Digital exploration of selected heavy metals using Random Forest and a set of environmental covariates at the watershed scale

Accurate prediction of spatial distribution of soil heavy metal in complex mining terrain using an improved machine learning method

Accurate prediction of heavy metals (HMs) spatial distribution in mining areas is crucial for pollution management. However, predicting the spatial distribution of HMs remains a significant challenge in mining areas with complex terrain and variable contaminant transport pathways. This study aims to optimize the spatial prediction of arsenic (As) distribution in the Shimen realgar mining area, the largest in Asia, by integrating machine learning models with kriging interpolation and feature selection techniques. The results show that the Random Forest (RF) model achieved the best performance in predicting soil As concentration, with an R2 of 0.84 for the test data. Incorporating environmental variables improved the spatial prediction accuracy, with RF (R2 = 0.76, RMSE = 24.68 mg/kg) and Random Forest Regression Kriging (RFRK) (R2 = 0.78, RMSE = 23.46 mg/kg) outperforming ordinary kriging and geographically weighted regression kriging. Importance analysis and recursive feature elimination further optimized the model, leading to a 5 % increase in R2 and a reduction of RMSE by 8 %-12.4 %. The optimized RFRK model accurately captured the spatial distribution of As in the mining area, revealing the outward diffusion pattern of As from the smelting plant. The findings highlight the critical role of feature selection in improving prediction accuracy in highly polluted and complex terrain regions, an aspect that has often been overlooked in previous studies. This study provides a practical framework for spatial prediction of contaminants in similar areas, enhancing the understanding of pollution distribution.

Evaluation of machine learning models for accurate prediction of heavy metals in coal mining region soils in Bangladesh

Coal mining soils are highly susceptible to heavy metal pollution due to the discharge of mine tailings, overburden dumps, and acid mine drainage. Developing a reliable predictive model for heavy metal concentrations in this region has proven to be a significant challenge. This study employed machine learning (ML) techniques to model heavy metal pollution in soils within this critical ecosystem. A total of 91 standardized soil samples were analyzed to predict the accumulation of eight heavy metals using four distinct ML algorithms. Among them, random forest model outer performed in predicting As (0.79), Cd (0.89), Cr (0.63), Ni (0.56), Cu (0.60), and Zn (0.52), achieving notable R squared values. The feature attribute analysis identified As-K, Pb-K, Cd-S, Zn-Fe2O3, Cr- Fe2O3, Ni-Al2O3, Cu-P, and Mn- Fe2O3 relationships resembled with correlation coefficients among them. The developed models revealed that the contamination factor for metals in soils indicated extremely high levels of Pb contamination (CF ≥ 6). In conclusion, this research offers a robust framework for predicting heavy metal pollution in coal mining soils, highlighting critical areas that require immediate conservation efforts. These findings emphasize the necessity for targeted environmental management and mitigation to reduce heavy metal pollution in mining sites.

Phragmites australis elevated concentrations of soil-bound heavy metals and magnetic particles in a typical urban plateau lake wetland, China

Vegetation change significantly altered the hydrological processes and soil erosion within riparian ecosystems. It is unclear how change in managed vegetation types affect the geochemical behavior of heavy metals (HMs) and magnetic particles in karst riparian areas. Two soil depths of 0-20 cm and 20-40 cm were taken in alien species Phragmites australis (P. australis), native species Juncus effuses and Schoenoplectus tabernaemontan in a typical urban plateau Lake wetland, Caohai lake, China. Low-frequency mass magnetic susceptibility (χLF), anhysteretic remanent susceptibility (χARM), isothermal remanent magnetization, Cd, Cr, Cu, Sb, Ni and Zn were determined. Compared with Juncus effuses and Schoenoplectus tabernaemontani, P. australis habitat had the higher values of HMs, χLF, χARM, and isothermal remanent magnetization in top-soils. Frequency-dependent magnetic susceptibility ranged from 4.84 % to 10.87 % in top-soils and 6.82 %-9.95 % in sub-soils, lithogenic/pedogenic factors mainly masked the contribution of anthropogenic factors to magnetic signal enhancement. The correlation between variations of Cu and Sb with χARM and isothermal remanent magnetization was found to be significant in top-soils, but not in sub-soils. P. australis tended to promote the enrichment of HMs and enhancement of magnetic signal, the impact of P. australis expansion on the distribution of soil HMs and magnetic particles in Caohai riparian wetland should be not disregarded.

