Underground water microplastic pollution prediction method based on machine learning

Abstract

The invention discloses an underground water micro-plastic pollution prediction method based on machine learning, and relates to the field of ecological environment, and the method comprises the steps: carrying out the unified grid and standardization processing of global micro-plastic observation data and multi-source environmental factor data; constructing and training a regression model and a classification model to predict a logarithm value of the microplastic concentration and a probability that the logarithm value exceeds a preset threshold value; carrying out feature importance analysis on the regression model and the classification model, and determining master control environment factors influencing micro-plastic pollution; according to the output probability of the classification model, a risk threshold is determined in combination with a diagnostic criterion, and a high-risk area is divided; applying a space mask to the high-risk area; carrying out coupling calculation on the micro-plastic pollution risk probability, population distribution data and underground water dependency data, and evaluating the number of potential exposed population grid by grid; and carrying out uncertainty quantification on the evaluation result of the number of potential exposed population. According to the method, the spatial distribution prediction of underground water micro-plastic pollution is realized.

Description

Technical Field

[0001] This invention belongs to the field of ecological environment, and in particular relates to a method for predicting groundwater microplastic pollution based on machine learning. Background Technology

[0002] In recent years, with the acceleration of industrialization and urbanization, microplastics (fibers, debris, etc.) can enter soil and groundwater through various pathways, including atmospheric deposition, irrigation recharge, and sludge and solid waste disposal. The soil-groundwater interface has gradually become an important sink for microplastics. Microplastic particles are retained due to sieving in the soil's pores and surface adhesion, and are resuspended and migrated under strong groundwater recharge or human disturbance such as pumping. This process spans multiple time and space scales and has a complex mechanism. After entering aquifers, they may also carry or release additives, posing potential ecological and health risks. Therefore, it is necessary to finely characterize the distribution and risk classification of microplastic particles on a global scale to provide a scientific basis for environmental governance and management.

[0003] Existing predictions mostly rely on traditional statistical and empirical models, which exhibit three limitations in the context of particulate microplastics: First, they generally assume spatial continuity and stability, making it difficult to characterize the patchy and abrupt distributions driven by climate, replenishment intensity, land use, and human activities. Second, they lack compatibility with data sources and types, making it difficult to integrate multi-source information such as remote sensing, socioeconomic data, and hydrogeological data. Furthermore, the sparse and inconsistent monitoring points weaken the generalization and comparability of the models. Third, most studies focus on concentration prediction, lacking exposure assessments and uncertainty characterization that couple the results with population distribution and groundwater dependence.

[0004] In emerging research trends, machine learning methods (such as random forests, support vector machines, and XGBoost) are being increasingly introduced into groundwater pollution risk prediction. These methods can handle high-dimensional, multi-source, and nonlinear input variables, and are theoretically better suited to the complex migration characteristics of microplastics. However, existing research largely focuses on heavy metals or persistent organic pollutants, with no application to microplastics. Furthermore, these methods have limitations: some models neglect the interpretation of interactions between variables, resulting in predictions that, while accurate, lack interpretability; some studies use limited training data points, leading to insufficient spatial representativeness and poor model generalization ability; and most existing methods remain at the concentration prediction level, lacking an exposure assessment framework that integrates population distribution and groundwater dependence.

[0005] Therefore, there is an urgent need to propose a machine learning-based method for predicting groundwater microplastic pollution. Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention proposes a machine learning-based method for predicting groundwater microplastic pollution. This method effectively overcomes the limitations of traditional approaches and enables spatial distribution prediction and risk exposure assessment of groundwater microplastic pollution at regional and even global scales. Consequently, it provides scientific support and technical tools for environmental management, drinking water safety assurance, and policy formulation.

