Intelligent screening method and system for water and soil environment assessment indexes in antimony mine area

Abstract

The application provides an intelligent screening method and system for water and soil environment evaluation indexes of antimony mine areas, and relates to the technical field of big data analysis. The method comprises the following steps: obtaining a water and soil environment monitoring original data set of an antimony mine area, and preprocessing the original data set; extracting candidate monitoring indexes from the preprocessed result data set, quantifying the contribution degree of each index to a target response representation quantity by using a contribution explanation model, and obtaining high importance monitoring indexes; combining a time series analysis model and a spatial correlation analysis model, identifying the time series evolution characteristics and spatial correlation characteristics of the high importance monitoring indexes, and obtaining time series characteristic results and spatial characteristic results corresponding to each high importance monitoring index; performing time-space collaborative screening on the high importance monitoring indexes, and obtaining an optimized key monitoring index set; verifying the key monitoring index set, and determining a target monitoring index set used for water and soil environment evaluation of the antimony mine area according to the verification result.

Description

Technical Field

[0001] This invention relates to the field of big data analysis technology, and in particular to an intelligent screening method and system for assessing water and soil environmental indicators in antimony mining areas. Background Technology

[0002] Antimony (Sb) is a toxic metalloid element. The processes of antimony mining, beneficiation, smelting, and tailings storage result in large quantities of antimony and its associated pollutants entering the surrounding soil, surface water, and groundwater, posing a continuous threat to ecological environment safety and human health. To effectively monitor the evolution of antimony pollution and implement targeted remediation, it is necessary to select key assessment indicators from a vast pool of environmental monitoring data that accurately reflect the pollution status and evolution trend. However, with the continuous improvement of automated environmental monitoring in antimony mining areas, the types of available monitoring indicators and the amount of monitoring data have increased explosively. How to quickly and accurately select representative and process-indicating core monitoring indicators under multi-indicator, high-dimensional, and large-sample conditions has become a pressing technical challenge in the field of water and soil environmental assessment for antimony mining areas.

[0003] Currently, existing technologies acquire and process data on mine ore content, groundwater level, soil pH, and mine geological and meteorological conditions in a time-series manner. Simultaneously, they acquire mine geological structure exploration data, perform three-dimensional spatial structure modeling, and monitor environmental risk factors at various points. Finally, they conduct a comprehensive environmental risk assessment and analysis based on the three-dimensional spatial model of the mine's geological structure distribution, generating decision-making recommendations for mine geological environmental risk management. This approach enables, to a certain extent, dynamic monitoring and comprehensive evaluation of the mine's geological environment across different time scales and spatial dimensions.

[0004] However, existing technologies rely on the spatiotemporal synchronous processing and three-dimensional modeling of existing monitoring indicators, which does not solve the problem of how to intelligently select key assessment indicators under the conditions of multiple indicators and massive monitoring data. The indicator selection process still relies heavily on experience and judgment, making it difficult to fully extract the spatiotemporal characteristics of antimony pollution evolution. At the same time, existing technologies lack the ability to identify complex nonlinear interactions between indicators, and tend to retain a large number of accompanying and redundant indicators, resulting in a lack of prominence of core information. Furthermore, they have poor adaptability and transferability in different mining areas and different pollution stages. Summary of the Invention

[0005] To address the shortcomings of existing technologies that rely on spatiotemporal synchronous processing and 3D modeling of existing monitoring indicators, which fail to solve the problem of intelligently selecting key assessment indicators under conditions of multiple indicators and massive monitoring data, the indicator selection process still relies heavily on empirical judgment, making it difficult to fully extract the spatiotemporal characteristics of antimony pollution evolution. Furthermore, existing technologies lack the ability to identify complex nonlinear interactions between indicators, easily retaining a large number of accompanying and redundant indicators, resulting in a lack of prominent core information and poor adaptability and transferability across different mining areas and pollution stages. Therefore, this invention provides an intelligent screening method and system for water and soil environment assessment indicators in antimony mining areas.

[0006] The technical solutions provided by the embodiments of the present invention are as follows: The first aspect of this invention provides an intelligent screening method for water and soil environment assessment indicators in antimony mining areas, comprising: S1: Obtain the original dataset of water and soil environment monitoring in the antimony mining area, and preprocess the original dataset to obtain the preprocessed result dataset; S2: Extract candidate monitoring indicators from the preprocessed dataset. Use a machine learning model to evaluate the importance of the candidate monitoring indicators in conjunction with the target response characterization in the soil and water environment of the antimony mining area. Use a contribution interpretation model to quantify the contribution of each indicator to the target response characterization and obtain highly important monitoring indicators. S3: Combining time series analysis model and spatial correlation analysis model, identify the time series evolution characteristics and spatial correlation characteristics of high-importance monitoring indicators to obtain the time series characteristic results and spatial characteristic results corresponding to each high-importance monitoring indicator; S4: Based on the temporal and spatial characteristics, high-importance monitoring indicators are selected through spatiotemporal collaborative screening to obtain a preferred set of key monitoring indicators; S5: Verify the set of key monitoring indicators and determine the set of target monitoring indicators to be used in the water and soil environment assessment of the antimony mining area based on the verification results.

[0007] The second aspect of this invention provides an intelligent screening system for water and soil environment assessment indicators in antimony mining areas, comprising: processor; The memory stores computer-readable instructions, which, when executed by the processor, implement the intelligent screening method for water and soil environmental assessment indicators in antimony mining areas as described in the first aspect.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, a machine learning model combined with a contribution explanation model is used to assess the importance and quantify the contribution of candidate monitoring indicators. This overcomes the limitations of traditional linear analysis, accurately identifies the nonlinear relationship between indicators and the evolution of antimony pollution, and effectively eliminates redundant and accompanying indicators. Simultaneously, by combining time-series analysis and spatial correlation analysis models, highly important monitoring indicators are synergistically identified and screened based on their temporal evolution and spatial correlation characteristics. This fully explores the phased change patterns of indicators over time and their propagation and diffusion characteristics in space. This significantly improves the ability to extract the spatiotemporal evolution characteristics of pollutants in the water and soil environment of antimony mining areas, reduces the reliance on empirical judgment for indicator screening, and ensures that the selected key monitoring indicators have stronger representativeness, stability, and cross-regional applicability under different pollution stages and mining area conditions. This effectively solves the technical problems of insufficient spatiotemporal feature extraction, lack of accuracy in indicator screening, and insufficient adaptability in existing technologies. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating an intelligent screening method for water and soil environment assessment indicators in antimony mining areas, provided as an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of the structure of an intelligent screening system for water and soil environment assessment indicators in an antimony mining area, provided as an embodiment of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0013] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0014] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0015] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0016] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0017] Reference manual attached Figure 1 The diagram shows a flowchart of an intelligent screening method for water and soil environment assessment indicators in antimony mining areas, provided by an embodiment of the present invention.

[0018] This invention provides an intelligent screening method for water and soil environmental assessment indicators in antimony mining areas. This method can be implemented using an intelligent screening device for water and soil environmental assessment indicators in antimony mining areas, which can be a terminal or a server. The processing flow of the intelligent screening method for water and soil environmental assessment indicators in antimony mining areas may include the following steps: S1: Obtain the original dataset of water and soil environment monitoring in the antimony mining area, and preprocess the original dataset to obtain the preprocessed result dataset.