Machine learning combined with geodetector to predict the spatial distribution of soil heavy metals in mining areas

An accurate understanding of the spatial distribution of soil heavy metals (HMs) is crucial for the effective prevention of soil pollution and remediation strategies. Traditional machine learning models often overlook the spatially stratified heterogeneity inherent to environmental data, which can impair predictive accuracy. Therefore, we combined the Geodetector model (GDM) with machine learning models. The factor detection results were used to screen covariates to consider the local spatial heterogeneity of model features. The interaction detection results were used to construct spatially stratified covariates to consider the spatially stratified heterogeneity of model features. The results showed that covariate screening largely avoided the introduction of redundant features. The constructed spatially stratified covariates improved the predictive performance of the model (both the R-squared (R2) and root mean square error (RMSE) of different models were optimized). Among these, the XGB model exhibited the best performance. Analysis of the factors influencing Pb and Cr revealed that the interaction between pH and NDVI was the main determinant of Pb spatial distribution (q = 0.3516, XGB Importance Score = 93). In contrast, the interaction between DEM and pH (q = 0.7156, XGB Importance Score = 121) as well as the distance to waste piles (q = 0.6390, XGB Importance Score = 66), were the main driving factors for the spatial distribution of Cr. The current work provides an improved approach for interrogating the factors that influence HM distribution in soil. This study offers valuable insights into the spatial distribution of soil HMs. The proposed methodology can be applied in future soil pollution assessments and environmental management strategies, thus contributing to more precise pollution prevention and remediation efforts.

Machine learning models with innovative outlier detection techniques for predicting heavy metal contamination in soils

Machine learning (ML) models for accurately predicting heavy metals with inconsistent outputs have improved owing to dataset outliers, which influence model reliability and accuracy. A comprehensive technique that combines machine learning and advanced statistical methods was applied to assess data outlier's effects on ML models. Ten ML models with three outlier detection methods predicted Cr, Ni, Cd, and Pb in Narayanganj soils. XGBoost with density-based spatial clustering of applications with noise (DBSCAN) improved model efficacy (R2). The R2 of Cr, Ni, Cd, and Pb was considerably enhanced by 11.11 %, 6.33 %, 14.47 %, and 5.68 %, respectively, indicating that outliers affected the model's HM prediction. Soil factors affected Cr (80 %), Ni (72.61 %), Cd (53.35 %), and Pb (63.47 %) concentrations based on feature importance. Contamination factor prediction showed considerable contamination for Cr, Ni, and Cd. LISA revealed Cd (55.4 %), Cr (49.3 %), and Pb (47.3 %) as the significant pollutant (p < 0.05). Moran's I index values for Cr, Ni, Cd, and Pb were 0.65, 0.58, 0.60, and 0.66, respectively, indicating strong positive spatial autocorrelation and clusters with similar contamination. Finally, this work successfully assessed the influence of data outliers on the ML model for soil HM contamination prediction, identifying crucial regions that require rapid conservation measures.

Determining the priority control sources of heavy metals in the roadside soils in a typical industrial city of North China

Heavy metals (HMs) in roadside soils (RS) are closely related to urban ecosystem and human health, but priority control sources and eco-health risks of HMs remain unclear. We explored pollution sources of HMs using positive matrix factorization (PMF), scanning electron microscopy with energy dispersive spectroscopy (SEM-EDS), and random forest (RF), and assessed their source-specific eco-health risks. The combination of PMF, SEM-EDS and RF indicated that pollution sources of HMs were coal combustion (27.99 %), construction materials (30.15 %), and traffic emissions (41.86 %). SEM-EDS revealed RS particles were identified by a combination of physical (spherical, porous rounded, crustal flake, crustal rounded and irregular particles) and elemental surface characteristics (O, C, Si, Al, Fe, Ca, Mg, K, Zn, Cu, and Mn). RF highlighted road network density, residential density, and industrial density had significant importance influence on pollution sources. Approximately 72.35 % of soil samples were at a low ecological risk with traffic emissions being the major contributor. Non-carcinogenic risks had a minimal effect, but carcinogenic risks were at a cautionary level with coal combustion being the highest contributor. Overall, coal combustion and traffic emissions were regarded as priority control sources of HMs. These findings provided effective guidance for soil pollution prevention and risk control.