[0007] To achieve the above objectives, this invention provides a machine learning-based method for predicting groundwater microplastic pollution, comprising: Training data is obtained by uniformly gridding and standardizing global-scale microplastic observation data and multi-source environmental factor data; Using the training data, a regression model and a classification model are constructed and trained to predict the logarithm of microplastic concentration and the probability of it exceeding a preset threshold, respectively. SHAP interpretability analysis was used to perform feature importance analysis on the regression and classification models to identify the main environmental factors affecting microplastic pollution. Based on the output probability of the classification model, and combined with diagnostic criteria, a risk threshold is determined, and high-risk areas are identified. A spatial mask is applied to the high-risk areas to exclude invalid areas and define hydrologically active land areas; Within the hydrologically active terrestrial area, the probability of microplastic pollution risk is coupled with population distribution data and groundwater dependence data for calculation. The number of potentially exposed people is assessed grid by grid and summarized by administrative unit. The uncertainty of the assessment results of the potential exposed population was quantified using the Monte Carlo simulation method.

[0008] Optional, the process of unified gridding and standardization includes: Unify the microplastic observation data and the multi-source environmental factor data to the same coordinate system and target grid with the same resolution; The microplastic observation data were subjected to logarithmic transformation; The geometric mean of multiple microplastic observation data within the same grid is taken as the representative value of that grid; the microplastic observation data of the grid is converted into binary labels according to a preset concentration threshold; Continuous environmental factor data are standardized, while categorical environmental factor data are one-hot encoded.

[0009] Optionally, the process of building and training regression and classification models includes: The regression model is constructed using the XGBoost algorithm, and its objective function is to minimize the regularized squared error loss. The classification model is constructed using the XGBoost algorithm, and its objective function is to minimize the weighted log loss function. The regression and classification models were trained using a hierarchical K-fold cross-validation strategy.

[0010] Optionally, the process of using SHAP interpretability analysis includes: Calculate the SHAP value of each environmental factor on the model prediction results; Based on the average absolute value of the SHAP of each environmental factor, the environmental factors with the highest contribution are selected as the main control environmental factors.

[0011] Optionally, the process of determining risk thresholds in conjunction with diagnostic criteria includes: Plot the ROC curve of the classification model and calculate the Youden index; The probability value that maximizes the Yoden index is used as the first risk threshold, or the probability value when the sensitivity equals the specificity is used as the second risk threshold.

[0012] Optionally, the process of applying a spatial mask includes: The high-risk areas were treated sequentially using marine masks, desert masks, and permafrost masks. After masking, only grids with a population greater than zero are retained to obtain the hydrologically active land area.

[0013] Optionally, the coupled computation process includes: The potential exposed population for each grid is obtained by multiplying the microplastic pollution risk probability, population size, and groundwater dependence ratio for each grid.

[0014] Optionally, the process of quantifying uncertainty includes: Random perturbations are applied to the aforementioned risk probability, population data, and groundwater dependence data; Based on the perturbed data, the coupling calculation and summarization steps are repeated a preset number of times; Based on the results of all repeated calculations, the mean and confidence interval of the potential exposed population were statistically obtained.

[0015] Technical advantages of this invention: This invention discloses a machine learning-based method for predicting microplastic pollution in groundwater. It employs a strategy of using marine / desert / permafrost regional masks and a uniform resolution grid to eliminate invalid samples and areas with low groundwater content. The grid is uniformly set so that NaN values ​​are displayed in gray. Spatial K-fold is introduced in sample construction to suppress spatial leakage during model prediction. Regression and classification models for microplastic prediction are performed simultaneously to obtain the predicted abundance and risk probability of microplastics in groundwater. Both models undergo SHAP interpretability analysis. Finally, the predicted risk probability of microplastics is overlaid with a world map population raster and groundwater dependence rate to calculate the number of exposed individuals per country, and confidence intervals are given using Monte Carlo methods. This significantly improves the accuracy, robustness, and policy usability of the assessment. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic flowchart of a machine learning-based method for predicting microplastic pollution in groundwater, according to an embodiment of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0019] like Figure 1 As shown, this embodiment provides a machine learning-based method for predicting groundwater microplastic pollution, including: Training data is obtained by uniformly gridding and standardizing global-scale microplastic observation data and multi-source environmental factor data; Using the training data, a regression model and a classification model are constructed and trained to predict the logarithm of microplastic concentration and the probability of it exceeding a preset threshold, respectively. SHAP interpretability analysis was used to perform feature importance analysis on the regression and classification models to identify the main environmental factors affecting microplastic pollution. Based on the output probability of the classification model, and combined with diagnostic criteria, a risk threshold is determined, and high-risk areas are identified. A spatial mask is applied to the high-risk areas to exclude invalid areas and define hydrologically active land areas; Within the hydrologically active terrestrial area, the probability of microplastic pollution risk is coupled with population distribution data and groundwater dependence data for calculation. The number of potentially exposed people is assessed grid by grid and summarized by administrative unit. The uncertainty of the assessment results of the potential exposed population was quantified using the Monte Carlo simulation method.

