Spatial distribution and driving factors of molybdenum in soil in the Pearl river delta, China

Summary

As a critical element for ecosystem functioning, the spatial distribution patterns and influencing mechanisms of molybdenum (Mo) remain poorly understood. The Pearl river delta (PRD), a heavily industrialized and densely populated region in China, faces significant ecological risks. This study utilizing soil geochemical survey data to identify the primary drivers of soil Mo distribution on PRD combining random forest and geodetector model. Mo concentrations in surface and deep soils range from 0.15 to 109.76 mg/kg and 0.14–43.68 mg/kg, with median values of 1.16 mg/kg and 1.31 mg/kg, respectively. Key factors influencing surface soil Mo include Si/Al, CoD (coastline distance), and TFe2O3, whereas deep soil Mo is primarily affected by Si/Al, Mn, and TFe2O3. There are varying degrees of synergy between the different driving factors. The study concludes that soil Mo in the PRD predominantly originates from natural sources, with parent material weathering and seawater erosion shaping its spatial distribution.

Graphical abstract

Highlights

• •The GDM and RF revealed potential effects of driving factors affecting Mo in soils
• •The weathering of soil parent materials is the main driving factors for soil Mo
• •The surface soil is affected by the sea-land interaction

Soil science; environmental science; pollution; soil ecotoxicology

Introduction

Molybdenum (Mo) is an essential trace element in biology.^1,2 It participates in nitrogen fixation in plants and metabolic processes in animals through Mo coenzymes.³ Additionally, it plays a crucial role in soil fertility, pollutant removal, and greenhouse gas emissions through nitrification and sulfur oxidation.⁴ Therefore, it is crucial for maintaining the normal functioning of ecosystems.

The spatial distribution of soil Mo reflects complex interactions between natural processes and human activities. Parent materials control Mo concentrations by determining soil mineralogical composition.⁵ For example, the enrichment of Mo in the soil is highly correlated with the distribution of carbonate rocks in karst regions.⁶ Weathering intensity further modulates Mo mobility, with differential enrichment patterns across pedogenic stages.^7,8 Terrain, climate and other environmental factors could also affect the distribution pattern of soil Mo.^9,10,11,12 Additionally, human activities are significant sources of soil Mo. For example, mining activities have caused soil Mo to exceed its background values in the Erdenet and Baganuur mining areas of Mongolia.¹³ Pesticides and fertilizers in agriculture had also significant impacts on the distribution of Mo in soil,^14,15 while industrial emissions contribute via atmospheric deposition and wastewater discharge.^13,16,17

There are many delta regions in the world, such as the Nile delta, the Mississippi delta, and so on.^18,19,20 These regions have unique geographical conditions and are also significant for agricultural cultivation, economic development and urban agglomeration. Diverse environmental problems have arisen in the delta regions due to natural processes and intense human activities.^21,22 Potential toxic elements in the soil of the delta have been paid much attention, and it has been generally believed that human activities exerted an important impact on trace elements in soil in these areas.

The driving factors for soil Mo vary in different regions and scales. Several studies have used different mathematical methods to assess the relative importance of these factors affecting soil Mo, including correlation analysis,⁶ principal component analysis²³ and machine learning.²⁴ Among machine learning methods, random forest (RF) algorithms have been widely applied in such analyses,²⁵ which offer advantages such as anti-overfitting, insensitivity to noise features, and unbiased error rate measurement.²⁴ The geodetector model (GDM) has been demonstrated to be an effective tool for the evaluation of linear and nonlinear effects arising from the interaction between two independent variables. This capability is attributed to its effectiveness in capturing the complex interactions among various influencing elements.²⁶ Therefore, by harnessing the complementary strengths of RF and GDM, we can effectively identify key factors and determine their individual influence using the former, while the latter allows us to deeply explore the interactions among these factors. This integrated approach transcends the analytical limitations of single-model studies, enabling a more comprehensive and systematic elucidation of the driving mechanisms behind complex geographical phenomena.²⁷