A geographically weighted neural network model for digital soil mapping of heavy metal copper in coastal cities

Assessing the spatial distribution of heavy metals in soil is essential for mitigating risks to human health and ensuring the sustainable use of soil resources. This study proposed a geographically weighted neural network (GWNN) model, leveraging deep learning and geographically weighted regression (GWR). The model was designed based on the GWR concept to address the spatial autocorrelation of copper (Cu) in soil and incorporated convolutional neural network (CNN) to capture the nonlinear relationships between Cu and environmental covariates. The GWNN model was compared with GWR, extreme gradient boosting (XGBoost), and artificial neural network (ANN) models. XGBoost was employed to select important environmental covariates and spatial autocorrelation of Cu concentrations was assessed using Moran's I. The model's performance was evaluated using 10-fold cross-validation, and prediction uncertainty was quantified with 100 bootstrap models. The results indicated that temperature covariates were the most significant predictors of soil Cu concentrations. The R2 values for Cu prediction accuracy were 0.60 for GWNN, 0.53 for ANN, 0.49 for XGBoost, and 0.32 for GWR. The spatial distribution of Cu showed a trend of higher concentrations in the north and lower concentrations in the south, consistent with spatial clusters identified by local Moran's I. The mean uncertainty of the 90 % confidence interval for GWNN was 16.49 %, closely aligning with XGBoost (15.44 %) and ANN (16.29 %) and significantly outperforming the GWR (18.25 %). Overall, the GWNN model demonstrated strong predictive accuracy and low uncertainty, offering an improved approach for digital soil mapping applications.

Quantitative expression of LNAPL pollutant concentrations in capillary zone by coupling multiple environmental factors based on random forest algorithm

The capillary zone plays a crucial role in migration and transformation of pollutants. Light nonaqueous liquids (LNAPLs) have become the main organic pollutant in soil and groundwater environments. However, few studies have focused on the concentration distribution characteristics and quantitative expression of LNAPL pollutants within capillary zone. In this study, we conducted a sandbox-migration experiment using diesel oil as a typical LNAPL pollutant, with the capillary zone of silty sand as the research object. The variation characteristics of LNAPL pollutants (total petroleum hydrocarbon) concentration and environmental factors (moisture content, electrical conductivity, pH, and oxidationreduction potential) were essentially consistent at different locations with the same height. These characteristics differed within range of 10.0-50.0 cm and above 60.0 cm from groundwater. A model for quantitative expression of concentrations was constructed by coupling multiple environmental factors of 968 sets-7744 data via random forest algorithm. The goodness of fit (R2) for both training and test sets was greater than 0.90, and the mean absolute percentage error (MAPE) was less than 16.00 %. The absolute values of relative errors in predicting concentrations at characteristic points were less than 15.00 %. The constructed model can accurately and quantitatively express and predict concentrations in capillary zone.

A hierarchical residual correction-based hyperspectral inversion method for soil heavy metals considering spatial heterogeneity

Promising hyperspectral remote sensing exhibits substantial potential in monitoring soil heavy metal (SHM) contamination. Nevertheless, the local spatial perturbation effects induced by environmental factors introduce considerable variability in SHM distribution. This engenders non-stationary relationship between SHM concentrations and spectral reflectance, posing challenges for accurate inversion of SHM globally. Addressing this gap, a novel Hierarchical Residual Correction-based Hyperspectral Inversion Method (HRCHIM) is proposed for SHM, considering their spatial heterogeneity. Initially, a global model is constructed using ground hyperspectral data to predict SHM concentration, capturing overarching contamination trends. Subsequently, four hierarchical levels, segmented by residual standard deviation (SD) intervals, identify critical environmental factors via Geodetector. These factors inform local residual correction models, refining global model predictions. HRCHIM aims to synergize global trends and local stochasticity to enhance prediction accuracy and interpretation of SHM spatial heterogeneity. Validated through a case study of a Cadmium(Cd)-contaminated mine area, six critical environmental factors were identified, exhibiting significant differences across hierarchical levels. By incorporating hierarchical correction models, HRCHIM demonstrated superior inversion performance compared to other conventional methods, achieving optimal prediction accuracies (Rv2 = 0.94, RMSEv = 0.21, and RPDv = 4.11). This innovative method can facilitate more precise and targeted strategies for preventing and controlling SHM contamination.

