Method and system for evaluating regional aquifer vulnerability

Abstract

This application discloses a method and system for assessing the vulnerability of regional aquifers, relating to the field of groundwater pollution risk assessment. The method includes: obtaining environmental impact factor values ​​for the area to be predicted; using a multi-source domain joint transfer learning model, based on the environmental impact factor values, obtaining the predicted groundwater nitrate concentration for the area to be predicted; classifying the vulnerability level based on the prediction results, and generating a regional aquifer vulnerability assessment map for the area to be predicted. This application enables efficient and accurate assessment of regional aquifer vulnerability.

Description

Technical Field

[0001] This application relates to the field of groundwater pollution risk assessment, and in particular to a method and system for assessing the vulnerability of regional aquifers. Background Technology

[0002] Currently, groundwater vulnerability assessment methods are mainly divided into three categories: overlay index methods (such as the DRASTIC model, which is a seven-parameter overlay index model considering groundwater depth, net recharge, aquifer media, soil media, topography, impact of vadose zone, and hydraulic conductivity), process numerical simulation methods, and statistical methods. While overlay index methods are simple and easy to implement, the determination of factor weights and scores is highly subjective and ignores the nonlinear relationships between factors. Process numerical simulation methods require a large amount of detailed hydrogeological parameters, resulting in complex models that are difficult to apply on large regional scales. Traditional statistical methods have high data requirements and struggle to handle high-dimensional nonlinear problems.

[0003] In recent years, machine learning methods such as random forests and support vector machines have been gradually applied to the field of groundwater nitrate prediction. However, most existing methods are limited to data modeling in a single region and lack the ability to predict across regions and scales. Specific problems include: (1) Nitrate monitoring data are mainly concentrated in a few agricultural and industrialized areas, while data in other regions are scarce, making it difficult to form a comprehensive spatial distribution map; (2) The environmental characteristics (climate, soil, land use, etc.) of different regions in Asia are significantly different, resulting in strong spatial heterogeneity in nitrate distribution, making it difficult to effectively extend models trained in a single region to other regions; (3) Existing transfer learning methods are mostly aimed at tasks such as image classification and lack a multi-source domain joint transfer learning framework for groundwater nitrate pollution assessment. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for evaluating the vulnerability of regional aquifers, which can achieve efficient and accurate evaluation of the vulnerability of regional aquifers.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for assessing the vulnerability of regional aquifers, including: Obtain the environmental impact factor values ​​for the area to be predicted; A multi-source domain joint transfer learning model was adopted to obtain the predicted nitrate concentration of groundwater in the area to be predicted based on the environmental impact factor values. Based on the prediction results, the vulnerability level is classified, and a regional aquifer vulnerability assessment map of the area to be predicted is generated. The construction process of the multi-source domain joint transfer learning model includes: Acquire groundwater nitrate concentration data and environmental impact factor data for the sample area; the sample monitoring data includes source area sample data and target area sample data; The sample monitoring data and environmental impact factor data are preprocessed to obtain processed sample monitoring data and processed environmental impact factor data, in order to construct a multi-source domain dataset; the multi-source domain dataset includes source domain datasets and target domain datasets; The first random forest model was constructed and trained based on the source domain dataset using the bootstrap resampling method to obtain the source domain basic model; The first importance score of environmental impact factors is determined based on the source domain basic model; A target domain random forest model is constructed based on the target domain dataset using the bootstrap resampling method. The second importance score of environmental impact factors was determined based on the target domain random forest model; The characteristic values ​​of environmental impact factors are determined based on the first importance score, the second importance score, and the environmental impact factor data; A joint training dataset was constructed based on multi-source domain datasets and feature values ​​of environmental impact factors. A second random forest model was constructed and trained based on the joint training dataset using the bootstrap resampling method, resulting in a multi-source domain joint transfer learning model.

[0006] Secondly, this application provides a regional aquifer vulnerability assessment system, comprising: The data acquisition module is used to acquire groundwater nitrate concentration sample monitoring data and environmental impact factor data in the sample area; the sample monitoring data includes source area sample data and target area sample data; the data acquisition module is also used to acquire environmental impact factor values ​​for the area to be predicted; The data processing module, connected to the data acquisition module, is used to preprocess the sample monitoring data and environmental impact factor data to obtain processed sample monitoring data and processed environmental impact factor data, so as to construct a multi-source domain dataset; the multi-source domain dataset includes a source domain dataset and a target domain dataset; The source domain model training module, connected to the data processing module, is used to construct and train a first random forest model based on the source domain dataset using a bootstrap resampling method to obtain the source domain basic model; the source domain model training module is also used to determine the first importance score of environmental impact factors based on the source domain basic model; The feature reweighting module, connected to the source domain model training module, is used to construct a target domain random forest model based on the target domain dataset using a bootstrap resampling method. The feature reweighting module is also used to determine the second importance score of environmental impact factors based on the target domain random forest model. Furthermore, the feature reweighting module is used to determine the feature values ​​of environmental impact factors based on the first importance score, the second importance score, and the environmental impact factor data. The multi-source domain joint transfer learning model training module is connected to the feature reweighting module and the data processing module, and is used to construct a joint training dataset based on the multi-source domain dataset and the feature values ​​of environmental impact factors; the multi-source domain joint transfer learning model training module is also used to construct and train a second random forest model based on the joint training dataset using the bootstrap resampling method to obtain the multi-source domain joint transfer learning model. The prediction and evaluation module is connected to the data acquisition module and the multi-source domain joint transfer learning model training module. It is used to obtain the predicted result of groundwater nitrate concentration in the area to be predicted based on the environmental impact factor value using the multi-source domain joint transfer learning model. The prediction and evaluation module is also used to classify the vulnerability level based on the prediction result and generate a regional aquifer vulnerability evaluation map of the area to be predicted.

[0007] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and system for assessing regional aquifer vulnerability. It constructs a multi-source domain dataset using source and target domain sample data of groundwater nitrate concentration and environmental impact factor data. Based on this dataset, it builds a source domain basic model and a target domain random forest model, further determining the first and second importance scores. This captures the complex relationship between nitrate concentration and environmental impact factors, achieving high-precision prediction across regions. Furthermore, it determines the characteristic values ​​of environmental impact factors and constructs a training dataset using the multi-source domain dataset. Finally, it constructs a multi-source domain joint transfer learning model, forming a multi-source domain joint transfer learning framework for groundwater nitrate pollution assessment, improving the interpretability of regional aquifer vulnerability assessment. The constructed multi-source domain joint transfer learning model can accurately predict nitrate concentration in the region to be predicted, generating a regional aquifer vulnerability assessment map for the region, achieving efficient and accurate assessment of regional aquifer vulnerability. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart of a regional aquifer vulnerability assessment method according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a random forest provided in an embodiment of this application; Figure 3 This is a schematic diagram showing the importance score ranking of environmental impact factors in a multi-source domain joint transfer learning model provided in an embodiment of this application. Figure 4 This is a schematic diagram of a regional aquifer vulnerability assessment system provided in an embodiment of this application. Detailed Implementation

[0010] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] This application aims to overcome the following technical problems existing in current groundwater nitrate pollution assessment technologies: First, there is the problem of uneven distribution and scarcity of monitoring data. Groundwater nitrate monitoring data in Asia are mainly concentrated in a few agriculturally intensive areas (such as the North China Plain), while data is severely lacking in other areas (such as parts of North Asia, Central Asia, Southeast Asia, and South Asia). Traditional machine learning models are unable to achieve accurate cross-regional predictions due to the lack of training samples in the target domain.

