Method for identifying groundwater pollution source based on three-dimensional fluorescence spectrum and neural network
By combining three-dimensional fluorescence spectroscopy with neural networks, and utilizing tensor decomposition and weighted neural networks, the problems of time-consuming, labor-intensive, and inaccurate methods in traditional approaches are solved, enabling rapid and accurate identification of groundwater pollution sources and continuous spatial distribution representation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TECH CENT FOR SOIL AGRI & RURAL ECOLOGY & ENVIRONMENT MINIST OF ECOLOGY & ENVIRONMENT
- Filing Date
- 2026-04-30
- Publication Date
- 2026-07-07
Smart Images
Figure CN122109040B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of groundwater pollution source identification, and in particular to a groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural networks. Background Technology
[0002] Groundwater is a vital freshwater resource for human survival, widely used for agricultural irrigation, industrial production, and residential life. Once groundwater is polluted, remediation is often lengthy and costly, and the extent of pollution is difficult to pinpoint. Therefore, effectively identifying groundwater pollution sources has become a crucial aspect of environmental monitoring and management.
[0003] Traditional methods for identifying groundwater pollution sources, such as source tracing modeling, isotope tracing, and geochemical analysis, while providing information on pollution sources to some extent, generally suffer from the following problems: First, they rely on extensive on-site sampling and testing, which is time-consuming and labor-intensive, making rapid response difficult, and the accuracy of source tracing is insufficient; second, model construction is complex, feature extraction depends on human experience, and has high requirements for data integrity and prior knowledge, often subject to various constraints in practical applications; third, for real-world scenarios with numerous pollutant types and complex pollution pathways, they often struggle to achieve efficient and accurate pollution source identification and address the problem of discontinuous spatial distribution representation. Therefore, there is an urgent need to invent a novel pollution source identification method with high sensitivity, high recognition rate, and high degree of automation.
[0004] Because three-dimensional fluorescence spectroscopy can capture the characteristics of trace organic pollutants and neural networks can mine nonlinear relationships in data, the pollution source identification model constructed by combining three-dimensional fluorescence spectroscopy with neural networks can achieve real-time pollution source identification. It is suitable for deployment in online monitoring systems and supports dynamic water quality monitoring. Compared with traditional physical modeling methods, neural networks rely on data-driven approaches, reducing the need for theoretical modeling of pollution diffusion processes. They can simultaneously identify multiple types of pollution sources and are suitable for actual groundwater environments with complex compositions and many interfering factors. It is a product of the cross-integration of environmental science, artificial intelligence, and spectral analysis, with significant theoretical value and application prospects, representing an important development direction for future intelligent environmental monitoring. Summary of the Invention
[0005] This invention provides a groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural networks to solve the technical problems of fluorescence feature analysis relying on subjective judgment and lacking objective quantitative decomposition methods; pollution source classification models lacking adaptive differentiation ability for the contribution importance of different fluorescence components; and the discretization, lack of continuity and intuitiveness in the spatial distribution representation of pollution sources.
[0006] The groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural networks of the present invention includes the following steps:
[0007] S1. Groundwater samples were collected at the sampling point and fluorescence spectroscopy was performed to obtain a three-dimensional fluorescence excitation-emission matrix. Tensor decomposition was performed on the three-dimensional fluorescence excitation-emission matrix to obtain the excitation spectrum intensity, emission spectrum intensity, and relative concentration score of the groundwater sample on each component.
[0008] S2. Based on the relative concentration scores of each component in the groundwater sample, a weighted single-hidden-layer feedforward neural network classification algorithm based on a pollution component gating mechanism is introduced. The algorithm includes a gating factor calculation stage, a hidden-layer activation vector calculation stage, and an output layer probability calculation and training optimization stage. In the gating factor calculation stage, a global statistical analysis is performed on the relative concentration scores of all components in the groundwater sample. For each component, the average relative concentration score and the standard deviation of the average relative concentration score are calculated. Based on the average relative concentration score and the standard deviation of the average relative concentration score of the component, a gating factor is generated. The gating factor is used in the hidden-layer activation vector calculation stage and the output layer probability calculation and training optimization stage to calculate the pollution source prediction results of the groundwater sample.
