Abstract
This study addresses the specific question of how the remediation of fluoride-contaminated farmland can quantify its impact on farmers’ income structure and its transmission to rural economic revitalization. To solve this problem, firstly, multi-source data were integrated to construct governance characteristic variables; then, the Light Gradient Boosting Machine (LightGBM) model was used to predict farmers’ multidimensional income, and SHapley Additive explanations (SHAP) values were employed to analyze the contribution and interaction effects of key factors. Finally, path analysis and Granger causality tests were used to verify the micro-to-macro transmission mechanisms. Experimental results shows that the governance intensity and non-agricultural skills are key driving factors, and the synergy between governance intensity and non-agricultural skills yields a substantial positive contribution to income structure optimization. Mediation analysis shows that the indirect path through income restructuring accounts for the majority of the total policy effect on rural economic revitalization. This study confirms that the hybrid framework integrating machine learning interpretability and econometric causal inference can effectively reveal the complex transmission mechanism through which environmental governance influences household-level responses and subsequently contributes to macro-level rural development, providing quantitative evidence and methodological support for the coordinated promotion of ecological restoration and rural revitalization.
Introduction
Fluoride contamination of agricultural soils arises from both natural geochemical processes and anthropogenic activities (). In regions with fluoride-rich parent rocks or prolonged irrigation with high-fluoride groundwater, soil fluoride concentrations can accumulate to phytotoxic levels, commonly ranging from 200 to over 1,000 mg/kg, far exceeding background levels of uncontaminated soils (typically <200 mg/kg) (; ; ). Industrial emissions, phosphate fertilizer application, and aluminum smelting further exacerbate localized contamination. To mitigate these risks, various remediation strategies have been explored, including chemical stabilization using amendments such as lime or biochar, soil washing with extracting agents, and phytoremediation using hyperaccumulator plants (; ). The effectiveness of these methods depends on site-specific factors such as soil pH, organic matter content, and the speciation of fluoride in the soil matrix (). Given the complex interplay between environmental conditions and remediation efficiency, the economic implications of such interventions on affected farming communities remain poorly quantified.
The remediation of fluoride-contaminated farmland is an important task for safeguarding food security and advancing the development of an ecological civilization in China. Under the macro-background of the rural revitalization strategy, such environmental remediation projects are not only ecological restoration actions, but also have the economic and social expectation of activating the endogenous development momentum of rural areas. However, traditional policy evaluations mostly focus on the environmental benefits or macroeconomic growth, and lack systematic and detailed empirical analysis on the key link of ‘how to accurately quantify and reshape the micro-income structure of farmers through remediation projects, and then transmit the changes in micro-behavior to the macro-revitalization of rural areas’ (). This “black box” of the “micro-macro” transmission mechanism makes it difficult to scientifically evaluate the comprehensive economic and social benefits of remediation policies, and also restricts the optimization design of subsequent differentiated and precise policy tools. Therefore, revealing the intrinsic mechanism of farmland remediation affecting farmers’ income structure and its driving path on rural economy has important theoretical and practical significance for promoting the coordinated development of ecological governance and rural revitalization ().
Existing studies on the economic impacts of environmental governance have mainly focused on macroeconomic indicators, such as regional GDP and total agricultural output, or on the average effects of governance interventions on farmers’ agricultural income. constructed a tailings dam failure scenario for the Jin Ding mining area in Southwest China, applied a geological environmental risk assessment model to assess the exposure, and used disability-adjusted life years to assess its health impact in order to analyze the economic feasibility of tailings dam risk management. Assess the economic and environmental benefits of industrial by-products replacing consumable resources, and used fly ash for stabilizing sludge and improving soil as an example, pointing out that reasonable site selection can maximize benefits, otherwise it will increase costs (). proposed a techno-economic evaluation approach for urban solid waste management based on the social cost-benefit analysis, incorporating external impacts, such as social and environmental factors into decision-making.
The global analysis showed that the cost of blue ecosystem restoration is optimum, its ecosystem service benefits generally significantly exceed the costs, making it economically feasible (). The typical watersheds in the Qinghai-Tibet Plateau as an example to simulate four ecological restoration schemes based on the land use change and assessed from the perspective of ecosystem services, providing a basis for ecological policy formulation (). The dynamic spatial model based on panel data found that water pollution significantly and more strongly inhibits agricultural economic growth in the long-term, and proposed that high-quality agricultural development and rural revitalization should be promoted through pollution reduction and greening (). proposed dynamic model based on the panel data of multiple economies found that income growth and economic integration drive agricultural emissions in the short term, while income growth, agricultural added-values and energy consumption are the main driving factors in the long-term. The dynamic asymmetric panel model to study the impact of Asian climate change on agricultural productivity and found that there is a long-term relationship between the two, and only CO2 emissions have a positive effect in the short-term but turn to negative effect during the long-duration ().
developed an improved cost-benefit analysis method to evaluate the benefits of different ecological restoration applications in the Three-North Shelterbelt Project, providing guidance for future ecological restoration. The cost-benefit analysis and value realization path of ecological restoration projects in China, analyzed the role and challenges of benefit transfer method in policy formulation, and explored potential ways of ecological restoration (). However, existing studies often treat environmental governance and income structure changes as relatively independent issues, or simply use linear models for simplification. They fail to fully reveal how the external shock of farmland consolidation drives the nonlinear reorganization of income sources and composition through complex interactions with the heterogeneous characteristics of farmers, and even less explore the specific path and intensity of this micro-restructuring process to overall rural economic revitalization.