[0012] Second, regional environmental heterogeneity leads to poor model generalization ability. The Asian continent has complex topography, diverse climate types (arctic, temperate, and tropical), and significant differences in land use patterns. Models trained on a single source domain (such as the North China Plain) cannot transfer the mapping relationship between environmental factors and nitrate concentration to target domains with vastly different climates, soils, and hydrogeological conditions (such as permafrost-covered North Asia, arid Central Asia, and tropical monsoon Southeast Asia), resulting in serious biases in prediction results.

[0013] Third, existing transfer learning methods are not well-suited for assessing nitrate pollution in groundwater. Current transfer learning techniques are primarily geared towards image classification and text analysis, lacking mechanisms for feature reweighting and sample weight optimization for geochemical spatial prediction, and thus failing to effectively address the issue of cross-regional variations in the importance of environmental factors.

[0014] Fourth, the models have poor interpretability and struggle to identify key driving factors. Most existing machine learning models are black-box models, unable to quantitatively reveal the marginal contribution of various environmental factors to nitrate concentration prediction and their regional differences, making it difficult to provide a scientific basis for targeted pollution prevention and control measures.

[0015] To make the objectives, features, and advantages of this application more apparent and understandable, the following detailed description of this application is provided in conjunction with the accompanying drawings and specific embodiments.

[0016] In one exemplary embodiment, such as Figure 1 As shown, a method for assessing the vulnerability of a regional aquifer is provided, including: Step 1: Obtain the environmental impact factor values ​​for the area to be predicted.

[0017] Step 2: Using a multi-source domain joint transfer learning model, based on the environmental impact factor values, the predicted groundwater nitrate concentration in the area to be predicted is obtained.

[0018] Step 3: Based on the prediction results, classify the vulnerability levels and generate a regional aquifer vulnerability assessment map for the area to be predicted.

[0019] The construction process of the multi-source joint transfer learning model in step 2 includes: S1, acquire groundwater nitrate concentration sample monitoring data and environmental impact factor data for the sample area. The sample monitoring data includes source area sample data and target area sample data.

[0020] S2 involves preprocessing the sample monitoring data and environmental impact factor data to obtain processed sample monitoring data and processed environmental impact factor data, thus constructing a multi-source domain dataset. The multi-source domain dataset includes source domain datasets and target domain datasets.

[0021] S3 employs a bootstrap resampling method to construct and train a first random forest model based on the source domain dataset, obtaining the source domain base model. The first importance score of environmental impact factors is then determined based on the source domain base model.

[0022] S4 employs a bootstrap resampling method to construct a target domain random forest model based on the target domain dataset. The second importance score of the environmental impact factors is determined based on the target domain random forest model. The eigenvalues ​​of the environmental impact factors are then determined based on the first importance score, the second importance score, and the environmental impact factor data.

[0023] S5. A joint training dataset is constructed based on multi-source domain datasets and environmental impact factor feature values. A bootstrap resampling method is used to construct and train a second random forest model based on the joint training dataset, resulting in a multi-source domain joint transfer learning model.

[0024] In one embodiment, groundwater nitrate concentration sample monitoring data is collected from the target area (i.e., the sample area). The sample monitoring data includes source domain sample data from data-rich areas and target domain sample data from data-scarce areas. The criteria for defining data-rich and data-scarce areas are based on the number of samples per unit area (i.e., sample density); data-rich areas are defined as areas with a sample density of not less than 10 samples per 10,000 km². 2 Furthermore, the samples should spatially cover sub-regions of the main environmental gradients within the area; data-scarce regions refer to areas with a sample density of less than 3 samples per 10,000 km². 2 Or sub-regions with scattered sample distribution that cannot independently support robust modeling; the aforementioned threshold (e.g., 10 samples / 10,000 km²) 2 3 per 10,000 km 2 The determination of the sample density is based on the basic requirements of the random forest model for the number of training samples. When the sample density is less than 3 samples / 10,000 km², the requirement is met. 2 At that time, random forest models struggle to effectively learn the complex mapping relationship between environmental impact factors and the target variable (i.e., groundwater nitrate concentration).

[0025] In this embodiment, the source domain sample data consists of 4,410 groundwater nitrate concentration monitoring data points from a certain region, and the target domain sample data consists of 2,818 sporadic monitoring data points from other regions.

[0026] Environmental impact factors include static environmental factors (factors reflecting the inherent properties of aquifers and remaining unchanged during the forecast period) and dynamic environmental factors (factors affected by climate change and human activities). In the embodiments provided in this application, a total of 21 impact factors are selected, including: static environmental factors (10): aquifer thickness, soil texture, topographic slope, porosity, permeability, land use type, distance from river, surface elevation (Digital Elevation Model, DEM), longitude (X), and latitude (Y); and dynamic environmental factors (11): groundwater storage (GWS), rainfall, soil electrical conductivity, soil moisture, baseflow-groundwater runoff, evaporation intensity, normalized difference vegetation index (NDVI), anthropogenic stress index, terrestrial water storage (TWS), water yield, and natural runoff recharge modulus.

[0027] Among them, groundwater storage is used to reflect the long-term changes in regional groundwater storage; water production is used to characterize the production capacity of surface water resources; aquifer thickness is used to reflect the size of groundwater storage space; soil texture is used to characterize the soil's ability to adsorb and migrate nitrates; and surface elevation is used to reflect the influence of topography on surface runoff and groundwater recharge.

[0028] In one embodiment, step S2 includes: unifying the format and transforming the coordinates of the sample monitoring data to obtain sample monitoring data in a unified format; identifying outliers in the unified format sample monitoring data, removing or correcting outliers to obtain processed sample monitoring data; converting the environmental impact factor data to GeoTIFF format and unifying it to the WGS84 geographic coordinate system to obtain environmental impact factor data in a unified format; spatially interpolating missing values ​​in the unified format environmental impact factor data using Kriging interpolation or inverse distance weighting to obtain interpolated environmental impact factor data; resampling the interpolated environmental impact factor data to a unified spatial resolution using bilinear interpolation or nearest neighbor interpolation to obtain resampled environmental impact factor data; extracting environmental impact factor values ​​at each monitoring well location from the resampled environmental impact factor data using nearest neighbor interpolation or bilinear interpolation using the latitude and longitude coordinates of the monitoring well locations; and combining the processed sample monitoring data with the corresponding environmental impact factor values ​​at each location to construct a multi-source domain dataset.

