Wetland nitrogen cycle optimization method and system based on machine learning

Abstract

The invention discloses a wetland nitrogen cycle optimization method and system based on machine learning, and belongs to the technical field of ecological engineering.The method comprises the steps that according to a comprehensive environment data set, a transverse lateral infiltration path and a vertical priority flow path are recognized, spatio-temporal evolution simulation is conducted through a cellular automaton model, and a nitrogen migration path is constructed; analyzing the interaction between the water flow driving flux and the underground water level on the path resistance, and determining the broadening degree of the transverse lateral seepage path; if the widening degree exceeds a preset threshold value, integrating the water body turbulence intensity data and the flow velocity shear effect value, and obtaining a bubble escape rate and a surface sediment disturbance frequency; calculating the nitrogen saturation and the spatial distribution of the sediment particle suspension amount, and obtaining the total re-release flux through flux overlay analysis; and constructing a nitrogen loss flux model in combination with the fluctuation rule, simulating the change of the nitrogen loss proportion, and determining the net nitrogen loss flux. According to the invention, fine monitoring of the dynamic change of wetland nitrogen can be realized.

Description

Technical Field

[0001] This invention belongs to the field of ecological engineering technology, specifically relating to a wetland nitrogen cycle optimization method and system based on machine learning. Background Technology

[0002] As an important ecosystem, wetlands play a crucial role in regulating water resources, purifying pollutants, and maintaining biodiversity. The migration and release of nitrogen in the soil directly affects the stability of wetland water quality and the surrounding environment, which is of great significance to regional ecological security.

[0003] Current research and monitoring of nitrogen in wetlands mostly rely on sampling analysis at fixed points or short-term water quality observations. These methods are difficult to capture the actual path of nitrogen at different soil depths over time, and are also difficult to reflect the continuous impact of water level fluctuations and water flow on the lateral diffusion and upward release of nitrogen, resulting in a significant deviation in the understanding of the true migration pattern of nitrogen.

[0004] Therefore, accurately quantifying the dynamic evolution of nitrogen migration pathways with water level changes and establishing their response relationship with repeated surface nitrogen re-release are prerequisites for precisely understanding the distribution and release patterns of nitrogen loss in wetlands. However, existing technologies struggle to achieve high-precision monitoring of these processes, which has become a key issue hindering the scientific management of wetland nitrogen cycles. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a wetland nitrogen cycle optimization method and system based on machine learning, thereby resolving the issues present in the background art.

[0006] To achieve the aforementioned objectives, this invention proposes a wetland nitrogen cycle optimization method based on machine learning, comprising: Collect multi-level environmental data of wetland soil profiles to generate a comprehensive environmental dataset; Based on the comprehensive environmental dataset, we identified lateral seepage paths and vertical priority flow paths through spatial overlay and temporal analysis, and constructed nitrogen migration paths by using a cellular automata model to simulate spatiotemporal evolution. Based on the analysis of nitrogen migration pathways, the interaction between water flow-driven flux and groundwater level on path resistance is used to determine the degree of widening of lateral seepage pathways. If the widening exceeds the preset threshold, the water turbulence intensity data and flow velocity shear effect value are integrated to obtain the bubble escape rate and surface sediment disturbance frequency. Based on the bubble escape rate and the surface sediment disturbance frequency, the spatial distribution of nitrogen saturation and sediment particle suspension were calculated, and the total re-release flux was obtained through flux superposition analysis. A nitrogen loss flux model was constructed by combining the fluctuation pattern. The changes in the nitrogen loss ratio were simulated under multiple rounds of water level fluctuations, and spatial mapping was performed to determine the net nitrogen loss flux.

[0007] Furthermore, a comprehensive environment dataset is generated, including the following steps: Monitoring points and sensors were set up in multiple soil profiles to form a sensor network. The redox potential sequence of each monitoring point in the target wetland soil profile was obtained through the sensor network, and the corresponding water level data and dissolved oxygen data were retrieved simultaneously. Time-aligned processing was performed on redox potential sequences, water level data, and dissolved oxygen data to generate multi-source time series data. Based on multi-source time series data, the potential gradient distribution data is obtained by calculating the ratio of the difference in redox potential values ​​between adjacent layers to the depth interval. Based on water level data and soil effective stress model, soil pore structure data is calculated. Based on potential gradient distribution data, the root penetration influence range data is extracted using a threshold segmentation algorithm. Potential gradient distribution data, soil pore structure data, and influence range data are used as a comprehensive environmental dataset.