[0019] Specifically, the collected data were organized into a raw dataset for monitoring the soil and water environment in the antimony mining area. This raw dataset underwent preprocessing, including data cleaning, missing value imputation, outlier handling, and standardization. The raw dataset includes, but is not limited to, antimony content, pH value, organic matter content, and moisture content in the soil; antimony concentration, pH value, conductivity, and temperature in groundwater; rainfall, air temperature, humidity, wind speed, wind direction, and sunshine duration from meteorological data; and normalized difference vegetation index, normalized difference water index, surface temperature, vegetation cover, albedo, humidity index, and bare soil index from remote sensing data.

[0020] Specifically, data cleaning is performed on the original dataset to standardize the data field structure and check whether the time information, spatial location information, monitoring values, and units of measurement are complete and correctly formatted. The `duplicated()` function of Python software is used to identify records in the data table that have identical content or the same specified fields, determining whether duplicate values ​​or records exist in the dataset. Duplicate records, logically conflicting records, and invalid records are then removed based on field logical relationships to obtain a basic, valid dataset.

[0021] The duplicated() function refers to a function in the Python data analysis library used to identify duplicate data.

[0022] Furthermore, missing data in the basic valid dataset are identified, and based on the temporal continuity, spatial adjacency, and historical observation availability of the missing data, they are filled in using adjacent time window data, adjacent spatial unit data, the most recent valid observation, or the mean of adjacent valid data, respectively. Source markers are added to the filling results to complete the missing value imputation.

[0023] Furthermore, outlier processing is performed on the completed data. For indicators that do not follow a normal distribution or are not suitable for outlier identification using the 3σ principle, box plots are used to identify outliers. Further, by combining temporal continuity, spatial consistency, and environmental-driven matching relationships, outlier samples are categorized into measurement outliers, long-term drift values, and environmental mutation values. Measurement outliers are removed or replaced, long-term drift values ​​are corrected, and environmental mutation values ​​are retained.

[0024] Among them, the 3σ principle refers to the outlier identification method in statistics based on the assumption of normal distribution.

[0025] Furthermore, the indicators after the above processing are standardized. For indicators with large numerical ranges or obvious skewness, a logarithmic transformation is first performed to compress the value range. Subsequently, the minimum, maximum, and fluctuation range of the processed data for each indicator are statistically analyzed, and a uniform scaling mapping is performed on each indicator to uniformly transform the values ​​of different indicators to the preset interval [0, 1], thereby eliminating the influence of dimensional differences on subsequent analysis results. At the same time, the original values, processed values, and processing parameter records are retained to form a preprocessed result dataset that can be used in subsequent screening, attribution, and validation steps.

[0026] S2: Extract candidate monitoring indicators from the preprocessed dataset. Use a machine learning model to evaluate the importance of the candidate monitoring indicators in conjunction with the target response characterization in the soil and water environment of the antimony mining area. Use a contribution interpretation model to quantify the contribution of each indicator to the target response characterization and obtain highly important monitoring indicators.

[0027] In this context, the machine learning model specifically refers to the random forest regression model. It is a non-linear algorithm based on ensemble learning that improves accuracy and stability by constructing multiple decision trees and combining their prediction results.

[0028] In one possible implementation, S2 specifically includes sub-steps S201 to S205: S201: Extract candidate monitoring indicators related to the evolution of water and soil environment in antimony mining areas from the preprocessed dataset, and construct a random forest regression model using the characterization information related to the evolution of antimony content as the response variable.

[0029] Specifically, during model construction, the preferred settings for the random forest regression model are: 500 decision trees, a maximum tree depth of 20, a minimum number of samples for node splits of 5, and a minimum number of samples for leaf nodes of 2. Bootstrap sampling is used to generate training subsets for each decision tree. Setting the number of decision trees to 500 ensures sufficient nonlinear fitting capability while reducing the impact of random fluctuations in individual trees on the overall results, resulting in more stable feature importance outputs. Setting the maximum tree depth to 20 ensures the model can depict the complex hierarchical relationships between water and soil environmental monitoring indicators in the antimony mining area, while avoiding overfitting to local noise due to excessive tree depth. Setting the minimum number of samples for node splits to 5 prevents unstable splits due to insufficient sample size and improves the statistical reliability of the splitting process. Setting the minimum number of samples for leaf nodes to 2 prevents the model from becoming overly sensitive to abnormal individuals due to leaf nodes corresponding to only a single sample. Bootstrap sampling is used to form differentiated training subsets from massive monitoring samples, enhancing the model's ensemble effect and improving its adaptability to data fluctuations in different regions and time periods.

[0030] Bootstrap sampling refers to a probability-based random resampling technique. In the random forest regression model of this invention, this method generates multiple training subsets of the same size but slightly different content from the original dataset by randomly drawing samples with replacement from the original training dataset.

[0031] Furthermore, the model's training data consists of a historical monitoring sample set. Each historical monitoring sample includes input feature data and target output data. The input feature data is a feature vector formed by organizing candidate monitoring indicators related to the evolution of the soil and water environment in the antimony mining area in a preset order for a specific monitoring time, monitoring point, or preset time window. This includes soil physicochemical indicator vectors, water physicochemical indicator vectors, associated pollutant indicator vectors, hydrological and meteorological indicator vectors, and spatial location-related indicator vectors. The target output data is the antimony pollution characterization result corresponding to the sample, preferably the antimony content characterization value in the corresponding medium, or the antimony content change characterization value within the preset time window. Spatial location-related indicators include at least one or more of the following parameters: monitoring point coordinates, elevation, slope, aspect, distance from the mining source area, distance from surface runoff channels, and groundwater flow direction.

[0032] Among them, the soil physicochemical index vector consists of multiple soil physicochemical index values ​​obtained from soil sample testing; the water body physicochemical index vector consists of multiple water body physicochemical index values ​​obtained from surface water or groundwater sample testing; the associated pollutant index vector consists of associated pollutant concentration values ​​related to the migration and evolution of antimony pollution; the hydro-meteorological index vector consists of hydro-meteorological data such as rainfall, water level, flow rate, and temperature in the corresponding period of the monitoring area; and the spatial location-related index vector consists of latitude, longitude, elevation, relative distance, and spatial adjacency parameters of the monitoring point.

[0033] Among them, the antimony content characterization value refers to the measured antimony concentration in the target medium or the antimony pollution intensity value after unified conversion. The antimony content change characterization value refers to the change, change rate or trend slope calculated based on the antimony content detection results at the beginning and end of the preset time window, or based on the antimony content detection results at multiple consecutive times within the window, and is used to characterize the degree of dynamic change in antimony content.

[0034] For example, during model training and validation, at least 70% of the sample data is allocated to the training set, with the remaining samples serving as the validation set. This ensures sufficient sample support during model training while retaining independent samples for testing model generalization performance. Furthermore, at least five iterations of cross-validation are used to comprehensively evaluate model stability and generalization ability. The coefficient of determination is used as the performance evaluation metric. The average coefficient of determination obtained from multiple validations was used as the criterion for model validity. The average coefficient of determination was preferred. The value should be no less than 0.9 to ensure that the constructed model accurately reflects the response relationship between environmental monitoring indicators and the evolution of antimony content, so that the subsequent indicator screening results based on feature importance have high credibility.

[0035] S202: Based on the random forest regression model, the importance of each candidate monitoring indicator is evaluated, and the importance evaluation results are obtained.

[0036] S203: Combining the SHAP value, analyze the contribution of the importance assessment results to each candidate monitoring indicator to explain the degree of contribution explanation.

[0037] Among them, the SHAP value (Shapley Additive Explanations) is a model interpretation method based on the Shapley value in game theory, used to quantify the contribution of each input feature to the prediction results of the machine learning model.

[0038] S204: The importance score is obtained by weighted summation based on the importance assessment results and contribution interpretation results.

[0039] S205: Retain monitoring indicators with importance scores higher than the preset threshold as high-importance monitoring indicators.

