can be primarily explained by the CHELSA Aridity Index (long-term), annual precipitation, and salinity grade. The global stocks for 2020–2022+ period for 0–30 cm depth interval are estimated at 461 Pg (Peta grams); the results further indicate that, in the last 25 years, the world has lost at least 11 Pg of SOC in the top soil. Suggestions are made on how to set up global permanent monitoring stations to accurately track land degradation and enable land restoration projects. The training dataset is available at https://doi.org/10.5281/zenodo.4748499 (Hengl and Gupta, 2025), while the resulting data products can be accessed at https://doi.org/10.5281/zenodo.15470431 (Consoli et al., 2025). Both datasets are released under a CC-BY license.
- 24 June 2025
- Published by Copernicus GmbH in EGUsphere
- Vol. 2025, 1-66
Open AccessOpenLandMap-soildb: global soil information at 30 m spatial resolution for 2000–2022+ based on spatiotemporal Machine Learning and harmonized legacy soil samples and observations
There is increasing interest in global dynamic soil information with changes in soil properties mapped over time and at high spatial resolution. Thanks to long-term, multi-temporal, and fine- and medium-resolution satellite missions such as Landsat, MODIS, Copernicus Sentinel and similar, it is possible to produce globally consistent predictions of key soil variables that match other 10–30 m spatial resolution global data sets. This paper describes data preparation, modeling, and production of OpenLandMap-soildb: global dynamic predictions of soil organic carbon content, soil organic carbon density, bulk density, soil pH in H2O, soil texture fractions (clay, sand and slit) and USDA subgroup soil types (USDA soil taxonomy subgroups) at 30 m spatial resolution based on spatiotemporal Machine Learning (Quantile Regression Random Forest with output predictions showing the mean plus the lower and upper prediction intervals of 68 % probability). To train the models, a large compilation of soil samples imported from legacy soil projects was used: 216,000 soil samples with soil carbon density (kg m-3), 408,000 soil samples with soil carbon content (g kg-1), 272,000 samples with soil pH in H2O, 363,000 samples with clay, silt, and sand (%), and 134,000 samples with bulk density oven dry (t m-3). Soil carbon and soil pH were mapped with 5 year time-intervals; soil texture fractions, bulk density, and soil types were mapped for recent years only. The cross-validation results indicate RMSE of 17.7 (kg m-3; 0.486 in log-scale) and CCC of 0.88 for SOC density, RMSE of 51.3 (g kg-1; 0.574 in log-scale) and CCC of 0.87 for SOC content, RMSE of 0.15 (t m-3) and CCC of 0.92 for bulk density of fine-earth, RMSE of 0.51 and CCC of 0.91 for soil pH, RMSE of 8.4 % and CCC of 0.87 for soil clay content, and RMSE of 12.6 % and CCC of 0.84 for soil sand content respectively. The most important variables for predicting soil organic carbon density (kg m-3) were: soil depth, Landsat-based uncalibrated Gross Primary Productivity (GPP), Normalized Difference Vegetation Index (NDVI) and CHELSA bioclimatic indices. The global distribution of soil pH can be primarily explained by the CHELSA Aridity Index (long-term), annual precipitation, and salinity grade. The global stocks for 2020–2022+ period for 0–30 cm depth interval are estimated at 461 Pg (Peta grams); the results further indicate that, in the last 25 years, the world has lost at least 11 Pg of SOC in the top soil. Suggestions are made on how to set up global permanent monitoring stations to accurately track land degradation and enable land restoration projects. The training dataset is available at https://doi.org/10.5281/zenodo.4748499 (Hengl and Gupta, 2025), while the resulting data products can be accessed at https://doi.org/10.5281/zenodo.15470431 (Consoli et al., 2025). Both datasets are released under a CC-BY license.
- Open Access
- 1 May 2025
- Published by Springer Nature in Research Square
Spatiotemporal prediction of Soil Organic Carbon Density in Europe (2000–2022) using Earth Observation and Machine Learning
The paper describes a comprehensive framework for soil organic carbon density (SOCD) (kg/m3) modeling and mapping, based on spatiotemporal Random Forest (RF) and Quantile Regression Forests (QRF). A total of 45,616 SOCD observations and various Earth Observation (EO) feature layers were used to produce 30m SOCD maps for the EU at four-year intervals (2000--2022) and four soil depth intervals (0-20cm, 20-50cm, 50-100cm, and 100-200cm). Per-pixel 95% probability prediction intervals (PIs) and extrapolation risk probabilities are also provided. Model evaluation indicates good overall accuracy (R2 = 0.63 and CCC = 0.76 for hold-out independent tests). Prediction accuracy varies by land cover, depth interval and year of prediction with accuracy the worst for shrubland and deeper soils 100--200cm. PI validation confirmed effective uncertainty estimation, though with reduced accuracy for higher SOCD values. Shapley analysis identified soil depth as the most influential feature, followed by vegetation, long-term bioclimate, and topographic features. While pixel-level uncertainty is substantial, spatial aggregation reduces uncertainty by approximately 66%. Detecting SOCD changes remains challenging but offers a baseline for future improvements. Maps, based primarily on topsoil data from cropland, grassland, and woodland, are best suited for applications related to these land covers and depths. We recommend that users interpret the maps in conjunction with local knowledge and consider the accompanying uncertainty and extrapolation risk layers. All data and code are available under an open license at https://doi.org/10.5281/zenodo.13754343 and https://github.com/AI4SoilHealth/SoilHealthDataCube/.
- 1 May 2025
- Published by Springer Nature in Research Square
Open AccessSpatiotemporal prediction of Soil Organic Carbon Density in Europe (2000–2022) using Earth Observation and Machine Learning
The paper describes a comprehensive framework for soil organic carbon density (SOCD) (kg/m3) modeling and mapping, based on spatiotemporal Random Forest (RF) and Quantile Regression Forests (QRF). A total of 45,616 SOCD observations and various Earth Observation (EO) feature layers were used to produce 30m SOCD maps for the EU at four-year intervals (2000--2022) and four soil depth intervals (0-20cm, 20-50cm, 50-100cm, and 100-200cm). Per-pixel 95% probability prediction intervals (PIs) and extrapolation risk probabilities are also provided. Model evaluation indicates good overall accuracy (R2 = 0.63 and CCC = 0.76 for hold-out independent tests). Prediction accuracy varies by land cover, depth interval and year of prediction with accuracy the worst for shrubland and deeper soils 100--200cm. PI validation confirmed effective uncertainty estimation, though with reduced accuracy for higher SOCD values. Shapley analysis identified soil depth as the most influential feature, followed by vegetation, long-term bioclimate, and topographic features. While pixel-level uncertainty is substantial, spatial aggregation reduces uncertainty by approximately 66%. Detecting SOCD changes remains challenging but offers a baseline for future improvements. Maps, based primarily on topsoil data from cropland, grassland, and woodland, are best suited for applications related to these land covers and depths. We recommend that users interpret the maps in conjunction with local knowledge and consider the accompanying uncertainty and extrapolation risk layers. All data and code are available under an open license at https://doi.org/10.5281/zenodo.13754343 and https://github.com/AI4SoilHealth/SoilHealthDataCube/.