[0008] Furthermore, the nitrogen migration pathway is constructed, including the following steps: Based on soil pore structure data and influence range data, lateral seepage paths are identified through spatial overlay. The comprehensive environmental dataset also includes real-time response sequences of seepage. By analyzing the time series, monitoring points with response time intervals less than a critical threshold are identified and connected to obtain the vertical priority flow path. The lateral seepage path is then fused with the vertical priority flow path to generate a preliminary migration path. Based on the initial migration path, the permeability adjustment coefficient is calculated and applied in conjunction with soil pore structure data to correct the initial migration path and obtain the adjusted migration path. Based on adjusting the migration path, a cellular automata model is used to simulate the spatiotemporal evolution and construct a dynamic nitrogen migration path.

[0009] Further, determining the extent of widening of the lateral seepage path includes the following steps: The water level fluctuation sequence was obtained, and combined with the nitrogen migration path, an autoregressive moving average model including exogenous variables was used to simulate and generate the distribution data of water flow-driven flux. Analyze water level fluctuation sequences and soil pore structure data to generate data on the impact of path resistance; Based on distribution and impact data, the degree of widening of the lateral seepage path is determined by calculating the widening index and accumulating the widening increment.

[0010] Further, obtaining the bubble escape rate and the surface sediment disturbance frequency includes the following steps: Three-dimensional velocity data of monitoring points are collected by acoustic Doppler current meter. The turbulence intensity data of water body is obtained by the velocity fluctuation component in the three-dimensional velocity data. The velocity shear effect value is obtained by analyzing the vertical velocity gradient in the three-dimensional velocity data. A weighted average method was used to integrate the water turbulence intensity data and velocity shear effect values ​​from the monitoring points to generate a comprehensive value. Based on the comprehensive values ​​from the monitoring points, the bubble escape rate and the surface sediment disturbance frequency are calculated.

[0011] Furthermore, the total re-release flux is obtained through flux superposition analysis, including the following steps: Calculate the nitrogen partial pressure and saturation at each monitoring point to generate saturation distribution data; Based on surface sediment disturbance frequency data, the particulate nitrogen release model was applied to calculate the suspended nitrogen at each monitoring point and obtain suspended nitrogen distribution data. By integrating saturation distribution data and suspended matter distribution data, the total re-release flux at each monitoring point under multiple water level fluctuations is calculated through flux overlay analysis.

[0012] Furthermore, determining the net nitrogen loss flux also includes the following steps: The surface oxide layer thickness data and sediment particle blockage data are integrated into the nitrogen loss flux model. Multiple rounds of water level fluctuation data are input to simulate and generate dynamic change data of nitrogen loss ratio. Using the parameter sensitivity analysis method, the surface oxide layer thickness data was set as a variable to calculate the nitrogen loss ratio corresponding to different surface oxide layer thickness data; Based on the data on the impact of sediment particle blockage, the dynamic change data of nitrogen loss ratio is iteratively adjusted by a correction function to generate the adjusted dynamic change data of nitrogen loss ratio. The calculated total re-release flux is combined with the adjusted dynamic change data of nitrogen loss ratio to calculate the net nitrogen loss flux.

[0013] Furthermore, determining the net nitrogen loss flux also includes the following steps: Based on the dynamic change data of nitrogen loss ratio, combined with the surface oxide layer thickness data and sediment particle blockage data, a distribution model of wetland surface nitrogen re-release is constructed. Multiple rounds of water level fluctuation data were input into the distribution model of nitrogen re-release from the wetland surface to simulate and generate time series data of nitrogen re-release flux at monitoring points. The time series data of nitrogen re-release flux were analyzed to extract amplitude and frequency, and to obtain the fluctuation characteristics data of nitrogen re-release.