[0029] Specifically, the preprocessing process is divided into two paths. Path 1: Preprocessing of sample monitoring data. Step S2-1a (format unification): Unify the format and transform the coordinates of the sample monitoring data (including source domain sample data and target domain sample data) (CSV format, WGS84 coordinate system) to obtain sample monitoring data in a unified format.

[0030] Step S2-1b (Outlier Handling): For sample monitoring data in a uniform format, outliers are identified and removed or corrected using box plot method (with a threshold of 1.5 times the interquartile range (IQR)) or the 3σ principle, resulting in processed sample monitoring data (including processed source domain sample data and processed target domain sample data).

[0031] Path 2: Preprocessing of environmental impact factor data.

[0032] Step S2-2a (Format unification and coordinate transformation): For environmental impact factor data in different formats (NC4, TIF, Shp, etc.), convert them into GeoTIFF format and unify them to the WGS84 geographic coordinate system to obtain environmental impact factor data in a unified format.

[0033] Step S2-2b (Missing Value Imputation): For environmental impact factor data in a uniform format, spatial missing values ​​are imputed using Kriging interpolation or inverse distance weighting to obtain imputed environmental impact factor data.

[0034] Step S2-2c (spatial resampling): Using bilinear interpolation or nearest neighbor method, the interpolated environmental impact factor data is resampled to a uniform spatial resolution (1km×1km in the embodiments provided in this application) to obtain the resampled environmental impact factor data.

[0035] Finally, based on the processed sample monitoring data and the resampled environmental impact factor data, the environmental impact factor values ​​at each monitoring well location are extracted using the latitude and longitude coordinates of the well locations through the nearest neighbor method or bilinear interpolation. The processed sample monitoring data are then combined with the corresponding environmental impact factor values ​​at each location to construct a multi-source domain dataset (including source domain dataset and target domain dataset).

[0036] In one embodiment, step S3 employs a bootstrap resampling method to construct and train a first random forest model based on the source domain dataset. The process of obtaining the source domain base model may include: extracting K sample subsets from the source domain dataset using a bootstrap resampling method; constructing K initial source domain decision trees based on the K sample subsets to build the first random forest model; determining the optimal hyperparameters of the first random forest model based on the source domain dataset using a grid search method combined with 5-fold cross-validation; and training the first random forest model based on the optimal hyperparameters and the source domain dataset to obtain the source domain base model. The source domain base model includes K source domain decision trees (obtained by training the first random forest model from the initial source domain decision trees).

[0037] The risk assessment of nitrate pollution in groundwater involves many uncertainties and nonlinear factors, and the random forest model often exhibits strong robustness and applicability in handling such problems. Forest and randomness are its two main attributes. The forest aspect is reflected in combining many decision trees to form a forest of hundreds or thousands of decision trees, then averaging or voting the results of each decision tree to output the final result. The randomness aspect is reflected in using bootstrap resampling to generate multiple new datasets from the original dataset, with each new dataset forming a decision tree, and randomly selecting some features as input to the model. The theoretical basis of the random forest algorithm stems from the collective decision-making mechanism in statistics, aggregating the prediction results of a large number of weak classifiers and continuously training and optimizing them to achieve the best results.

[0038] The following are the basic principles of random forest: (1) Bootstrap resampling generates multiple equal-sized bootstrap sample sets from the original dataset through sampling with replacement. The decision tree model is trained using each dataset containing duplicate samples. Some original samples may not be selected. The unselected samples do not cause so-called waste, but can be used for subsequent model verification and evaluation. (2) When generating the nodes of each decision tree, some features are randomly selected to form a candidate subset (usually less than the total number of features), and the optimal splitting feature is selected from them to improve the difference between trees. (3) Each decision tree grows fully based on the dual randomness of samples and features (bootstrap sample set + random feature subset) (usually without pruning), and the generalization ability of the overall model is improved through differential training. (4) Under the collective decision-making mechanism, each basic learner will output the prediction results independently, and finally the average of the prediction results of all subtrees is taken as the output of the collective model through the aggregation strategy. The structure of random forest is as follows Figure 2 As shown.

[0039] In ensemble learning methods, the mathematical expressions of each basic model can be uniformly defined as: In the formula, Let y represent the prediction of the m-th decision tree for the input sample, where m represents the m-th decision tree and y represents the input sample. This represents the total number of leaf nodes in the m-th decision tree. Indicates an indicator function, Let represent the predicted value of the nth leaf node in the m-th decision tree. Let represent the sample set of the nth leaf node in the m-th decision tree. The prediction function of multiple decision trees can be expressed as: In the formula, M represents the number of decision trees.

[0040] The random forest model encompasses the following three layers of meaning: ① Single decision tree: A binary decision tree constructed using the Classification and Regression Tree (CART) algorithm. Each internal node corresponds to an environmental impact factor and its splitting threshold, and each leaf node stores the mean of the target variable (i.e., nitrate concentration sample monitoring data) in that subset. ② Random forest model: An ensemble model structure composed of K decision trees trained using bootstrap resampling and random feature selection. This structure contains all node information, splitting rules, and leaf node predictions of the K decision trees, representing a persistently storable and loadable model output. ③ Model output: For any input data, it is fed into the K decision trees. Each decision tree outputs a prediction value according to its node splitting rules. The final value obtained by taking the arithmetic mean of the K predictions is the random forest model's prediction output for that data.

[0041] For example, a random forest regression algorithm is used to construct and train a first random forest model using a source domain dataset (e.g., monitoring data of 4410 samples from a certain region and corresponding environmental impact factor data). Specifically, a bootstrap resampling method is used to randomly select K sample subsets with replacement from the source domain dataset (K is the number of source domain decision trees, K=150 in this embodiment). The number of samples in each subset is the same as the number of samples in the original source domain dataset (achieved through sampling with replacement, the same sample may be selected multiple times; some original samples may not be selected and are treated as out-of-bag samples). An initial source domain decision tree (regression tree) is trained independently on each sample subset. When splitting at each node of each initial source domain decision tree, the initial source domain decision tree is split from all... One environmental impact factor ( Randomly select from =21) Candidate factors (e.g., or From this, the optimal splitting feature and splitting threshold that maximize the reduction of the mean square error are selected. In all embodiments provided in this application, the feature refers to the environmental impact factor.

[0042] During training, the initial source domain decision trees grow fully without pruning. This results in the first random forest model. The arithmetic mean of the predictions from the K initial source domain decision trees is used as the regression output of the first random forest model. For any new input sample, it is fed into each of the K initial source domain decision trees, and each tree outputs a prediction value. The arithmetic mean of the predictions from the K initial source domain decision trees is used as the prediction output value of the first random forest model for that sample (i.e., the logarithmically transformed prediction value of nitrate concentration).