However, the existing literature remains limited in systematically integrating the full logical chain linking environmental intervention, micro-level household responses, and macro-level development outcomes, and in adequately capturing the nonlinear and heterogeneous mechanisms involved. To address these limitations, recent studies have begun to introduce more advanced analytical approaches, particularly machine-learning and artificial-intelligence-based methods, to improve prediction, mechanism identification, and policy evaluation. The application of deep learning modeling tools, artificial neural networks (ANN), in multi-objective modeling, using to simulate the removal effect of arsenic and fluoride (F) in the electrocoagulation process, based on parameters, such as current density, pH value, time frequency, and initial concentration (). An innovative approach to improve the accuracy and efficiency of quantitative analysis of aluminum and F ions in aqueous solutions based on fluorescence detection, in order to meet the requirement of environmental monitoring and risk assessment (). A hybrid artificial intelligence machine learning framework for high-throughput screening of metal-organic frameworks as radon adsorbents, aiming to effectively separate radon from indoor air and provide guidance for its strategic design (). However, most of these applications remain at the level of prediction or feature importance ranking.
Therefore, this study aims to construct a hybrid analytical framework integrating machine learning and econometrics to systematically reveal the impact mechanism and transmission path of F-contaminated farmland remediation on farmers’ income structure and rural economic revitalization. Specifically, the study first integrates multi-source data, uses the LightGBM model to jointly predict the impact of remediation on farmers’ multidimensional income, and uses SHAP values to analyze the contribution and interaction effects of each driving factor to identify the key mechanisms and heterogeneity of micro-level impacts. Furthermore, by constructing a mediation effect model of the pathway through which remediation policies affect income-structure optimization and subsequently influence rural economic indicators and Granger causality tests, the study empirically verifies and quantifies the transmission path and effect of micro-level income structure adjustment to macro-level rural economic development. Through these steps, this study not only aims to accurately quantify the microeconomic effects of remediation policies but also strives to open the ‘process black box’ of its role in driving comprehensive rural revitalization, providing a scientific decision-making basis for improving the comprehensive economic and social benefits of environmental governance policies.
Methodology
Analytical framework and causal hierarchy
This study defines the inputs and outputs of each analytical stage. At the micro level, remediation characteristics, household endowments, village conditions, and baseline income indicators are used as model inputs, while the income-structure optimization index and related income components are treated as outputs. At the macro level, township-level income-structure indicators are used to examine their links with agricultural output growth, non-agricultural employment growth, and the rural economic revitalization index. This design connects policy intervention, household response, and macroeconomic outcomes within a unified analytical framework. The first stage: causal effect estimation (baseline identification). The second stage: mechanism exploration and attribution (characteristic importance and interactions). The third stage: macro-level transmission verification (path confirmation). This structured approach ensures that each method addresses a specific, well-defined question within a coherent causal hierarchy, from “Is the policy effective?” to “How does it work, and to whom?” and then to “Does micro-adjustment promote macro-level development?”.
The combined use of LightGBM, SHAP, Granger causality testing, and path analysis is necessary because the study involves prediction, interpretation, and transmission verification. LightGBM captures nonlinear and heterogeneous income responses, SHAP explains the contribution of each factor, and the econometric models verify whether micro-level changes are associated with macro-level rural development. This integrated framework provides more comprehensive evidence than conventional linear analysis alone.
Multidimensional data collection and governance feature quantification
Initially, regarding data collection, this study constructed a multidimensional panel dataset covering environmental background conditions, policy interventions, micro-level household characteristics, and macro-level development outcomes. The data primarily adapted from the different sources, i.e., initially, administrative and project archives data, including soil sampling data from the survey of F-contaminated farmland in the study area, a list of specific remediation projects (including implementation scope, technical routes, funding and construction periods), and basic information, such as the confirmed farmland area and collective economy of all sample villages, obtained from local ecological and environmental bureaus, and agricultural and rural affairs bureaus. Second, farmer follow-up survey data, conducting a follow-up survey of 200 households located in the typical remediation project area and adjacent control areas.
The questionnaire systematically covered family population structure and human capital, land assets and productive fixed assets, agricultural production inputs and outputs, non-agricultural employment and business activities, and details of various incomes and expenditures. Third, macro-socioeconomic statistics, collecting time-series data on total agricultural output, secondary and tertiary industry output, resident population, and non-agricultural employment in the sample townships from county-level statistical yearbooks and township annual reports. Fourth, spatial and remote sensing-assisted data, combined with high-resolution satellite imagery, were used to calculate the accessibility of sample villages and their distance from towns/markets.