[0009] S3. Based on the known location of the sampling point and the pollution source prediction results of the groundwater sample, calculate the probability that the spatial location to be estimated belongs to the target pollution source type, and generate a continuous spatial probability distribution field.
[0010] Preferably, the three-dimensional fluorescence excitation-emission matrix is decomposed by tensor analysis using parallel factor analysis to output the excitation spectrum intensity, emission spectrum intensity, and relative concentration scores of the groundwater sample for each component.
[0011] Preferably, in the hidden layer activation vector calculation stage:
[0012] Based on the relative concentration scores of all components in a single groundwater sample, a column vector of component concentrations for the groundwater sample is constructed; based on the gating factors of all components, a column vector of gating factors is constructed; based on the column vectors of component concentrations and gating factors of the groundwater sample, combined with the hidden layer weight matrix and the hidden layer bias vector, the hidden layer activation vector of the groundwater sample is calculated.
[0013] Preferably, during the prediction process in the output layer probability calculation and training optimization phase:
[0014] By introducing an output layer weight matrix, the hidden layer activation vector of the groundwater sample is mapped to the pollution source score. Combined with the output layer bias vector, the pollution source prediction probability of the groundwater sample is calculated. The pollution source category corresponding to the maximum value of the pollution source prediction probability of the groundwater sample is taken as the pollution source prediction result of the groundwater sample.
[0015] Preferably, during the training process of output layer probability calculation and training optimization:
[0016] Based on the relative concentration scores of the components in the groundwater sample, the pollution intensity weight of the groundwater sample is calculated. Based on the predicted probability of the pollution source and the pollution intensity weight of the groundwater sample, combined with the true category, a weighted cross-entropy loss function is constructed and iteratively optimized to update the hidden layer weight matrix, hidden layer bias vector, output layer weight matrix and output layer bias vector until the loss converges.
[0017] Preferably, the probability that the spatial location to be estimated belongs to the target pollution source type is calculated by the following method:
[0018] Iterate through all known sampling points, and based on the pollution source prediction results of groundwater samples, combine the type matching indicator function to retain sampling points that are determined to belong to the target pollution source type. For each retained sampling point, combine the spatial Euclidean distance between the known sampling point and the spatial location to be estimated and the pollution intensity weight of the groundwater sample to calculate the probability that the spatial location to be estimated belongs to the target pollution source type.
[0019] The beneficial effects of the technical solution of the present invention are:
[0020] 1. The three-dimensional fluorescence excitation-emission matrix of groundwater samples was obtained by a three-dimensional fluorescence spectrometer, and tensor decomposition was performed by parallel factor analysis. This effectively decomposes complex high-dimensional spectral data into excitation spectrum intensity, emission spectrum intensity, and relative concentration score with clear physical meaning. This not only improves the efficiency of data dimensionality reduction, but also enables the fine extraction of spectral characteristics of pollutants, which facilitates subsequent analysis and pollution feature identification.
[0021] 2. By designing a weighted single-hidden-layer feedforward neural network based on a pollution component gating mechanism, the statistical characteristics of component concentration are utilized to dynamically generate gating factors, thereby strengthening key pollution features and suppressing secondary features. This effectively enhances the feature discrimination ability of the neural network in complex pollution scenarios, reduces noise interference, and improves the generalization ability of classification.
[0022] 3. By introducing a weighted cross-entropy loss function based on the total intensity of groundwater sample components during the output layer probability calculation and training optimization phase, groundwater samples with higher pollution levels are given higher weights, which can focus more on heavily polluted areas and improve the sensitivity to high-risk pollution sources.
[0023] 4. In the prediction process of output layer probability calculation and training optimization, the probability of any spatial location belonging to the target pollution source type is estimated by using the known sampling point location and the pollution source prediction results of groundwater samples, combined with the distance inverse weighting strategy. Finally, a continuous spatial probability distribution field is generated, avoiding subjective assumptions about the specific location of the pollution source, and has strong objectivity and adaptability. Attached Figure Description
[0024] Figure 1 This is a flowchart of the groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural networks described in this invention. Detailed Implementation
[0025] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0027] The specific scheme of the groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural network provided by the present invention will be described in detail below with reference to the accompanying drawings.