[0043] The source domain dataset was randomly divided into training and test sets in a 7:3 ratio. Stratified sampling was used to ensure similar distribution of nitrate concentration monitoring data in the training and test sets. Specifically, all samples in the source domain dataset were first grouped according to nitrate concentration levels (low, medium, high). Then, samples were randomly drawn from each group in a 7:3 ratio and assigned to the training and test sets, respectively. This method aims to ensure that the proportion of samples at each concentration level in the training and test sets is consistent with the original source domain dataset (in this example, low salt 39.62%, medium salt 43.40%, high salt 16.98%), avoiding significant deviations in the proportion of samples at a certain concentration level from the overall distribution due to uneven sample distribution, thus preventing biased or unreliable model evaluation results. The training set is used for cross-validation during the training and hyperparameter optimization of the first random forest model; the test set is used for the final performance evaluation after the first random forest model training is completed. This test set does not participate in any stage of the first random forest model training or hyperparameter selection to ensure the objectivity of the evaluation results.

[0044] After constructing the first random forest model, its hyperparameters are optimized. Specifically, a grid search algorithm combined with 5-fold cross-validation is used to optimize the hyperparameters of the first random forest model based on the training set. The candidate parameter range includes: , For each set of candidate hyperparameters, a random forest model is temporarily trained on four sets of training data in each fold cross-validation, and R is calculated on one set of reserved validation data. 2 R value, 5-fold cross-validation 2 The average value is used as the performance index for this set of hyperparameters. Based on the cross-validation results, the hyperparameter combination with the largest performance index is selected as the optimal hyperparameters. In this embodiment, the optimal hyperparameters for the first random forest model are: , .

[0045] After determining the optimal hyperparameters, use these optimal hyperparameters to train the first random forest model on the training set, and obtain the trained first random forest model as the base model of the source domain.

[0046] In one embodiment, the process of determining the first importance score of an environmental impact factor based on a source domain basic model may include: obtaining the Gini importance score of the environmental impact factor based on the reduction in Gini impurity when all nodes of the source domain decision tree in the source domain basic model split; and normalizing the Gini importance score to obtain the first importance score.

[0047] For example, the Gini importance (or ranking importance) of each environmental impact factor can be extracted from the source domain base model. For each environmental impact factor, the reduction in Gini impurity resulting from splitting at all nodes of all source domain decision trees is accumulated. Specifically, suppose that in the k-th source domain decision tree, feature f reduces the Gini impurity from [value missing] at the s-th node split. Reduce to Then the reduction in Gini impurity contributed by this node is The Gini importance score of feature f , Let f represent the set of all nodes in the k-th source domain decision tree that are split using feature f. After calculating the importance of all features, the Gini importance scores of each feature are normalized so that the sum of the Gini importance scores of all features equals 1. The normalized Gini importance score of each feature is the first importance score of that feature. Furthermore, the ranking importance is calculated as follows: randomly shuffle the values ​​of a feature in the test set, and calculate the prediction performance of the first random forest model before and after the shuffle (e.g., R²). 2 The magnitude of the decrease ( ). The larger the decrease, the more important the feature.

[0048] In one embodiment, the same algorithm structure as step S3 above is used (i.e., random forest regression algorithm, bootstrap resampling method, number of decision trees, random selection of features and splitting rules for full growth), but since the sample size of the target domain sample data is small (2818 in the embodiment provided in this application), all data in the target domain dataset is used as training data.

[0049] The optimal hyperparameters of the first random forest model determined above are directly used. , This eliminates the need to repeat grid search and 5-fold cross-validation. This is because hyperparameters reflect the structural properties of the algorithm (such as the number of decision trees and random seeds), rather than properties specific to a particular dataset. The optimal hyperparameters searched on the training set of the source domain can be directly transferred and used.

[0050] Using the optimal hyperparameters, a random forest model is trained on the target domain dataset, resulting in the target domain random forest model.

[0051] Using the same Gini importance calculation method (or permutation importance calculation method) as the above-mentioned first importance score determination process, the second importance scores of each environmental impact factor are extracted from the target domain random forest model. Since the sample size of the target domain data is small and a separate test set was not defined, if permutation importance is to be calculated, out-of-bag samples from the random forest are used to replace the test set for feature shuffling and accuracy reduction calculations.

[0052] In one embodiment, the process of determining the characteristic values ​​of environmental impact factors based on the first importance score, the second importance score, and environmental impact factor data includes: determining the reweighting coefficient of the environmental impact factor based on the ratio of the first importance score and the second importance score; and determining the characteristic values ​​of the environmental impact factor based on the reweighting coefficient and the environmental impact factor data.

[0053] For example, the reweighting coefficient for each feature is calculated based on the ratio of its first importance score to its second importance score. In the formula, For the first The reweighting coefficients of each feature, For the first The second importance score of each feature, For the first The first importance score of each feature. To prevent coefficient explosion caused by a zero or excessively small denominator, a protection mechanism is set: when season (i.e., the preset value).

[0054] Multiply the original values ​​of the environmental impact factors (determined based on the processed environmental impact factor data) by the corresponding reweighting coefficients to obtain the reweighted characteristic values ​​(i.e., the characteristic values ​​of the environmental impact factors): In the formula, For the first The original values ​​of each feature, For the first The eigenvalues ​​obtained by reweighting each feature.

[0055] After reweighting, features that were relatively important in the target domain but relatively unimportant in the source domain (such as soil texture, human activity pressure index, and terrain slope) are magnified by a larger factor, and their values ​​are significantly increased; while features that were relatively unimportant in the target domain are magnified by a smaller factor or not magnified at all.

[0056] In one embodiment, the reweighted coefficients for some features are calculated as follows: soil texture 37.00, anthropogenic stress index 20.00, topographic slope 17.00, soil electrical conductivity 8.70, distance from river 7.60, groundwater storage 5.50, terrestrial water storage 3.00, permeability 2.00, porosity 1.50, water yield 1.50, normalized difference in vegetation index 1.08, evaporation intensity 1.06, aquifer thickness 1.14, and rainfall 0.53. These coefficients reflect the degree of difference in importance between the source and target domains for each environmental impact factor.

[0057] In one embodiment, the process of constructing a joint training dataset based on the multi-source domain dataset and environmental impact factor feature values ​​in step S5 includes: replacing the processed environmental impact factor data in the multi-source domain dataset with environmental impact factor feature values ​​to obtain a new source domain dataset and a new target domain dataset. A target domain validation set is constructed based on the new target domain dataset. A training dataset is constructed based on the new target domain dataset and the new source domain dataset. Multiple candidate weight combinations are determined based on the training dataset. For each candidate weight combination, a bootstrap resampling method is used to construct a temporary random forest model based on the training dataset. The candidate weight combination corresponding to the temporary random forest model with the smallest mean squared error on the target domain validation set is taken as the optimal weight combination. A joint training dataset is constructed based on the optimal weight combination, the new source domain dataset, and the new target domain dataset.

[0058] For example, the reweighted feature values ​​replace the original values ​​of environmental impact factors in the source and target domain datasets, forming new source and target domain datasets. During the training of the second random forest model, since each feature has been scaled according to the difference in importance between domains, the second random forest model naturally pays more attention to features with higher importance in the target domain when splitting nodes—because the numerical range of these features has been magnified, their variance contribution is larger, and they are more likely to become the optimal splitting features under the splitting conditions. This achieves the goal of reducing the feature distribution difference between the source and target domains and improving the generalization ability of the second random forest model in the target domain. It makes the second random forest model pay more attention to features with higher importance in the target domain and lower importance in the source domain during training, thereby reducing the feature distribution difference between the source and target domains.

