Intelligent identification and monitoring method for seawater intrusion interface based on multi-parameter gradient characteristics

Abstract

The application discloses a kind of based on multi-parameter gradient feature seawater intrusion interface intelligent identification monitoring method, it is related to marine environmental monitoring technical field, comprising: layout monitoring section and monitoring well site, deployment multi-parameter sensor group synchronous acquisition conductivity, chloride ion concentration, temperature and water level time series data;Calculate vertical and horizontal direction multi-parameter gradient, constitute multi-parameter gradient feature vector;Interface identification model is constructed with FixMatch as framework, output the classification result of each monitoring horizon and determine salt fresh water interface position;Adopt multivariate state estimation method to set adaptive dynamic threshold, determine interface significant migration and trigger update;Calculate seawater mixing ratio to divide intrusion degree grade, and interface position is predicted;Three conditions of interface migration, grade upgrade and forecast out of bounds are comprehensively issued graded early warning information and visualized display.The application realizes the intelligent identification and dynamic monitoring of seawater intrusion interface.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring technology, and more specifically, to a method for intelligent identification and monitoring of seawater intrusion interfaces based on multi-parameter gradient features. Background Technology

[0002] Seawater intrusion is a common geological environmental problem, mainly caused by excessive extraction of groundwater or sea-level rise, leading to seawater entering aquifers and causing ecological problems such as groundwater quality deterioration and soil salinization. Currently, seawater intrusion monitoring mainly relies on traditional monitoring wells and manual sampling analysis. Traditional monitoring wells include single wells and cluster monitoring wells, the latter of which involves deploying multiple wells at different depths for different aquifer groups within the same site. However, this method has the following shortcomings: first, it requires a large number of boreholes, occupies a large area, is difficult to coordinate, and has high construction and maintenance costs; second, the frequency of manual sampling is low, making it difficult to capture the dynamic migration of the brackish water interface; and third, it relies heavily on a single conductivity parameter for judgment, which is easily affected by factors such as temperature and native salinity, resulting in insufficient accuracy.

[0003] Electrical conductivity is widely used for interface identification due to the significant electrical differences between fresh and brackish water. Studies have shown that electrical conductivity exhibits a clear step-like jump near the brackish water interface. However, relying solely on electrical conductivity gradients has limitations: firstly, electrical conductivity is influenced by various environmental factors, and anomalies may not necessarily correspond to modern seawater intrusion; secondly, gradient thresholds vary considerably across different hydrogeological units and seasons, lacking adaptive judgment capabilities; and thirdly, single parameters cannot provide quantitative indicators such as chloride ion concentration and seawater mixing ratio, making it difficult to meet the needs of graded assessment of intrusion severity. To address these issues, some studies have attempted to introduce multi-parameter monitoring, such as chloride ion concentration, salinity, water level, and water temperature, but existing systems are mostly deployed in a decentralized manner, with inconsistent parameter acquisition times and frequencies, making data fusion difficult. Furthermore, brackish water interface determination still relies on human experience, lacking intelligent identification methods based on multi-parameter gradient characteristics, thus hindering automatic interface location, dynamic tracking, and real-time evaluation.

[0004] Therefore, there is an urgent need to develop an intelligent identification and monitoring method for seawater intrusion interfaces that integrates multi-parameter gradient features, in order to overcome the shortcomings of existing technologies in terms of cost, accuracy, and intelligence, and to provide technical support for the prevention and control of seawater intrusion. Summary of the Invention

[0005] To address the aforementioned technical challenges, this invention proposes an intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features. This method improves interface identification accuracy by constructing multi-parameter gradient feature vectors, solves the problem of scarce labeled samples through the FixMatch semi-supervised framework, achieves adaptive dynamic threshold determination through multivariate state estimation, and realizes hierarchical early warning by integrating multiple triggering conditions, thus providing reliable technical support for the prevention and control of seawater intrusion.

[0006] This invention provides an intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features, comprising the following steps: Step S1: Set up monitoring sections in the target area, set up monitoring wells on the monitoring sections, set up several monitoring layers along the vertical direction of the monitoring wells, deploy multi-parameter sensor groups at each monitoring layer, and synchronously collect and acquire conductivity, chloride ion concentration, temperature and water level time series data of each layer. Step S2: Based on the collected multi-parameter data, calculate the vertical and horizontal conductivity gradients, chloride ion concentration gradients, and temperature gradients to form a multi-parameter gradient feature vector. Step S3: Construct an interface recognition model based on FixMatch framework, take the multi-parameter gradient feature vector as the model input, output the classification results of each monitoring layer belonging to freshwater layer, transition layer or saline layer, and determine the location of freshwater and saline-salt water interface according to probability distribution. Step S4: Using the multivariate state estimation method, a memory matrix is ​​constructed based on historical data of normal working conditions. The current multi-parameter gradient feature vector is matched with the memory matrix to calculate the residual. An adaptive dynamic threshold is set according to the statistical distribution of the residual. When the residual exceeds the threshold, it is determined that the interface has undergone significant migration and an update is triggered. Step S5: Calculate the seawater mixing ratio based on the identified freshwater and brackish water interface locations and measured chloride ion concentrations, classify the degree of intrusion, and predict the interface location based on multi-period interface location data in the time series. Step S6: When the interface is determined to have migrated significantly and the change in the position of the brackish water interface exceeds the preset threshold, the intrusion level is upgraded, or the predicted interface position crosses the preset key control section, a graded early warning message is automatically issued and displayed visually.

[0007] In this scheme, the construction of the three-dimensional monitoring network in step S1 specifically includes: Hydrogeological surveys were conducted to divide the target area into freshwater zone, brackish water transition zone, and saline water zone along the inland to ocean direction. Monitoring sections are set up along the normal direction of the coastline, and monitoring wells are set up in the freshwater area, the brackish water transition zone and the brackish water area for each section. Based on the monitoring depth requirements, a single-hole mixed structure or a single-hole multi-layer structure is selected. For a single-hole multi-layer structure, a multi-level complete monitoring well method is adopted, in which the monitoring pipes of each level are radially distributed in the borehole according to the depth sequence, and hydraulic separation between each monitoring layer is achieved by using a water stopper. Multiple monitoring layers are set vertically within each monitoring well, with the spacing between layers gradually increasing towards the sea. Each monitoring pipe extends independently to the wellhead, which is equipped with sensor connectors and a data transmission terminal.