[0040] It should be noted that those skilled in the art can set the size of the preset threshold according to actual needs, and this invention does not limit it.

[0041] Specifically, the trained random forest regression model, the feature vectors of candidate monitoring indicators for the samples to be explained, and the background sample set are used as inputs to the SHAP value parsing method to obtain the contribution value of each candidate monitoring indicator to the model's prediction results. The contribution value is obtained by calculating the prediction increment caused by each candidate monitoring indicator under different feature combinations and summing them with reference to the model's baseline prediction output.

[0042] It should be noted that a positive contribution value indicates that the corresponding candidate monitoring indicator has a positive driving effect on the prediction result, while a negative contribution value indicates that the corresponding candidate monitoring indicator has a negative inhibiting effect on the prediction result. The larger the absolute value of the contribution value, the higher the degree of influence of the corresponding candidate monitoring indicator on the prediction result. Preferably, the Tree SHAP algorithm is used to analyze the random forest regression model to obtain the SHAP contribution value of each candidate monitoring indicator at the single-sample level. The importance score of the candidate monitoring indicator is obtained by statistically analyzing the absolute values ​​of the SHAP values ​​of the corresponding candidate monitoring indicators in all samples. This score is used to characterize the overall contribution of the indicator to the change in the response variable. (Importance score of the indicator) The calculation is as follows:

[0043] in, Indicates the first Importance scores for each candidate monitoring indicator Represents the total number of samples. Indicates the first In the nth sample The SHAP contribution value of each candidate monitoring indicator. The importance score of the candidate monitoring indicator. If the score is greater than or equal to the threshold, it is retained as a high-importance monitoring indicator. The importance score of the candidate monitoring indicator is... Indicators below a certain threshold are discarded as low-contribution indicators. In this way, while retaining the main explanatory information, redundant indicators with low contribution are removed, resulting in a selection of high-importance monitoring indicators. This provides a more refined set of indicators for subsequent time-series analysis, spatial correlation identification, and attribution verification. Preferably, the number of high-importance monitoring indicators is preset to be more than half the number of candidate monitoring indicators. In one specific embodiment, when there are 28 candidate monitoring indicators, there are 15 high-importance indicators.

[0044] Among them, the Tree SHAP algorithm refers to a dedicated and efficient implementation of the SHAP value parsing method for tree models.

[0045] In this embodiment of the invention, a nonlinear importance assessment and contribution quantification of water and soil environmental monitoring indicators in antimony mining areas are achieved by introducing a random forest regression model combined with the SHAP value analysis method. Compared with traditional linear correlation analysis or empirical judgment, this step can accurately identify the complex response relationship between each candidate monitoring indicator and the evolution of antimony content, effectively eliminate redundant and accompanying indicators, and avoid retaining multiple indicators reflecting the same pollution process simultaneously. Simultaneously, by obtaining an importance score through weighted summation and screening high-importance monitoring indicators, the data dimensionality is significantly reduced while retaining core explanatory information, making the screening results more refined, stable, and targeted. This provides a high-quality data foundation for subsequent spatiotemporal evolution feature identification and collaborative screening of key indicators.

[0046] S3: Combining time series analysis model and spatial correlation analysis model, identify the time series evolution characteristics and spatial correlation characteristics of high-importance monitoring indicators to obtain the time series characteristic results and spatial characteristic results corresponding to each high-importance monitoring indicator.

[0047] In one possible implementation, the temporal evolution feature recognition process in S3 is specifically as follows: For the high-importance monitoring indicator sequences, a time series model is performed using a bidirectional long short-term memory network. Based on the output of the bidirectional long short-term memory network, the time series evolution features and key time period features corresponding to each high-importance monitoring indicator are extracted.

[0048] Specifically, in order to identify the evolution patterns and key driving periods of the soil and water environment in the antimony mining area at different stages, a bidirectional long short-term memory network was used to perform time-series modeling on the selected high-importance monitoring index sequences.

[0049] The BiLSTM network structure consists of 5 layers: an input layer, a bidirectional LSTM layer 1, a bidirectional LSTM layer 2, an attention layer, and a fully connected output layer.

[0050] BiLSTM (Bidirectional Long Short-Term Memory) is a special recurrent neural network structure that consists of two LSTM layers in opposite directions.

[0051] For example, the input layer receives time-series input data. To cover a relatively complete continuous observation window, the time step is set to 12. When monitoring data is collected monthly, it corresponds to a calendar year, which is beneficial for capturing the periodic effects of seasonal changes, rainfall infiltration, tailings release, and groundwater migration. The feature dimension is set to 15 because preliminary screening has been completed through random forest and SHAP analysis, and the 15 retained indicators can cover the main driving factors while controlling input complexity.

[0052] It should be noted that the bidirectional LSTM layer 1 contains 64 hidden units, which are used to simultaneously extract the forward and backward dependency information of the time series and output the hidden state sequence corresponding to each time step, so that the subsequent attention layer can perform time weight allocation.

[0053] It should be noted that the bidirectional LSTM layer 2 contains 32 hidden units, which are used to further compress and refine the temporal features extracted from the previous layer, reduce redundant information and retain key evolutionary features.

[0054] The attention layer is used to calculate the contribution weight of different time steps to the output, thereby highlighting the environmental change characteristics of key periods.

[0055] Furthermore, the fully connected output layer receives the temporal feature vector extracted by the bidirectional LSTM layer and weighted and converged by the attention layer. This vector characterizes the comprehensive temporal response characteristics of each candidate monitoring indicator to antimony concentration changes within a preset time window. Based on this temporal feature vector, the fully connected output layer outputs the corresponding antimony concentration prediction results, including: soil antimony concentration prediction results, surface water antimony concentration prediction results, or groundwater antimony concentration prediction results. During training, the model with the lowest validation loss is selected as the target model for subsequent temporal evolution feature extraction, key time period identification, and scene evolution pattern analysis.

[0056] The model's input consists of a multivariate time series sample comprising 15 highly important monitoring indicators over 12 consecutive time steps. For any sample, the input can be represented as a 12×15 temporal feature matrix, where each row represents the value of one of the 15 highly important monitoring indicators at a given time step, and each column represents the change sequence of a particular highly important monitoring indicator over the 12 consecutive time steps. The model's training data includes input samples constructed based on historical data and corresponding output labels. Specifically, the input sample is a 12×15 temporal feature matrix composed of 15 highly important monitoring indicators over 12 consecutive time steps, and the output label is the measured antimony concentration value at the target prediction time corresponding to that input sample. During model execution, the temporal feature matrix is ​​input into a BiLSTM network for temporal feature extraction, weighted by an attention layer, and then input into a fully connected output layer to obtain the antimony concentration prediction result. Model training is then completed based on the error between the predicted result and the measured value.

[0057] In one possible implementation, the spatial association feature recognition process in S3 is specifically as follows: A spatial correlation map among monitoring points in the antimony mining area was constructed, and a graph convolutional network (GCN) model was established to learn features from the spatial correlation map, thereby obtaining the spatial embedding representation corresponding to each monitoring point.

[0058] Among them, the Graph Convolutional Network (GCN) model refers to a deep learning model specifically designed for processing graph-structured data.

[0059] Based on spatial embedding representation, we identify the clustering characteristics, propagation characteristics, and spatial response characteristics of each high-importance monitoring indicator in the spatial dimension.

[0060] Specifically, based on the spatial relationships between monitoring points in the antimony mining area, a graph neural network is used to construct a spatial relationship map to identify the clustering characteristics, propagation patterns, and potential redundant points of different monitoring indicators in the spatial dimension. The nodes represent the regular monitoring points within the monitoring area, including at least soil, surface water, and groundwater monitoring points. The node feature vector consists of the normalized values ​​of the high-importance monitoring indicators at the corresponding monitoring point, used to characterize the environmental state characteristics of that monitoring point.