[0014] The present invention also provides a wetland nitrogen cycle optimization system based on machine learning, which is used to implement the above-described method, and the system includes: The data acquisition module collects multi-level environmental data of wetland soil profiles and generates a comprehensive environmental dataset. The first analysis module identifies lateral seepage paths and vertical priority flow paths based on a comprehensive environmental dataset through spatial overlay and temporal analysis, and uses a cellular automata model to simulate spatiotemporal evolution and construct nitrogen migration paths. The second analysis module analyzes the interaction between the nitrogen migration path analysis water flow-driven flux and the groundwater level on the path resistance to determine the degree of widening of the lateral seepage path. If the widening degree exceeds the preset threshold, it integrates water turbulence intensity data and flow velocity shear effect value to obtain the bubble escape rate and surface sediment disturbance frequency. The loss calculation module calculates the spatial distribution of nitrogen saturation and sediment particle suspension based on the bubble escape rate and surface sediment disturbance frequency. Through flux superposition analysis, the total re-release flux is obtained. Combined with the fluctuation law, a nitrogen loss flux model is constructed to simulate the change of nitrogen loss ratio under multiple rounds of water level fluctuations and perform spatial mapping to determine the net nitrogen loss flux.

[0015] The beneficial effects of this invention are as follows: This invention generates a comprehensive environmental dataset by collecting multi-level environmental data from wetland soil profiles. Based on this dataset, two key pathways—lateral seepage and vertical preferential flow—are identified through spatial overlay and temporal analysis. A cellular automata model is then used to simulate the spatiotemporal evolution, thus dynamically constructing the complex migration pathways of nitrogen in the soil. Next, the interaction between water flow-driven flux and groundwater level on the constructed pathways is analyzed to determine the extent of pathway widening. When widening exceeds a threshold, hydrodynamic data is further integrated to address the impact of hydrological changes on the physical morphology of nitrogen migration channels, accurately acquiring the resulting bubble escape rate and sediment disturbance frequency. Subsequently, the spatial distribution of nitrogen saturation and particulate suspension is calculated based on these rates and frequencies, and the total re-release flux is obtained by combining flux overlay analysis. Finally, a nitrogen loss flux model is constructed to simulate and map the net nitrogen loss flux, thereby completing a quantitative assessment of the entire chain from nitrogen migration to final gas loss. Of particular note is that this invention solves the key problem in the prior art that it is difficult to capture the actual path of nitrogen in the soil over time and cannot reflect the dynamic impact of water level changes on the lateral diffusion and surface re-release of nitrogen, leading to a bias in the understanding of the true migration law of nitrogen. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a wetland nitrogen cycle optimization method based on machine learning according to the present invention. Figure 2 This is an overall flowchart of the present invention; Figure 3 This is a schematic diagram of the nitrogen loss ratio sensitivity analysis of the present invention; Figure 4 This is a schematic diagram illustrating the dynamic changes in total nitrogen re-release flux according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] It is understood that the terms "first," "second," etc., used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this application, a first script may be referred to as a second script, and similarly, a second script may be referred to as a first script.

[0019] like Figure 1 and Figure 2 As shown, a machine learning-based wetland nitrogen cycle optimization method includes: S1: Collect multi-level environmental data of wetland soil profiles to generate a comprehensive environmental dataset.

[0020] In one specific embodiment, generating a comprehensive environment dataset includes the following steps: Monitoring points and sensors were set up in multiple soil profiles to form a sensor network. The redox potential sequence of each monitoring point in the target wetland soil profile was obtained through the sensor network, and the corresponding water level data and dissolved oxygen data were retrieved simultaneously.

[0021] Specifically, in the target wetland area, a grid-like wireless sensor network containing multiple sensors is deployed, with sensors buried layer by layer along the vertical soil profile, such as at depths of 10 cm, 30 cm, and 50 cm below the surface. The sensors collect data at high frequency and transmit the data in real time via wireless technologies such as LoRa or NB-IoT. This data is then integrated to form a multi-layered redox potential sequence. Water level and dissolved oxygen data corresponding to the study period are then retrieved from the database of the local hydrological and environmental monitoring station.

[0022] Time-aligned processing was performed on the redox potential sequence, water level data, and dissolved oxygen data to generate multi-source time series data.

[0023] Since the redox potential sensor performs real-time high-frequency sampling, while water level data and dissolved oxygen data in surface water may be low-frequency or non-uniform, a dynamic time warping algorithm is first used to align these three time series to address the issue of inconsistent sampling frequencies. The process of using the dynamic time warping algorithm to align the time series data is existing technology and will not be described here.

[0024] Based on multi-source time series data, potential gradient distribution data is obtained by calculating the ratio of the difference in redox potential values ​​between adjacent layers to the depth interval. Soil pore structure data is then calculated based on water level data and the effective soil stress model.