[0008] In this scheme, the multi-parameter gradient feature extraction in step S2 specifically includes: The conductivity, chloride ion concentration, temperature and water level time series data collected from each monitoring layer are preprocessed. The preprocessing includes temperature compensation correction and identification of tidal or pumping disturbances based on water level data, and marking or removing disturbed data. For each monitoring well, a depth sequence of each monitoring layer is established, and the vertical gradients of conductivity, chloride ion concentration, and temperature between adjacent monitoring layers within each monitoring well are calculated. For the same monitoring section, the monitoring wells are sorted along the inland to ocean direction, and the horizontal gradients of conductivity and chloride ion concentration at the same standard depth layer between adjacent monitoring wells are calculated. The gradient symbols within each depth interval are arranged in depth order to construct a symbol sequence as a gradient symbol pattern feature; the correlation coefficient between the conductivity gradient and the chloride ion concentration gradient within the same depth interval is calculated to construct a consistency index feature; the ratio of the vertical gradient to the horizontal gradient on the same standard depth layer is calculated to construct a multi-scale fusion feature; the rate of change of each gradient feature within a preset time period is calculated to construct a time-series rate of change feature. The calculated basic gradient features, gradient sign pattern features, consistency index features, multi-scale fusion features, and time series change rate features are concatenated in sequence to form a multi-parameter gradient feature vector and then output.

[0009] In this solution, step S3, which involves intelligent interface recognition based on the FixMatch semi-supervised framework, specifically includes: We construct an interface recognition model based on FixMatch, which includes a teacher network and a student network, and use unlabeled multi-parameter gradient feature vectors as unlabeled samples. Strong enhancement and weak enhancement operations are applied to the unlabeled samples respectively. The strong enhancement includes random noise injection, scaling transformation and channel dropping. The weak enhancement includes slight translation and scaling. The weakly enhanced unlabeled samples are input into the teacher network to generate pseudo-labels and confidence scores. Different confidence thresholds were set for freshwater, transition, and saline water layers. The comprehensive confidence of each pseudo-label was calculated based on the original confidence of the model, the temporal consistency score, and the spatial consistency score. Pseudo-labels with a comprehensive confidence score exceeding the corresponding category dynamic threshold were selected for training. Oversampling and feature space interpolation enhancement are performed on samples of the transition layer category to generate synthetic samples to expand the coverage of the transition layer in the feature space, and a re-labeling mechanism is triggered for uncertain samples at the category boundary. The enhanced unlabeled samples and labeled samples are input into the student network together. The student network parameters are updated by weighted sum of supervised loss and unsupervised consistency loss. The teacher network parameters are updated by exponential moving average of the student network parameters. The multi-parameter gradient feature vectors collected in real time are input into the trained student network, which outputs the classification results and probability distribution of each monitoring layer as freshwater layer, transition layer or saline layer, and determines the vertical depth position and horizontal cross-sectional position of the freshwater and saline water interface through linear interpolation based on the classification results.

[0010] In this scheme, step S3, which determines the location of the brackish water interface based on the probability distribution, specifically includes: The probability distribution of each monitored layer belonging to freshwater layer, transition layer and saline layer was extracted, and a vertical probability profile of a single well was constructed in depth order, and the probability profile was smoothed. Based on a preset probability threshold, the freshwater-transition interface and the transition-saline interface are initially located. A probability-weighted interpolation method is used to locate the interface with sub-layer precision between adjacent monitoring layers and obtain the interface depth. Several standard depth layers were selected, and the classification probabilities of each monitoring well at the same depth were vertically interpolated to construct a horizontal probability profile. The horizontal positions of the freshwater-transition horizontal interface and the transition-saline water horizontal interface were determined by the probability weighted interpolation method. The vertical interface positioning results of each monitoring well and the horizontal interface positioning results of each standard depth layer are fused to generate a two-dimensional discrete interface point set, and a continuous and smooth two-dimensional interface curve is generated by fitting the spline interpolation method. The interface curves at different time points are overlaid and analyzed to calculate the rate of change of the interface depth at a fixed horizontal position and the rate of change of the interface horizontal position at a fixed depth.

[0011] In this scheme, step S4, the dynamic threshold adaptive determination based on multivariate state estimation, specifically includes: Normal operating condition data are selected from historical monitoring data, and the multi-parameter gradient feature vectors of all monitoring wells and all monitoring layers at each time point are concatenated into observation vectors. Representative observation vectors are selected to construct a memory matrix. The estimated vector is calculated using a sparse coding method. The current observation vector is sparsely decomposed in the space spanned by the memory matrix, and the sparse representation coefficients are obtained. The product of the memory matrix and the sparse representation coefficients is used as the estimated vector. Calculate the residual and squared prediction error between the observed vector and the estimated vector, and set a basic threshold based on the statistical distribution of the squared prediction error under normal operating conditions; Based on recent normal operating data, the current season, the current tidal phase, and the type of hydrogeological unit, the basic threshold is dynamically and adaptively adjusted, and three levels of thresholds are set: warning threshold, early warning threshold, and alarm threshold. The anomaly index is obtained by calculating the relative position of the current squared prediction error with each level threshold. When the anomaly index exceeds the level corresponding to the alarm threshold and continues to exceed it for multiple consecutive time points, it is determined to be an abnormal state, triggering the re-identification of the interface position.

[0012] In this solution, step S5, which involves classifying the intrusion severity level, specifically includes: Extract the measured values ​​of chloride ion concentration in the monitoring layer near the brackish water interface, determine the chloride ion concentration in the freshwater endmember and the chloride ion concentration in the seawater endmember, and calculate the seawater mixing ratio based on the mass conservation of chloride ion concentration. The classification standards are set based on chloride ion concentration, with chloride ion concentration from low to high corresponding to no invasion level, slight invasion level, moderate invasion level and severe invasion level. The classification standards are also set based on seawater mixing ratio, with mixing ratio from low to high corresponding to no invasion level, slight invasion level, moderate invasion level and severe invasion level. Calculate the intrusion distance between the current horizontal position of the brackish water interface and the historical reference position, and the depth change between the current vertical depth of the interface and the historical reference depth. Based on the intrusion distance and the depth change, set a classification standard for the interface position. The chloride ion concentration level, seawater mixing ratio level, and interface location level are weighted and integrated to calculate a comprehensive invasion severity score, and the final invasion severity level is determined based on the comprehensive score.

[0013] In this scheme, step S5 involves predicting the location of the predicted interface, specifically including: Based on historical monitoring results, a time series of the location of the brackish water interface was constructed in chronological order, and the time series was smoothed. A linear trend prediction model is adopted to calculate the annual rate of change by fitting a linear trend line of interface position change over time, and to predict the future interface position based on the product of the current interface position and the annual rate of change. The seasonal decomposition forecasting method is adopted to decompose the time series of interface location into trend, seasonal and random terms. The trend term is extrapolated and the seasonal term is cyclically repeated and then superimposed with the trend term to obtain the forecast value. A time-series prediction model is constructed using a long short-term memory network. The model takes the historical interface position sequence as input and learns the long- and short-term dependencies of interface migration through the forget gate, input gate and output gate structure to predict the future interface position. The results of linear trend prediction, seasonal decomposition prediction and long short-term memory network prediction are weighted and integrated to output the predicted value of the interface position at a future set time point.