In terms of quantifying governance characteristics, the core lies in transforming the key intervention of ‘governance policy’ into a set of observable, measurable, and analyzable characteristic variables. Based on cross-validation of project archives and farmer surveys, this study constructed three dimensions of governance characteristic variables, i.e., governance intensity, applying ‘project governance funding per unit of cultivated land area’ as the core metric, and standardizing and synthesizing it with the physical engineering quantities of the governance project to form a continuous variable; technology adoption and participation depth, based on actual farmer behavior; and intervention timing and continuity, generating the variable ‘number of years of governance implementation’ based on the project initiation year. To systematically and intuitively demonstrate the complete technical process from multi-source data collection to the final construction of the analysis dataset, this study drew a data integration technology roadmap shown in Figure 1.
FIGURE 1
Construction of key variable feature engineering
The construction of explained variable (farmers’ income structure) is the foundation of this section. Specifically, based on farmers’ annual income and expenditure details, we calculated net income from agricultural operations, income from non-agricultural work, property income, and transfer income. To capture the changes in structure, we further constructed three key indicators, such as the proportion of income sources, including the proportion of agricultural income and the proportion of non-agricultural income , to measure the degree of income diversification; the income stability index , calculated based on the inverse of the coefficient of variation of agricultural income over three consecutive years; and the comprehensive index for income structure optimization , which combines the aforementioned proportions and stability indicators through principal component analysis (PCA) to form continuous variable reflecting the overall optimization of the income structure. To enhance the intuitiveness and reproducibility of the indicator construction process, it presents the construction process of the income structure optimization index in a structured manner (Figure 2).
FIGURE 2
Let the standardized index vector be , and its covariance matrix be . Solve the characteristic Equation 1:
Where, is the -th eigenvalue (arranged in descending order, ≥ ≥), and is the corresponding unit eigenvector. The extracted first principal component (PC1), i.e., as defined in this study, is a linear combination of the original variables, and its weight is given by the eigenvector corresponding to the largest eigenvalue , as shown in Equation 2.
The index contains the largest portion of the variation information in the original dataset, thus enabling an effective and objective comprehensive measurement of the optimization state of the income structure and serving as a core mediating variable in subsequent path analysis (). For the constructed index, the first principal component (PC1) exhibited an eigenvalue of 2.14, explaining 71.3% of the total variance in the three income-structure indicators. The Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy was 0.68, exceeding the recommended threshold of 0.6, indicating acceptable factorability.
To achieve a refined measurement of the multi-dimensional policy intervention of rectification, we abandoned simple binary treatment variables and constructed three levels of features, i.e., intervention intensity, technological penetration, and process participation. These features together constitute a multi-level policy characterization from “whether it exists” to “how strong it is” and “how deep it is” Furthermore, to effectively control confounding effects and test heterogeneity, systematically constructed control and moderating variables covering farmers, households, and villages. All continuous variables were standardized before being included in the model. Table 1 summarizes the core variables associated in this study, their specific measurement methods, descriptive statistics, and preliminary theoretical expectations of their impact.
TABLE 1
| Variable category | Variable name | Measurement (Unit) | Sample mean (SD) | Expected direction |
|---|---|---|---|---|
| Dependent variable | Income structure optimization index | PCA composite index (0–1) | 0.53 (0.21) | - |
| Core explanatory variables (policy intervention) | Governance intensity index | Remediation investment per unit area, standardized (0–1) | 0.68 (0.22) | + |
| Technology adoption type | Categorical: 0 = non-adoption (41%), 1 = conventional (35%), 2 = biological/chemical (24%) | - | + | |
| Participation level | Average household labor days invested per year (days) | 12.5 (10.3) | + | |
| Key control variables (household endowment) | Land scale | Total contracted cropland area, mu/ha conversion noted as 1 mu = 0.0667 ha, 7.2 mu, approximately 0.48 ha, SD = 5.8 mu, approximately 0.39 ha | 7.2 (5.8) | +/− |
| Human capital | Average years of schooling of laborers (years) | 8.6 (2.5) | + | |
| Non-farm skills | Binary: 1 = possessing vocational certificate(s) (58%), 0 = none (42%) | - | + | |
| Fixed asset value | Value of productive fixed assets (10k CNY, 2021 price) | 3.4 (4.1) | + |
Summary of core feature variables, measurement approaches, descriptive statistics, and expected effects.
Training of a joint prediction model for farmers’ income structure based on LightGBM
The LightGBM model uses policy-intervention variables, household characteristics, village contextual factors, and baseline income indicators as inputs, with the income-structure optimization index as the main output. LightGBM was adopted because the effects of farmland remediation on farmers’ income may be nonlinear and heterogeneous, and may involve interactions among policy intensity, household capacity, and local conditions. Compared with conventional linear models, it can better capture threshold effects and complex feature interactions. This model learns the optimal tree set by minimizing the following regularization objective function , as shown in Equation 3.
Where, n is the sample size, is the loss function that measures the difference between the predicted value and the true value , is the total number of trees, and is the regularization term used to control the complexity of the -th tree , thereby effectively preventing overfitting.