[0025] For sensors deployed at different depths within the soil profile, redox potential values ​​at adjacent depths are extracted at each identical time point. The difference between these two potential values ​​is then divided by the vertical depth interval between them to calculate the redox potential gradient for that depth range. Subsequently, spatial interpolation methods such as Kriging are used to expand the discrete gradient point data into a continuous gradient distribution covering the entire target wetland area. This calculation is repeated for all sensor pairs at adjacent depths, ultimately generating the potential gradient distribution data.

[0026] In addition, by using aligned water level sequence data and pre-input soil texture parameters, a soil pore structure data estimation model based on the effective stress principle is applied. By performing point-by-point calculations on the water level data throughout the entire time series, the time series data of soil pore structure data is finally calculated and generated.

[0027] The soil pore structure estimation model is a hybrid model integrating classical soil mechanics theory and machine learning correction. The core of this model is based on Terzaghi's effective stress principle, which states that the effective stress of soil equals the total stress minus the pore water pressure. In this invention, the total stress at a specific depth can be considered constant, while the pore water pressure is directly calculated from real-time monitored water level data. When using the model for calculations, the dynamic change sequence of pore water pressure is calculated based on the changes in water level sequence data, thereby obtaining the change sequence of effective stress. Then, the classical unidirectional consolidation theory is applied to estimate the compression or rebound of the soil skeleton caused by water level fluctuations, thus obtaining the soil pore structure data.

[0028] Based on potential gradient distribution data, the root penetration influence range data is extracted using a threshold segmentation algorithm.

[0029] Potential gradient distribution data, soil pore structure data, and influence range data are used as a comprehensive environmental dataset.

[0030] To extract the influence range of root infiltration, a threshold segmentation method is applied to the potential gradient distribution data. Specifically, a gradient threshold is first set based on botanical knowledge, and grid areas exceeding this threshold are identified as areas of strong influence dominated by root activity. By calculating the area or equivalent radius of these areas, quantified data on the influence range of root infiltration can be obtained.

[0031] For example, setting the gradient threshold to 50 mV / m, the threshold segmentation algorithm traverses each grid region and compares its gradient value with the preset gradient threshold. If the gradient value of a grid region is greater than 50 mV / m, the grid region is marked as 1, indicating that it belongs to the root active influence zone; conversely, if the gradient value is less than or equal to 50 mV / m, it is marked as an uninfluenced zone. After the traversal is completed, all grid regions marked as 1 together constitute the area of ​​root infiltration influence.

[0032] Finally, at each time step, the values ​​of each indicator are assigned to their corresponding monitoring points. In this way, a comprehensive environmental dataset containing data reflecting the soil aerobic / anaerobic interface, water channels, and the intensity of biological activity can be obtained.

[0033] S2: Based on the comprehensive environmental dataset, the lateral seepage path and the vertical priority flow path are identified through spatial overlay and temporal analysis, and the spatiotemporal evolution is simulated using a cellular automata model to construct the nitrogen migration path.

[0034] In one specific embodiment, constructing a nitrogen migration pathway includes the following steps: Based on soil pore structure data and influence range data, lateral seepage paths are identified through spatial overlay.

[0035] The rate of change of soil pore structure data at monitoring points is calculated by dividing the difference in soil pore structure data between adjacent time points by the time interval. If the rate of change of soil pore structure data at a monitoring point is greater than a preset positive threshold, such as 2%, then that monitoring point is marked as a pore expansion zone. Subsequently, spatial overlay analysis is performed on the pore expansion zone and the affected area data. If a monitoring point is simultaneously identified as both a pore expansion zone and a root infiltration affected area, then the monitoring area covered by the monitoring point is determined to be a high-permeability area caused by the combined effects of root activity and hydrological changes. All identified high-permeability areas are connected to form a lateral seepage path.

[0036] The comprehensive environmental dataset also includes real-time response sequences of seepage. By analyzing the time series, monitoring points with response intervals less than a critical threshold are identified and connected to obtain the vertical priority flow path. The lateral seepage path is then fused with the vertical priority flow path to generate a preliminary migration path.

[0037] To identify vertically preferential flow paths, it is first necessary to construct and analyze real-time response sequences of infiltration from a comprehensive environmental dataset. A real-time response sequence is a dataset formed by the significant changes in monitoring values ​​of multiple sensors deployed at different depths on the same vertical profile over time after a water infiltration event such as rainfall or irrigation. The real-time response sequence reflects the dynamic process of water infiltration from shallow to deep layers.