[0028] See attached document Figure 1 The diagram illustrates a flowchart of a groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural networks, according to an embodiment of the present invention. The method includes the following steps:
[0029] S1. Groundwater samples were collected at the sampling point and fluorescence spectroscopy was performed to obtain a three-dimensional fluorescence excitation-emission matrix. Tensor decomposition was performed on the three-dimensional fluorescence excitation-emission matrix to obtain the excitation spectrum intensity, emission spectrum intensity, and relative concentration score of the groundwater sample on each component.
[0030] Within the study area, sampling points were scientifically deployed based on the distribution of potential pollution sources, groundwater flow direction, hydrogeological conditions, and historical pollution records. The geographic coordinates of each sampling point were recorded using GPS equipment. ,in, This system represents the groundwater sample index and records auxiliary information such as sampling depth and elevation to ensure the accuracy of subsequent spatial tracing. A dedicated groundwater sampler, such as a submersible pump or peristaltic pump, is used to extract groundwater samples from monitoring wells or boreholes. During sampling, contact with air is avoided to prevent oxidation. The collected groundwater samples are immediately placed in brown glass or polyethylene bottles and stored in the dark at a low temperature (4°C). Within 24 hours, fluorescence spectra are scanned using a three-dimensional fluorescence spectrometer to obtain the three-dimensional fluorescence excitation-emission matrix (EEM). The specific expression of the three-dimensional fluorescence excitation-emission matrix is as follows:
[0031] ,
[0032] in, Indicates the first A three-dimensional fluorescence excitation-emission matrix (EEM) of a groundwater sample is used to store the raw fluorescence intensity data of the groundwater sample; Indicates the first A groundwater sample at the excitation wavelength and emission wavelength The original fluorescence intensity value; Indicates the excitation wavelength index; Indicates the number of samples at the excitation wavelength; Indicates the transmit wavelength index; Indicates the number of samples at the transmitted wavelength;
[0033] Furthermore, parallel factor analysis was used to perform tensor decomposition on the three-dimensional fluorescence excitation-emission matrix to obtain the excitation spectrum intensity, emission spectrum intensity, and relative concentration score of each groundwater sample for each component. The formulas are as follows:
[0034] ,
[0035] in, This represents the summation over all independent fluorescent components; The number of components is represented by the iterative optimization process of the parallel factor analysis algorithm; Indicates the fluorescent component index; Indicates the first Each component at the excitation wavelength The intensity of the excitation spectrum below; Indicates the first Each component at the emission wavelength The emission spectrum intensity below; Indicates the first The first groundwater sample The relative concentration scores of each component; Indicates the first A groundwater sample at the excitation wavelength and emission wavelength The fitting residuals are obtained through the iterative optimization process of the parallel factor analysis algorithm.
[0036] S2. Based on the relative concentration scores of each component in the groundwater sample, a weighted single-hidden-layer feedforward neural network classification algorithm based on a pollution component gating mechanism is introduced. The algorithm includes a gating factor calculation stage, a hidden-layer activation vector calculation stage, and an output layer probability calculation and training optimization stage. In the gating factor calculation stage, a global statistical analysis is performed on the relative concentration scores of all components in the groundwater sample. For each component, the average relative concentration score and the standard deviation of the average relative concentration score are calculated. Based on the average relative concentration score and the standard deviation of the average relative concentration score of the component, a gating factor is generated. The gating factor is used in the hidden-layer activation vector calculation stage and the output layer probability calculation and training optimization stage to calculate the pollution source prediction results of the groundwater sample.
[0037] The weighted single-hidden-layer feedforward neural network classification algorithm based on the pollutant component gating mechanism consists of three main stages: gating factor calculation stage, hidden layer activation vector calculation stage, and output layer probability calculation and training optimization stage.