All variables derived from the preceding feature-construction procedure were used as input features in the LightGBM model. In terms of model output settings, in order to directly reflect the impact of policies on income structure, we adopted a multi-output regression setting. In terms of technical implementation, we randomly divided the integrated panel dataset into a training set (70%), validation set (15%) and a test set (15%) according to the farmer ID. The performance of the final model is evaluated by the root mean square error (RMSE) and the coefficient of determination (R2) on the test set ().
Given the moderate sample size (n = 200 households) relative to the dimensionality of the feature space, careful attention was paid to model regularization and validation to ensure the reliability of the SHAP interpretations. The LightGBM model incorporates built-in L1 and L2 regularization terms (as specified in Equation 3) to penalize model complexity. Furthermore, we implemented early stopping based on the validation set performance (monitoring RMSE with a patience of 50 rounds) to prevent overfitting to the training data. To ensure the stability of the feature importance rankings derived from SHAP, we performed a 5-fold cross-validation procedure and report the average absolute SHAP values across folds. The consistency of the feature ranking across resampled data partitions confirms that the identified key drivers (governance intensity, non-agricultural skills) are robust to sampling variation and not artifacts of model over-parameterization.
Attribution analysis of income influence factors based on SHAP value
SHAP analysis was used to improve the interpretability of the LightGBM results. Using the trained model and its input features, SHAP identifies the relative importance, contribution direction, and interaction effects of key variables, thereby clarifying how remediation-related and household-level factors influence the predicted income-structure optimization index. The specific analysis process includes two progressive levels. First, we conduct global importance analysis and contribution direction identification. It calculates the average absolute SHAP value of each feature on all samples, which serves as an objective indicator to measure the global influence of the feature, and rank the features accordingly. For the ith sample, the SHAP value of its feature is calculated according to the Shapley value of cooperative game theory, and its expression is Equation 4.
Where, is the total number of features, is a subset of features, and represents the expected prediction of the model when only features in subset are used. The value fairly distributes the marginal contribution of feature to the difference between the model prediction and the baseline prediction , when it is added to all possible feature alliances, satisfying , thus achieving an exact additive decomposition of the prediction results. More importantly, by plotting the distribution of the SHAP values of each feature for all sample points, we can visualize the direction of influence of each feature: if most of the SHAP values of a feature are positive, it indicates that it usually has a positive contribution to income improvement; otherwise, it is negative. This allows us to determine whether, for example, the enhancement of “health risk perception” is generally associated with the reduction in income prediction .
Secondly, we delve into the nonlinear patterns and key interaction effects of feature influence. As shown in Figure 3, it uses two deep analysis techniques as SHAP dependency and interaction effect analysis. Figure 3a and b shows the schematic diagram of SHAP dependency and SHAP interaction effect analysis. By plotting SHAP dependency diagrams for specific features, we can observe how the SHAP value (i.e., marginal contribution) responds when the feature value changes. For example, analyzing the dependency diagram of “governance intensity” may reveal a significant threshold effect or diminishing marginal returns. To identify interactions, we further examine how the SHAP value of one feature, such as “non-agricultural skills” changes with the value of another feature (such as governance intensity). When the SHAP value of “non-agricultural skills” is significantly larger when “governance intensity” is higher. It indicates a positive synergistic interaction between the two, meaning that the policy effect is amplified among highly skilled farmers. Identifying this interaction effect is key to understanding the heterogeneity of policy impact.
FIGURE 3
Path analysis model verification of the effects of farmland improvement on rural economy
Township-level panel data were further used to test whether micro-level income-structure optimization is linked to macro-level rural development. In the Granger causality and path analysis models, income-structure optimization serves as the key mediating variable, while agricultural output growth, non-agricultural employment growth, and the rural economic revitalization index are treated as macro-level outcomes. Table 2 presents the descriptive statistics and correlation coefficient matrix of the core variables at the township level. To demonstrate the causal time-series nature of the transmission direction, we employed a Panel Vector Autoregression (VAR) model and Granger causality tests. At the township level, variables, such as the “average income structure optimization index for farmers,” “agricultural output growth rate,” and “non-agricultural employment growth rate” were incorporated into a dynamic system. Through Granger causality tests, we can determine whether the lagged term of “income structure optimization” statistically significantly contributes to predicting the current “agricultural output growth” and “non-agricultural employment growth,” and vice versa. If the test results support that the former is a Granger cause of the latter, and the inverse relationship does not hold, then preliminary causal evidence for the transmission direction of ‘micro-driven macro’ from the perspective of time series predictive power is provided.
TABLE 2
| Variable | Mean | Std. Dev | −1 | −2 | −3 | −4 |
|---|---|---|---|---|---|---|
| Income structure optimization index | 0.53 | 0.21 | 1 | | | |
| Agricultural output growth rate (%) | 5.2 | 2.1 | 0.42*** | 1 | | |
| Non-agricultural employment growth rate (%) | 3.8 | 1.6 | 0.38*** | 0.25 | 1 | |
| Rural economic revitalization index | 0 | 1 | 0.61*** | 0.67*** | 0.58*** | 1 |
Description of statistical analysis and correlation matrix.