[0038] To accurately capture this process, the starting time of the rainfall or irrigation event must first be determined as the trigger signal for analysis. This can be achieved by accessing real-time rainfall data from a weather station or start / stop records of farmland irrigation systems. Once the event time is determined, the system will use this as a baseline zero point and track the response times of sensors at different depths. If the actual response delay time is less than the baseline response time multiplied by a preset discount factor (e.g., 0.8), it is determined that a vertical priority flow channel exists at the location of that deep sensor. Finally, the monitoring points above and below the points with the vertical priority flow channel are connected. The connected monitoring points form a channel for rapid water infiltration, i.e., the vertical priority flow path.

[0039] The baseline response time is theoretically calculated using Darcy's law or Richards' equation, combined with the average permeability coefficient of the soil matrix and the vertical distance between sensors. The identified vertical preferential flow path is then fused with the previously generated lateral seepage path in three-dimensional space to generate a preliminary migration path containing both lateral and longitudinal migration channels. The average permeability coefficient is a static baseline value pre-determined or estimated for the target wetland area; it represents the macroscopic average hydraulic conductivity of the soil as a whole in the absence of preferential flow influence. This value is typically determined through in-situ testing or laboratory experimental analysis.

[0040] Based on the initial migration path, a permeability adjustment coefficient is calculated and applied using soil pore structure data to correct the initial migration path and obtain an adjusted migration path.

[0041] For each monitoring point in the initial migration path, the calculated soil pore structure data is extracted again. Based on empirical formulas in soil hydraulics, such as the Kozeny-Kaman equation, the soil pore structure data at the path location is converted into a permeability adjustment coefficient. The permeability adjustment coefficient quantifies the change in permeability of a specific migration path due to factors such as water level fluctuations and root activity. The specific formula of the Kozeny-Kaman equation is existing technology and will not be given here. Subsequently, the permeability adjustment coefficient is applied to the initial migration path to adjust the path's distribution characteristics. The adjustment methods include: for lateral seepage paths, increasing the influence width or weight of monitoring points according to the permeability adjustment coefficient; for vertical priority flow paths, enhancing the vertical connectivity probability of monitoring points according to the permeability adjustment coefficient.

[0042] Specifically, the adjustment can be made as follows: In the initial migration path, each monitoring point has a basic influence width, which, based on actual surveys, can be set to 1 meter. The adjusted influence width will be calculated using the following formula: New influence width = Basic influence width × Permeability adjustment coefficient. Taking a permeability adjustment coefficient of 1.2 as an example, the new influence width is 1 meter × 1.2 = 1.2 meters. This means that compared to the initial migration path, the adjusted migration path expands the influence range of the monitoring point on the surrounding area by 20%.

[0043] The probability of water flow transfer between adjacent vertical monitoring points is determined by the basic connectivity probability. The adjusted vertical connectivity probability is calculated using the following formula: New connectivity probability = 1 - (1 - basic connectivity probability) / permeability adjustment coefficient. Taking the aforementioned permeability adjustment coefficient of 1.2 as an example, if the basic connectivity probability is 0.6, then the new connectivity probability is 1 - (1 - 0.6) / 1.2 ≈ 0.67.

[0044] Through this adjustment, the original, rather rough preliminary migration path is given physical properties that can reflect the differences in local hydraulic conduction capacity.

[0045] Based on adjusting the migration path, a cellular automata model is used to simulate the spatiotemporal evolution and construct a dynamic nitrogen migration path.