[0014] In this solution, step S6 automatically issues tiered early warning information, specifically including: Calculate the horizontal change between the current horizontal position and the historical reference position of the brackish water interface, as well as the vertical change between the current vertical depth and the historical reference depth. Set differentiated multi-level migration thresholds according to different hydrogeological units, compare the calculated changes with the multi-level migration thresholds, and determine the warning level triggered by interface migration. The current intrusion level is compared with the historical intrusion level to monitor the level escalation and determine the warning level triggered by the level escalation based on the escalation magnitude. Based on the interface position prediction results, determine whether the predicted interface position at a future set time point will cross the preset key control section, estimate the boundary crossing time, and determine the warning level triggered by the predicted boundary crossing based on the boundary crossing certainty and the distance of the boundary crossing time. The system comprehensively evaluates three triggering conditions: interface migration trigger, level upgrade trigger, and prediction out-of-bounds trigger. When multiple triggering conditions are met simultaneously, the highest level is taken as the final warning level.

[0015] In this solution, the visualization demonstration described in step S6 specifically includes: Using the electronic map as the base map, overlay the coastline layer, administrative division layer, and key control section layer, and mark the location of each monitoring well. Two-dimensional profile diagrams were drawn for each monitoring section to show the monitoring layer classification results, the fitted freshwater-transition interface curves, and the transition-saline water interface curves for each monitoring well location. It supports selecting multiple historical time points and overlaying the interface curves from different periods on the same cross-section, using different colors to distinguish the curves from different periods; When an alert is triggered, the affected area will be highlighted on the map according to the corresponding color based on the alert level, and an alert information window will pop up on the monitoring system interface.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a three-dimensional seawater intrusion monitoring network by deploying monitoring sections and layered monitoring wells, and simultaneously collecting time-series data on conductivity, chloride ion concentration, temperature, and water level using multi-parameter sensor groups. This overcomes the limitations of traditional single-well monitoring, which suffers from limited coverage and incomplete information. By calculating multi-parameter gradients in the vertical and horizontal directions and constructing multi-parameter gradient feature vectors, it integrates multi-dimensional information such as conductivity gradient, chloride ion concentration gradient, and temperature gradient, thus overcoming the technical bottleneck of insufficient identification accuracy caused by the reliance on a single conductivity parameter in traditional methods.

[0017] This invention constructs a semi-supervised learning model based on the FixMatch framework, jointly training it with a small number of labeled samples and a large number of unlabeled samples. This addresses the problems of scarce labeled samples and high labeling costs in the field of seawater intrusion monitoring. Simultaneously, it introduces a category-aware dynamic threshold adjustment and a few-sample category enhancement mechanism, effectively improving the recognition accuracy of few-sample categories such as transition zones. An adaptive dynamic threshold is set using a multivariate state estimation method, and a memory matrix is ​​constructed based on historical data under normal operating conditions to achieve sensitive detection of abnormal states, avoiding the shortcomings of fixed thresholds that cannot adapt to different hydrogeological units and seasonal changes. Finally, a graded early warning system is triggered by three conditions: interface migration judgment, intrusion degree level upgrade, and predicted boundary crossing. This realizes the transformation from passive monitoring to proactive early warning, providing reliable technical support for seawater intrusion prevention and disaster mitigation decision-making. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0019] Figure 1 A flowchart of a method for intelligent identification and monitoring of seawater intrusion interfaces based on multi-parameter gradient features is shown. Figure 2 The flowchart for multi-parameter gradient feature extraction is shown; Figure 3 A flowchart of interface intelligent recognition based on the FixMatch semi-supervised framework is shown; Figure 4 A flowchart for automatically issuing graded early warning information is shown. Detailed Implementation

[0020] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0022] like Figure 1 As shown, this embodiment provides an intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features, including: Step S1: Set up monitoring sections in the target area, set up monitoring wells on the monitoring sections, set up several monitoring layers along the vertical direction of the monitoring wells, deploy multi-parameter sensor groups at each monitoring layer, and synchronously collect and acquire conductivity, chloride ion concentration, temperature and water level time series data of each layer. Step S2: Based on the collected multi-parameter data, calculate the vertical and horizontal conductivity gradients, chloride ion concentration gradients, and temperature gradients to form a multi-parameter gradient feature vector. Step S3: Construct an interface recognition model based on FixMatch framework, take the multi-parameter gradient feature vector as the model input, output the classification results of each monitoring layer belonging to freshwater layer, transition layer or saline layer, and determine the location of freshwater and saline-salt water interface according to probability distribution. Step S4: Using the multivariate state estimation method, a memory matrix is ​​constructed based on historical data of normal working conditions. The current multi-parameter gradient feature vector is matched with the memory matrix to calculate the residual. An adaptive dynamic threshold is set according to the statistical distribution of the residual. When the residual exceeds the threshold, it is determined that the interface has undergone significant migration and an update is triggered. Step S5: Calculate the seawater mixing ratio based on the identified freshwater and brackish water interface locations and measured chloride ion concentrations, classify the degree of intrusion, and predict the interface location based on multi-period interface location data in the time series. Step S6: When the interface is determined to have migrated significantly and the change in the position of the brackish water interface exceeds the preset threshold, the intrusion level is upgraded, or the predicted interface position crosses the preset key control section, a graded early warning message is automatically issued and displayed visually.

[0023] It should be noted that, through hydrogeological surveys, the characteristics of groundwater flow field, spatial distribution patterns of aquifers, orientation of ancient river channels, and information on groundwater mineralization zoning in the target area are obtained. The target area is divided into freshwater zone, brackish water transition zone, and saline water zone along the inland to ocean direction. Among them, the freshwater zone is the area with mineralization less than 1 g / L, the brackish water transition zone is the area with mineralization between 1 g / L and 3 g / L, and the saline water zone is the area with mineralization greater than 3 g / L.

[0024] Monitoring sections are set up along the coastline normal direction, with 3 to 5 monitoring wells set up sequentially from inland to ocean along the section line. Among them, at least one well is set up in the freshwater area as a background reference well, 2 to 3 wells are set up in the brackish water transition zone to increase the intensity of monitoring the interface migration process, and at least one well is set up in the brackish water area as an intrusion end-member reference well. The spacing between adjacent wells is set according to the principle of increasing density in the transition zone, with a spacing of 5 to 10 km near the interface and a spacing of 15 to 25 km away from the interface. Based on the complexity of the geological structure, aquifer thickness, monitoring depth requirements, and available site area of ​​the target area, either a single-well hybrid structure or a single-well multi-layer structure is selected. The single-well hybrid structure is suitable for shallow monitoring scenarios with well depths of 10 to 40 meters, while the single-well multi-layer structure is suitable for deep stratified monitoring scenarios with well depths of 40 to 55 meters where independent water samples from different aquifers need to be obtained. For monitoring wells using a single-well multi-layer structure, a multi-stage complete monitoring well structure is adopted: the longest first-stage monitoring tube is placed at the center of the borehole and lowered first, with the remaining monitoring tubes arranged in descending order of depth around the central tube. The monitoring pipes are arranged radially, with a minimum spacing of 50 mm between adjacent monitoring pipes; the borehole diameter is no less than 450 mm, gradually decreasing in size after 2 to 3 diameter changes; water stoppers are installed at the bottom boundary of each monitoring layer, using expanded rubber or grouting to ensure no hydraulic connection occurs between layers; the filter pipe section of each monitoring layer corresponds to the aquifer location, and the filter pipe adopts a bridge-type filter slot or wire-wound structure; multiple monitoring layers are set vertically in each monitoring well, with the spacing between layers gradually increasing towards the sea; each monitoring pipe extends independently to the wellhead, and the wellhead is equipped with sensor connectors and a data transmission terminal.