[0061] Furthermore, the Euclidean distance between any two monitoring points is calculated based on the spatial coordinates of each monitoring point, and the graph edge weights are determined by combining the reciprocal of the Euclidean distance with an exponential decay term.

[0062] Among them, the closer the monitoring points are, the larger the edge weight is, and the farther the monitoring points are, the smaller the edge weight is. The distance attenuation scale parameter is adaptively determined based on the statistical results of the distance between monitoring points in the monitoring area to form a spatial association map that reflects the spatial proximity relationship of monitoring points.

[0063] Furthermore, after constructing the spatial association graph, a two-layer graph convolutional network (GCN) model is established to learn features from the spatial association graph, including a first graph convolutional layer and a second graph convolutional layer. Specifically, the second graph convolutional layer performs weighted fusion of the node's own features and the features of its neighboring nodes based on the topological relationships and edge weights between nodes in the spatial association graph, and maps the fused features to the output features. Preferably, the output dimension of the first graph convolutional layer is set to 16 to extract richer spatial association features. The output dimension of the second graph convolutional layer is set to 8 to further compress redundant information and retain key spatial pattern features based on the extraction results of the previous layer, so as to obtain a more compact spatial embedding representation. The 8-dimensional features output by the second graph convolutional layer are the spatial embedding representation of the corresponding node. Ideally, the ReLU activation function is used for nonlinear mapping after the output of each graph convolutional layer. Output values ​​less than 0 are compressed to 0, while output values ​​greater than or equal to 0 remain unchanged, thereby improving the discriminative power of key spatial features and reducing interference from invalid features.

[0064] The training data for the graph convolutional network model includes input training data and output training data. The input training data consists of the spatial correlation graph and node feature matrix corresponding to historical graph samples, while the output training data consists of the measured or labeled antimony concentration of the target medium at each monitoring point. During model training, the antimony concentration at each monitoring point is predicted based on the input training data, and the model parameters are updated according to the error between the predicted and measured values. After training, the 8-dimensional node feature vector output from the second graph convolutional layer is used as the spatial correlation feature output for each monitoring point.

[0065] It should be noted that, to reduce the model's excessive reliance on local sample noise and local topological features during training, Dropout layers are set between or after the graph convolutional layers. While keeping the rest of the network structure and training parameters unchanged, training and validation are repeated at different Dropout rates, comparing the validation loss, generalization error, and result fluctuations under each candidate value. Preferably, the Dropout rate is set between 0.2 and 0.5. When the Dropout rate is less than a preset threshold, the model's ability to suppress overfitting is relatively limited. When the Dropout rate is not lower than the preset threshold, the model can still retain the main spatial correlation features while suppressing overfitting. Therefore, in this embodiment, the preset threshold is 0.3 to improve the model's generalization ability and the stability of spatial pattern recognition results.

[0066] Specifically, after identifying the temporal characteristics and spatial correlations of high-importance monitoring indicators, these indicators are further comprehensively ranked to obtain the final set of key monitoring indicators. Specifically, regarding the first... For each highly important monitoring indicator, based on the output of the attention layer of a bidirectional LSTM network, the contribution weight of the indicator at different time steps is statistically calculated and defined as the temporal contribution of the indicator. Simultaneously, based on the output of the graph convolutional network, the spatial embedding representation of the index at each monitoring point is statistically analyzed and defined as the spatial centrality of the index. To eliminate the influence of different dimensions and numerical ranges, the time-series contribution of each highly important monitoring indicator was analyzed. and spatial centrality Normalization is performed to obtain the corresponding normalized temporal contribution and normalized spatial centrality. Based on this, the first... Spatiotemporal composite score of key monitoring indicators :

[0067] in, Indicates the first Normalized time-series contribution of each highly important monitoring indicator. Indicates the first Normalized spatial centrality of a number of highly important monitoring indicators. This represents the weighting coefficient of the time series contribution.

[0068] It should be noted that, considering the typically distinct stages and lags in the evolution of the water and soil environment in antimony mining areas, the order and intensity of responses of different pollution indicators vary significantly across different evolutionary stages such as rainfall infiltration, tailings disturbance, groundwater migration, and surface accumulation. Therefore, when comprehensively evaluating highly important monitoring indicators, temporal contribution is given a higher weight than spatial centrality. Simultaneously, to ensure that pollution diffusion paths, spatial aggregation characteristics, and correlation patterns between monitoring points can still be effectively represented, spatial centrality retains its corresponding weight. Therefore, it is preferable to assign a higher weight to temporal contribution than to spatial centrality. By comparing the coverage of key monitoring indicator screening results for major pollution evolution characterization information and the stability of subsequent identification results under different weight settings, this embodiment is determined... The preferred value is 0.6. This value highlights the dominant role of temporal evolution characteristics in indicator selection while also considering the auxiliary role of spatial propagation characteristics in pollution process identification, thereby improving the consistency between the key monitoring indicator selection results and the actual pollution evolution process. After obtaining the spatiotemporal comprehensive scores of each highly important monitoring indicator, they are sorted from highest to lowest score, and the highly important monitoring indicators with the highest scores (previously a preset number of digits) are selected as the set of key monitoring indicators. Preferably, the preset number of digits is 8.

[0069] In this embodiment of the invention, a bi-branch spatiotemporal analysis model combining BiLSTM and GCN is constructed to achieve a synergistic characterization of highly important monitoring indicators in both temporal and spatial dimensions. At the temporal level, the bidirectional propagation and attention mechanisms of BiLSTM are utilized to accurately capture the temporal response characteristics and key driving periods of antimony pollution at different evolutionary stages. At the spatial level, GCN is used to learn features from the spatial correlation map of monitoring points, effectively identifying pollutant accumulation hotspots and diffusion paths. Finally, by fusing temporal contribution and spatial centrality to construct a comprehensive score for indicator ranking, the selected set of key monitoring indicators can more comprehensively and realistically reflect the dynamic evolution of the soil and water environment in antimony mining areas.

[0070] S4: Based on the temporal and spatial characteristics, high-importance monitoring indicators are selected through spatiotemporal collaborative screening to obtain a preferred set of key monitoring indicators.

[0071] In this embodiment of the invention, by constructing a spatiotemporal comprehensive score that integrates temporal contribution and spatial centrality, the highly important monitoring indicators are weighted, sorted, and screened, thereby achieving synergistic optimization in the spatiotemporal dimension.

[0072] In one possible implementation, after S4 and before S5, the following is also included: For each unified analysis sample, time-series feature determination, spatial feature determination, and spatiotemporal attribution analysis are performed on each key monitoring indicator in the key monitoring indicator set. The time-series feature determination results, spatial feature determination results, and spatiotemporal attribution results corresponding to each key monitoring indicator are output for the unified analysis sample.

[0073] Specifically, by combining temporal contribution and spatial centrality, the key monitoring indicators corresponding to the unified analysis sample are subjected to temporal characteristic determination, spatial characteristic determination, and spatiotemporal attribution analysis.

[0074] Each unified analysis sample corresponds to a specific time label and spatial label, and includes information related to key monitoring indicators corresponding to that time label and spatial label, as well as hydrological and meteorological indicators, spatial location-related indicators, and associated response information related to their changes. Instead of simply classifying samples based on the absolute values ​​of a single key monitoring indicator, the analysis combines the changes in key monitoring indicators across the time and spatial dimensions to determine the abnormal change characteristics of the current unified analysis sample, and further analyzes the sources of importance for each key monitoring indicator. In this embodiment, a unified analysis sample refers to an analysis unit formed by uniformly aligning key monitoring indicator data, hydrological and meteorological indicators, spatial location-related indicators, and associated response information using the same time label and spatial label as indexes. The associated response information may include one or more of the following: vegetation change response, sediment enrichment response, groundwater quality linkage response, surface runoff anomaly response, and biotoxicity response.