The Pearl river delta (PRD) is a highly urbanized and industrialized area in China, known for its high heavy metal concentration and serious ecological risks.^21,23,28,29 The primary sources of elements in soils include natural processes such as chemical weathering and parent material erosion. Furthermore, intensive human activities had the potential to lead to non-point source pollution and localized soil contamination in soils.²³ However, little attention has been paid to the influence of catchment-scale geological units, topography, and climatic conditions on the distribution of trace elements in soils. Moreover, studies investigating the role of these factors in shaping the spatial distribution patterns of trace elements (particularly Mo) and their driving factors remain even scarcer. Therefore, this study chose the PRD as a typical delta to investigate the spatial distribution pattern of Mo in soil and its driving factors by using a large amount of geochemical survey data of surface and deep soil. The objectives of this research were to (1) analyze the spatial patterns of Mo distribution in surface and deep soil, (2) assess the relative contributions of natural factors, human activities, and soil properties to Mo concentration in soil, and (3) use GDM and RF to elucidate the impact of various drivers on Mo distribution patterns.

Results

Descriptive statistics of Mo concentration in soil

Mo concentration in surface soil ranges from 0.15 to 109.76 mg/kg. The average and median values of Mo concentration in surface soil are 1.54 and 1.16 mg/kg respectively. The coefficient of variation (CV) is 153.96%. Mo concentration in deep soil ranges from 0.14 to 43.68 mg/kg. The average and median values of Mo concentration in deep soil are 4.76 and 1.31 mg/kg. The CV is 114.97%. Mo concentrations in both surface and deep soil are highly variable (Table S1).

The median Mo concentration in surface soil of the study area is higher than that in surface soil of China (0.67 mg/kg),²⁸ global crustal abundance (0.6 mg/kg),³⁰ Australian surface soil (0.2 mg/kg),³¹ European agricultural soil (0.42 mg/kg),³² and the USA surface soil (0.78 mg/kg).³³ It is also higher than the mean concentration of surface soil in the Nile delta of Egypt (0.18 mg/kg),²⁰ the soil environmental background value (0.58 mg/kg)²² and the soil geochemical baseline (0.62 mg/kg)²² in the Yangtze river delta of China.

It could be preliminarily indicated that soil parent materials and land use types have significant impacts on soil Mo (Figure 1). Parent materials directly determine the mineral and particle composition of the soil and consequently influence the distribution of soil elements.³⁴ As shown in Figure 1A, the median values of Mo concentration in surface soil with different parent materials are in the following order: granite rocks (1.25 mg/kg) > quaternary sediments (1.20 mg/kg) > carbonate rocks (1.15 mg/kg) > metamorphic rocks (1.09 mg/kg) ≈ clastic rocks (1.06 mg/kg). The median concentrations of Mo in deep soil with different parent materials are as follows: granite rocks (1.52 mg/kg) > carbonate rocks (1.34 mg/kg) > clastic rocks (1.29 mg/kg) ≈ metamorphic rocks (1.25 mg/kg) ≈ quaternary sediments (1.24 mg/kg). In general, Mo concentration in soil developed from granite rocks is significantly higher than those from other soil parent materials. The Mo concentration in surface soil is lower than in corresponding deep soil. We use the ratio of surface soil to deep soil to estimate element enrichment and depletion from different parent materials. From this ratio, it can be seen that the highest median value is 1.02 in soil formed by quaternary sediments, while the median concentrations of other parent materials are similar to the whole study area, and all are less than 1.00 (Figure 1B).

Figure 1.

Box diagrams of soil Mo

(A) Mo concentration in different SP.

(B) The ratio of surface soil Mo to deep soil Mo in different SP.

(C) Mo concentration in different LU.

(D) The ratio of surface soil Mo to deep soil Mo in different LU. SP-soil parent material; LU-land use type; k-migration coefficient; The surface soil Mo concentration in the ratio is the average Mo concentration of four surface soils adjacent to the deep soil point.