[0025] A multi-parameter sensor array acquires raw monitoring data from each monitoring layer synchronously, including four parameters: conductivity, chloride ion concentration, temperature, and water level. The time-series data of conductivity, chloride ion concentration, temperature, and water level collected from each monitoring layer are preprocessed. This preprocessing includes temperature compensation correction and identification of tidal or pumping disturbances based on water level data, with disturbed data being marked or removed.

[0026] like Figure 2As shown, for each monitoring well, a depth sequence for each monitoring layer is established, and the vertical gradients of conductivity, chloride ion concentration, and temperature between adjacent monitoring layers within each monitoring well are calculated. For the same monitoring section, the monitoring wells are sorted along the inland-to-ocean direction, and the horizontal gradients of conductivity and chloride ion concentration at the same standard depth layer between adjacent monitoring wells are calculated. The gradient symbols within each depth interval are arranged in depth order to construct a symbol sequence as a gradient symbol pattern feature, reflecting the vertical variation of the parameters, including all-positive mode, all-negative mode, positive-to-negative transition mode, and negative-to-positive transition mode. The symbol pattern feature can provide additional structural information on the basis of a single numerical gradient, effectively distinguishing between seawater intrusion and native saline water distribution. The correlation coefficient between conductivity gradient and chloride ion concentration gradient within the same depth range is calculated to construct a consistency index feature. When the consistency index feature is close to 1, it indicates that conductivity and chloride ion concentration change synchronously, the water chemical composition is stable, and the gradient change is mainly caused by changes in salinity, indicating a modern seawater intrusion process. When the consistency index feature is close to 0 or negative, it indicates that conductivity and chloride ion concentration change asynchronously, possibly due to interference from other ions (such as sulfates and carbonates) or a complex mixing process, indicating paleoseawater residue or anthropogenic pollution. The ratio of vertical gradient to horizontal gradient at the same standard depth layer is calculated to construct a multi-scale fusion feature. The multi-scale fusion feature can simultaneously characterize the vertical structural features and horizontal distribution features of the brackish water interface. The rate of change of each gradient feature within a preset time period is calculated to construct a time-series rate of change feature. The time-series gradient rate of change extends the static gradient feature into a dynamic feature, which can capture the evolution trend of the seawater intrusion process and provide time-dimensional information for early warning. The calculated basic gradient features, gradient sign pattern features, consistency index features, multi-scale fusion features, and time series change rate features are concatenated in sequence to form a multi-parameter gradient feature vector and then output.

[0027] It should be noted that the constructed recognition model is based on the FixMatch semi-supervised learning framework, integrating a pseudo-label noise suppression module and a geospatial constraint module. For example... Figure 3 As shown, an interface recognition model based on the FixMatch framework is constructed, comprising a teacher network and a student network. Both networks employ the same neural network structure, taking multi-parameter gradient feature vectors as input and outputting probability distributions for three categories (freshwater layer, transition layer, and saline layer). The network structure uses a lightweight multilayer perceptron or a one-dimensional convolutional neural network. Unlabeled multi-parameter gradient feature vectors are used as unlabeled samples, and the number of unlabeled samples is far greater than that of labeled samples.

[0028] Strong and weak augmentation operations are applied to unlabeled samples. Strong augmentation includes random noise injection, scaling, and channel discarding. Weak augmentation includes slight translation and scaling. The weakly augmented unlabeled samples are then input into the teacher network to generate pseudo-labels and confidence scores. Specifically, weak augmentation includes adding Gaussian noise with a mean of zero and a variance equal to 1% to 5% of the standard deviation of the eigenvalues, multiplied by a random scaling factor in the range of 0.95 to 1.05. Strong augmentation includes adding random noise following a Gaussian distribution, randomly setting 10% to 30% of the feature dimension values ​​to zero, linearly interpolating the feature vectors of two different samples with random weights, and adding adversarial perturbations along the gradient ascent direction. Weak augmentation is used to generate stable pseudo-labels, while strong augmentation is used to improve the robustness of the model.

[0029] Differentiated confidence thresholds were set for freshwater, transition, and saline layers. The comprehensive confidence of each pseudo-label was calculated based on the model's original confidence, temporal consistency score, and spatial consistency score. Pseudo-labels with a comprehensive confidence exceeding the corresponding category's dynamic threshold were selected for training. The temporal consistency score was calculated as the proportion of category consistency between the current sample pseudo-label and the prediction results of that layer at the most recent K historical moments. The spatial consistency score was calculated as the proportion of category consistency between the current sample pseudo-label and the prediction results of adjacent spatial locations (similar depth layers at adjacent well locations in the same cross-section). The confidence threshold for each category was dynamically adjusted based on the number of selected pseudo-labels during training. When the number of selected samples for a certain category was significantly lower than the historical average, the threshold for that category was automatically lowered; conversely, when the number of selected samples was too high, the threshold was appropriately increased to control the noise ratio.

[0030] Oversampling and feature space interpolation enhancement are performed on samples of the transition layer category. Feature vectors from two transition layer samples are input, and difference coefficients are randomly sampled from the Beta distribution. These difference coefficients and the two feature vectors are used to generate synthetic samples to expand the coverage of the transition layer in the feature space. A re-labeling mechanism is triggered for uncertain samples at the category boundaries. Specifically, in the unsupervised prediction results of the samples, the probability of the largest category is below a high threshold (e.g., 0.85) but above a low threshold (e.g., 0.55), and the probability difference between the largest and second-largest categories is less than a preset difference threshold. Samples meeting these conditions are considered to have ambiguous boundaries and high re-labeling value. Boundary sample mining is performed every K training rounds. The mined boundary samples are sorted according to their uncertainty scores, and a preset number of samples are selected for re-labeling.

[0031] The strongly enhanced unlabeled samples and labeled samples are input into the student network. The student network parameters are updated using a weighted sum of supervised and unsupervised consistency losses. The loss weight for each category is inversely proportional to the number of samples in that category, with transition layer categories receiving higher loss weights. The supervised loss uses the cross-entropy loss function, and only labeled samples are calculated. For each labeled sample, the cross-entropy between the model's predicted probability distribution and the one-hot encoded true label is calculated, and the mean of all labeled samples is taken. The unsupervised consistency loss uses the mean squared error or cross-entropy loss function, and only unlabeled samples that pass the confidence screening are calculated. For each selected unlabeled sample, the difference between the student network's prediction for strongly enhanced samples and the pseudo-label generated by the teacher network for weakly enhanced samples is calculated. The teacher network parameters are updated using the exponential moving average of the student network parameters. The multi-parameter gradient feature vectors collected in real time are input into the trained student network, which outputs the classification results and probability distributions for each monitoring layer belonging to the freshwater layer, transition layer, or saline layer. Based on the classification results, the vertical depth and horizontal cross-sectional position of the freshwater-saline interface are determined through linear interpolation.