Mapping global distributions of clay-size minerals via soil properties and machine learning techniques

Clay-size mineral is a vital ingredient of soil that influences various environment behaviors. It is crucial to establish a global distribution map of clay-size minerals to improve the recognition of environment variations. However, there is a huge gap of lacking some mineral contents in poorly accessible remote areas. In this work, machine learning (ML) approaches were conducted to predict the mineral contents and analyze their global abundance changes through the relationship between soil properties and mineral distributions. The average content of kaolinite, illite, smectite, vermiculite, chlorite, and feldspar were predicated to be 28.69 %, 22.30 %, 12.42 %, 5.43 %, 5.03 %, and 1.44 % respectively. Model interpretation showed that topsoil bulk density and drainage class were the most significant factors for predicting all six minerals. It could be seen from the feature importance analysis that bulk density notably reflected the distribution of 2:1 layered minerals more than that of 1:1 mineral. High drainage favored secondary minerals development, while low drainage was more benefited for primary minerals. Moreover, the content variation of different minerals aligned with the distribution of corresponding soil properties, which affirmed the accuracy of established models. This study proposed a new approach to predict mineral contents through soil properties, which filled a necessary step of understanding the geochemical cycles of soil-related processes.

Modelling Eucalyptus globulus spatial distribution in the upper Blue Nile basin using multi spectral Sentinel-2 and environmental data

Eucalyptus plantations are widespread in the highlands of northern Ethiopia. The species has been used for centuries for various purposes. However, there are controversies surrounding the species with excessive soil nutrient and water consumption. Modelling the spatial distribution of the species is fundamental to understand its ecological and hydrological effects in the region for policy inputs. Therefore, the purpose of this study is to develop a model for mapping the spatial distribution of Eucalyptus globulus. We used the spectral bands of Sentinel-2 data, vegetation indices, and environmental data as predictor variables and three machine learning algorithms (Random Forest, Support Vector Machine, and Boosted Regression Trees) to model the current distribution of Eucalyptus globulus. Eleven of the twenty-five predictor variables were filtered using a variance inflation factor (VIF). 419 in situ georeferenced data points were used for training, and validating the models. The area under the curve (AUC), kappa statistic (K), true skill statistic (TSS), Root Mean Squared Error and coefficient of determination (R2) were used to validate the models' performance. The model validation metrics confirmed the highest performance of Random Forest. The prediction map of Random Forest revealed that Eucalyptus globulus was fairly detected in non-Eucalyptus globulus woody vegetation (R2 = 0.86, P < 0.001; RMSE = 0.31). We found that the Green Normalized Difference Vegetation Index and environmental variables, such as elevation and distance from the road, were the most important predictor variables in explaining the distribution of Eucalyptus globulus. Our findings demonstrate that machine learning algorithms with Sentinel-2 spectral bands and vegetation indices compounded with environmental data can effectively model the spatial distribution of Eucalyptus globulus.

Quantifying the impact of factors on soil available arsenic using machine learning

Arsenic (As) can accumulate in edible plant parts and thus pose a serious threat to human health. Identifying the contributions of various factors to soil available As is crucial for evaluating environmental risks. However, research quantitatively assessing the importance of soil properties on available As is scarce. In this study, we utilized 442 datasets covering total As, available As, and properties of farmland soils. The five machine learning models were employed to predict soil available As content, and the model with the best predictive performance was selected to calculate the importance of soil properties on available As and interpret the model results. The Random Forest model exhibited the best predictive performance, with R2 for the test set of dryland and paddy fields being 0.83 and 0.82 respectively, while also outperforming other machine learning models in terms of accuracy. Concurrently, evaluating the contribution of soil properties to soil available As revealed that increases in soil total arsenic, pH, organic matter (OM), and cation exchange capacity (CEC) led to higher soil available As content. Among these factors, soil total As had the greatest impact, followed by CEC. The influence of pH on soil available As was greater in dryland compared to OM, while in paddy fields, it was smaller than OM (p < 0.01). Sensitivity analysis results indicated that reducing soil total As content had the greatest effect on available As. In both dryland and paddy field soils, reducing soil total As had the most pronounced effect on available As, leading to reductions of 10.09% and 8.48%, respectively. Therefore, prioritizing the regulation of soil total As and CEC is crucial in As contamination management practices to alter As availability in farmland soils.