[0038] In the gating factor calculation stage, a global statistical analysis is performed on the relative concentration scores of components in all groundwater samples. Specifically, the average relative concentration score of each component in all groundwater samples is calculated, and the median is extracted from the average relative concentration scores of all components as the threshold value. At the same time, the standard deviation of the average relative concentration score is calculated to quantify the dispersion of the contribution intensity among components. Based on the standard deviation of the average relative concentration score, the steepness parameter of the gating function is determined. The larger the dispersion, the smoother the transition of the gating function; the smaller the dispersion, the steeper the transition of the gating function, ensuring that the gating factor can effectively distinguish important components from minor components under different pollution scenarios. The deviation between the average relative concentration score of each component and the threshold value is mapped to the interval of 0 to 1 using a sigmoid function to generate the gating factor. When the gating factor value is close to 1, it indicates that the component contribution is significantly higher than the central level and belongs to the key pollution feature. When the gating factor value is close to 0, it indicates that the component contribution is lower than the central level and belongs to the background or noise feature.
[0039] The formula for calculating the gating factor is:
[0040] ,
[0041] in, Indicates the first The gating factor for each component is used to dynamically weight the characteristic importance of the components; Indicates the first The average relative concentration score of each component is calculated using the following formula: , Indicates the number of groundwater samples; The threshold value is automatically calculated by taking the median of the average relative concentration scores of all components. ; This represents the gate function steepness parameter, used to control the steepness of the Sigmoid transition region, calculated by the standard deviation of the average relative concentration scores of all components. set up, This ensures that when the dispersion of component intensity distribution is greater, the transition of the gate function is smoother, and when the dispersion is smaller, the transition is steeper.
[0042] In the hidden layer activation vector calculation stage, forward propagation is performed individually for each groundwater sample. For a single groundwater sample, the component concentration column vector and the gating factor column vector are multiplied element-wise. This element-wise multiplication achieves dynamic selective weighting of the relative concentration score vector. This result is then multiplied item-by-item by the hidden layer weight matrix to obtain a linear combination. The linear combination result is then added to the hidden layer bias vector and processed by a nonlinear activation function to generate the hidden layer activation vector for the groundwater sample. The component concentration column vector of the groundwater sample is constructed based on the relative concentration scores of all components in a single groundwater sample, and the gating factor column vector is constructed based on the gating factors of all components. The formula for the hidden layer activation vector of the groundwater sample is expressed as follows:
[0043] ,
[0044] in, Indicates the first The activation vector of a groundwater sample in the hidden layer (layer 1) of a neural network, i.e., the hidden activation vector of the groundwater sample, with dimension . , This represents the number of hidden layer nodes. It is an activation function used to introduce nonlinearity, avoid gradient vanishing, and promote sparse activation; This represents the hidden layer weight matrix, with dimension 1. ; , indicating the first The component concentration column vector of each groundwater sample, with dimension [missing information]. ; This indicates element-wise multiplication; , represents the column vector of gating factors, with dimension . ; This represents the hidden layer bias vector, with dimension . ; Indicates transpose;
[0045] In the prediction process during the output layer probability calculation and training optimization phase, the hidden layer activation vector of the groundwater sample is multiplied by the output layer weight matrix, and then the output layer bias vector is shifted and adjusted. The output layer bias vector is used to compensate for prior biases between different pollution source categories. Finally, the linear transformation is converted into a probability distribution using the softmax function, ensuring that the sum of the probabilities of all categories is 1, thus obtaining the predicted probability of the pollution source in the groundwater sample. The pollution source category corresponding to the maximum predicted probability of the groundwater sample is taken as the predicted pollution source result for the groundwater sample, as expressed in the following formula:
[0046] ,
[0047] in, Indicates the first Pollution source prediction results for one groundwater sample; This represents the index operation corresponding to the maximum value, used to select the pollution source category with the highest probability. Indicates the pollution source category index; The function is used to convert a linear transformation into a probability distribution; This represents the output layer weight matrix, with dimension 1. The first layer weight matrix of the output layer Travelogue , dimension , used to map hidden layer activation vectors to the first Scores for different types of pollution sources This represents the total number of pollution source categories, derived from known pollution source information provided by local environmental protection departments, and is not limited here. This represents the output layer bias vector, with dimension . , and its first element As the first Baseline bias in the scores of pollution source categories;
[0048] During the training process of output layer probability calculation and training optimization, weighted cross-entropy loss is used as the training objective. For each groundwater sample, negative log loss is calculated only on the probability component corresponding to the true category and multiplied by the pollution intensity weight. The pollution intensity weight is obtained based on the normalization of the total intensity of the groundwater sample components, so that groundwater samples with higher pollution levels contribute more to the weighted cross-entropy loss. The true category is determined by sampling and labeling of known pollution source areas or laboratory pollutant detection, and is used as the input of the supervised learning training label.