[0075] Specifically, to avoid misjudging random fluctuations within a single analysis period as abnormal trends, the key monitoring indicators are first assessed for their temporal characteristics corresponding to time labels and their spatial characteristics corresponding to spatial labels.

[0076] Furthermore, in determining the time-series characteristics, the three consecutive analysis periods preceding the current analysis period are used as reference periods. For the same key monitoring indicator at the same monitoring point, the absolute value of the indicator change between adjacent periods within these three consecutive analysis periods is calculated, and the average of the absolute values ​​of change, or a combination of the average and the dispersion statistic, is used to form the basic fluctuation. The basic fluctuation is then multiplied by a preset coefficient to obtain the time-series characteristic determination threshold corresponding to the current analysis period. The preset coefficient can be determined based on historical normal sample statistics, verification results, or empirical calibration results.

[0077] Furthermore, the values ​​of key monitoring indicators in the current analysis period are compared with the average values ​​of the previous three consecutive analysis periods: when the deviation is greater than the time series judgment threshold, it is determined to be an increasing state. When the deviation is less than the negative of the time series judgment threshold, it is determined to be a decreasing state. When the absolute value of the deviation is less than or equal to the time series judgment threshold, it is determined to be a stable state. Further, when the same key monitoring indicator at the same monitoring point deviates significantly in the same direction from the average value of the previous three analysis periods in two consecutive analysis periods, and the deviation in each period is greater than the corresponding time series characteristic judgment threshold, it is determined to be a continuous abnormal state. Specifically, if the deviation is positive, it is determined to be a continuous increasing abnormality. If the deviation is negative, it is determined to be a continuous decreasing abnormality. In the status output, the determination of a continuous abnormal state has a higher priority than the increasing, decreasing, or stable states in a single analysis period.

[0078] Specifically, in spatial feature determination, the monitoring area is divided into source area units, path units, and peripheral units, and the proportion of key monitoring indicators within each spatial unit that are within the preset normal range and the abnormal range is compared. When the proportion of key monitoring indicators within the abnormal range in the source area unit is higher than that in the path unit and the peripheral unit, it is determined to be a concentrated abnormal state in the source area. When the proportion of key monitoring indicators within the preset abnormal range in the path unit or the peripheral unit increases compared to the previous analysis period, and the abnormal distribution continuously shifts outward along the predetermined propagation path, it is determined to be a path-spreading state. When the same unified analysis sample simultaneously meets both the concentrated abnormal state in the source area and the path-spreading state, a combined primary and secondary approach is used to output the spatial feature determination results, where the path-spreading state is retained as the primary spatial feature state, and the concentrated abnormal state in the source area is retained as the secondary spatial feature state.

[0079] The preset normal range for key monitoring indicators is a normal fluctuation range determined before the start of the current analysis period, obtained based on statistical results of historical normal monitoring data, regional background values, standard limits, or empirical calibration results. The predetermined propagation path direction is determined based on the topographic slope aspect, surface runoff direction, groundwater flow direction, or historical pollution migration paths of the study area.

[0080] It should be noted that after completing the determination of temporal and spatial characteristics, the spatiotemporal attribution analysis of each key monitoring indicator is further conducted by combining the hydrological and meteorological indicators, spatial location-related indicators, and associated response information in the unified analysis sample.

[0081] Specifically, a rule-based determination method is adopted. When the temporal contribution of a key monitoring indicator is higher than a preset temporal contribution threshold, and its spatial centrality is lower than a preset spatial centrality threshold, while the temporal characteristic state corresponding to the indicator shows an increasing state or a continuous abnormal state, it is determined that it is mainly driven by temporal changes. When the spatial centrality of a key monitoring indicator is higher than a preset spatial centrality threshold, and its spatial characteristic state shows a concentrated abnormal state in the source region or an expansion state along the path, it is determined that it is mainly driven by spatial propagation. When a key monitoring indicator simultaneously meets the aforementioned temporal change-driven condition and spatial propagation-driven condition, it is determined that it originates from the combined effect of temporal change and spatial propagation.

[0082] The preset time-series contribution threshold and the preset spatial centrality threshold can be determined based on the quantiles of historical sample distributions, empirical calibration results, or verification results.

[0083] In this embodiment of the invention, by performing temporal and spatial feature determination, as well as spatiotemporal attribution analysis on each unified analysis sample, the limitations of traditional methods that rely solely on absolute numerical values ​​of indicators for simple classification are overcome. At the temporal level, judgment thresholds are dynamically set based on historical fluctuations, and continuous abnormal states are identified, effectively avoiding misjudgments caused by random fluctuations within a single analysis period. At the spatial level, the pollution source and diffusion direction are accurately located through unit division of source area, path, and periphery, and identification of abnormal expansion trajectories. Finally, attribution analysis is performed by combining temporal contribution and spatial centrality, clarifying whether the abnormal changes of each key monitoring indicator originate from time-driven, spatial propagation, or a combination of both. This provides an interpretable basis for the subsequent construction of dynamic risk baselines and early warning determination, significantly improving the accuracy and predictability of water and soil environmental evolution monitoring in antimony mining areas.

[0084] S5: Verify the set of key monitoring indicators and determine the set of target monitoring indicators to be used in the water and soil environment assessment of the antimony mining area based on the verification results.

[0085] In one possible implementation, S5 specifically includes: The key monitoring indicators selected were validated, including cross-validation based on time windows, permutation tests, and generalization tests.

[0086] Specifically, a sliding time window cross-validation method is used to evaluate the stability of the importance ranking of key monitoring indicators. The time window width and step size are set, and corresponding time windows are generated on the complete monitoring time series. Preferably, the time window width is set to 36 months, the step size is 6 months, and the corresponding time window is 13. Within each time window, random forest modeling and SHAP importance calculation are repeatedly performed based on the samples within the window, and the importance ranking results of the key monitoring indicators under that time window are recorded. Subsequently, a rank matrix is ​​constructed based on the ranking results of the key monitoring indicators under each time window, and the Kendall coordination coefficient is calculated based on the rank and dispersion of each key monitoring indicator in different time windows, denoted as . This is used to quantify the consistency in the ranking of the importance of key monitoring indicators across different time windows. Among them, The value ranges from 0 to 1, with a larger value indicating higher sorting consistency.

[0087] Kendall's consistency coefficient is a nonparametric statistic used to measure the consistency of ranking results for the same group of objects by multiple raters or multiple evaluation indicators.

[0088] Furthermore, a test statistic is constructed based on the Kendall consistency coefficient, and the significance test probability is calculated using the chi-square distribution under the corresponding degrees of freedom, denoted as . .when When the value is less than the preset significance level, it indicates that the consistency of the ranking is statistically significant. In a specific embodiment, When, corresponding This indicates that the ranking of key monitoring indicators is highly consistent across different time windows and exhibits good temporal stability.

[0089] Furthermore, after completing the sliding time window cross-validation, the permutation test was used to verify the statistical significance of the selected key monitoring indicators.