Accurate prediction of spatial distribution of soil potentially toxic elements using machine learning and associated key influencing factors identification: A case study in mining and smelting area in southwestern China

Accurate prediction of spatial distribution of potentially toxic elements (PTEs) is crucial for soil pollution prevention and risk control. Achieving accurate prediction of spatial distribution of soil PTEs at a large scale using conventional methods presents significant challenges. In this study, machine learning (ML) models, specially artificial neural network (ANN), random forest (RF), and extreme gradient boosting (XGB), were used to predict spatial distribution of soil PTEs and identify associated key factors in mining and smelting area located in Yunnan Province, China, under the three scenarios: (1) natural + socioeconomic + spatial datasets (NS), (2) NS + irrigation pollution index (IPI) datasets, (3) NS + IPI + deposition (DEPO) datasets. The results highlighted the combination of NS+IPI+DEPO yielded the highest predictive accuracy across ML models. Particularly, XGB exhibited the highest performance for As (R2 =0.7939), Cd (R2 =0.6679), Cu (R2 =0.8519), Pb (R2 =0.8317), and Zn (R2 =0.7669), whereas RF performed the best for Ni (R2 =0.7146). The feature importance and Shapley additive explanation (SHAP) analysis revealed that DEPO and IPI were the pivotal factors influencing the distribution of soil PTEs. Our findings highlighted the important role of DEPO in spatial distribution prediction of soil PTEs, which has often been ignored in previous studies.

Geochemical characteristics, source analysis, influencing factors, and reserves of soil Selenium in Wuming, Guangxi, China

Although selenium (Se) reserves are crucial for the development and exploitation of Se-rich resources in karst soil areas, research on these reserves is still limited. A comprehensive study was conducted in a typical karst region known for its Se richness. A total of 12,547 surface soil samples, 134 deep soil samples, and 60 soil profiles from various locations were systematically collected. The findings showed that the Se content in the surface soil ranged from 0.073 to 9.04 mg/kg, with a baseline level of 0.84 mg/kg. This underscores the high background level and moderate variability in the region. Surface soil Se exhibited a notable positive correlation with deep soil Se, and an inverse correlation with pH (p < 0.01). One-way analysis of variance indicated that land formations and soil structure were the primary determinants affecting the concentration of Se in the topsoil (p = 0.000), with parent rock type and land-use type following closely (p = 0.003). In addition, the study included an investigation of soil Se variations with depth using 60 soil profiles. Through this analysis, it was revealed that Se content exhibited an exponential change with depth. Multiple integrations were employed to derive formulas for calculating Se reserves in the 0-200 cm depth range. Following these calculations, the estimations of Se stockpile across diverse types of source materials, varieties of soils, and land management methods were determined, highlighting the findings using a passive construction. This paper lays the groundwork for advancing the extraction and application of Se.

Prediction of adsorption of metal cations by clay minerals using machine learning

Adsorption of heavy metals by clay minerals occurs widely at the solid-liquid interface in natural environments, and in this paper, the phenomenon of adsorption of Cd2+, Cu2+, Pb2+, Zn2+, Ni2+ and Co2+ by montmorillonite, kaolinite and illite was simulated using machine learning. We firstly used six machine learning models including Random Forest(R), Extremely Forest(E), Gradient Boosting Decision Tree(G), Extreme Gradient Boosting(X), Light Gradient Boosting(LGB) and Category Boosting(CAT) to feature engineer the metal cations and the parameters of the minerals, and based on the feature engineering results, we determined the first order hydrolysis constant(log K), solubility product constant(SPC), and higher hydrolysis constant (HHC) as the descriptors of the metal cations, and site density(SD) and cation exchange capacity(CEC) as the descriptors of the clay minerals. After comparing the predictive effects of different data cleaning methods (pH50 method, Box method and pH50-Box method) and six model combinations, it was finally concluded that the best simulation results could be achieved by using the pH 50-Box method for data cleaning and Extreme Gradient Boosting for modelling (RMSE = 4.158 %, R2 = 0.977). Finally, model interpretation was carried out using Shapley explanation plot (SHAP) and partial dependence plot(PDP) to analyse the potential connection between each input variable and the output results. This study combines machine learning with geochemical analysis of the mechanism of heavy metal adsorption by clay minerals, which provides a different research perspective from the traditional surface complexation model.

Predicting the distribution coefficient of cesium in solid phase groups using machine learning

The migration and retention of radioactive contaminants such as 137Cesium (137Cs) in various environmental media pose significant long-term storage challenges for nuclear waste. The distribution coefficient (Kd) is a critical parameter for assessing the mobility of radioactive contaminants and is influenced by various environmental conditions. This study presents machine-learning models based on the Japan Atomic Energy Agency Sorption Database (JAEA-SDB) to predict the Kd values for Cs in solid phase groups. We used three different machine learning models: random forest (RF), artificial neural network (ANN), and convolutional neural network (CNN). The models were trained on 14 input variables from the JAEA-SDB, including factors such as the Cs concentration, solid-phase properties, and solution conditions, which were preprocessed by normalization and log-transformation. The performances of the models were evaluated using the coefficient of determination (R2) and root mean squared error (RMSE). The RF, ANN, and CNN models achieved R2 values greater than 0.97, 0.86, and 0.88, respectively. We also analyzed the variable importance of RF using an out-of-bag (OOB) and a CNN with an attention module. Our results showed that the environmental media, initial radionuclide concentration, solid phase properties, and solution conditions were significant variables for Kd prediction. Our models accurately predict Kd values for different environmental conditions and can assess the environmental risk by analyzing the behavior of radionuclides in solid phase groups. The results of this study can improve safety analyses and long-term risk assessments related to waste disposal and prevent potential hazards and sources of contamination in the surrounding environment.