[0020] Furthermore, the process of unifying gridding and standardizing processing includes: Unify the microplastic observation data and the multi-source environmental factor data to the same coordinate system and target grid with the same resolution; The microplastic observation data were subjected to logarithmic transformation; The geometric mean of multiple microplastic observation data within the same grid is taken as the representative value of that grid; the microplastic observation data of the grid is converted into binary labels according to a preset concentration threshold; Continuous environmental factor data are standardized, while categorical environmental factor data are one-hot encoded.

[0021] Furthermore, the process of building and training regression and classification models includes: The regression model is constructed using the XGBoost algorithm, and its objective function is to minimize the regularized squared error loss. The classification model is constructed using the XGBoost algorithm, and its objective function is to minimize the weighted log loss function. The regression and classification models were trained using a hierarchical K-fold cross-validation strategy.

[0022] Furthermore, the process of using SHAP interpretability analysis includes: Calculate the SHAP value of each environmental factor on the model prediction results; Based on the average absolute value of the SHAP of each environmental factor, the environmental factors with the highest contribution are selected as the main control environmental factors.

[0023] Furthermore, the process of determining risk thresholds in conjunction with diagnostic criteria includes: Plot the ROC curve of the classification model and calculate the Youden index; The probability value that maximizes the Yoden index is used as the first risk threshold, or the probability value when the sensitivity equals the specificity is used as the second risk threshold.

[0024] Furthermore, the process of applying a spatial mask includes: The high-risk areas were treated sequentially using marine masks, desert masks, and permafrost masks. After masking, only grids with a population greater than zero are retained to obtain the hydrologically active land area.

[0025] Furthermore, the coupled computation process includes: The potential exposed population for each grid is obtained by multiplying the microplastic pollution risk probability, population size, and groundwater dependence ratio for each grid.

[0026] Furthermore, the process of quantifying uncertainty includes: Random perturbations are applied to the aforementioned risk probability, population data, and groundwater dependence data; Based on the perturbed data, the coupling calculation and summarization steps are repeated a preset number of times; Based on the results of all repeated calculations, the mean and confidence interval of the potential exposed population were statistically obtained.

[0027] Specifically, the implementation of this embodiment includes: This embodiment uses 0.1° grids as basic computational units, integrating observational data with environmental factors such as climate, hydrogeology, and land use. Data alignment, standardization, and feature construction are completed under the same resolution and coordinate system. A regression-classification dual-channel learning framework is employed to output the logarithmic prediction and overthreshold probability of concentration, respectively. The controlling factor is determined based on model interpretability, and robustness is verified. Risk thresholds are determined according to diagnostic criteria, and spatial masks such as ocean, desert, and permafrost are used to define the effective domain. Within the effective domain, risk probability is coupled pixel-by-pixel with population raster and groundwater dependence, and exposure is summarized by administrative unit. Uncertainty is quantified using Monte Carlo simulation and other methods to form verifiable and traceable assessment results. The pipeline includes the following steps: Step (1) Data standardization and gridding: Unit and scale unification: The target grid was set to 0.1°, and the data was unified to the same coordinate system and resolution; a robust logarithmic transformation was performed on the groundwater microplastic concentration. (1); in, C This represents the raw concentration data of microplastics within the grid; y reg This represents the concentration data of microplastics in the grid after logarithmic transformation.