Different land-use patterns may significantly influence soil characteristics, thereby altering the spatial distribution of soil elements.³⁵ The median values of Mo in surface soil for different land use types are as follows: building lot (1.32 mg/kg) > paddy field (1.16 mg/kg) ≈ forest land (1.15 mg/kg) > dry land (1.05 mg/kg) ≈ grassland (1.04 mg/kg). Deep soil is slightly different from surface soil. The median values of Mo concentration in deep soil are as follows: grassland (1.54 mg/kg) > forest land (1.41 mg/kg) > dry land (1.30 mg/kg) > paddy field (1.24 mg/kg) ≈ building lot (1.22 mg/kg) (Figure 1C). Based on the ratio of surface soil to deep soil, it can be seen that the highest median concentration is 1.10 in construction land, followed by 1.00 in paddy fields. The ratios of construction land and paddy soil are both higher than the ratio in the whole study area (0.95), indicating that they may have other surface input sources. The ratio of other land use types is less than the whole study area. Grassland has the lowest median value (0.89) and the highest loss (Figure 1D).

Spatial distribution of Mo concentration in soil

The spatial distribution pattern of Mo in surface soil was shown in Figure 2. Mo concentration in surface soil is relatively high in coastal areas, mainly in Shenzhen, Zhuhai, and the southern part of Huizhou. In addition, Mo concentration in surface soil is also relatively high in the northern part of the study area, around Guangzhou and Dongguan (Figure 2A). The spatial distribution of Mo in deep soil is similar to that in surface soil (Figure 2B), which indicates that the main sources of Mo in surface soil and deep soil are consistent. The geochemical maps of other main parameters (such as Si/Al, TFe₂O₃, and Mn) are shown in Figure S1.

Figure 2.

Geochemical map of Mo

(A) Surface soil.

(B) Deep soil.

Identification of dominant driving factors

Random forest model

Figure 3 shows dominant driving factors of Mo concentration in soil suggested by the RF model. Si/Al has the greatest impact on the distribution of Mo (relative importance is 30.38%), followed by coastal distance (CoD, 9.85%), TFe₂O₃ (8.65%), Mn (7.94%), mean annual temperature (MAT, 6.55%), PM_2.5 (6.29%), soil parent material (SP, 5.87%), elevation (EL, 5.31%), and S (5.26%). pH, organic matter (OM), road distance (RoD), river distance (RiD), and land use (LU) have the least impact on Mo distribution (Figure 3A). Compared with surface soil, the impacts of CoD and PM_2.5 on Mo concentration in deep soil are greatly reduced. The importance ranking of other factors of Mo concentration in deep soil is similar to surface soil. In summary, Si/Al, CoD, and TFe₂O₃ are the main controlling factors affecting Mo concentration in surface soil in the PRD (Figure 3B). Si/Al, Mn, and TFe₂O₃ are the main controlling factors affecting Mo concentration in deep soil in the PRD.

Figure 3.

The relationship between Mo concentration and its driving factors

(A) Variable importance score of surface soil.

(B) Variable importance score of deep soil.

(C) Spearman correlation value of surface soil.

(D) Spearman correlation value of deep soil. CoD-coastal distance; MAT-mean annual temperature; SP-soil parent material; EL-elevation; OM-Soil organic matter; RoD-road distance; RiD-river distance; LU-land use type.

Pearson correlation

To ensure data accuracy, soil Mo concentrations exceeding ±3 standard deviations from the mean values were removed as outliers prior to statistical analysis.³⁶ 8,371 surface soil samples and 2,757 deep soil samples were left for correlation analysis. The results show that Mo concentration in surface soil is significantly negatively correlated with Si/Al (r = −0.585, p < 0.01) and CoD (r = −0.207, p < 0.01) and significantly positively correlated with TFe₂O₃ (r = 0.305, p < 0.01), Mn (r = 0.265, p < 0.01), and S (r = 0.224, p < 0.01) (Figure 3C). Mo concentrations in deep soil are significantly negatively correlated with Si/Al (r = −0.602, p < 0.01) and MAT (r = −0.202, p < 0.01) and significantly positively correlated with TFe₂O₃ (r = 0.348, p < 0.01) and EL (r = 0.219, p < 0.01) (Figure 3D).