[0046] A cellular automata model is used to simulate the dynamic evolution of nitrogen migration paths. First, the target wetland area is discretized into a three-dimensional cellular mesh. Then, the adjusted migration path generated in the previous step is used as the initial state of the cellular automata. Depending on whether a cell is on the adjusted migration path, each cell is assigned a path or non-path state, and a corresponding permeability coefficient value is assigned to cells in the path state. Next, an evolution rule for the cell state is defined, driven by time series data of inferred soil pore structure and root infiltration influence range. For example, the evolution rule can be set as follows: if the number of surrounding path-state cells and the sum of their permeability coefficients exceed a critical threshold for a non-path-state cell, and the soil pore structure data of this non-path-state cell is increasing, then it has a certain probability of transitioning to a path state in the next time step. This probability is the transition probability, calculated as follows: First, a transformation potential score is calculated for each non-path cell. This score quantifies the likelihood of the cell transforming into a path. The transformation potential score S = W1 × (neighborhood influence) + W2 × (self-condition), where neighborhood influence is the sum of the permeability coefficients of all path-state cells in the surrounding neighborhood, and self-condition is the rate of change of the cell's own soil pore structure data. This term is positive only when the soil pore structure data is increasing, indicating that its physical structure is conducive to the formation of a new permeation path. W1 and W2 are preset weighting coefficients used to balance the importance of neighborhood influence and self-condition. These weights can be set through expert knowledge or obtained through model calibration. For example, W1 is 0.4 and W2 is 0.6. Then, the transformation potential score is mapped to a transition probability between 0 and 1 using the Sigmoid function. Key parameters such as weights and thresholds in this method can be determined using machine learning methods such as genetic algorithms to achieve the best fit between the simulation results and the measured data.

[0047] In one embodiment, the critical threshold is randomly generated. That is, at each time step of the cellular automaton, the transition probability of all non-path cells added to the soil pore structure data is calculated, and a random number between 0 and 1 (the critical threshold) is generated. If the transition probability is greater than the critical threshold, the cell transitions to a path state in the next time step; otherwise, it remains in a non-path state. This is because while the formation of water flow paths follows definite physical laws, at the microscopic scale, the arrangement of soil particles, the distribution of organic matter, and the activity of microorganisms are all highly random. These unpredictable factors may promote or hinder the formation of a new path at some point. Therefore, the critical threshold in this embodiment is randomly generated. In other embodiments, the critical threshold can also be a pre-set fixed value.

[0048] By iteratively executing these evolutionary rules at consecutive time steps, a dynamic change model of nitrogen migration paths that can simulate path widening, contraction, or migration can be constructed, providing a dynamic evolutionary data foundation for subsequent accurate analysis of migration resistance of different paths.

[0049] S3: Based on the analysis of nitrogen migration pathways, the interaction between water flow-driven flux and groundwater level on path resistance is used to determine the degree of widening of the lateral seepage path.

[0050] In one embodiment, determining the degree of widening of the lateral seepage path includes the following steps: The water level fluctuation sequence was obtained, and combined with the nitrogen migration path, an autoregressive moving average model including exogenous variables was used to simulate and generate the distribution data of water flow-driven flux.

[0051] First, water level data for the study period is acquired from wetland hydrological monitoring stations or field-deployed sensors to form a water level fluctuation sequence. The constructed nitrogen migration path is then run to simulate and generate nitrogen migration path distribution data at the same time step as the water level fluctuation sequence. This nitrogen migration path distribution data is a spatiotemporal data sequence containing spatial information such as the location, width, and permeability coefficient of the migration path at each time point. Next, an autoregressive moving average model incorporating exogenous variables is used for time series analysis. In the autoregressive moving average model, the water level fluctuation sequence is used as input to the autoregressive and moving average terms, while the time-varying path permeability coefficient from the nitrogen migration path distribution data is used as an exogenous variable input. The model output predicts the water level value at a specific location at a future time, thereby calculating the hydraulic gradient between points. Finally, based on a variant of Darcy's law, combining the permeability coefficient as an exogenous variable input with the hydraulic gradient predicted by the model, the distribution data of the flow-driven flux is calculated.

[0052] In a preferred embodiment, constructing an autoregressive moving average model includes the following steps: First, the obtained water level fluctuation sequence is subjected to an enhanced Dickey-Fowler test for stationarity. If the sequence is non-stationary, it is subjected to first-order or multiple-order differencing until it becomes stationary; the number of differencing operations is determined as the integral order of the model. Next, the autocorrelation function plot and partial autocorrelation function plot of the stationary sequence are analyzed, and the autoregressive order and moving average order are determined in conjunction with the Akaike Information Criterion. Based on the determined autoregressive order, moving average order, and integral order, an autoregressive moving average model is constructed.

[0053] It is important to note that the flow-driven flux is not a direct output of the autoregressive moving average model, but rather a secondary calculation based on its prediction results. In this invention, the flow-driven flux Q is calculated using a variant of Darcy's law: Q = K × i, where the permeability coefficient K is taken from the nitrogen migration path distribution data as an exogenous variable, and the hydraulic gradient i is calculated based on the water level difference at different future spatial points predicted by the autoregressive moving average model.