[0049] The weighted cross-entropy loss formula is expressed as follows:
[0050] ,
[0051] in, This represents the weighted cross-entropy loss, which serves as the training objective. Indicates the first The pollution intensity weight of each groundwater sample is expressed by the formula: , Indicates the first The first groundwater sample The strength of each component; It is a category matching indicator function, only applicable to the true category. equal to pollution source category The value is 1 if it is true, and 0 otherwise, and the range of values is {0,1}. Indicates the first The groundwater sample belongs to the first Predicted probability of pollution sources of this type ;
[0052] Using existing gradient descent optimizers, such as Adam or stochastic gradient descent, the gradient of the weighted cross-entropy loss with respect to all learnable parameters, including the hidden layer weight matrix, hidden layer bias vector, output layer weight matrix, and output layer bias vector, is calculated. Then, using the chain rule, the error signal is propagated back through the output layer to the input layer, updating each learnable parameter to reduce the total loss until convergence. Convergence is considered achieved when the relative rate of change of the weighted cross-entropy loss function between two consecutive iterations falls below a preset threshold, at which point training is terminated. The preset threshold ranges from [value missing]. to .
[0053] S3. Based on the known location of the sampling point and the pollution source prediction results of the groundwater sample, calculate the probability that the spatial location to be estimated belongs to the target pollution source type, and generate a continuous spatial probability distribution field.
[0054] Based on the known locations of sampling points and the pollution source prediction results of groundwater samples, a distance-inverse weighting strategy is adopted to estimate the probability of belonging to the target pollution source type at any unknown location within the study area, thereby generating a continuous spatial probability distribution field of the target pollution source type.
[0055] For any spatial location to be estimated within the study area, all known sampling points are traversed, and the spatial Euclidean distance between the spatial location to be estimated and each sampling point is calculated one by one. The closer the sampling point is, the greater its influence on the current estimated location, and the farther the sampling point is, the smaller its influence on the current estimated location. For a specific pollution source category of interest to the user, i.e., the target pollution source type, only sampling points that are determined to belong to the target pollution source type after being predicted by the neural network are retained. For each retained sampling point, two weighting factors are introduced: one is the reciprocal of the distance between the sampling point and the spatial location to be estimated; the other is the pollution intensity weight.
[0056] Multiply the pollution intensity weights of all retained sampling points by the category matching indicator function, then multiply by the inverse of their respective distances and sum them to obtain the numerator. Simultaneously, sum the inverse distances of all sampling points individually to obtain the denominator. Divide the numerator by the denominator to obtain the conditional probability estimate of the current spatial location belonging to the target pollution source type, i.e., the probability that the spatial location to be estimated belongs to the target pollution source type. The value ranges from 0 to 1, and is expressed by the following formula:
[0057] ,
[0058] in, Representing coordinates The conditional probability estimate of a location belonging to the target pollution source type, i.e., the probability that the spatial location to be estimated belongs to the target pollution source type; This indicates that only those identified as belonging to the target pollution source type after being predicted by the neural network are included. The sampling points are weighted; It is a category matching indicator function, used to match the first category. The system determines whether a groundwater sample belongs to the target pollution source type; if it does, the value is 1, otherwise it is 0. Indicates the spatial location to be estimated and the first The spatial Euclidean distance of each sampling point;
[0059] The generation of continuous spatial probability distribution fields significantly improves the spatial continuity and intuitiveness of pollution source identification. The fusion of pollution intensity information enhances the sensitivity to heavily polluted areas. It does not require assuming the exact location of pollution sources, is objective, and is easy to implement in engineering, thus having strong practical value and promising prospects for promotion.