Among the main influencing factors, Si/Al has a negative effect on Mo in both surface soil and deep soil. TFe₂O₃ has a positive effect on Mo in both surface soil and deep soil. CoD has a negative effect on Mo in surface soil. Mn has a positive effect on Mo in deep soil. Other factors also have positive or negative effects on soil Mo, but the degree of influence is relatively low. It is noteworthy that the importance score of Mn in deep soil is high (15.87%), but the correlation between Mo and Mn in deep soil is low (r = 0.162, p < 0.01). The impact of Mn on Mo in deep soil may be multifaceted and cannot be concluded through simple linear analysis. The importance scores of some factors (such as S in the surface soil, pH in the deep soil, etc.) are low, but their correlation with Mo in soil is high (more than 0.2).

Geodetector modeling

The spatial distribution of soil elements results from complex interactions among multiple interdependent factors. These factors often exhibit synergistic or antagonistic effects rather than isolated influences. To properly evaluate these relationships, we employed GDM’s interaction detector to analyze the combined effects of key driving factors on surface soil Mo concentrations (Figure 4).³⁷

Figure 4.

Interactions among the main driving factors of Mo in soil

(A) Surface soil.

(B) Deep soil.

CoD-coastal distance; MAT-mean annual temperature; SP-soil parent material; EL-elevation; OM-Soil organic matter; RoD-road distance; RiD-river distance; LU-land use type.

Si/Al, as a single driving factor, contributes the most to the spatial distribution of Mo concentrations in surface soil (Figure 4A). The maximum contribution of the interaction also includes Si/Al and CoD [q(Si/Al∩CoD) = 0.392]. The contribution between Si/Al and other factors is above 0.30, far greater than the contribution of other interactions. The findings demonstrate that Si/Al interactions with other variables enhance Mo enrichment in the study area’s surface soil. Additionally, synergistic effects between SP∩TFe₂O₃[q(SP∩TFe₂O₃) = 0.238] and CoD∩MAT [q(CoD∩MAT) = 0.200] combinations substantially influence Mo accumulation patterns. The interaction detector results of driving factors on Mo concentrations in deep soil are very similar to those in surface soil, indicating a good inheritance relationship between the surface soil and the deep soil (Figure 4B). Si/Al also contributes the most to the spatial distribution of deep soil Mo. Unlike in surface soil, the interaction between Mn and pH significantly promotes the accumulation of Mo in deep soil [q (Mn∩pH) = 0.212]. Based on the interaction detector results among Si/Al, CoD, SP, and other factors, it can be concluded that the weathering of parent material, seawater erosion, and subsequent changes in soil physicochemical properties are the dominant driving factors in the study area.

Discussion

Effects of parent material weathering on Mo in soil

It was suggested that the weathering of parent material was the main driving factor of Mo concentrations in soil. Mo concentrations in soil developed from granite rocks are significantly higher than those from other soil parent materials. Mo mainly exists in alkaline feldspar (potassium feldspar series) and biotite among the main minerals in granite.^38,39 Among accessory minerals, the Mo concentrations in Fe-Ti minerals are the highest, especially in titanite, niobite, magnetite, and ilmenite, which can reach tens or even hundreds of times Mo concentrations in the main minerals.⁴⁰

Mo mainly occurs in the +4-valence form in granite rocks, and Mo⁴⁺ replaces major elements with similar ionic radii (mainly Al³⁺, Fe³⁺, Ti⁴⁺, and some Si⁴⁺) in these primary minerals.³⁸ As the degree of weathering increases, primary mica and feldspar tend to dissolve. K, Ca, Na, and Mg are eluviated into the water. Parts of these elements are transformed into clay minerals such as montmorillonite and illite.^6,41 Fe, Mn, and Al accumulate in residual soil and form secondary stable minerals such as goethite and gibbsite.⁴² These secondary minerals formed by weathering can adsorb Mo through various surface complexes such as edge sharing and angle sharing.⁴³ Fe-Ti minerals with high Mo concentration have strong weathering resistance and can be easily preserved in soil.^44,45 Finally, soil with high Mo concentration was formed.