[0059] 20% of the new target domain dataset is randomly selected as the target domain validation set, and the remaining 80% of the new target domain dataset is merged with the new source domain dataset to construct the training dataset (as training data for the weight search stage). Stratified sampling is used during the partitioning to ensure that the sample proportion of each nitrate concentration level in the target domain validation set is consistent with the level distribution proportion in the new target domain dataset. While keeping the sample weight of the source domain data in the training dataset fixed at 1.0, the sample weight values ​​of the target domain data in the training dataset are traversed in the range of 1.0 to 5.0 with a step size of 0.5 to form multiple sets of candidate weight combinations (9 sets in this embodiment). For each set of candidate weight combinations, the following temporary process is executed: (1) Based on the current candidate weight combination (source domain weight 1.0 and the current candidate weight of the target domain), the same algorithm structure as step S3 above (i.e., random forest regression algorithm, bootstrap resampling method, number of decision trees, random selection of features and splitting rules of full growth) is used to construct a temporary random forest model using the training dataset. (2) The temporary random forest model is used to predict on the target domain validation set, and the mean squared error (MSE) is calculated. This temporary random forest model is only used to evaluate the performance of the current candidate weight combination during the weight search process. It is discarded after the search is complete and not used as the model output for subsequent steps. The MSE values ​​of the nine candidate weight combinations are compared, and the candidate weight combination corresponding to the temporary random forest model with the smallest mean squared error on the target domain validation set is selected as the optimal weight combination. After debugging, the optimal weight combination determined in this embodiment is: source domain sample weight = 1.0, target domain sample weight = 2.5. The optimal weight combination is combined with the new source domain dataset and the new target domain dataset to construct a joint training dataset.

[0060] In one embodiment, in the joint training dataset, when the ratio of the number of samples in the new source domain dataset to the number of samples in the new target domain dataset is greater than 3:1, a random undersampling method is used to reduce the number of samples in the new source domain dataset, or a synthetic minority class oversampling technique is used to increase the number of samples in the new target domain dataset, until the ratio of the number of samples in the new source domain dataset to the number of samples in the new target domain dataset is less than or equal to 3:1. In the embodiment provided in this application, there are 4410 source domain samples and 2818 target domain samples, with a ratio of approximately 1.56:1, which does not exceed the 3:1 threshold; therefore, the balancing process in this embodiment is not performed.

[0061] In one embodiment, step S5 employs a bootstrap resampling method to construct and train a second random forest model based on the joint training dataset. The process of obtaining the multi-source domain joint transfer learning model includes: extracting J data subsets from the joint training dataset using a bootstrap resampling method; constructing J initial joint decision trees based on the J data subsets to build the second random forest model; determining the optimal hyperparameters of the second random forest model based on the joint training dataset using a grid search method combined with 5-fold cross-validation; and training the second random forest model using a weighted random forest regression algorithm based on the optimal hyperparameters of the second random forest model and the joint training dataset to obtain the multi-source domain joint transfer learning model. The multi-source domain joint transfer learning model includes J joint decision trees (obtained by training the second random forest model from the initial joint decision trees).

[0062] The construction process of the second random forest model has a similar algorithm structure to step S3 above, namely, it uses the same random forest regression algorithm: ① Data subsets are extracted with replacement using bootstrap resampling. ② Each initial joint decision tree is fully grown without pruning. ③ When splitting a node, some features are randomly selected as candidates. ④ The average of the prediction results from multiple trees is taken as the output. The difference is that step S3 uses the source domain dataset (containing 4410 samples in the embodiment provided in this application), while this step uses the merged joint training dataset (containing 7228 samples in the embodiment provided in this application). The number of decision trees K in step S3 is 150, while the number of decision trees J in this step is 160.

[0063] The hyperparameter optimization method is the same as that of the first random forest model, both employing grid search combined with 5-fold cross-validation. The range of candidate parameters is also consistent with that of the first random forest model. , Following the grid search in the hyperparameter determination process of the first random forest model, a second round of refinement search was performed near the optimal hyperparameters of the first random forest model (targeting...). The search is performed with a step size of 10 between 150 and 200. However, due to the different training data (the joint training dataset is larger than the source domain dataset), the cross-validation results are different. The optimal hyperparameters of the second random forest model obtained in this step are: , .

[0064] Based on the determined optimal hyperparameters of the second random forest model, a weighted random forest regression algorithm is used to train the second random forest model on the joint training dataset. Specifically, the environmental impact factor feature values ​​in the joint training dataset are used as input, and the processed nitrate concentration values ​​in the joint training dataset are logarithmically transformed and used as the output target. The weighted random forest regression algorithm is then used for training.

[0065] In the construction of each initial joint decision tree, the split gain of each node is calculated as follows: For any environmental impact factor (candidate feature) and its split point (candidate threshold) within its value range, the difference between the weighted mean square error of the parent node before splitting and the weighted average of the weighted mean square errors of the left and right child nodes after splitting (i.e., the split gain) is calculated. The feature-threshold combination that maximizes this gain is selected as the optimal splitting rule for that node. The weighted mean square error incorporates sample weights into the calculation, making the prediction error of high-weight samples have a greater impact on the splitting decision.

[0066] After the second random forest model is trained, a multi-source domain joint transfer learning model containing 160 joint decision trees is obtained, which can be persistently saved for subsequent spatial prediction.

[0067] In the standard random forest regression algorithm, the optimal splitting feature and threshold are selected by maximizing the split gain (i.e., the reduction in impurity) when splitting a node. For regression tasks, impurity is typically measured using mean squared error (MSE). The weighted random forest algorithm incorporates sample weights into the MSE calculation, allowing high-weight samples to account for a larger proportion of the total error. The weighted random forest algorithm is implemented by including sample weights in the mean squared error calculation when calculating the split gain for each node, thus giving high-weight samples a greater influence on the splitting decision.

[0068] In one embodiment, the implementation phase of applying a multi-source joint transfer learning model to a region to be predicted is described. Step 2 includes: determining an environmental impact factor raster layer for the region to be predicted based on environmental impact factor values; obtaining a raster stack covering the region to be predicted based on the environmental impact factor raster layer; and using a multi-source joint transfer learning model to perform pixel-by-pixel prediction on the raster stack to obtain the predicted groundwater nitrate concentration value for each pixel in the raster stack, thereby obtaining the predicted groundwater nitrate concentration for the region to be predicted.

[0069] Collect environmental impact factor values ​​for the region to be predicted (e.g., the entire Asian continent), and determine the raster layer of environmental impact factors for the region to be predicted. After preprocessing (i.e., format unification, coordinate transformation, missing value imputation, and spatial processing in steps S2-2a to S2-2c above), a raster stack covering the entire region to be predicted is formed.