[0028] Same-grid summary: If multiple observations exist within the same grid, the geometric mean is selected to represent that grid. (2); in, This represents the geometric mean of observations from multiple points within the same grid. K This indicates the number of microplastic observation data points within the grid. This represents the raw concentration observation data of the k-th microplastic in the same grid.

[0029] Category label construction: The threshold is configurable. In this example, 10.4 items / L is selected as the grid label threshold. (3); in, y cls Indicates grid label, C This represents the raw concentration data of microplastics within the grid. C th This indicates the grid label threshold.

[0030] Environmental factor registration: Continuous variables such as climate, soil, topography, groundwater, land cover, plastic emissions, and socioeconomic factors are resampled to the target grid and spatially registered with the sample points; categorical variables maintain coding consistency. Missing values ​​are imputed using KNN; continuous features are standardized using Z-scores, and categorical features are encoded using one-hot encoding to avoid sparse high-dimensional data.

[0031] Step (2) Integrated regression and classification modeling (mainly using XGBoost): Based on the aligned data obtained in step (1), two machine learning models are established: one is a regression model, used to predict the logarithmic transformation value of groundwater microplastic concentration; the other is a classification model, used to estimate the probability that the observed microplastic value exceeds a preset threshold. Through comparison, XGBoost is preferred as the core learner.

[0032] Regression model (logarithmic concentration prediction): Using the XGBoost regressor (objective=reg:squarederror), minimizing the regularized squared error: (4); in, L reg This represents the regularized squared error loss function; n Indicates the number of training samples; Indicates the first i The true label of each sample; Model for the first i Predicted values ​​for each sample; This represents a regularization term used to penalize model complexity and prevent overfitting. γ The penalty coefficient represents the number of leaf nodes; T This represents the total number of leaf nodes in all trees of the model; λ Represents the L2 regularization coefficient; This represents the L2 norm of the weight vector, which is the "sum of squares" of the weights of all leaf nodes.

[0033] Classification model (probability prediction exceeding the threshold): Employs an XGBoost binary classifier (objective=binary:logistic), assuming: (5); in, This indicates the probability that a pixel exceeds a set microplastic threshold. x i Indicates the first i Feature vectors of each sample; f ( x i ) represents the prediction function of the XGBoost binary classifier.

[0034] For class imbalance, the iterative loss of the model is logarithmic loss plus regularization: (6); in, L cls Represents the classification loss function of the classification model; n Indicates the number of training samples; Indicates the first i The true label of each sample; This indicates the probability that a pixel exceeds a set microplastic threshold. This represents a regularization term used to penalize model complexity and prevent overfitting.

[0035] Training strategy: The collected microplastic data were uniformly divided into training and test sets in an 8:2 ratio; hierarchical k-fold cross-validation and randomized grids were used with Bayesian search for optimal parameters, including max_depth, learning_rate, n_estimators, subsample, colsample_bytree, min_child_weight, reg_alpha, reg_lambda, etc.

[0036] Step (3) Interpretability and Feature Selection (SHAP): The Shapley Additive Interpretation (SHAP) method is used to perform additivity decomposition on the output of the trained microplastic pollution prediction model to quantify the marginal contribution of each feature to microplastic pollution and select the main control factors accordingly. Model prediction writing: (7); in, This represents the predicted value of the microplastic pollution prediction model for sample x; This represents the global average predicted value; m represents the total number of features used by the model. Indicates the firstj Each feature for the sample x Individual contributions to the prediction results. The top-10 non-zero variables that make a substantial contribution to the model are retained by sorting by the absolute value of the average SHAP, for robust modeling and mechanism interpretation.

[0037] Step (4) Determination of diagnostic thresholds and delineation of high-risk areas: ROC and Youden Index: (8); (9); Sensitivity = Specificity threshold: (10); in, This is the Youden's Index. t The decision threshold for the classification model; TPR ( t The true positive rate, also known as sensitivity, is the percentage of positive results. TNR ( t The true negative rate, also known as specificity, is the percentage of true negatives. To maximize the Youden index J ( t The optimal threshold obtained by ")"; The threshold for making "sensitivity = specificity".