Carbonate rocks are different from other soil parent materials and have a two-stage soil forming model.^46,47 During the carbonate dissolution stage, most of the carbonate components in the bedrock have already been dissolved, resulting in a significant reduction in the volume of the bedrock.⁴⁸ Mo is present in the remaining aluminosilicate and Fe minerals, resulting in relatively less Mo leaching.⁴⁹ The later stage of weathering into soil is similar to the silicate rocks. Secondary clay minerals and Fe minerals re-adsorb dissolved Mo elements.^50,51 Secondary minerals can play a dominant role in soil Mo enrichment in some carbonate rock areas.^6,52 The superposition of the dissolution stage and the soil formation stage results in higher Mo concentration in soil from carbonate rocks.

Quaternary sediments, clastic and metamorphic rocks are mainly composed of non-Mo enriched minerals like quartz, resulting in low Mo concentrations in soil from these parent materials.^5,29,53 The Quaternary sediments in the Pearl river delta were formed by river transport and accumulation. The influence of leaching deposition on it is minimal.⁵⁴ Therefore, there was little difference in Mo concentrations between surface soil and deep soil from Quaternary sediments.

There may be slight differences in Mo concentrations between deep soil and surface soil. The study area is located on the southern edge of China’s subtropical zone. It is hot and rainy throughout the year. The soil has undergone strong weathering and clayization effects.²³ The clayization process is mainly characterized by leaching deposition.⁵⁵ Soil clay particles migrate downwards, increasing the Si/Al ratio of the surface soil and reducing Mo adsorption.⁵⁴ The Mo concentrations in surface soil are lower than that in deep soil. In summary, the distinct material compositions and pedogenic processes inherent to different parent materials fundamentally govern the distribution and interplay of elements in soils. This intrinsic mechanism is clearly illustrated by the relationships between Mo and key soil physicochemical properties such as Si/Al, Fe, and Mn.

Effects of seawater erosion on Mo in soil

Under surface oxidation conditions, MoO₄²⁻ has a high solubility and is easy to transport in solution or migrate in colloidal form.⁵⁶ The surface soil was affected by weathering leaching and surface runoff, and MoO₄²⁻ gradually migrates toward the ocean or downwards.⁵⁷ This leads to a lower Mo concentration in surface soil than in deep soil. Seawater can cause changes in soil pH, salinity, and redox conditions in sea-land interaction areas.^58,59 The results in the previous context indicate that pH has little effect on Mo in soil, so pH is not the way that seawater affects soil Mo in the study area. Salinity of soil off the coastline reflects the oceanic source of elements. Under the influence of waves, tides and other factors, elements in seawater can exchange with nearby sediments or soil, allowing them to remain in the soil.⁵⁶ Na is a typical oceanic source element. During seawater intrusion, Na undergoes cation exchange. Na in seawater can replace Ca and Mg in the soil and then remain in the soil.⁵⁸ Assuming that the ocean is a major source of soil Mo in the study area, Na is a representative element of seawater. The spatial distribution of Mo and Na of soil in coastal areas should be similar. We have plotted boxplots of Mo concentration of surface soil at different coastal distances (Figure 5). it can be seen that the Mo concentration in surface soil gradually decreases and stabilizes around 60 km as the coastal distance increases, indicating that the influence distance of seawater on soil Mo is about 60 km. However, Na and Mo in the Shenzhen area are not similar on the geochemical map within 60 km of the coast (Figure 6). Mo forms MoS₂ with S at +4 valence under reducing conditions, which is fixed by Fe, Mn, and OM in soil.⁶⁰ Mo exists in the form of MoO₄²⁻ under oxidizing condition, which is easily adsorbed by secondary minerals or OM in soil. The redox condition in the sea-land interaction region is not static, and there will be alternating changes. These alternating changes cause repeated dissolution and co-precipitation of Fe and Mn oxides, forming Fe-Mn nodules. Mo is enriched in Fe-Mn nodules. Therefore, seawater affects the redox condition of soil and thus affects the occurrence form and concentration of Mo in soil (especially in surface soil).

Figure 5.

Box diagrams in surface soil from different CoD

(A) Mo.

(B) Na₂O (CoD-coastal distance).

Figure 6.