The Rural Economic Revitalization Index is a comprehensive indicator derived from the agricultural output growth rate, non-agricultural employment growth rate, and per capita disposable income growth rate using Principal Component Analysis (PCA), and has been standardized to a mean of 0 and a standard deviation of 1. This table lists the Pearson correlation coefficients. *, **, and *** indicate significance at the 10, 5, and 1% significance levels, respectively.
Secondly, to more accurately quantify the magnitude of the effects at each stage of the transmission path, especially to decompose the direct and indirect effects, we further constructed a structural equation model (SEM). In this model, it sets intensity of governance policies (measured by the average governance intensity within townships) as an exogenous variable, index of optimization of farmers’ income structure as a mediating variable, and level of rural economic revitalization (synthesized from indicators, such as agricultural output and non-agricultural employment through PCA) as an endogenous outcome variable. The model simultaneously estimates the coefficients of two paths, i.e., one is the direct path from governance policies to rural economy, and the other is the indirect path mediated by income structure optimization. By calculating the confidence interval of the indirect effect through repeated sampling using the bootstrap method, we can determine whether the mediating effect is significant. Finally, the model outputs the standardized coefficients of the direct, indirect, and total effects, thereby accurately quantifying the proportion of micro-income structure adjustment in the macro-level driving effect of policies.
Results and Discussion
Experimental setup
The experimental design includes three links in total, such as first, based on the construction of income structure optimization index and inter-group comparison analysis, the net impact of the rectification policy on the micro-economic structure of farmers is analyzed. Second, based on the trained prediction model and SHAP attribution method, the driving factors of income change and their complex nonlinear interactions are analyzed deeply. Finally, through constructing panel econometric model based on township level, the empirical test is conducted to test whether and how the micro-level income structure optimization transfers to and promotes the macro-rural economic development. All above analysis are based on the unified sample and variable system, so the logical consistency of whole paper and conclusions from mechanism identification to causal verification are maintained well.
Experimental analysis
Income structure optimization index calculation experiment
This study constructed a complete evaluation experiment with three levels of effect confirmation, mechanism analysis, and transmission at the macro level. First, the net effect of the policy was estimated using propensity score matching-differences-differences method. Second, attribution analysis based on SHAP values was used to explore the key influencing factors and their heterogeneous effects. Finally, a multi-level mediation effect model was used to test the transmission path from micro-level income optimization to macro-level rural development, and to systematically evaluate the overall economic benefits of the remediation policy.
Figure 4a illustrates the overall improvement effect of F-contaminated farmland remediation policy on the income structure optimization index of farmers and the significant differences between groups. Figure 4b further reveals the heterogeneous distribution of the policy effect among farmers with different land sizes and education levels. Based on the empirical survey data collected from 200 farming households, the F-contaminated farmland remediation policy significantly optimized farmers’ income structure. The income structure optimization index of the remediation group (0.652 ± 0.097) was significantly higher than control group (0.405 ± 0.072), with an absolute difference of 0.247 (t = 20.43, p < 0.001), reflecting the strong intervention effect set in the simulated data.
FIGURE 4
The multidimensional improvement in income structure is manifested in 15.2% increase in agricultural income stability, an 8.7% increase in the growth rate of non-agricultural income, and 12.3% improvement in the diversification of income sources. Heterogeneity analysis showed that the policy effect varied among different groups. Farmers with medium-sized landholdings, defined as 4–7 mu, approximately 0.27–0.47 ha, and higher education levels (high school or above) benefited the most, with an optimization index difference of 0.290, which verified that the index can effectively capture the differentiated characteristics of the policy impact.
SHAP value contribution ranking and interaction effect identification experiment
To investigate the heterogeneous mechanisms underlying the average treatment effect identified earlier, we examined the feature contributions and interaction effects derived from the LightGBM model and SHAP framework. Figure 5 presents the importance ranking of the key factors influencing the optimization of farmers’ income structure, as well as the conditional effects of governance intensity under varying levels of non-agricultural skills.
FIGURE 5
Figure 5a ranks the variables according to their average absolute SHAP values, where larger values indicate stronger relative influence on the model prediction of the income-structure optimization index. Therefore, the SHAP values should be interpreted mainly in relative rather than absolute terms. In this context, governance intensity has the largest SHAP value, indicating that remediation investment and implementation strength are the most influential predictors of farmers’ income-structure optimization in the LightGBM model. Non-agricultural skills rank second, suggesting that households with stronger non-farm employment capacity are more likely to translate remediation opportunities into diversified and more stable income sources. Figure 5b further shows how the effect of governance intensity varies across households with different levels of non-agricultural skills; the horizontal axis represents governance intensity, the vertical axis represents its contribution to the predicted income-structure optimization index, and the color gradient reflects the level of non-agricultural skills. The upward and more pronounced pattern among households with stronger non-farm skills indicates a reinforcing interaction: remediation policies are more effective when farmers have the capacity to participate in non-agricultural employment or related income-generating activities. Thus, the SHAP results suggest that policy design should not rely only on increasing remediation intensity, but should also combine environmental governance with vocational training, employment support, and differentiated assistance for households with different livelihood capacities. These findings highlight the complex, multi-dimensional associations between farmland consolidation policies and income outcomes in the studied sample, thereby offering a quantitative reference for informing differentiated policy design.