[0032] It should be noted that after the FixMatch semi-supervised learning model completes the classification and identification of each monitoring layer, the model outputs the probability distribution of each monitoring layer belonging to the freshwater layer, transition layer, or saline layer.

[0033] To extract the probability distribution of each monitored stratum belonging to freshwater, transitional, and saline water layers within a single monitoring well, a vertical probability profile for each well is constructed in depth order and then smoothed. A Gaussian kernel-weighted moving average method is used to perform a neighborhood-weighted average of the probability values ​​at each depth point. The smoothing window size is adaptively determined based on the spacing between the monitored strata; a smaller window is used when the strata are densely packed, and a larger window is used when the strata are sparsely packed. The smoothed probability profile eliminates local anomalies and reflects the overall trend of the regional probability distribution.

[0034] Initial localization of the freshwater-transition interface and the transition-saline interface is performed based on a preset probability threshold. A probability-weighted interpolation method is then used to perform sub-stratum-precision interface localization between adjacent monitoring layers to obtain the interface depth. In the sub-stratum localization of the freshwater-transition interface, a depth is set... The probability of being in a freshwater layer is ,depth The probability of the transition layer is The interface is located between the two, with the probability increasing rapidly closer to the transition layer. A probability-weighted interpolation method is used: interface depth. Interpolation weights Calculated based on probability changes. When and When both are relatively high, the interface is closer to locations with drastic probability changes. The weight calculation formula comprehensively considers both the absolute value of the probability and the probability gradient, making the interface localization sensitive to probability changes. In sub-layer localization of the transition-saline water interface, a depth is set... Probability of transition layer ,depth The probability of being in a saline aquifer is The same probability-weighted interpolation method was used to calculate the sub-layer depth at the transition-saline interface. The accuracy of sub-layer localization depends on the smoothness of the probability profile and the probability difference between adjacent layers.

[0035] Several standard depth layers were selected, and horizontal probability profiles were constructed by vertical interpolation of the classification probabilities at the same depth for each monitoring well. A probability-weighted interpolation method was used to determine the horizontal positions of the freshwater-transitional horizontal interface and the transition-saline horizontal interface. Specifically, the position where the probability of the transition layer changes from below a threshold to above a threshold was used as the horizontal position of the freshwater-transitional horizontal interface, and the position where the probability of the saline layer changes from below a threshold to above a threshold was used as the horizontal position of the transition-saline horizontal interface. The vertical interface positioning results of each monitoring well and the horizontal interface positioning results of each standard depth layer were fused to generate a two-dimensional discrete interface point set. A continuous and smooth two-dimensional interface curve was generated by fitting the points using spline interpolation. The interface curves at different time points were overlaid and analyzed to calculate the rate of change of interface depth at a fixed horizontal position and the rate of change of horizontal position at a fixed depth. The migration rate provides a quantitative indicator for assessing the development trend of seawater intrusion.

[0036] Based on continuous multi-period monitoring data, time series of morphological feature parameters of the interface curve are extracted. These morphological feature parameters include the average dip angle, maximum curvature and its horizontal position, average transition zone thickness, maximum interface depth and its horizontal position, and the position of the horizontal front of the interface. According to the variation pattern of the time series of these morphological feature parameters, a change point detection algorithm is used to automatically identify the transition points of the evolutionary stages, dividing the seawater intrusion evolution process into an initial intrusion stage, a rapid development stage, a stable advancement stage, a decline and recovery stage, and an oscillation and fluctuation stage. The initial intrusion stage is characterized by an average dip angle less than a first dip angle threshold and a horizontal front position... Located within the first distance threshold; the rapid development stage is characterized by the average dip angle increasing at a rate exceeding the first rate threshold and the horizontal front advancing inland at a rate exceeding the second rate threshold; the stable advancement stage is characterized by the average dip angle stabilizing and the horizontal front advancing at an approximately constant rate; the decline and recovery stage is characterized by the average dip angle decreasing and the horizontal front retreating towards the ocean; the oscillation and fluctuation stage is characterized by morphological characteristic parameters exhibiting periodic fluctuations without a significant monotonic trend; outputting the current evolution stage of seawater intrusion and its corresponding confidence level, and adaptively adjusting the sensitivity of the early warning strategy in subsequent steps according to the evolution stage.

[0037] It should be noted that during seawater intrusion monitoring, it is necessary to determine in real time whether the current monitoring status deviates from normal operating conditions, thereby triggering the re-identification of the interface location. Traditional fixed threshold methods are difficult to adapt to the normal fluctuation range under different hydrogeological units and seasonal conditions. This step introduces a multivariate state estimation method, constructing a memory matrix based on historical monitoring data under normal operating conditions. By matching the current observation vector with the memory matrix, the residuals are calculated, and an adaptive dynamic threshold is set based on the statistical distribution characteristics of the residuals.

[0038] Normal operating condition data is selected from historical monitoring data. Normal operating condition refers to a state where there is no significant seawater intrusion in the monitoring area and the brackish water interface is within a stable fluctuation range. The multi-parameter gradient feature vectors of all monitoring wells and all monitoring layers at each time point are concatenated into an observation vector. Data during extreme weather events (such as storm surges and torrential rains), data during periods of human disturbance (such as the operation of nearby pumping wells), and data during sensor malfunctions or calibration periods are excluded. Representative observation vectors are selected to construct a memory matrix. For example, cluster analysis is performed on the normal operating condition data, and the center vector of each cluster is selected as the memory vector. A sparse coding method is used to calculate the estimated vector. The current observation vector is sparsely decomposed in the space spanned by the memory matrix. L1 norm regularization transforms the sparse decomposition problem into a LASSO problem. The coordinate descent method or orthogonal matching pursuit method is used to solve for the sparse representation coefficients. The product of the memory matrix and the sparse representation coefficients is used as the estimated vector. The sparse coding method can automatically select the most representative memory vector, suppress noise interference, and improve the accuracy and robustness of the estimation.