[0090] Specifically, feature samples are reconstructed based on key monitoring indicators. While maintaining the prediction target, sample partitioning method, and random forest parameter configuration, the random forest regression model is retrained to obtain a validation model for permutation testing. For each key monitoring indicator, while keeping other key monitoring indicators and target output unchanged, the correspondence between the key monitoring indicator and each sample is randomly shuffled. The shuffled samples are then input into the validation model, and the coefficient of determination is recalculated on the same validation set. The difference between the coefficient of determination of the validation model before and after the permutation is defined as the model performance loss of that indicator under this permutation. After repeating the random permutation a preset number of times, the permutation loss distribution corresponding to that indicator is obtained.

[0091] Furthermore, the proportion of permutations in the permutation loss that are greater than or equal to the actual model performance loss is defined as the probability value corresponding to the permutation test. When the actual model performance loss corresponding to each key monitoring indicator falls within the preset high-loss region of its respective permutation loss distribution, the corresponding... When the probability is less than the preset probability threshold, it indicates that the key monitoring indicators are all statistically significant and can be used as key monitoring indicators in the monitoring of soil and water environment evolution in antimony mining areas. Preferably, when the preset number of times is 1000, the preset high-loss area is the first... ,correspond .

[0092] Furthermore, a subset of independent samples, randomly reserved from the total sample and not used in training, is designated as the validation dataset. These samples are input into the model constructed above, and the model outputs a predicted value for each sample. The predicted values ​​are then compared with the actual sample data to calculate the coefficient of determination, which quantifies the model's fit on the independent validation dataset. The threshold for the coefficient of determination is determined based on the prediction consistency requirements on the independent validation data and the stability requirements of subsequent time-series analysis, spatial analysis, and attribution validation. A coefficient of determination is calculated if and only if the coefficient of determination on the independent validation dataset is... If the coefficient of determination is greater than the preset coefficient of determination, the indicator is considered to have good generalization ability and is thus deemed effective. If the coefficient of determination... If the value is less than or equal to a preset determination coefficient, it is considered invalid. In one specific embodiment, the preset determination coefficient is... Set it to 0.85.

[0093] It should be noted that when the significance test probability values ​​are simultaneously satisfied in different time windows... The probability value corresponding to the permutation test is less than the preset significance level. The coefficient of determination of the independent validation set is less than the preset probability threshold. When the coefficient of determination is greater than the preset coefficient of determination, the key monitoring indicators show high consistency in ranking across different time windows, have statistical significance, and possess good generalization ability. The system determines that the indicator is the target monitoring indicator and identifies the target monitoring indicator set used for the water and soil environment assessment of the antimony mining area.

[0094] In this embodiment of the invention, a triple validation mechanism—sliding time window cross-validation, permutation test, and generalization test—is used to comprehensively evaluate the selected key monitoring indicators. Cross-validation utilizes the Kendall coefficient of consistency to quantify the consistency of indicator importance ranking across different time periods, ensuring the temporal stability of the screening results. The permutation test verifies whether each indicator has a statistically significant independent contribution by randomly shuffling the correspondence between indicators and targets, avoiding misselection due to accidental correlations. The generalization test uses independent samples to validate the model's predictive performance, ensuring the transferability of indicators across different datasets. This triple validation mechanism collectively guarantees the stability, significance, and generalization ability of the final target monitoring indicator set under different time windows and sample conditions, significantly improving the reliability and engineering application value of the screening results.

[0095] In one possible implementation, steps S6 to S8 are included after S5: S6: Based on the target monitoring indicator set, construct a dynamic risk baseline for key monitoring indicators corresponding to a unified analysis sample.

[0096] In one possible implementation, S6 specifically includes sub-steps S601 and S602: S601: For each unified analysis sample, based on its corresponding time label, spatial label, set of key monitoring indicators, time series characteristic judgment results, spatial characteristic judgment results, and spatiotemporal attribution results, construct a dynamic risk baseline for the key monitoring indicators corresponding to the unified analysis sample.

[0097] Specifically, after completing the determination of temporal features, spatial features, and spatiotemporal attribution, for each unified analysis sample, based on its corresponding time label, spatial label, set of key monitoring indicators, temporal feature determination results, spatial feature determination results, and spatiotemporal attribution results, a dynamic risk baseline for the key monitoring indicators corresponding to the unified analysis sample is constructed, and an early warning determination is performed accordingly.

[0098] It should be noted that the dynamic risk baseline is not a fixed threshold uniform across the entire region, but rather determined separately for different spatial units, different analysis stages, different abnormal change states, and different driving conditions. For any key monitoring indicator, reference samples with similar or identical spatial unit attributes, analysis stage attributes, hydrological and meteorological conditions, and temporal characteristics to the current unified analysis sample are selected from historical normal state samples. When the number of historical normal state samples is insufficient, samples verified as low-risk states can be used as auxiliary reference samples. Then, based on the reference samples, the normal fluctuation center and normal fluctuation range of the key monitoring indicator under the corresponding state are statistically analyzed and determined as the dynamic risk baseline of the key monitoring indicator under the current unified analysis sample. The normal fluctuation center can be determined using the mean, median, or weighted mean, and the normal fluctuation range can be determined using standard deviation intervals, quantile intervals, interquartile ranges, or empirical tolerance intervals.

[0099] The temporal characteristic states include at least rising, falling, stable, and persistent anomalous states. Spatial unit attributes include at least source area units, path units, and peripheral units. Analysis stage attributes include at least the seasonal stage, hydrological stage, or continuous change stage to which the current analysis period belongs. Spatial characteristic states include at least concentrated anomalous states in the source area and states spreading along the path. Similar hydrological and meteorological conditions include at least one or more of the following: rainfall level, groundwater level change, temperature level, and evaporation conditions. Through the above methods, the dynamic risk baseline can reflect the natural fluctuation characteristics and reasonable change boundaries of key monitoring indicators under different regions, stages, and anomalous change states.

[0100] S602: For any key monitoring indicator, select a baseline reference sample set from the same or similar spatial unit attributes, the same or similar analysis stage attributes, similar hydrological and meteorological conditions, and historical normal or low-risk state samples. Based on the baseline reference sample set, statistically analyze the normal fluctuation center and normal fluctuation range of the key monitoring indicator under the corresponding state, and determine it as the dynamic risk baseline of the key monitoring indicator under the current unified analysis sample.

[0101] Specifically, after obtaining the dynamic risk baselines for each key monitoring indicator, the actual state of each key monitoring indicator in the current unified analysis sample is compared with its respective dynamic risk baseline. Early warning judgments are then made based on the degree of deviation, persistence, spatial expansion, and multi-indicator consistency. The degree of deviation characterizes how much the current key monitoring indicator exceeds the dynamic risk baseline. The persistence characterizes whether the deviation persists throughout the continuous analysis period. The spatial expansion characterizes whether the deviation extends from the source unit to the path unit or peripheral unit, or whether it propagates continuously along a predetermined propagation path. Multi-indicator consistency is characterized by the proportion of key monitoring indicators exhibiting abnormal deviations in the same direction within the same unified analysis sample to the total number of key monitoring indicators. When this proportion exceeds a preset consistency threshold, it is determined to be an abnormality in multi-indicator consistency. In one specific embodiment, the preset consistency threshold is set to 50%.

[0102] In this embodiment of the invention, a dynamic risk baseline is constructed, overcoming the limitations of traditional fixed threshold assessment methods. This baseline, tailored to different spatial units, analysis stages, temporal characteristics, and driving conditions, selects reference samples with similar spatiotemporal attributes from historical normal samples, statistically analyzing their normal fluctuation centers and ranges to achieve differentiated and adaptive baseline construction. By comparing the actual state with this dynamic baseline and combining the degree of deviation, persistence, spatial expansion, and consistency of multiple indicators, early warning judgments are made, significantly improving the accuracy of anomaly identification and the targeting of early warnings, effectively avoiding misjudgments caused by regional differences or stage changes.