Granger causality test of macroeconomic driving effect
To verify whether the optimization of farmers’ income structure can serve as leading factor driving rural economic development, this experiment constructed a panel vector autoregression model at the township level. Through the Granger causality test, it was statistically determined whether the “farmers” income structure optimization index’ was a Granger cause of the “agricultural output growth rate” and the “non-agricultural employment growth rate,” thus empirically testing the causal direction and path of transmission from micro-level behavior to macro-level results. The Granger causality test results show at 5% significance level, the optimization of farmers’ income structure is a Granger cause of agricultural output and non-agricultural employment growth (P-values, 0.028 and 0.013), while reverse causal relationships were not established. This statistically confirms the unidirectional causal transmission mechanism from micro-level optimization to macro-level development (Table 3). While Granger causality indicates temporal precedence and predictive utility, it should be interpreted as evidence consistent with—but not definitive proof of—structural causal transmission. These findings provide time-series support for the hypothesized direction of influence running from micro-level income restructuring to subsequent macro-level development indicators.
TABLE 3
| Null hypothesis (H0) | Lag | F-statistic | p-value | Conclusion |
|---|---|---|---|---|
| Income structure optimization does not granger-cause agricultural output growth | 2 | 4.215 | 0.028 | Reject H0€ |
| Agricultural output growth does not granger-cause income structure optimization | 2 | 1.245 | 0.312 | Cannot reject H0€ |
| Income structure optimization does not granger-cause non-agricultural employment growth | 2 | 5.123 | 0.013 | Reject H0€ |
| Non-agricultural employment growth does not granger-cause income structure optimization | 2 | 0.893 | 0.428 | Cannot reject H0€ |
Analysis of Granger causality test results.
The quantitative decomposition of the path effect further reveals the strength of this transmission process. The total effect of the rectification policy on rural economic revitalization is 0.428, of which the indirect effect (0.248) accounts for 57.9% of the total effect. This indicates that the policy effect is mainly achieved through the intermediate path of optimizing farmers’ income structure, highlighting its core channel role in macro-driving (Table 4).
TABLE 4
| Pathway | Direct effect | Indirect effect | Total effect | Significance (p-value) |
|---|---|---|---|---|
| Remediation policy → rural economic revitalization | 0.18 | 0.248 | 0.428 | <0.001 |
| Remediation policy → income structure optimization | 0.412 | – | 0.412 | 0.007 |
| Income structure optimization → rural economic revitalization | 0.602 | – | 0.602 | <0.001 |
Analysis of the transmission path of economic benefits of rectification policy.
This experiment, through rigorous econometric testing, confirms that the optimization of farmers’ income structure is a statistical cause of the increase in rural agricultural output and the expansion of non-agricultural employment. Path analysis decomposes the estimated policy association, indicating that over half of the observed relationship between farmland consolidation policies and rural economic revitalization is accounted for by the mediating path of income structure optimization. While these estimates derive from observational data and rely on the specified model assumptions, they provide correlational evidence aligned with the policy logic that micro-level income adjustments may serve as an important channel for broader rural development.
Experimental discussion
This study establishes a cross-scale analytical chain linking micro-level behavioral adjustments to macro-level rural development. The findings indicate that remediation reshapes farmers’ income structure through factor reallocation and demand transmission, thereby generating downstream effects on the rural economy. This transmission pathway indicates that policy intervention first shapes micro-level household responses and then contributes to macro-level rural development, providing a micro-behavioral foundation for understanding rural ecological revitalization.
Policy effects are heterogeneous and nonlinear, interacting with household endowments and local conditions. Notably, SHAP analysis reveals a synergistic predictive contribution of 0.152 when high governance intensity coincides with non-agricultural skills, and heterogeneity results show that households with medium landholdings (4–7 mu) and higher education benefit most. These empirical patterns yield three actionable implications: (1) remediation programs should be bundled with targeted non-farm skill training to amplify synergies; (2) support should be prioritized toward identified high-benefit groups rather than uniformly applied; and (3) local adaptation of the remediation–training package is essential given the context-specificity of effect magnitudes. Although the quantitative estimates are tied to the study region, the hybrid analytical framework—integrating machine learning interpretability with multi-level validation—is transferable to other geographical and pollution contexts. Future research should refine differentiated intervention strategies by further identifying key target groups and optimizing policy combinations.