[0070] The raster stack can be converted into a dot matrix and predicted using a multi-source domain joint transfer learning model; or a pixel-by-pixel prediction method can be directly used to predict using a multi-source domain joint transfer learning model.

[0071] In this embodiment, a pixel-by-pixel prediction method is directly adopted: for each pixel location in the raster stack, the values ​​of the environmental impact factors at that location are extracted from multiple bands (the number of environmental impact factors is equal to the number of environmental impact factors), forming a multi-dimensional vector (the dimension is equal to the number of environmental impact factors, which is 20 in the embodiment provided in this application). This multi-dimensional vector is input into a multi-source domain joint transfer learning model. The model internally calls the standard predict function—that is, the multi-dimensional vector is fed into each joint decision tree in the model. Each joint decision tree outputs a predicted value according to its node splitting rules. The arithmetic mean of the predicted values ​​of all joint decision trees is taken to obtain the predicted result of the groundwater nitrate concentration at that pixel location (the logarithmic transformation value of the concentration). The prediction result is inversely transformed using an exponential function to obtain the predicted value of the groundwater nitrate concentration (mg / L, expressed as NO3) at each pixel location. - count).

[0072] The `predict` function is not a structural component of the multi-source domain joint transfer learning model itself, but rather a general interface method name in the field of machine learning / deep learning, referring to the operation of using a trained model to perform forward computation on new data and obtain the output result.

[0073] After obtaining the predicted groundwater nitrate concentration at each pixel location, vulnerability levels are assigned. Based on preset nitrate concentration thresholds, the predicted values ​​are divided into five levels: extremely low risk (0-20 mg / L), low risk (20-50 mg / L), medium risk (50-70 mg / L), high risk (70-90 mg / L), and extremely high risk (90-120 mg / L). The classification results are then output as a map, generating a vulnerability assessment map. This map can be rendered using software such as ArcGIS or QGIS (Quantum Geographic Information System). The vulnerability assessment map can use a layered color scheme: extremely low risk (green), low risk (light green), medium risk (yellow), high risk (orange), and extremely high risk (red). Basic geographic information such as regional boundaries, water systems, and cities can be overlaid for easier interpretation and application.

[0074] In an exemplary embodiment, the performance of the multi-source domain joint transfer learning model constructed in this application is evaluated and verified. Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²) are used. 2 As an evaluation indicator, the calculation formula is: ; ; ; ; in, This represents the actual nitrate concentration value. The predicted value is obtained by transforming the predicted result output by the model through an exponential function. This represents the average of the actual nitrate concentration values. This represents the number of samples in the test set.

[0075] On the same test set (i.e., the test set obtained by partitioning the source domain dataset), calculate the evaluation metrics (MSE, RMSE, MAE, R) for the following three models respectively. 2 To verify the effectiveness of the multi-source domain joint transfer learning model: (1) Source domain base model (used to evaluate the prediction performance without cross-regional knowledge transfer). (2) Source domain-target domain joint model (used to evaluate the prediction effect of simply merging data without domain adaptation). (3) Multi-source domain joint transfer learning model.

[0076] The source-target domain joint model was constructed as a control model for the ablation experiment during the joint training process in step S5 above. The source and target domain datasets were directly merged without feature reweighting or adjustment of sample weights between the source and target domains, using the same hyperparameters as the source domain base model. , The source domain-target domain joint model is obtained by training the model.

[0077] The four evaluation metrics mentioned above were calculated and compared on the same test set for the source domain base model, the source domain-target domain joint model, and the multi-source domain joint transfer learning model. By comparing the four evaluation metrics of the three models, we can verify: ① whether the source domain data has the basic value of transfer to the target domain (comparing the source domain base model and the multi-source domain joint transfer learning model); ② whether the feature reweighting and sample weight optimization proposed in this application have a substantial improvement on cross-regional prediction performance (comparing the source domain-target domain joint model and the multi-source domain joint transfer learning model).

[0078] Importance scores (Gini importance or permutation importance) of each environmental impact factor are extracted from a multi-source domain joint transfer learning model. For example... Figure 3 As shown, by sorting the samples from highest to lowest scores, key driving factors are identified. This not only reveals the extent to which each environmental factor influences the model's predicted nitrate concentration but also allows for in-depth analysis of the key causes of nitrate pollution. Further interpretability analysis is conducted using the SHAP (SHapley Additive exPlanations) method: SHAP values ​​for each feature of each sample are calculated, and SHAP summary and dependency plots are generated. This provides a technical foundation for subsequently quantitatively revealing the positive and negative correlations and nonlinear relationships between each feature and nitrate concentration.

[0079] The multi-source domain joint transfer learning model adjusts the sample weights and feature weights based on the source-target domain joint model, resulting in better adaptability in the target domain. In the multi-source domain joint transfer learning model, water yield (derived from long-term interannual water yield composite data products from 1981 to 2022) has the highest importance, reaching 25.83%. Longitude (X) has the second highest importance, reaching 21.24%. Latitude (Y) has the third highest importance, reaching 13.08%. Topographic height (i.e., surface elevation) has the fourth highest importance, reaching 0.0806. Rainfall has the fifth highest importance, reaching 7.28%. Aquifer thickness has the sixth highest importance, reaching 6.34%. Other factors, such as natural runoff, river distance, and soil moisture, have lower importance and are therefore not discussed in detail.

[0080] Water yield reflects changes in water resources over long timescales and has a significant long-term impact on nitrate concentrations. Changes in water yield cause changes in nitrate concentrations in groundwater. Increased water yield may lead to more nitrate entering groundwater, while decreased water yield may lead to increased nitrate concentrations. Latitude and longitude reflect different geographical locations; changes in latitude and longitude cause differences in the natural environment, and these environmental changes can lead to changes in nitrate concentrations. Differences in topographic elevation (DEM) affect surface water flow and groundwater recharge in a region. Higher elevations lead to increased surface runoff, affecting nitrate migration and distribution. Precipitation directly affects the amount of nitrate leaching and infiltration into groundwater. In areas with high precipitation, nitrates are more likely to enter the groundwater system; while in areas with low precipitation, the risk of nitrate leaching is lower. Changes in aquifer thickness affect the storage and transport of groundwater within aquifers. Thicker aquifers may increase nitrate storage, while thinner aquifers may reduce its impact.

[0081] Analysis of the top six most important factors reveals that water yield, longitude, and latitude are the main factors influencing nitrate concentration. These factors reflect changes in water resources and geographical location, thus affecting altered climate conditions and land use variations. Aquifer thickness, precipitation, and topographic elevation (DEM) further influence groundwater storage, transport, and nitrate leaching and migration. Therefore, when formulating nitrate pollution control measures, it is necessary to comprehensively consider the impact of these factors and adopt different strategies for different regions to effectively manage and protect groundwater resources.