[0054] By analyzing water level fluctuation sequences and soil pore structure data, data on the impact of path resistance are generated.

[0055] Path resistance is a physical quantity characterizing the hindering effect of soil pore structure on water flow and solute transport; its variation is influenced by both water level and the characteristics of the path itself. In practice, a change sequence composed of water level fluctuation sequences and soil pore structure data is extracted. By calculating the cross-correlation function between the water level fluctuation sequences and the soil pore structure change sequences, the response intensity and time delay of the influence of water level changes on changes in soil pore structure data can be determined.

[0056] This invention employs a cross-correlation function to analyze historical water level fluctuation sequences and soil pore structure change data sequences. The cross-correlation function determines whether water level changes drive pore structure changes, or vice versa, by observing whether the correlation peak occurs with a positive or negative delay. The specific delay time at which the peak occurs is then determined, which is the most significant average response lag time between the two variables. This lag time is a key parameter for constructing an effective predictive model because it indicates which historical water level data should be used to predict path resistance at the current moment.

[0057] Based on the correlation analysis above, a linear or nonlinear regression model incorporating a time delay term is constructed. Water level data with a time delay is used as the independent variable, and the corresponding soil pore structure data or its derived path resistance value is used as the dependent variable. The path resistance value used as training data is specifically the reciprocal of the permeability coefficient. Finally, the regression model is trained and its parameters are fitted using historical datasets to obtain a regression model that can predict changes in path resistance based on changes in water level. The regression model is used at each time step to dynamically generate path resistance impact data based on real-time water level data.

[0058] Based on distribution and impact data, the degree of widening of the lateral seepage path is determined by calculating the widening index and accumulating the widening increment.

[0059] The obtained data on flow-driven flux distribution and the dynamic impact data of path resistance are integrated to calculate the specific degree of lateral seepage path widening. First, a widening index is defined as: Widening Index = Flow-driven flux value / Path resistance value. The physical meaning of this formula is that areas with high flux and low resistance are more prone to path widening. Then, monitoring points at each time step are iterated through, and the corresponding flow-driven flux distribution value and the dynamic impact value of path resistance for each monitoring point are substituted into the formula to calculate the widening index for that unit. Finally, the calculated widening index is multiplied by the initial width of the monitoring point to obtain the widening increment of the monitoring point at the current time step. The specific degree of lateral seepage path widening is obtained by summing the widening increments of all time steps.

[0060] S4: If the widening exceeds the preset threshold, integrate the water turbulence intensity data and flow velocity shear effect value to obtain the bubble escape rate and surface sediment disturbance frequency.

[0061] In one embodiment, obtaining the bubble escape rate and the surface sediment disturbance frequency includes the following steps: The three-dimensional velocity data of the monitoring points are collected by an acoustic Doppler current meter. The turbulence intensity data of the water body is obtained by the velocity fluctuation component in the three-dimensional velocity data, and the velocity shear effect value is obtained by analyzing the vertical velocity gradient in the three-dimensional velocity data.

[0062] When the lateral seepage path widens beyond a preset threshold, it signifies that the groundwater flow has shifted from diffuse, slow seepage to concentrated, rapid channel flow. This leads to a concentrated upward discharge of nitrogen-laden groundwater into the wetland surface water in a localized area. This concentrated discharge creates strong local hydrodynamic disturbances, and the strong upward flow physically scours and disturbs surface sediments, causing sediment particles to be resuspended and releasing particulate nitrogen attached to them. Furthermore, the disturbance alters the pressure and mass transfer conditions at the sediment-water interface, potentially accelerating the explosive release of nitrogen gas produced by denitrification from the sediments in the form of bubbles. This sudden nitrogen re-release process is difficult to capture using traditional steady-state models, and without quantification, it will lead to significant biases in the assessment of the wetland's total nitrogen budget, particularly nitrogen loss flux. To address this issue, this embodiment proposes the following steps.

[0063] First, an acoustic Doppler current meter was deployed near the bottom of the wetland water body to collect three-dimensional velocity data at monitoring points in real time with a high sampling frequency. Based on the collected three-dimensional instantaneous velocity data, it was decomposed into average velocity and fluctuating velocity using the Reynolds decomposition method. The Reynolds decomposition method is a statistical method used to analyze turbulence, and its specific details are not described here. The root mean square of the velocity fluctuation component was calculated to obtain the turbulence intensity data of the water body, which is used to characterize the turbulent energy inside the water body. At the same time, by moving the acoustic Doppler current meter probe, the average velocity at multiple points 5 cm, 10 cm, and 20 cm above the vertical profile near the bottom was acquired. By applying the logarithmic law of wall velocity to the velocity profile data of these points, the velocity shear effect value was calculated.