[0060] In summary, a groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural networks has been completed.
[0061] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.
[0062] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0063] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural networks, characterized in that, Includes the following steps: S1. Groundwater samples were collected at the sampling point and fluorescence spectroscopy was performed to obtain a three-dimensional fluorescence excitation-emission matrix; Tensor decomposition was performed on the three-dimensional fluorescence excitation-emission matrix to obtain the excitation spectrum intensity, emission spectrum intensity, and relative concentration scores of the groundwater sample for each component. S2. Based on the relative concentration scores of groundwater samples on each component, a weighted single hidden layer feedforward neural network classification algorithm based on the pollution component gating mechanism is introduced. The algorithm includes a gating factor calculation stage, a hidden layer activation vector calculation stage, and an output layer probability calculation and training optimization stage. During the gating factor calculation phase, a global statistical analysis was performed on the relative concentration scores of components in all groundwater samples. For each component, the average relative concentration score and the standard deviation of the average relative concentration score were calculated. A gating factor is generated based on the average relative concentration score of the component and the standard deviation of the average relative concentration score of the component. The gating factor is used in the hidden layer activation vector calculation stage and the output layer probability calculation and training optimization stage to calculate the pollution source prediction results of the groundwater sample. S3. Based on the known location of the sampling point and the pollution source prediction results of the groundwater sample, calculate the probability that the spatial location to be estimated belongs to the target pollution source type, and generate a continuous spatial probability distribution field.
2. The groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural network according to claim 1, characterized in that, The three-dimensional fluorescence excitation-emission matrix was decomposed using parallel factor analysis to output the excitation spectrum intensity, emission spectrum intensity, and relative concentration scores of the groundwater sample for each component.
3. The groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural network according to claim 1, characterized in that, In the hidden layer activation vector calculation stage: Based on the relative concentration scores of all components in a single groundwater sample, a column vector of component concentrations for the groundwater sample is constructed; based on the gating factors of all components, a column vector of gating factors is constructed. Based on the column vectors of component concentrations and gating factors of groundwater samples, combined with the hidden layer weight matrix and hidden layer bias vector, the hidden layer activation vector of groundwater samples is calculated.
4. The groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural network according to claim 3, characterized in that, During the prediction process in the output layer probability calculation and training optimization phase: By introducing an output layer weight matrix, the hidden layer activation vector of the groundwater sample is mapped to the pollution source score. Combined with the output layer bias vector, the pollution source prediction probability of the groundwater sample is calculated. The pollution source category corresponding to the maximum value of the pollution source prediction probability of the groundwater sample is taken as the pollution source prediction result of the groundwater sample.
5. The groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural network according to claim 4, characterized in that, During the training process, specifically in the output layer probability calculation and training optimization phases: The pollution intensity weight of the groundwater sample is calculated based on the relative concentration scores of its components. Based on the predicted pollution source probability and pollution intensity weight of groundwater samples, combined with the true category, a weighted cross-entropy loss function is constructed and iteratively optimized to update the hidden layer weight matrix, hidden layer bias vector, output layer weight matrix, and output layer bias vector until the loss converges.
6. The groundwater pollution source identification method based on three-dimensional fluorescence spectroscopy and neural network according to claim 4, characterized in that, The probability that the spatial location to be estimated belongs to the target pollution source type is calculated using the following method: Iterate through all known sampling points, and based on the pollution source prediction results of groundwater samples, combine the type matching indicator function to retain the sampling points that are determined to belong to the target pollution source type; For each retained sampling point, the probability that the spatial location to be estimated belongs to the target pollution source type is calculated by combining the known Euclidean distance between the sampling point and the spatial location to be estimated and the pollution intensity weight of the groundwater sample.