[0082] In conjunction with the content of the above embodiments, this application has the following beneficial effects: 1. Significantly improves cross-regional prediction accuracy. Through multi-source domain joint transfer learning, a data-rich region (4410 samples) is jointly modeled with other data-scarce regions (2818 samples). Feature reweighting and sample weight optimization techniques are employed to effectively address the issues of data scarcity and regional heterogeneity. Experimental results show that the mean squared error (MSE) of the multi-source domain joint transfer learning model is 62.27, a 39.6% reduction compared to the source domain base model (MSE=103.03); the mean absolute error (MAE) is 4.30, a 24.1% reduction compared to the source domain base model (MAE=5.67); and the coefficient of determination reaches 0.923, significantly outperforming the source-target domain joint model (R²). 2 =0.89). This demonstrates that the improved method of this application can more accurately capture the complex nonlinear relationship between nitrate concentration and environmental factors, achieving high-precision prediction across regions.

[0083] 2. Correcting systematic biases in the source domain basic model. This application effectively corrects the prediction bias of the source domain basic model in regions with vastly different environmental characteristics. For example, the source domain basic model, based on data from one region, incorrectly predicts another region (e.g., permafrost areas) as an extremely high-risk area (90-120 mg / L). However, the method in this application comprehensively considers the characteristics of sporadic data from the target domain, correcting the prediction result for this region to a low-risk area (20-50 mg / L), which highly matches the actual situation (cold climate, permafrost development, and sparse agricultural activity). This correction verifies the effectiveness of feature reweighting and sample weight optimization for domain adaptation.

[0084] 3. Identification and Quantification of Key Driving Factors. Through feature importance analysis, the top six key driving factors influencing groundwater nitrate concentration in Asia and their contributions were quantitatively identified: water yield (25.83%), longitude (21.24%), latitude (13.08%), topographic elevation (8.06%), rainfall (7.28%), and aquifer thickness (6.34%). This result reveals that water yield (reflecting long-term changes in water resources) is the most important controlling factor, while latitude and longitude play a significant role by influencing climate zones and land use types. This provides a clear scientific basis for developing differentiated nitrate pollution control measures for different regions—high-risk areas should prioritize water resource management and agricultural non-point source pollution control.

[0085] 4. Generate high-precision, high-reliability vulnerability distribution maps. A vulnerability distribution map of aquifers covering all of Asia (spatial resolution 1 km) was generated. The prediction results show good consistency with existing independent research findings. The prediction maps in this application can provide intuitive and scientific decision support for groundwater resource management and pollution prevention and control.

[0086] 5. Provides a complete end-to-end solution. The regional aquifer vulnerability assessment method provided in this application supports fully automated processing from raw data input to vulnerability distribution map generation, has good operability and repeatability, and is easy to promote and apply.

[0087] In summary, this application overcomes the key technical challenges of data scarcity, strong regional heterogeneity, and poor cross-regional generalization ability in the prior art, and provides a new technical solution that is efficient, accurate, and interpretable for the assessment of regional aquifer vulnerability.

[0088] Based on the same inventive concept, this application also provides a system for implementing the aforementioned regional aquifer vulnerability assessment method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations of one or more embodiments of the regional aquifer vulnerability assessment system provided below can be found in the limitations of the regional aquifer vulnerability assessment method described above, and will not be repeated here.

[0089] In one exemplary embodiment, such as Figure 4 As shown, a regional aquifer vulnerability assessment system is provided, comprising: The data acquisition module is used to acquire groundwater nitrate concentration sample monitoring data and environmental impact factor data for the sample area. The sample monitoring data includes source area sample data and target area sample data. The data acquisition module is also used to acquire environmental impact factor values ​​for the area to be predicted.

[0090] The data processing module, connected to the data acquisition module, is used to preprocess the sample monitoring data and environmental impact factor data to obtain processed sample monitoring data and processed environmental impact factor data, in order to construct a multi-source domain dataset. The multi-source domain dataset includes source domain datasets and target domain datasets.

[0091] The source domain model training module, connected to the data processing module, is used to construct and train a first random forest model based on the source domain dataset using a bootstrap resampling method, thus obtaining the source domain base model. The source domain model training module is also used to determine the first importance score of environmental impact factors based on the source domain base model.

[0092] The feature reweighting module, connected to the source domain model training module, is used to construct a target domain random forest model based on the target domain dataset using a bootstrap resampling method. The feature reweighting module is also used to determine the second importance score of environmental impact factors based on the target domain random forest model. Furthermore, the feature reweighting module is used to determine the feature values ​​of environmental impact factors based on the first importance score, the second importance score, and the environmental impact factor data.

[0093] The multi-source domain joint transfer learning model training module, connected to the feature reweighting module and the data processing module, is used to construct a joint training dataset based on the multi-source domain dataset and the feature values ​​of environmental impact factors. This module also employs a bootstrap resampling method to construct and train a second random forest model based on the joint training dataset, resulting in the multi-source domain joint transfer learning model.

[0094] The prediction and evaluation module, connected to the data acquisition module and the multi-source domain joint transfer learning model training module, is used to obtain the predicted groundwater nitrate concentration of the area to be predicted based on environmental impact factor values ​​using the multi-source domain joint transfer learning model. The prediction and evaluation module is also used to classify the vulnerability level based on the prediction results and generate a regional aquifer vulnerability assessment map of the area to be predicted.

[0095] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0096] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0097] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0098] It should be noted that the data involved in this application (including but not limited to data used for analysis, data stored, data displayed, etc.) are all information and data that have been fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for assessing the vulnerability of regional aquifers, characterized in that, include: Obtain the environmental impact factor values ​​for the area to be predicted; A multi-source domain joint transfer learning model was adopted to obtain the predicted nitrate concentration of groundwater in the area to be predicted based on the environmental impact factor values. Based on the prediction results, vulnerability levels are classified, and a regional aquifer vulnerability assessment map of the area to be predicted is generated. The construction process of the multi-source domain joint transfer learning model includes: Acquire groundwater nitrate concentration data and environmental impact factor data for the sample area; the sample monitoring data includes source area sample data and target area sample data; The sample monitoring data and environmental impact factor data are preprocessed to obtain processed sample monitoring data and processed environmental impact factor data, in order to construct a multi-source domain dataset; the multi-source domain dataset includes source domain datasets and target domain datasets; The first random forest model was constructed and trained based on the source domain dataset using the bootstrap resampling method to obtain the source domain basic model; The first importance score of environmental impact factors is determined based on the source domain basic model; A target domain random forest model is constructed based on the target domain dataset using the bootstrap resampling method. The second importance score of environmental impact factors was determined based on the target domain random forest model; The characteristic values ​​of environmental impact factors are determined based on the first importance score, the second importance score, and the environmental impact factor data; A joint training dataset was constructed based on multi-source domain datasets and feature values ​​of environmental impact factors. A second random forest model was constructed and trained based on the joint training dataset using the bootstrap resampling method, resulting in a multi-source domain joint transfer learning model.