Geochemical map of Na2O in surface soil

Conclusion

This study analyzed the characteristics and spatial distribution of soil Mo in the Pearl river delta region, and used GDM and RF models to identify the main driving factors affecting soil Mo. The spatial distribution of Mo in surface soil and in deep soil is similar. Soil with rich Mo is mainly distributed in the southern coastal part, northern part, and central part of the study area. The surface soil Mo concentration is lower than the corresponding deep soil. The soil Mo is mainly influenced by natural factors (parent materials, pedogenic process, and sea-land interaction). The composition and stability of primary minerals that make up the parent materials are the most important driving factors. Secondary clay minerals and Fe-Mn minerals in soil could adsorb and fix Mo. Sea-land interaction could exert influence on the occurrence form Mo in soil by changing the redox condition of soil. Si/Al, TFe₂O₃, Mn, and CoD were identified as the main influencing factors of Mo in soil. There are varying degrees of synergy between the different driving factors. These factors interact with each other and finally form the overall distribution pattern of soil Mo in the Pearl river delta.

Limitations of the study

Although progress has been made in research on the driving factors of soil Mo, the following limitations remain in this study. First, this study primarily focuses on the horizontal distribution of Mo in surface soil (0–20 cm) and deep soil (150–180 cm) at a local scale, but does not investigate the vertical migration patterns of Mo in soil profiles or its micro-scale occurrence mechanisms. Second, although the RF model exhibits strong predictive performance, its “black box” nature limits the in-depth interpretation of the geochemical mechanisms driving soil Mo. Moreover, this method has inherent limitations in capturing the spatial heterogeneity of geochemical parameters, which may lead to underestimation of the contributions of certain categorical variables. Third, this study was conducted only in the PRD region; therefore, the generalizability of the findings needs to be validated in other delta regions through more extensive comparative studies.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Qingye Hou (qingyehou@cugb.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All data reported in this study will be shared by the lead contact upon request.
• •This article does not report any original code.
• •Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (41773019) and the National Key Research and Development Program of China (2017YFD0800304). The author sincerely thanks the editors and reviewers for putting in considerable time and effort for this study.

Author contributions

Conceptualization, Z.W. and Q.H.; methodology, Z.W., Y.Y., C.Z., and X.M.; formal analysis, Z.W. and Y.Y.; writing – original draft, Z.W.; writing – review and editing, Z.W., Q.H., Y.Y., Z.Y., T.Y., and D.L.; visualization, Z.W., C.Z., and X.M.; project administration, Q.H.; investigation, Y.Y. and J.W.; supervision, Z.Y., T.Y., and D.L.; data curation, X.L.; resources, J.W.

Declaration of interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this study.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
Microsoft Excel 2019	Microsoft Corporation	https://www.microsoft.com/
IBM SPSS Statistics 22.0	International Business Machines Corporation	https://www.ibm.com/
ArcGIS 10.2	Environmental Systems Research Institute Inc	https://www.esri.com/
The Random Forest Modeling	Python Software Foundation	https://www.python.org/
Geodetector modeling	Chinese Academy of Sciences	http://geodetector.cn/
CorelDRAW 2024	Corel Corporation	https://www.coreldraw.com/
OriginPro 2024	OriginLab	https://www.originlab.com/

Method details

Study area

The PRD is located in the central part of Guangdong Province. It includes big cities such as Guangzhou, Shenzhen, Zhuhai, etc. (Figure 7). The topography of the region is predominantly characterized by the extensive alluvial plain of the Pearl River Delta, which is circumscribed by low-lying mountains and hills to the west, north, and east. The region is traversed by several major rivers, including the Dongjiang, Xijiang, Beijiang, and Tanjiang Rivers. These rivers ultimately empty into the South China Sea.⁶¹ The PRD has experienced rapid urbanization and industrialization in the past few decades. The road network in the region is well-developed, and the air particulate matter is mainly fine particulate matter (PM2.5).⁶² The parent material of the study area is mainly composed of granite rocks, quaternary sediments, and clastic rocks, accompanied by a small amount of carbonate and metamorphic rocks. The predominant land use types consist of paddy fields, dry land, and forest land, accompanied by a small amount of building lot and grassland^21,23,28 (Figure 8).

Figure 7.

Map of the study area

Figure 8.