Conclusion and future research directions
This study provides empirical evidence that the remediation of F-contaminated farmland is associated with significant optimization of farmers’ income structure and predictive of improvements in rural economic indicators. Remediation intensity and non-agricultural skills emerge as two salient factors, exhibiting a synergistic pattern of association with income outcomes. Furthermore, temporal analysis indicates that micro-level income restructuring precedes growth in agricultural output and non-agricultural employment, consistent with a bottom-up transmission pathway. It indicates that micro-level optimization can effectively transmit to macro-level development. Given that the study is based on a typical regional sample of 200 households from a single fluoride-affected area, the specific empirical magnitudes reported here require verification in diverse socioeconomic contexts; however, the analytical framework itself is methodologically transferable. While the combination of machine learning interpretability with econometric causal inference offers a novel approach, the modest sample size warrants cautious interpretation of higher-order interaction effects. Moreover, the relatively short temporal span of the current panel limits the capacity to fully assess the long-term dynamic effects of remediation, which may unfold gradually over extended periods. Future studies with larger, longitudinal micro-datasets spanning longer time horizons would enhance the statistical power to detect more granular heterogeneous treatment effects and would permit a more comprehensive evaluation of the evolving relationship between environmental governance and rural economic revitalization. Additionally, the quantitative framework remains limited in fully capturing complex social network interactions and the institutional dynamics underlying governance outcomes; integrating qualitative evidence such as stakeholder interviews or field observations would provide valuable contextual depth to complement the statistical patterns identified here. Therefore, the study could be further extended to comparative analyses of different types of contaminated farmland. Long-term tracking data over extended periods could also be considered to further explore the dynamic and potentially lagged relationship between environmental governance and rural revitalization.
Statements
Data availability statement
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.
Ethics statement
Ethical review and approval was not required for the study on human participants in accordance with the local legislation and institutional requirements. Written informed consent from the [patients/ participants OR patients/participants legal guardian/next of kin] was not required to participate in this study in accordance with the national legislation and the institutional requirements.
Author contributions
YT: Conceptualization, Investigation, Supervision, Software, Writing – original draft, Resources, Visualization, Data curation, Project administration, Formal Analysis, Methodology. YZ: Formal Analysis, Visualization, Data curation, Resources, Project administration, Validation, Software, Investigation, Methodology, Supervision, Writing – original draft, Conceptualization. FW: Writing – review and editing, Methodology, Formal Analysis, Software, Data curation, Investigation, Resources. YH: Supervision, Methodology, Data curation, Software, Resources, Writing – review and editing, Formal Analysis.
Funding
The author(s) declared that financial support was not received for this work and/or its publication.
Acknowledgments
The authors thank to College of Economic and Management, Shenyang Agricultural University, Shenyang, China Beijing Equity Exchange, Administrative and Judicial Business Service Center, Beijing, and Jiangxi Police Institute, Nanchang, China for providing necessary facilities.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
1
AhnY.PandiK.ChoD. W.ChoiJ. (2021). Feasibility of soil washing agents to remove fluoride and risk assessment of fluoride-contaminated soils. J. Soils Sediments21 (8), 2770–2777. 10.1007/s11368-020-02808-8
2
ChenQ.WangX.LiQ.ChenJ.ZhuL.WangL.et al (2025). Soil fluoride enrichment process and the possible adaptation prevention principle in coal-burning fluorosis area in southwest China. Sci. Rep.15 (1), 971. 10.1038/s41598-024-84381-5
3
HadasE.MingelgrinU.FineP. (2021). Economic cost–benefit analysis for the agricultural use of sewage sludge treated with lime and fly ash. Int. J. Coal Sci. Technol.8 (5), 1099–1107. 10.1007/s40789-021-00439-z
4
IsingomaC. L.TalukdarA.ChakrabortyS.BhattacharyaS. (2025). Bioremediation of fluoride: potential of bacteria, plants, and plant-microbe interactions in fluoride removal from the environment. Geomicrobiol. J.42 (9), 818–838. 10.1080/01490451.2025.2526602
5
LiC.LiG. (2021). Impact of China’s water pollution on agricultural economic growth: an empirical analysis based on a dynamic spatial panel lag model. Environ. Sci. Pollut. Res. Int.28 (6), 6956–6965. 10.1007/s11356-020-11079-2
6
LiM.LiuS.LiuY.SunY.WangF.DongS.et al (2021). The cost–benefit evaluation based on ecosystem services under different ecological restoration scenarios. Environ. Monit Assess.193 (7), 398–405. 10.1007/s10661-021-09188-7