A remote sensing-based strategy for mapping potentially toxic elements of soils: Temporal-spatial-spectral covariates combined with random forest

The selection of predictor variables is a crucial issue in building a digital mapping model of potentially toxic elements (PTEs) in soil. Traditionally, the predictor variables for mapping models of soil PTEs have been chosen from sets of spatial parameters or spectral parameters derived from geographical environmental data. However, the enrichment of soil PTEs exhibits significant variations in both spatial and temporal dimensions, with the temporal dimension often being overlooked in the selection of predictor variables for digital mapping models. This limitation hampers the robustness and generalizability of the models. Therefore, multi-source geographical data were used in this study to determine three temporal indices for characterizing the enrichment process of soil PTEs in temporal dimensions, and additionally to construct the temporal-spatial-spectral (TSS) covariate combinations. The random forest (RF) algorithm was used to map soil PTEs at a regional scale. Results showed that: (1) When using spatial parameters or spectral parameters as predictor variables and measured Pb content as the dependent variable, the values of the model performance indicator RPIQ (ratio of performance to inter-quartile range) were 2.66 and 2.27, respectively. After incorporating the temporal parameters into the model input, values of RPIQ for the RF model reached 3.55 (using spatial-temporal covariates) and 3.21 (using spectral-temporal covariates), representing performance improvements of 33.46% and 41.41%, respectively. (2) The RF model constructed with the temporal-spatial-spectral covariates achieved satisfactory mapping accuracy (R2 = 0.85; RMSE = 0.80 mg kg-1; RPIQ = 4.09). (3) The soil Pb content in the western and northeastern regions was relatively high, while the remaining areas exhibited lower Pb levels, mainly due to industrial activities. (4) The mapping results of Pb obtained in this study were superior to other mapping methods, such as ordinary kriging, artificial neural networks, and multivariate linear regression methods. The soil PTE mapping technique employed in this study that combined TSS covariates with the RF provided an effective methodological approach for preventing soil pollution, controlling environmental risk, and improving soil management.

Simulation and prediction for the spatial heterogeneity of soil selenium bioavailability at different stratigraphic scales

Stratigraphic lithology strongly influences the spatial heterogeneity of soil available selenium (ASe), however, it is often neglected in regional simulation. Therefore, taking the Jiangjin District, where the soil is richer in selenium (Se), as the research area, the changes of soil ASe at different spatial scales have been simulated by combining Geodetector and three popular models (Multiple linear regression (MLR), Random forest (RF) and BP neural network (BPN)). The results showed that modelling with 'Formation' as the spatial scale could reduce the influence of stratum lithology difference on the spatial heterogeneity of soil ASe and improve the model's prediction accuracy. Compared with the MLR (R2 = 0.52, root mean squares error (RMSE) = 13.217 μg kg-1) and BPN (R2 = 0.55, RMSE = 13.79 μg kg-1), the RF (R2 = 0.67, RMSE = 10.85 μg kg-1) exhibited higher R2 and smaller RMSE, and the simulation effect of soil ASe is the best in the Middle Jurassic Shaximiao Formation (J2s). The outcomes of variable importance analysis revealed that soil total selenium (TSe) and soil organic matter (SOM) were the imperative factors for predicting ASe. The scenario simulation prediction showed that in the next 40 years, due to the combined influence of SOM and pH, the content of ASe in soil developed in the J2s would decrease from 40.8 μg kg-1 to 37.8 μg kg-1, a 7.8 percent drop. The main areas of soil ASe loss were in the western farming areas. The ASe content in dry land and paddy fields decreased by 12.0% and 4.9%, respectively. Therefore, long-term agricultural production activities would lead to soil ASe loss. The present results could provide a new scheme for the simulation and prediction of regional soil ASe, which is helpful for scientific planning, utilization of selenium-rich soil resources, and development of regional agricultural economy.