Spatial distribution of the potential driving factors in the study area

Sampling and chemical analysis

The National Multi-Purpose Regional Geochemical Survey (NMPRGS), launched by the China Geological Survey (CGS) in 1999.⁶³ In this project, a regular grid was employed to ensure spatial representativeness.⁶⁴ Considering the impact of human activities and the depth of the parent material, this study selected 0–20 cm and 150–180 cm as the sampling depths for surface soil and deep soil, respectively.^5,23 Surface soil samples (0–20 cm depth) were collected at a high sampling density of 1 sample/km,² totalling 10,907 samples across the study area. Additionally, deep soil samples (150–180 cm depth) were obtained at a lower sampling resolution (1 sample/4 km²), resulting in 2,798 deep soil samples for analysis in the study area (Figure S2). Surface soils were obtained through the collection of between 3 and 5 subsamples, each collected from within a 50 m radius of the designated sampling site. It is important to note that no sample was collected within 100 m of a major road or railway, nor from areas that were visibly contaminated or at the time of fertilising. In urban environments, sampling sites were strategically selected from established locations such as historical gardens, educational institutions, and residential zones, while deliberately excluding recently disturbed or artificially deposited soils. During collection, visible organic debris (e.g., plant residues) and coarse rock fragments were carefully removed by hand. The collected samples underwent natural air-drying before mechanical processing, which involved disaggregation using a wooden mallet followed by sieving through a 2-mm mesh nylon screen to obtain homogenized particles. These prepared samples were subsequently homogenized to create composite specimens for chemical characterization.⁶³

Soil Mo was analyzed by inductively coupled plasma-mass spectrometry (ICP-MS). Briefly, 0.1 g of powdered soil sample was mixed with 2 mL HClO₄, 5 mL HF, and 10 mL HNO₃ in a crucible, covered with a lid, and heated on an electric hot plate (150°C for 1 h). The crucible lid was then removed and the temperature increased to 240°C for decomposition. After decomposition was completed, the solution was cooled and transferred to a polyethylene test tube, added ultrapure water to measure the solution to 50 mL, and then analyzed by ICP-MS (Model X-SERIES II; Thermo Fisher Scientific., USA). For SiO₂, Al₂O₃, Mn, S, and TFe₂O₃ (total iron oxides), the soil was pressed into a pellet using a polythene ring and hydraulic press, followed by X-ray fluorescence spectrometry (PW2440, Philips Co., Dutch). Soil organic matter (OM) was determined by volumetric analysis (VOL), while pH was measured using a potentiometric method.

To ensure the accuracy and precision of the analytical results, four Chinese National Standard Reference Samples were included in each batch of sample analysis. A total of 1,723 such reference samples were incorporated for quality control purposes. The analytical accuracy ranged from −11.20% to 13.74%, and the precision ranged from 1.67% to 7.92%. For further details on the sampling and analytical procedures, refer to Dou et al.²¹ and Wang et al.²³

Data sources

The geochemical datasets for both surface and deep soils across the PRD region were obtained from the Guangdong Geological Bureau of China and the national geological archives of China. The types of parent material of soil in the study area were modified according to Dou et al.²¹ Land use type data were obtained from the Resource and Environmental Science Data Center, Chinese Academy of Sciences (RESDC, http://www.resdc.cn). Road distance (RoD), river distance (RiD), and coastline distance (CoD) were computed through the ArcGIS 10.2 spatial analysis function to calculate the main roads (including national highways and expressways), rivers, and coastline in the study area. Climatic parameters, specifically mean annual temperature (MAT), were obtained from the Resource and Environmental Science Data Center (RESDC) of the Chinese Academy of Sciences (http://www.resdc.cn). Topographic elevation (EL) information was derived from the digital elevation model (DEM) acquired from the China Geospatial Data Cloud platform (http://www.gscloud.cn). The data on fine particulate matter (PM2.5) was sourced from the Guangdong Department of Ecology and Environment (GDEE, http://gdee.gd.gov.cn) (Figure 8). Interpolate these factors to various sampling points using spatial analysis tools in ArcGIS 10.2.