[0039] The residuals and squared prediction errors between the observed and estimated vectors are calculated, and a basic threshold is set based on the statistical distribution of the squared prediction error under normal operating conditions. The basic thresholds are dynamically and adaptively adjusted according to recent normal operating condition data, the current season, the current tidal phase, and the type of hydrogeological unit. Specifically, the statistical distribution of the residuals and the thresholds are updated on a rolling basis based on recent normal operating condition data for a preset period. The statistical distribution of the residuals and the thresholds under normal operating conditions in different seasons are calculated respectively. The tidal phase at the current moment is determined based on tidal forecast data, and the residual distribution and thresholds under different tidal phases are calculated respectively. The normal fluctuation amplitudes of different hydrogeological units differ significantly. Due to the mixing of fresh and brackish water, the parameter fluctuation amplitudes are large in the transition zone; the fluctuation amplitudes are relatively small in the freshwater and brackish water areas. Memory matrices and thresholds are established for different hydrogeological units. Three threshold levels—warning, early warning, and alarm—are set. An anomaly index is obtained by calculating the relative position of the current squared prediction error to each threshold level. When the anomaly index exceeds the level corresponding to the alarm threshold and continues to exceed it for multiple consecutive time points, an anomaly is identified, triggering interface location re-identification. Simultaneously, the current observation vector and its corresponding interface location information are included in the candidate set of the memory matrix for subsequent memory matrix updates. Furthermore, the contribution of each component of the residual vector is used to preliminarily determine the main source of the anomaly. If the conductivity gradient and chloride ion concentration gradient components contribute the most to the residual, the anomaly is likely related to seawater intrusion interface migration; if the water level component contributes the most, the anomaly may be related to changes in recharge conditions.

[0040] It should be noted that after completing the intelligent identification and location of the brackish water interface, it is necessary to quantitatively assess and classify the degree of seawater intrusion, and predict the future migration trend of the interface based on historical monitoring data.

[0041] Measured chloride ion concentrations were extracted from monitoring layers near the freshwater-brine interface to determine the chloride ion concentrations in the freshwater and seawater end-members. The freshwater end-member chloride ion concentration represents the background value of natural freshwater unaffected by seawater intrusion, determined by the statistical mean of historical chloride ion concentration data from monitoring wells in freshwater areas far from the coastline and unaffected by seawater intrusion. The seawater end-member chloride ion concentration represents the chloride ion concentration of intruding seawater, determined by the statistical mean of chloride ion concentration data from monitoring wells in nearshore brackish water areas. The seawater mixing ratio is defined as the volume percentage of the intruding seawater in the mixed water body. Based on the principle of mass conservation of chloride ion concentration, the chloride ion concentration of the mixed water body is equal to the weighted average of the freshwater and seawater end-member chloride ion concentrations according to the mixing ratio. The formula for calculating the seawater mixing ratio is as follows: , The measured chloride ion concentration near the interface. This refers to the concentration of chloride ions in the freshwater terminal phase. The concentration of chloride ions at the seawater end-member.

[0042] The classification standards are based on chloride ion concentration, with concentrations ranging from low to high corresponding to no invasion, slight invasion, moderate invasion, and severe invasion. For example, a chloride ion concentration below 250 mg / L is considered no invasion, between 250 mg / L and 1000 mg / L is slight invasion, between 1000 mg / L and 2000 mg / L is moderate invasion, and above 2000 mg / L is severe invasion. The classification standards are also based on seawater mixing ratio, with mixing ratios ranging from low to high corresponding to no invasion, slight invasion, moderate invasion, and severe invasion. For example, a mixing ratio below 0.5% is no invasion, between 0.5% and 15% is slight invasion, between 15% and 30% is moderate invasion, and above 30% is severe invasion. Since background chloride ion concentrations in groundwater vary across regions, the classification thresholds should be adjusted appropriately based on regional hydrogeological conditions. For areas with high primary groundwater chloride ion concentrations, the thresholds for each level should be increased accordingly. Calculate the intrusion distance between the current horizontal position of the brackish water interface and the historical reference position, as well as the depth change between the current vertical depth of the interface and the historical reference depth. Based on the intrusion distance and the depth change, set the interface position classification standard. Weight and fuse the chloride ion concentration level, seawater mixing ratio level, and interface position level to calculate a comprehensive intrusion degree score. Determine the final intrusion degree level based on the comprehensive score.

[0043] Based on historical monitoring results, a time series of the brackish water interface position is constructed chronologically. The time series includes the vertical depth and horizontal position of the interface at each monitoring time point, and the time series is smoothed. For scenarios where the interface position changes steadily over time, a linear trend prediction model is used. The annual rate of change is calculated by fitting a linear trend line of the interface position over time, and the future interface position is predicted based on the product of the current interface position and the annual rate of change. The linear prediction formula is: Future interface position = Current interface position + Annual rate of change × Prediction period. For scenarios significantly affected by seasonal factors (such as precipitation, evaporation, and groundwater extraction), a seasonal decomposition prediction method is used. The interface position time series is decomposed into a trend term, a seasonal term, and a random term. The trend term reflects the overall direction of long-term changes in the interface position and is extracted using local weighted regression. The seasonal term reflects the periodic fluctuations within the year and is obtained by calculating the average value and variation pattern of the same month over the years. The random term reflects random fluctuations that cannot be explained by the trend and seasonal factors and is usually assumed to be white noise. The predicted value is obtained by extrapolating the trend term, cyclically repeating the seasonal term, and superimposing it with the trend term. A time-series prediction model is constructed using a long short-term memory network. The model takes the historical interface position sequence as input and learns the long- and short-term dependencies of interface migration through forget gate, input gate, and output gate structure to predict the future interface position. The results of linear trend prediction, seasonal decomposition prediction, and long short-term memory network prediction are weighted and integrated to output the predicted value of the interface position at a future set time point.

[0044] It should be noted that after completing the identification of the brackish water interface, the classification of the degree of intrusion, and the prediction of migration trends, the monitoring results need to be comprehensively analyzed. When abnormal changes are detected or potential risks are predicted, graded early warning information is automatically issued, and the current status and future trends are displayed intuitively through visualization.

[0045] like Figure 4As shown, the horizontal change between the current horizontal position and the historical reference position of the brackish water interface, as well as the vertical change between the current vertical depth and the historical reference depth, are calculated. The magnitude of the change in the horizontal position reflects the distance of seawater intrusion inland or retreat towards the ocean, while the magnitude of the change in the vertical depth reflects the magnitude of seawater intrusion upward or retreat downward. The changes are calculated in absolute value form, and the direction of change is also recorded. Differentiated multi-level migration thresholds are set according to different hydrogeological units. The first-level threshold reflects the upper limit of the normal fluctuation range of the interface, usually determined by multiplying the standard deviation of the interface position change in the same historical period by a coefficient. The second-level threshold reflects the critical value of significant interface migration, and the third-level threshold reflects the critical value of severe interface migration. The calculated changes are compared with the multi-level migration thresholds to determine the warning level triggered by interface migration. The current intrusion level is compared with the historical intrusion level to monitor the level escalation. The warning level triggered by the level escalation is determined based on the escalation magnitude. A single-level escalation triggers a medium-level warning, while multiple-level escalations trigger the highest-level warning. When the level remains unchanged or is downgraded, no warning is triggered alone; however, it can be determined in conjunction with other triggering conditions.

