TY - JOUR

T1 - Explainable machine learning models for predicting topsoil metals and oxides

AU - Сулейманов, Азамат Русланович

AU - Низамутдинов, Тимур Ильгизович

AU - Поляков, Вячеслав Игоревич

AU - Шевченко, Евгений Викторович

AU - Динкелакер, Наталья Владимировна

AU - Петрович, Марко

AU - Стосик, Лазарь

AU - Адельмурзина, Ильгиза

AU - Абакумов, Евгений Васильевич

PY - 2025/10/14

Y1 - 2025/10/14

N2 - Purpose: Interpretable machine learning (ML) models can help researchers and policymakers make well-informed decisions about soil and environmental protection. This study aimed to predict metals (including potential toxic elements) and their oxides in a topsoil using ML techniques with their subsequent interpretation. Materials and methods: We demonstrated it using a dataset containing eleven elements, including As, Co, Cr, Fe2O3, MnO, Ni, Pb, Sr, TiO2, V and Zn collected from mineral and organic soils. The prediction of each element’s concentration was based on a model incorporating other elements, soil properties (pH, SOC content and stock, bulk density), vegetation condition (as measured by NDVI), and land use type (abandoned and pristine soils, peatlands) as predictors. The Shapley Additive explanations (SHAP) approach, a technique from game theory, was used for interpretation of the ML models. Results and discussion: Cross-validation revealed the accurate prediction of most elements. Among them, vanadium (MAE = 3.72 mg/kg, RMSE = 4.59 mg/kg, R2 = 0.95, MEC = 0.93) and TiO2 (MAE = 0.03%, RMSE = 0.04%, R2 = 0.94, MEC = 0.93) were predicted with the highest accuracy. While elements served as primary predictors in ML models, the SHAP analysis allowed us to identify the specific contribution (positive or negative) of each element, as well as soil property and land use type. For instance, soil parameters and land use type exhibited distinct contributions to the prediction of higher or lower concentrations of specific elements, reflecting the different pedological and geochemical processes across mineral and organic soils. Furthermore, the method provided thresholds that indicate the levels at which predictors exert a positive or negative influence on the outputs. Conclusion: Interpretable ML models are essential for understanding the complex relationships between soil elements and properties, paving the way for more accurate predictions and informed environmental management decisions. • Topsoil metals and oxides were modeled using machine learning techniques. • Models were interpreted by Shapley Additive explanations approach. • Positive or negative contributions of predictors were examined for each element.

AB - Purpose: Interpretable machine learning (ML) models can help researchers and policymakers make well-informed decisions about soil and environmental protection. This study aimed to predict metals (including potential toxic elements) and their oxides in a topsoil using ML techniques with their subsequent interpretation. Materials and methods: We demonstrated it using a dataset containing eleven elements, including As, Co, Cr, Fe2O3, MnO, Ni, Pb, Sr, TiO2, V and Zn collected from mineral and organic soils. The prediction of each element’s concentration was based on a model incorporating other elements, soil properties (pH, SOC content and stock, bulk density), vegetation condition (as measured by NDVI), and land use type (abandoned and pristine soils, peatlands) as predictors. The Shapley Additive explanations (SHAP) approach, a technique from game theory, was used for interpretation of the ML models. Results and discussion: Cross-validation revealed the accurate prediction of most elements. Among them, vanadium (MAE = 3.72 mg/kg, RMSE = 4.59 mg/kg, R2 = 0.95, MEC = 0.93) and TiO2 (MAE = 0.03%, RMSE = 0.04%, R2 = 0.94, MEC = 0.93) were predicted with the highest accuracy. While elements served as primary predictors in ML models, the SHAP analysis allowed us to identify the specific contribution (positive or negative) of each element, as well as soil property and land use type. For instance, soil parameters and land use type exhibited distinct contributions to the prediction of higher or lower concentrations of specific elements, reflecting the different pedological and geochemical processes across mineral and organic soils. Furthermore, the method provided thresholds that indicate the levels at which predictors exert a positive or negative influence on the outputs. Conclusion: Interpretable ML models are essential for understanding the complex relationships between soil elements and properties, paving the way for more accurate predictions and informed environmental management decisions. • Topsoil metals and oxides were modeled using machine learning techniques. • Models were interpreted by Shapley Additive explanations approach. • Positive or negative contributions of predictors were examined for each element.

KW - Heavy metals

KW - Machine learning

KW - Potential toxic elements

KW - SHAP

KW - Shapley values

KW - Soil

UR - https://www.mendeley.com/catalogue/89911fae-5d69-3f7b-b34f-1ed55880c004/

U2 - 10.1007/s11368-025-04143-2

DO - 10.1007/s11368-025-04143-2

M3 - Article

JO - Journal of Soils and Sediments

JF - Journal of Soils and Sediments

SN - 1439-0108

ER -