Random forest modeling

Random forest (RF) is an ensemble learning algorithm introduced by Breiman.⁶⁵ This machine learning approach integrates multiple randomly generated classification and regression trees to enhance predictive performance and model robustness. The methodology operates by constructing numerous decision trees through bootstrap sampling of training data and random feature selection during node splitting, effectively reducing overfitting risks while maintaining model stability. The RF training process involves four key steps: (1) Bootstrap sampling to select data subsets for individual tree construction; (2) Random feature selection at each node split to ensure tree diversity; (3) Iterative generation of multiple decision trees; and (4) Aggregation of predictions through majority voting (for classification) or averaging (for regression) to produce the final output. Compared to conventional modeling techniques, RF offers distinct advantages, including high predictive accuracy, efficient handling of high-dimensional data, robust performance across diverse datasets, and inherent resistance to overfitting.^23,66 Its ability to manage complex feature spaces without extensive preprocessing further contributes to its widespread adoption in machine learning applications.

RF algorithm has demonstrated strong performance in evaluating the spatial variability of soil constituents and identifying key driving factors.^25,67,68 In the current research, an RF-based approach was employed to analyze the relative importance of 14 environmental variables in governing Mo distribution patterns within soil systems. The significance indices of each factor were normalized to sum up to 100%, allowing for a comprehensive evaluation of their contributions.

Geodetector modeling

GDM was used to measure spatially stratified heterogeneity and to identify the governing force of a responding variable. This method operates on the fundamental principle that significant causal relationships between variables manifest through comparable spatial distribution patterns.²⁶ The GDM offers distinct advantages for examining spatial covariation among environmental factors. Our analysis employed two core analytical components: the factor detector and the interaction detector. The factor detector quantifies explanatory power by computing the variance contribution of predictor X to response variable Y. The GDM identifies the spatial variability of Y and how X accounts for the spatial configurations of Y. The q statistic, which spans from 0 to 1, indicates the degree of spatially stratified heterogeneity observed in Y^5,26:

$$\text{SSW}=\sum _{i=1}^n{\mathrm{N}}_i{\mathrm{\sigma }}_i^{2}i=1,2,3,\dots ,n$$ (Equation 1)

$$\text{SST}={\mathrm{\sigma }}_i^2$$ (Equation 2)

$$q=1-\frac{SSW}{SST}$$ (Equation 3)

In the context of this study, the independent variable X is divided into strata labeled as i = 1, 2, …, n. The study region encompasses N units, with each stratum h containing N_i units. The variance within each stratum is represented by σ_i², while the overall population variance is denoted by σ². The within-strata sum of squares is indicated by SSW, and the total sum of squares is denoted by SST. The parameter q quantifies the level of explanation that the independent variable X provides for the dependent variable Y⁵.

The interaction detector stands as a key strength of GDM compared to other statistical analysis techniques, as it assesses the ability of interactions between driving factors X1 and X2 to either boost or reduce their explanatory power for the dependent variable Y. This evaluation includes determining whether the effects of these factors on Y are mutually independent. The interaction analysis may yield five characteristic patterns: (1) nonlinear synergy when q(X1∩X2) exceeds the sum of individual effects; (2) independent effects when the combined q-value equals the sum; (3) bidirectional enhancement where the interaction term falls between the maximum individual effect and their sum; (4) mutual weakening when the combined effect lies between minimum and maximum individual effects; or (5) nonlinear suppression when the interaction term is lower than either individual effect. The open-source GDM software package is publicly available at http://www.geodetector.org.^26,69

Quantification and statistical analysis

The statistical preprocessing was conducted using Microsoft Excel 2019 (Microsoft Corporation, USA) and IBM SPSS Statistics 22.0 (International Business Machines Corporation, USA). Spatial analysis of Mo concentrations and associated variables was performed in ArcGIS 10.2 (Environmental Systems Research Institute Inc., USA), while the random forest modeling was executed using Python 3.7 (Python Software Foundation, Wilmington, DE, USA). Continuous variables were discretized through the equal interval classification approach.⁵ GDM was used for interaction analysis. All figure representations were generated using CorelDRAW 2024 (Corel Corporation, Canada) and OriginPro 2024 (OriginLab, USA).

Published: July 8, 2026