2. The method for assessing the vulnerability of regional aquifers according to claim 1, characterized in that, Preprocessing of sample monitoring data and environmental impact factor data yields processed sample monitoring data and processed environmental impact factor data, enabling the construction of a multi-source domain dataset, including: The sample monitoring data is formatted and coordinate transformed to obtain sample monitoring data in a unified format; outlier identification is performed on the unified format sample monitoring data, and outliers are removed or corrected to obtain processed sample monitoring data. Environmental impact factor data were uniformly converted to GeoTIFF format and then unified to the WGS84 geographic coordinate system to obtain environmental impact factor data in a unified format. Missing values ​​in the unified format environmental impact factor data were spatially imputed using Kriging interpolation or inverse distance weighting to obtain imputed environmental impact factor data. The imputed environmental impact factor data were then resampled to a unified spatial resolution using bilinear interpolation or the nearest neighbor method to obtain resampled environmental impact factor data. Using the latitude and longitude coordinates of the monitoring well locations, the environmental impact factor values ​​at each location are extracted from the resampled environmental impact factor data through the nearest neighbor method or bilinear interpolation method; the processed sample monitoring data are then combined with the corresponding environmental impact factor values ​​at each location to construct a multi-source domain dataset.

3. The method for assessing the vulnerability of regional aquifers according to claim 1, characterized in that, Using the bootstrap resampling method, a first random forest model is constructed and trained based on the source domain dataset to obtain the source domain base model, including: K sample subsets are extracted from the source domain dataset using a self-sampling resampling method; K initial source domain decision trees are constructed based on K sample subsets to build the first random forest model; The optimal hyperparameters of the first random forest model are determined based on the source domain dataset by using a grid search method combined with 5-fold cross-validation. Based on the optimal hyperparameters of the first random forest model and the source domain dataset, the first random forest model is trained to obtain the source domain basic model; the source domain basic model includes K source domain decision trees.

4. The method for assessing the vulnerability of regional aquifers according to claim 3, characterized in that, The first importance score of environmental impact factors is determined based on the source domain-based model, including: Based on the reduction in Gini impurity when all nodes of all source domain decision trees in the source domain basic model split, the Gini importance score of the environmental impact factor is obtained. The Gini importance score is normalized to obtain the first importance score.

5. The method for assessing the vulnerability of regional aquifers according to claim 1, characterized in that, The characteristic values ​​of environmental impact factors are determined based on the first importance score, the second importance score, and the environmental impact factor data, including: The reweighting coefficient of the environmental impact factor is determined based on the ratio of the first importance score to the second importance score; The characteristic values ​​of the environmental impact factors are determined based on the reweighted coefficients and the environmental impact factor data.

6. The method for assessing the vulnerability of regional aquifers according to claim 1, characterized in that, A joint training dataset is constructed based on multi-source domain datasets and environmental impact factor feature values, including: The environmental impact factor data processed in the multi-source domain dataset is replaced with the environmental impact factor feature values ​​to obtain a new source domain dataset and a new target domain dataset. A target domain validation set is constructed based on the new target domain dataset; a training dataset is constructed based on the new target domain dataset and the new source domain dataset; Based on the training dataset, multiple candidate weight combinations are determined; for each candidate weight combination, a temporary random forest model is constructed based on the training dataset using a bootstrap resampling method; the candidate weight combination corresponding to the temporary random forest model with the smallest mean square error on the target domain validation set is taken as the optimal weight combination. The joint training dataset is constructed based on the optimal weight combination, the new source domain dataset, and the new target domain dataset.

7. The method for assessing the vulnerability of regional aquifers according to claim 6, characterized in that, In the joint training dataset, when the ratio of the number of samples in the new source domain dataset to the number of samples in the new target domain dataset is greater than 3:1, a random undersampling method is used to reduce the number of samples in the new source domain dataset or a synthetic minority class oversampling technique is used to increase the number of samples in the new target domain dataset, until the ratio of the number of samples in the new source domain dataset to the number of samples in the new target domain dataset is less than or equal to 3:

1.

8. The method for assessing the vulnerability of regional aquifers according to claim 1, characterized in that, A bootstrap resampling method is used to construct and train a second random forest model based on the joint training dataset, resulting in a multi-source domain joint transfer learning model, including: J data subsets are extracted from the joint training dataset using a self-sampling resampling method; J initial joint decision trees are constructed based on J subsets of data to build the second random forest model; The optimal hyperparameters of the second random forest model are determined based on the joint training dataset by using a grid search method combined with 5-fold cross-validation. Based on the optimal hyperparameters of the second random forest model and the joint training dataset, the second random forest model is trained using a weighted random forest regression algorithm to obtain the multi-source domain joint transfer learning model; the multi-source domain joint transfer learning model includes J joint decision trees.

9. The method for assessing the vulnerability of regional aquifers according to claim 1, characterized in that, A multi-source domain joint transfer learning model was used to obtain the predicted groundwater nitrate concentration in the area to be predicted, based on environmental impact factor values. The predicted results include: Based on the environmental impact factor values, determine the environmental impact factor raster layer for the area to be predicted; Based on the environmental impact factor raster layer, a raster stack covering the area to be predicted is obtained; The multi-source domain joint transfer learning model is used to perform pixel-by-pixel prediction on the grid stack to obtain the predicted value of groundwater nitrate concentration for each pixel in the grid stack, so as to obtain the predicted result of groundwater nitrate concentration in the area to be predicted.

10. A regional aquifer vulnerability assessment system, characterized in that, include: The data acquisition module is used to acquire groundwater nitrate concentration sample monitoring data and environmental impact factor data in the sample area; the sample monitoring data includes source area sample data and target area sample data; the data acquisition module is also used to acquire environmental impact factor values ​​for the area to be predicted; The data processing module, connected to the data acquisition module, is used to preprocess the sample monitoring data and environmental impact factor data to obtain the processed sample monitoring data and processed environmental impact factor data, so as to construct a multi-source domain dataset. Multi-source domain datasets include source domain datasets and target domain datasets; The source domain model training module, connected to the data processing module, is used to construct and train a first random forest model based on the source domain dataset using a bootstrap resampling method to obtain the source domain basic model; the source domain model training module is also used to determine the first importance score of environmental impact factors based on the source domain basic model; The feature reweighting module, connected to the source domain model training module, is used to construct a target domain random forest model based on the target domain dataset using a bootstrap resampling method. The feature reweighting module is also used to determine the second importance score of environmental impact factors based on the target domain random forest model. Furthermore, the feature reweighting module is used to determine the feature values ​​of environmental impact factors based on the first importance score, the second importance score, and the environmental impact factor data. The multi-source domain joint transfer learning model training module is connected to the feature reweighting module and the data processing module, and is used to construct a joint training dataset based on the multi-source domain dataset and the feature values ​​of environmental impact factors; the multi-source domain joint transfer learning model training module is also used to construct and train a second random forest model based on the joint training dataset using the bootstrap resampling method to obtain the multi-source domain joint transfer learning model. The prediction and evaluation module is connected to the data acquisition module and the multi-source domain joint transfer learning model training module. It is used to obtain the predicted result of groundwater nitrate concentration in the area to be predicted based on the environmental impact factor value using the multi-source domain joint transfer learning model. The prediction and evaluation module is also used to classify the vulnerability level based on the prediction result and generate a regional aquifer vulnerability evaluation map of the area to be predicted.

Images