[0103] S7: Based on the dynamic risk baseline of key monitoring indicators, compare the actual status of each key monitoring indicator in the current unified analysis sample with its respective dynamic risk baseline, and make early warning judgments in combination with deviation, persistence, spatial expansion status and consistency of multiple indicators.

[0104] Specifically, the early warning determination is carried out according to the following rules: When the deviation of each key monitoring indicator from its respective dynamic risk baseline is within the normal fluctuation range, and there is no sustained abnormality, spatial expansion, or consistent abnormality of multiple indicators, it is determined to be in a stable state. When at least one key monitoring indicator deviates from its dynamic risk baseline beyond the normal fluctuation range, but the deviation has not been sustained continuously and has not shown significant spatial expansion, it is determined to be in a state of concern. When one or more key monitoring indicators deviate from their respective dynamic risk baselines for multiple consecutive analysis periods, it is determined to be in an early warning state. When multiple key monitoring indicators deviate from the dynamic risk baseline beyond the normal fluctuation range for multiple consecutive analysis periods, and simultaneously meet the conditions of sustained abnormality, enhanced spatial expansion, and consistent abnormality of multiple indicators, it is determined to be in a high-risk early warning state.

[0105] For example, multiple consecutive analysis periods refer to two or three consecutive analysis periods. Significant deviation refers to a deviation exceeding 1.5 times or more of the upper or lower limit of the dynamic risk baseline fluctuation range. Consistent anomalies among multiple indicators refer to at least two or more key monitoring indicators showing abnormal deviations in the same direction in the same unified analysis sample, and the corresponding abnormal states being synchronous in time or space.

[0106] It should be noted that, to ensure consistency between the early warning judgment and the judgment results of S6, the system further differentiates the judgment focus under different states based on the temporal feature judgment results, spatial feature judgment results, and spatiotemporal attribution results: When the temporal feature judgment result is a continuous abnormal state and the spatial feature judgment result is a concentrated abnormal state in the source region, increased attention is paid to the continuous increase of key monitoring indicators within the source region unit. When the spatial feature judgment result is a path-spreading state, increased attention is paid to the continuous increase of key monitoring indicators in the path unit and peripheral units, and their outward movement along the predetermined propagation path. When the temporal feature judgment result is a stable state and the spatial feature judgment result does not show obvious expansion, the focus is on maintaining background stability, reducing the sensitivity to short-term slight fluctuations in anomaly judgment.

[0107] In this embodiment of the invention, a tiered early warning judgment mechanism is constructed based on a dynamic risk baseline. Risk levels are divided into four levels—stable state, watchful state, early warning state, and high-risk early warning state—according to the degree of deviation, persistence, spatial expansion, and consistency of multiple indicators. It not only focuses on short-term deviations of single indicators but also comprehensively considers whether the deviation is continuous, whether it expands along a spatial path, and whether multiple indicators are synchronously abnormal, effectively avoiding false alarms caused by single-point or single-event fluctuations. Simultaneously, the early warning judgment also incorporates differentiated processing based on temporal feature judgment results, spatial feature judgment results, and spatiotemporal attribution results, employing different focuses for different scenarios such as persistent anomalies in the source region and expansion along a path, significantly improving the accuracy and scenario adaptability of the early warning.

[0108] S8: Based on the early warning judgment result, perform joint correction to obtain the correction result.

[0109] In one possible implementation, S8 specifically refers to: Based on the early warning judgment result, subsequent actual monitoring results, and verification results, determine whether feedback correction is triggered. If so, output the correction result through the feedback correction model, and based on the correction result, jointly correct the key monitoring indicator screening conditions, time series characteristic judgment conditions, spatial characteristic judgment conditions, spatiotemporal attribution analysis parameters, dynamic risk baseline reference sample range, and early warning judgment conditions to obtain the correction result.

[0110] The verification results should include at least the stability verification results, the statistical significance verification results, and the transfer applicability verification results.

[0111] Specifically, after S7 completes the construction of dynamic risk baseline and early warning determination, in order to improve the stability and applicability of the key monitoring indicator screening results, time series characteristic determination results, spatial characteristic determination results, spatiotemporal attribution results, and dynamic risk baseline construction results, S8 is added after S7 to form a closed-loop update mechanism that is interconnected before and after.

[0112] Specifically, the system first determines whether to trigger feedback correction based on the early warning judgment result, subsequent actual monitoring results, cross-validation results, permutation test results, and generalization test results. After triggering feedback correction, the system outputs the corresponding correction result through the feedback correction model. The inputs to the feedback correction model include: the set of key monitoring indicators for the current round, time-series feature judgment results, spatial feature judgment results, spatiotemporal attribution results, dynamic risk baseline results, early warning judgment results, cross-validation results, permutation test results, generalization test results, subsequent actual monitoring results, and corresponding parameter configurations, including the weights of key monitoring indicators, time-series feature judgment thresholds, continuous anomaly judgment conditions, spatial feature judgment parameters, spatiotemporal attribution parameters, dynamic risk baseline construction parameters, and early warning judgment rule parameters.

[0113] The output of the feedback correction model includes: correction results of the key monitoring indicator set, correction results of the weights of key monitoring indicators, correction results of time series feature judgment parameters, correction results of spatial feature judgment parameters, correction results of spatiotemporal attribution parameters, correction results of the dynamic risk baseline reference sample range, and correction results of early warning judgment rules.

[0114] Furthermore, the training data for the feedback correction model includes multiple feedback correction input samples constructed based on historical operational data, and correction labels corresponding to each feedback correction input sample. The correction labels are used to characterize the correction actions or corrected parameter results to be taken in the corresponding historical scenario. They are preferably determined based on subsequent actual monitoring results and can be supplemented and corrected by combining historical validation results or manual annotation results. The subsequent actual monitoring results include at least the actual changes, spatial expansion results, and final risk status of the key monitoring indicators corresponding to the current unified analysis sample within the subsequent preset quantitative analysis period.

[0115] It should be noted that after obtaining the correction results output by the feedback correction model, the system updates the set of key monitoring indicators, time series feature judgment parameters, spatial feature judgment parameters, spatiotemporal attribution parameters, dynamic risk baseline reference sample range, and early warning judgment rules, and uses the updated parameters and rules for the re-judgment of subsequent unified analysis samples.

[0116] Furthermore, the revisions should include at least one or more of the following: prioritization of the selection of key monitoring indicator sets; revision of the judgment threshold or continuous anomaly judgment conditions in time-series feature determination; revision of spatial unit comparison conditions or expansion judgment conditions in spatial feature determination; revision of the weighting of each key monitoring indicator in spatiotemporal attribution analysis; revision of the selection range of the baseline reference sample set during dynamic risk baseline construction; and revision of the judgment conditions for deviation degree, persistence status, spatial expansion status, and consistency of multiple indicators in early warning determination.

[0117] Through feedback correction in S8, the key monitoring indicator set, time-series characteristic judgment results, spatial characteristic judgment results, spatiotemporal attribution results, dynamic risk baseline, and early warning judgment results can be continuously updated based on subsequent verification results. This enables the entire antimony mining area's water and soil environment assessment indicator screening and early warning process to form a closed-loop optimization, improving the applicability, stability, and interpretability of the intelligent screening method for antimony mining area water and soil environment assessment indicators in different regions, at different stages, and under different abnormal change states.