[0046] Based on the interface location prediction results, it is determined whether the predicted interface location at a future set time point will exceed the preset key control section. The key control section refers to the spatial boundary that is of great significance to water supply security, ecological protection, or economic and social development. For example, the boundary of a water source is the boundary line of a groundwater drinking water source protection area; the boundary of an ecological protection area is the boundary line of an important wetland, nature reserve, or other ecologically sensitive area; and the emergency warning line is the warning line set according to the regional disaster prevention and mitigation plan, which serves as the control section for early warning. Based on the predicted trend line, the time required to travel from the current interface location to the key control section location is calculated. The warning level for predicted boundary crossing is determined based on the certainty of the boundary crossing and the proximity of the crossing time. Three triggering conditions—interface migration triggering, level escalation triggering, and predicted boundary crossing triggering—are comprehensively evaluated. When multiple triggering conditions are simultaneously met, the highest level is taken as the final warning level, establishing a three-tiered warning system: blue, yellow, and red. A blue warning indicates abnormal signs or potential risks of seawater intrusion, requiring enhanced monitoring and attention. A yellow warning indicates significant changes in seawater intrusion or a short-term risk of boundary crossing, requiring preventative measures. A red warning indicates severe changes in seawater intrusion or that it has crossed the key control section, requiring immediate countermeasures. Warning information is generated based on the warning level, including the warning level, warning time, triggering reason, warning area, current status description, predicted trend, estimated boundary crossing time, and recommended measures. This information is then sent in tiers according to the warning level's established sending strategy.

[0047] Based on an electronic map, the system overlays coastline, administrative division, and key control section layers, marking the locations of each monitoring well. It supports interactive display of detailed monitoring well information. For each monitoring section, a two-dimensional profile is generated, showcasing the monitoring layer classification results for each well location, the fitted freshwater-transition interface curve, and the transition-saline water interface curve. Multiple historical time points can be selected to overlay interface curves from different periods on the same profile, using different colors to distinguish them. When an alert is triggered, the affected area is highlighted on the map according to the alert level, and an alert information window pops up on the monitoring system interface, with arrows indicating the predicted migration direction and speed of the interface on the map or profile. A comprehensive monitoring dashboard is designed, centrally displaying key indicator cards, alert status indicators, trend charts, and a list of recent alerts. Monitoring reports are automatically generated periodically, and report export and distribution are supported.

[0048] The second embodiment of the present invention provides a computer-readable storage medium, which includes a program for a method for intelligent identification and monitoring of seawater intrusion interfaces based on multi-parameter gradient features. When the program for intelligent identification and monitoring of seawater intrusion interfaces based on multi-parameter gradient features is executed by a processor, it implements the steps of the method for intelligent identification and monitoring of seawater intrusion interfaces based on multi-parameter gradient features.

[0049] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] 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 changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent identification and monitoring of seawater intrusion interfaces based on multi-parameter gradient features, characterized in that, Includes the following steps: Step S1: Set up monitoring sections in the target area, set up monitoring wells on the monitoring sections, set up several monitoring layers along the vertical direction of the monitoring wells, deploy multi-parameter sensor groups at each monitoring layer, and synchronously collect and acquire conductivity, chloride ion concentration, temperature and water level time series data of each layer. Step S2: Based on the collected multi-parameter data, calculate the vertical and horizontal conductivity gradients, chloride ion concentration gradients, and temperature gradients to form a multi-parameter gradient feature vector. Step S3: Construct an interface recognition model based on FixMatch framework, take the multi-parameter gradient feature vector as the model input, output the classification results of each monitoring layer belonging to freshwater layer, transition layer or saline layer, and determine the location of freshwater and saline-salt water interface according to probability distribution. Step S4: Using the multivariate state estimation method, a memory matrix is ​​constructed based on historical data of normal working conditions. The current multi-parameter gradient feature vector is matched with the memory matrix to calculate the residual. An adaptive dynamic threshold is set according to the statistical distribution of the residual. When the residual exceeds the threshold, it is determined that the interface has undergone significant migration and an update is triggered. Step S5: Calculate the seawater mixing ratio based on the identified freshwater and brackish water interface locations and measured chloride ion concentrations, classify the degree of intrusion, and predict the interface location based on multi-period interface location data in the time series. Step S6: When the interface is determined to have migrated significantly and the change in the position of the brackish water interface exceeds the preset threshold, the intrusion level is upgraded, or the predicted interface position crosses the preset key control section, a graded early warning message is automatically issued and displayed visually.

2. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S1, the construction of the three-dimensional monitoring network, specifically includes: Hydrogeological surveys were conducted to divide the target area into freshwater zone, brackish water transition zone, and saline water zone along the inland to ocean direction. Monitoring sections are set up along the normal direction of the coastline, and monitoring wells are set up in the freshwater area, the brackish water transition zone and the brackish water area for each section. Based on the monitoring depth requirements, a single-hole hybrid structure or a single-hole multi-layer structure is selected. For the single-hole multi-layer structure, a multi-level complete monitoring well method is adopted, in which the monitoring pipes of each level are radially distributed in the borehole according to the depth sequence, and the hydraulic separation between each monitoring layer is achieved by the water stopper. Multiple monitoring layers are set vertically in each monitoring well, and the spacing between the layers gradually increases towards the sea. Each monitoring pipe extends independently to the wellhead, and the wellhead is equipped with sensor connectors and data transmission terminals.

3. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S2, multi-parameter gradient feature extraction, specifically includes: The conductivity, chloride ion concentration, temperature and water level time series data collected from each monitoring layer are preprocessed. The preprocessing includes temperature compensation correction and identification of tidal or pumping disturbances based on water level data, and marking or removing disturbed data. For each monitoring well, a depth sequence of each monitoring layer is established, and the vertical gradients of conductivity, chloride ion concentration, and temperature between adjacent monitoring layers within each monitoring well are calculated. For the same monitoring section, the monitoring wells are sorted along the inland to ocean direction, and the horizontal gradients of conductivity and chloride ion concentration at the same standard depth layer between adjacent monitoring wells are calculated. The gradient symbols within each depth interval are arranged in depth order to construct a symbol sequence as a gradient symbol pattern feature; the correlation coefficient between the conductivity gradient and the chloride ion concentration gradient within the same depth interval is calculated to construct a consistency index feature; the ratio of the vertical gradient to the horizontal gradient on the same standard depth layer is calculated to construct a multi-scale fusion feature; the rate of change of each gradient feature within a preset time period is calculated to construct a time-series rate of change feature. The calculated basic gradient features, gradient sign pattern features, consistency index features, multi-scale fusion features, and time series change rate features are concatenated in sequence to form a multi-parameter gradient feature vector and then output.

4. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S3, the intelligent recognition of the interface based on the FixMatch semi-supervised framework, specifically includes: We construct an interface recognition model based on FixMatch, which includes a teacher network and a student network, and use unlabeled multi-parameter gradient feature vectors as unlabeled samples. Strong enhancement and weak enhancement operations are applied to the unlabeled samples respectively. The strong enhancement includes random noise injection, scaling transformation and channel discarding. The weak enhancement includes slight translation and scaling. The weakly enhanced unlabeled samples are input into the teacher network to generate pseudo-labels and confidence scores. Different confidence thresholds were set for freshwater, transition, and saline water layers. The comprehensive confidence of each pseudo-label was calculated based on the original confidence of the model, the temporal consistency score, and the spatial consistency score. Pseudo-labels with a comprehensive confidence score exceeding the corresponding category dynamic threshold were selected for training. Oversampling and feature space interpolation enhancement are performed on samples of the transition layer category to generate synthetic samples to expand the coverage of the transition layer in the feature space, and a re-labeling mechanism is triggered for uncertain samples at the category boundary. The enhanced unlabeled samples and labeled samples are input into the student network together. The student network parameters are updated by weighted sum of supervised loss and unsupervised consistency loss. The teacher network parameters are updated by exponential moving average of the student network parameters. The multi-parameter gradient feature vectors collected in real time are input into the trained student network, which outputs the classification results and probability distribution of each monitoring layer as freshwater layer, transition layer or saline layer, and determines the vertical depth position and horizontal cross-sectional position of the freshwater and saline water interface through linear interpolation based on the classification results.

5. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S3 involves determining the location of the freshwater-saltwater interface based on the probability distribution, specifically including: The probability distribution of each monitored layer belonging to freshwater layer, transition layer and saline layer was extracted, and a vertical probability profile of a single well was constructed in depth order, and the probability profile was smoothed. Based on a preset probability threshold, the freshwater-transition interface and the transition-saline interface are initially located. A probability-weighted interpolation method is used to locate the interface with sub-layer precision between adjacent monitoring layers and obtain the interface depth. Several standard depth layers were selected, and the classification probabilities of each monitoring well at the same depth were vertically interpolated to construct a horizontal probability profile. The horizontal positions of the freshwater-transition horizontal interface and the transition-saline water horizontal interface were determined by the probability weighted interpolation method. The vertical interface positioning results of each monitoring well and the horizontal interface positioning results of each standard depth layer are fused to generate a two-dimensional discrete interface point set, and a continuous and smooth two-dimensional interface curve is generated by fitting the spline interpolation method. The interface curves at different time points are overlaid and analyzed to calculate the rate of change of the interface depth at a fixed horizontal position and the rate of change of the interface horizontal position at a fixed depth.

6. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S4, the dynamic threshold adaptive determination based on multivariate state estimation, specifically includes: Normal operating condition data are selected from historical monitoring data, and the multi-parameter gradient feature vectors of all monitoring wells and all monitoring layers at each time point are concatenated into observation vectors. Representative observation vectors are selected to construct a memory matrix. The estimated vector is calculated using a sparse coding method. The current observation vector is sparsely decomposed in the space spanned by the memory matrix, and the sparse representation coefficients are obtained. The product of the memory matrix and the sparse representation coefficients is used as the estimated vector. Calculate the residual and squared prediction error between the observed vector and the estimated vector, and set a basic threshold based on the statistical distribution of the squared prediction error under normal operating conditions; Based on recent normal operating data, the current season, the current tidal phase, and the type of hydrogeological unit, the basic threshold is dynamically and adaptively adjusted, and three levels of thresholds are set: warning threshold, early warning threshold, and alarm threshold. The anomaly index is obtained by calculating the relative position of the current squared prediction error with each level threshold. When the anomaly index exceeds the level corresponding to the alarm threshold and continues to exceed it for multiple consecutive time points, it is determined to be an abnormal state, triggering the re-identification of the interface position.

7. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S5 involves classifying the level of intrusion, specifically including: Extract the measured values ​​of chloride ion concentration in the monitoring layer near the brackish water interface, determine the chloride ion concentration in the freshwater endmember and the chloride ion concentration in the seawater endmember, and calculate the seawater mixing ratio based on the mass conservation of chloride ion concentration. The classification standards are set based on chloride ion concentration, with chloride ion concentration from low to high corresponding to no invasion level, slight invasion level, moderate invasion level and severe invasion level. The classification standards are also set based on seawater mixing ratio, with mixing ratio from low to high corresponding to no invasion level, slight invasion level, moderate invasion level and severe invasion level. Calculate the intrusion distance between the current horizontal position of the brackish water interface and the historical reference position, and the depth change between the current vertical depth of the interface and the historical reference depth. Based on the intrusion distance and the depth change, set a classification standard for the interface position. The chloride ion concentration level, seawater mixing ratio level, and interface location level are weighted and integrated to calculate a comprehensive invasion severity score, and the final invasion severity level is determined based on the comprehensive score.

8. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S5 involves predicting the location of the predicted interface, specifically including: Based on historical monitoring results, a time series of the location of the brackish water interface was constructed in chronological order, and the time series was smoothed. A linear trend prediction model is adopted to calculate the annual rate of change by fitting a linear trend line of interface position change over time, and to predict the future interface position based on the product of the current interface position and the annual rate of change. The seasonal decomposition forecasting method is adopted to decompose the time series of interface location into trend, seasonal and random terms. The trend term is extrapolated and the seasonal term is cyclically repeated and then superimposed with the trend term to obtain the forecast value. A time-series prediction model is constructed using a long short-term memory network. The model takes the historical interface position sequence as input and learns the long- and short-term dependencies of interface migration through the forget gate, input gate and output gate structure to predict the future interface position. The results of linear trend prediction, seasonal decomposition prediction, and long short-term memory network prediction are weighted and integrated to output the predicted value of the interface position at a future set time point.

9. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, Step S6 automatically issues tiered early warning information, specifically including: Calculate the horizontal change between the current horizontal position and the historical reference position of the brackish water interface, as well as the vertical change between the current vertical depth and the historical reference depth. Set differentiated multi-level migration thresholds according to different hydrogeological units, compare the calculated changes with the multi-level migration thresholds, and determine the warning level triggered by interface migration. The current intrusion level is compared with the historical intrusion level to monitor the level escalation and determine the warning level triggered by the level escalation based on the escalation magnitude. Based on the interface position prediction results, determine whether the predicted interface position at a future set time point will cross the preset key control section, estimate the boundary crossing time, and determine the warning level triggered by the predicted boundary crossing based on the boundary crossing certainty and the distance of the boundary crossing time. The system comprehensively evaluates three triggering conditions: interface migration trigger, level upgrade trigger, and prediction over-limit trigger. When multiple triggering conditions are met simultaneously, the highest level is taken as the final warning level.

10. The intelligent identification and monitoring method for seawater intrusion interfaces based on multi-parameter gradient features according to claim 1, characterized in that, The visualization process described in step S6 specifically includes: Using the electronic map as the base map, overlay the coastline layer, administrative division layer, and key control section layer, and mark the location of each monitoring well. Two-dimensional profile diagrams were drawn for each monitoring section to show the monitoring layer classification results, the fitted freshwater-transition interface curves, and the transition-saline water interface curves for each monitoring well location. It supports selecting multiple historical time points and overlaying the interface curves from different periods on the same cross-section, using different colors to distinguish the curves from different periods; When an alert is triggered, the affected area will be highlighted on the map according to the corresponding color based on the alert level, and an alert information window will pop up on the monitoring system interface.

Images