[0038] Indicator definitions (TP, TN, FP, FN are elements of the confusion matrix): (11); (12); (13); (14); Among these, Sensitivity refers to the proportion of individuals who correctly identify a positive result; Specificity refers to the proportion of individuals who actually identify a negative result; PPV refers to the proportion of individuals who were correctly identified as positive; NPV refers to the proportion of individuals who were correctly identified as negative. TP indicates true positive; FN indicates false negative; TN indicates true negative; FP indicates false positive.

[0039] Select and High-risk areas can be designated; sensitivity analysis can be conducted in parallel with alternative concentration thresholds.

[0040] Step (5) Spatial mask (hydrologically active terrestrial area): A composite mask is constructed on a unified grid, and three types of areas are successively removed: ocean, desert and permafrost. Only pixels with population > 0 are retained to obtain the "hydrologically active land pixel set", which is used to normalize the effective denominator of area and exposure statistics.

[0041] Step (6) Population Exposure Assessment (Pixel-level → National-level): Pixel-level exposure count (combining urban / rural binary masking and groundwater dependence in various countries): (15); Among them, Risk ij The output probability is the classification channel, Population is the population size in the grid, and GWDep is the output probability. ij This represents the groundwater dependence ratio assigned by country and urban / rural category. After calculation, a national-level aggregation is performed: pixel-level results are topologically aggregated and sorted by country boundaries, outputting the total exposure for each country.

[0042] Step (7) Uncertainty Quantification (Monte Carlo): Perform N (preferably N=2000) perturbation simulations to give the national average and 95% confidence interval: (16); (17); (18); (19); in, This is the risk value after random perturbation; The original risk value; The random disturbance term of risk; refer to Follows a mean of 0 and a variance of The normal distribution; The population size after random perturbation; This represents the population size within the original grid. This refers to the random disturbance term in the population. For the random disturbance term of the population The vector follows a normal distribution with a mean of 0 and a variance of . ; The groundwater dependence ratio after disturbance; The original groundwater dependence ratio, assigned by country and urban / rural category; Refers to evenly distributed markings; The disturbance ratio parameter for groundwater dependence; The number of exposed individuals is obtained after perturbation simulation.

[0043] Step (8) Model Evaluation: return: (20); (twenty one); in, R 2 The coefficient of determination measures how well a model fits the data. y i Indicates the first i The actual value of each sample; Indicates the first i The model prediction value for each sample; This represents the average of the actual values ​​of all samples. n It represents the total number of samples; RMSE represents the root mean square error, which reflects the average error between the predicted and actual values.

[0044] Categories: ROC-AUC, Accuracy, Precision, Recall, F1, where: (twenty two); (twenty three); (twenty four); (25); Among them, Accuracy represents precision; Precision represents accuracy; Recall represents recall; F1 represents F1 score, which is the harmonic mean of precision and recall; TP indicates true positive; FN indicates false negative; TN indicates true negative; FP indicates false positive.

[0045] The purpose of this embodiment is to propose a machine learning-based method for predicting and assessing groundwater microplastic pollution risks, overcoming the shortcomings of existing technologies in areas such as data scarcity, insufficient variable integration, limited prediction accuracy, and lack of population exposure quantification. Current research on groundwater microplastics is mostly limited to single-point or regional monitoring, lacking the technical means to support large-scale or even global predictions. Furthermore, the modeling process often fails to fully consider the combined influence of multiple spatial variables such as climate conditions, hydrogeological characteristics, land use patterns, and socioeconomic factors, leading to insufficient representativeness and reliability of the prediction results. Therefore, this embodiment, within a unified data processing and modeling framework, systematically integrates multi-dimensional environmental factors such as climate, hydrogeology, land use, and socioeconomic factors. Advanced machine learning algorithms are used to capture the complex nonlinear laws governing the migration and spatial distribution of groundwater microplastics, thereby improving prediction accuracy and model generalization ability. Simultaneously, this embodiment introduces variable importance analysis, which clarifies the contribution of different environmental factors to groundwater microplastic pollution, thereby enhancing the model's interpretability and compensating for the shortcomings of existing methods in the interpretation and validation stages. Furthermore, this embodiment combines model predictions with population distribution data and groundwater usage rates to achieve a quantitative estimate of the number of potentially contaminated individuals and the level of risk. This not only reveals pollution hotspots but also provides a basis for public health risk assessment. Through the aforementioned technical means, this embodiment effectively overcomes the limitations of traditional methods, enabling spatial distribution prediction and risk exposure assessment of groundwater microplastic pollution at regional and even global scales. This provides scientific support and technical tools for environmental management, drinking water safety assurance, and policy formulation.