[0118] In this embodiment of the invention, a feedback correction mechanism is introduced to achieve closed-loop optimization of the screening and early warning process for water and soil environmental assessment indicators in antimony mining areas. Based on the early warning judgment results, subsequent actual monitoring results, and verification results such as cross-validation, substitution testing, and generalization testing, it is determined whether feedback correction is triggered. Once triggered, the feedback correction model jointly corrects the screening conditions for key monitoring indicators, time-series and spatial characteristic judgment parameters, spatiotemporal attribution analysis parameters, the reference sample range for dynamic risk baseline, and the early warning judgment rules. By feeding the correction results back to the front-end steps and using them for the re-judgment of subsequent unified analysis samples, a continuously iterative closed-loop update mechanism is formed, significantly improving the applicability, stability, and interpretability of the entire intelligent screening method in different regions, different pollution stages, and different abnormal change states.

[0119] Reference manual attached Figure 2 The diagram shows a schematic of the structure of an intelligent screening system for water and soil environment assessment indicators in antimony mining areas provided by the present invention.

[0120] This invention also provides an intelligent screening system 20 for water and soil environment assessment indicators in antimony mining areas, applied to the aforementioned intelligent screening method for water and soil environment assessment indicators in antimony mining areas, comprising: Processor 201.

[0121] The memory 202 stores computer-readable instructions. When the computer-readable instructions are executed by the processor 201, the intelligent screening method for water and soil environmental assessment indicators in the antimony mining area, as described in the method embodiment, is implemented.

[0122] The intelligent screening system 20 for water and soil environment assessment indicators in antimony mining areas provided by this invention can execute the intelligent screening method for water and soil environment assessment indicators in antimony mining areas described above and achieve the same or similar technical effects. To avoid duplication, this invention will not elaborate further.

[0123] It should be understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0124] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0125] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0126] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0127] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0128] It should be understood that, in various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0129] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0130] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0131] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0133] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0134] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the intelligent screening method for water and soil environmental assessment indicators in antimony mining areas as described in the method embodiments.

[0136] The present invention provides a computer-readable storage medium that can realize the steps and effects of the intelligent screening method for water and soil environment assessment indicators in antimony mining areas as described in the above method embodiments. To avoid repetition, the present invention will not repeat the details.

[0137] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0138] The following points need to be explained: (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0139] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0140] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0141] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligently screening water and soil environment assessment indicators in antimony mining areas, characterized in that, include: S1: Obtain the original dataset of water and soil environment monitoring in the antimony mining area, and preprocess the original dataset to obtain the preprocessed result dataset; S2: Extract candidate monitoring indicators from the preprocessed result dataset, evaluate the importance of the candidate monitoring indicators by using a machine learning model and combining the target response characterization quantity in the antimony mining area's soil and water environment, and quantify the contribution of each indicator to the target response characterization quantity using a contribution interpretation model to obtain highly important monitoring indicators. Specifically, S2 includes: S201: Extract the candidate monitoring indicators related to the evolution of the water and soil environment in the antimony mining area from the preprocessed dataset, and construct a random forest regression model using the characterization information related to the evolution of antimony content as the response variable; S202: Based on the random forest regression model, the importance of each candidate monitoring indicator is evaluated to obtain the importance evaluation results; S203: Combining the SHAP value, analyze and interpret the contribution of the importance assessment results to each candidate monitoring indicator to obtain the contribution interpretation results; S204: Based on the importance assessment results and the contribution interpretation results, a weighted summation is performed to obtain the importance score; S205: Retain the monitoring indicators whose importance scores are higher than a preset threshold as the high-importance monitoring indicators; S3: Combining the time series analysis model and the spatial correlation analysis model, the time series evolution characteristics and spatial correlation characteristics of the high-importance monitoring indicators are identified to obtain the time series characteristic results and spatial characteristic results corresponding to each high-importance monitoring indicator; Specifically, the temporal evolution feature recognition process in S3 is as follows: For the sequence of high-importance monitoring indicators, a time series model is performed using a bidirectional long short-term memory network, and the time series evolution features and key time period features corresponding to each high-importance monitoring indicator are extracted based on the output of the bidirectional long short-term memory network. Specifically, the spatial association feature recognition process in S3 is as follows: A spatial correlation map between monitoring points in the antimony mining area is constructed, and a graph convolutional network (GCN) model is established to learn features of the spatial correlation map to obtain the spatial embedding representation corresponding to each monitoring point. Based on the spatial embedding representation, the clustering characteristics, propagation characteristics, and spatial response characteristics of each high-importance monitoring indicator in the spatial dimension are identified; S4: Based on the temporal and spatial feature results, the highly important monitoring indicators are spatiotemporally screened to obtain a preferred set of key monitoring indicators; S5: Verify the set of key monitoring indicators and determine the set of target monitoring indicators to be used in the water and soil environment assessment of the antimony mining area based on the verification results.

2. The intelligent screening method for water and soil environment assessment indicators in antimony mining areas according to claim 1, characterized in that, Specifically, S5 is: The key monitoring indicators selected through screening are validated, including cross-validation based on time windows, permutation tests, and generalization tests.

3. The intelligent screening method for water and soil environment assessment indicators in antimony mining areas according to claim 1, characterized in that, The process includes the following steps after S4 and before S5: For each unified analysis sample, time-series feature determination, spatial feature determination, and spatiotemporal attribution analysis are performed on each key monitoring indicator in the set of key monitoring indicators, and the time-series feature determination results, spatial feature determination results, and spatiotemporal attribution results corresponding to each key monitoring indicator are output for the unified analysis sample.

4. The intelligent screening method for water and soil environment assessment indicators in antimony mining areas according to claim 1, characterized in that, Following S5, it also includes: S6: Based on the target monitoring indicator set, construct a dynamic risk baseline for key monitoring indicators corresponding to the unified analysis sample; S7: Based on the dynamic risk baseline of the key monitoring indicators, compare the actual status of each key monitoring indicator in the current unified analysis sample with its respective dynamic risk baseline, and perform early warning judgment in combination with deviation, continuous status, spatial expansion status and consistency of multiple indicators. S8: Based on the early warning judgment result, perform joint correction to obtain the correction result.

5. The intelligent screening method for water and soil environment assessment indicators in antimony mining areas according to claim 4, characterized in that, S6 specifically includes: S601: For each unified analysis sample, based on its corresponding time label, spatial label, set of key monitoring indicators, time series characteristic judgment result, spatial characteristic judgment result and spatiotemporal attribution result, construct the dynamic risk baseline of the key monitoring indicators corresponding to the unified analysis sample. S602: For any key monitoring indicator, a baseline reference sample set is selected from the same or similar spatial unit attributes, the same or similar analysis stage attributes, similar hydrological and meteorological conditions, and historical normal state samples or low-risk state samples. Based on the baseline reference sample set, the normal fluctuation center and normal fluctuation range of the key monitoring indicator under the corresponding state are statistically analyzed and determined as the dynamic risk baseline of the key monitoring indicator under the current unified analysis sample.

6. The intelligent screening method for water and soil environment assessment indicators in antimony mining areas according to claim 4, characterized in that, Specifically, S8 is: Based on the early warning judgment results, subsequent actual monitoring results, and verification results, determine whether to trigger feedback correction; If so, the correction result is output through the feedback correction model, and the key monitoring indicator screening conditions, time series characteristic judgment conditions, spatial characteristic judgment conditions, spatiotemporal attribution analysis parameters, dynamic risk baseline reference sample range and early warning judgment conditions are jointly corrected based on the correction result to obtain the correction result. The verification results include at least stability verification results, statistical significance verification results, and migration applicability verification results.

7. An intelligent screening system for water and soil environment assessment indicators in antimony mining areas, characterized in that, include: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement an intelligent screening method for water and soil environmental assessment indicators in antimony mining areas as described in any one of claims 1 to 6.

Images