7
LiuY.ChengX.LiW. (2021). Agricultural chemicals and sustainable development: the agricultural environment kuznets curve based on spatial panel model. Environ. Sci. Pollut. Res. Int.28 (37), 51453–51470. 10.1007/s11356-021-14294-7
8
LiuS.DongY.WangF.LiuH.YuL. (2023). Cost-benefit research and potential solutions of ecological restoration programs in China. Transact Earth. Environ. Sustain1 (1), 68–79. 10.1177/2754124X221140800
9
MbasabireP.MurindangaboY. T.BromJ.ByukusengeP.UfitikireziJ. D.UwihanganyeJ.et al (2024). Remediation of heavy metal-contaminated soils using phosphate-enriched sewage sludge biochar. Sustainability17 (16), 7345. 10.3390/su17167345
10
Medina-MijangosR.De AndrésA.Guerrero-Garcia-RojasH.Segui-AmorteguiL. (2021). A methodology for the technical-economic analysis of municipal solid waste systems based on social cost-benefit analysis with a valuation of externalities. Environ. Sci. Pollut. Res. Int.28 (15), 18807–18825. 10.1007/s11356-020-09606-2
11
MikkonenH. G.Van de GraaffR.MikkonenA. T.ClarkeB. O.DasikaR.WallisC. J.et al (2018). Environmental and anthropogenic influences on ambient background concentrations of fluoride in soil. Environ. Pollut.242, 1838–1849. 10.1016/j.envpol.2018.07.083
12
MyagmarK.DarkhijavB.RenchinT.ChultemD. (2023). Cost–benefit analysis for riverbank erosion control approaches in the steppe area. Environ. Dev. Sustain25 (9), 9251–9266. 10.1007/s10668-022-02433-0
13
NguyenC. P.LeT. H.SchinckusC.SuT. D. (2021). Determinants of agricultural emissions: panel data evidence from a global sample. Environ. Dev. Econ.26 (2), 109–130. 10.1017/s1355770x20000315
14
OzdemirD. (2022). The impact of climate change on agricultural productivity in Asian countries: a heterogeneous panel data approach. Environ. Sci. Pollut. Res.29 (6), 8205–8217. 10.1007/s11356-021-16291-2
15
PrabhuS. M.YusufM.AhnY.ParkH. B.ChoiJ.AminM. A.et al (2023). Fluoride occurrence in environment, regulations, and remediation methods for soil: a comprehensive review. Chemosphere324, 138334. 10.1016/j.chemosphere.2023.138334
16
RenJ.WangS.BiK.ChengM.LiuC.ZhouL.et al (2023). Machine learning-enabled framework for high-throughput screening of MOFs: application in radon/indoor air separation. ACS Appl. Mater Interfaces15 (1), 1305–1316. 10.1021/acsami.2c19207
17
ShajiE.SarathK.SantoshM.KrishnaprasadP.AryaB.BabuM. S. (2024). Fluoride contamination in groundwater: a global review of the status, processes, challenges, and remedial measures. Geosci. Front.15 (2), 101734. 10.1016/j.gsf.2023.101734
18
SinghG.KumariB.SinamG.KumarN.MallickS. (2018). Fluoride distribution and contamination in the water, soil and plants continuum and its remedial technologies, an Indian perspective– a review. Environ. Pollut.239, 95–108. 10.1016/j.envpol.2018.04.002
19
Stewart‐SinclairP. J.KleinC. J.BatemanI. J.LovelockC. E. (2021). Spatial cost–benefit analysis of blue restoration and factors driving net benefits globally. Conserv. Biol.35 (6), 1850–1860. 10.1111/cobi.13742
20
SuoX.CaoS. (2021). China’s three north shelter forest program: cost–benefit analysis and policy implications. Environ. Dev. Sustain23 (10), 14605–14618. 10.1007/s10668-021-01260-z
21
SuttieJ. S.DickieR.ClayA. B.NielsenP.MahanW. E.BaumannD. P.et al (1987). Effects of fluoride emissions from a modern primary aluminum smelter on a local population of white-tailed deer (Odocoileus virginianus). J. Wildl. Dis.23 (1), 135–143. 10.7589/0090-3558-23.1.135
22
ThakurA.DharmendraD. (2025). Ex-situ remediation of arsenic-fluorinated water using electrocoagulation: neural network multi-objective modelling and metaheuristic optimisation. J. Appl. Electrochem55 (6), 1439–1454. 10.1007/s10800-024-02250-5
23
XuZ.ItoL.dos MuchangosL. S.TokaiA. (2023). Health risk assessment and cost–benefit analysis of agricultural soil remediation for tailing dam failure in jinding mining area, SW China. Environ. Geochem Health45 (6), 3759–3775. 10.1007/s10653-022-01445-z
24
YanY.YangJ.GuoY.YangJ.WanX.ZhaoC.et al (2022). Potential evaluation of different intercropping remediation modes based on remediation efficiency and economic benefits–a case study of arsenic-contaminated soil. Int. J. Phytoremed.24 (1), 25–33. 10.1080/15226514.2021.1920571
25
ZhangQ.LiX.YuL.WangL.WenZ.SuP.et al (2025). Machine learning-assisted fluorescence visualization for sequential quantitative detection of aluminum and fluoride ions. J. Environ. Sci.149 (10), 68–78. 10.1016/j.jes.2024.01.023
Summary
Keywords
farmer income structure, farmland consolidation, fluoride pollution, machine learning algorithms, rural economic revitalization
Citation
Tan Y, Zhang Y, Wan F and Han Y (2026) Study based on machine learning on the impact of fluoride-contaminated farmland remediation on farmers’ income structure and rural economic revitalization. Front. Environ. Sci. 14:1805611. doi: 10.3389/fenvs.2026.1805611
Received
09 February 2026
Revised
29 April 2026
Accepted
29 April 2026
Published
18 May 2026
Volume
14 - 2026
Edited by
Veeriah Jegatheesan, RMIT University, Australia
Reviewed by
Sevda Kuşkaya, Erciyes University, Türkiye
Ziyang Zhou, South China University of Technology, China
Updates
Copyright
© 2026 Tan, Zhang, Wan and Han.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Yan Zhang, sz7986494boy@163.com
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.