[0046] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A machine learning-based method for predicting microplastic pollution in groundwater, characterized in that, include: Training data is obtained by uniformly gridding and standardizing global-scale microplastic observation data and multi-source environmental factor data; Using the training data, a regression model and a classification model are constructed and trained to predict the logarithm of microplastic concentration and the probability of it exceeding a preset threshold, respectively. SHAP interpretability analysis was used to perform feature importance analysis on the regression and classification models to identify the main environmental factors affecting microplastic pollution. Based on the output probability of the classification model, and combined with diagnostic criteria, a risk threshold is determined, and high-risk areas are identified. A spatial mask is applied to the high-risk areas to exclude invalid areas and define hydrologically active land areas; Within the hydrologically active terrestrial area, the probability of microplastic pollution risk is coupled with population distribution data and groundwater dependence data for calculation. The number of potentially exposed people is assessed grid by grid and summarized by administrative unit. The uncertainty of the assessment results of the potential exposed population was quantified using the Monte Carlo simulation method.

2. The machine learning-based groundwater microplastic pollution prediction method as described in claim 1, characterized in that, The process of unified gridding and standardization includes: Unify the microplastic observation data and the multi-source environmental factor data to the same coordinate system and target grid with the same resolution; The microplastic observation data were subjected to logarithmic transformation; The geometric mean of multiple microplastic observation data within the same grid is taken as the representative value of that grid; the microplastic observation data of the grid is converted into binary labels according to a preset concentration threshold; Continuous environmental factor data are standardized, while categorical environmental factor data are one-hot encoded.

3. The machine learning-based groundwater microplastic pollution prediction method as described in claim 1, characterized in that, The process of building and training regression and classification models includes: The regression model is constructed using the XGBoost algorithm, and its objective function is to minimize the regularized squared error loss. The classification model is constructed using the XGBoost algorithm, and its objective function is to minimize the weighted log loss function. The regression and classification models were trained using a hierarchical K-fold cross-validation strategy.

4. The machine learning-based groundwater microplastic pollution prediction method as described in claim 1, characterized in that, The process of using SHAP interpretability analysis includes: Calculate the SHAP value of each environmental factor on the model prediction results; Based on the average absolute value of the SHAP of each environmental factor, the environmental factors with the highest contribution are selected as the main control environmental factors.

5. The machine learning-based groundwater microplastic pollution prediction method as described in claim 1, characterized in that, The process of determining risk thresholds based on diagnostic criteria includes: Plot the ROC curve of the classification model and calculate the Youden index; The probability value that maximizes the Yoden index is used as the first risk threshold, or the probability value when the sensitivity equals the specificity is used as the second risk threshold.

6. The machine learning-based groundwater microplastic pollution prediction method as described in claim 1, characterized in that, The process of applying a spatial mask includes: The high-risk areas were treated sequentially using marine masks, desert masks, and permafrost masks. After masking, only grids with a population greater than zero are retained to obtain the hydrologically active land area.

7. The machine learning-based groundwater microplastic pollution prediction method as described in claim 1, characterized in that, The process of coupled computation includes: The potential exposed population for each grid is obtained by multiplying the microplastic pollution risk probability, population size, and groundwater dependence ratio for each grid.

8. The machine learning-based groundwater microplastic pollution prediction method as described in claim 1, characterized in that, The process of quantifying uncertainty includes: Random perturbations are applied to the aforementioned risk probability, population data, and groundwater dependence data; Based on the perturbed data, the coupling calculation and summarization steps are repeated a preset number of times; Based on the results of all repeated calculations, the mean and confidence interval of the potential exposed population were statistically obtained.

Images