[0064] A weighted average method was used to integrate the water turbulence intensity data and velocity shear effect values ​​from the monitoring points to generate a comprehensive value.

[0065] Multiply the water turbulence intensity data value at the same moment by a preset weight of 0.6, and multiply the water velocity shear effect value by a preset weight of 0.4. Then sum the products of the two to generate a comprehensive value that can comprehensively reflect the intensity of hydrodynamic disturbance. The preset weights can be determined based on experiments.

[0066] Based on the comprehensive values ​​from the monitoring points, the bubble escape rate and the surface sediment disturbance frequency are calculated.

[0067] A shear stress threshold is set, pre-determined based on the physical properties of the target wetland surface sediments. Regions with a composite value exceeding the shear stress threshold are identified as areas of intense hydrodynamic disturbance. Within these regions, the time series of bubble escape rates is calculated using gas flux measurement chambers deployed on the sediment surface, combined with Fick's first law, which describes the diffusion phenomenon. The specific formula is: bubble escape rate = gas diffusion coefficient * nitrogen concentration at the sediment-water interface. The nitrogen diffusion coefficient can be found in standard physicochemical handbooks. Simultaneously, the time series of surface sediment disturbance frequencies is obtained by statistically analyzing the number of times the composite value exceeds the shear stress threshold per unit time.

[0068] S5: Based on the bubble escape rate and the surface sediment disturbance frequency, calculate the spatial distribution of nitrogen saturation and sediment particle suspension, respectively, and obtain the total re-release flux through flux superposition analysis.

[0069] In one specific embodiment, the total re-release flux is obtained through flux superposition analysis, including the following steps: Calculate the nitrogen partial pressure and saturation at each monitoring point to generate saturation distribution data.

[0070] This embodiment calculates the spatial distribution of nitrogen saturation, which serves as the basis for dissolved nitrogen release, and the spatial distribution of sediment particle suspension, which serves as the basis for particulate nitrogen release, thereby laying a data foundation for subsequent flux superposition analysis and accurate calculation of total nitrogen re-release flux.

[0071] To quantify the flux of dissolved nitrogen escaping in gaseous form, this embodiment first calculates the spatial distribution of nitrogen saturation, which forms the basis for its release. Specifically, for each monitoring point, based on the bubble escape rate at that point, the nitrogen partial pressure at the sediment-water interface is calculated using the two-film theory. The two-film theory is a model describing cross-phase mass transfer processes, assuming that there exists a static thin film on each side of the gas-liquid or liquid-solid interface, through which matter can only be transferred via molecular diffusion. The rate of matter transfer (i.e., flux J) is proportional to the total mass transfer resistance across these two films and the concentration difference across the interface. Finally, based on Henry's Law, the nitrogen solubility is calculated in relation to its partial pressure in the gas phase. The ratio of the nitrogen partial pressure to the nitrogen saturation partial pressure under the total pressure of the deep water body is used as the calculated nitrogen saturation. Henry's Law is a fundamental physical law describing the solubility of gases in liquids. It states that at a given temperature, the solubility of a gas is proportional to its partial pressure in the gas phase. After traversing all monitoring points and completing the calculations, saturation distribution data covering all monitoring points within the entire target wetland area can be generated.

[0072] Based on surface sediment disturbance frequency data, a particulate nitrogen release model was applied to calculate the suspended nitrogen content at each monitoring point, thereby obtaining suspended nitrogen distribution data.

[0073] To quantify the particulate nitrogen flux released due to sediment disturbance, this embodiment first calculates the suspended sediment particles, which form the basis for its release. Specifically, a particulate nitrogen release model is applied, using the bed shear stress obtained from comprehensive numerical conversion as input to estimate the suspended sediment particles. The particulate nitrogen release model is based on classical sediment dynamics theory, using the bed shear stress obtained from comprehensive numerical conversion as the main input, combined with the critical initiation shear stress and erosion rate parameters of the sediment to calculate the suspended particles. For example, when the bed shear stress exceeds the critical value, the suspension rate can be expressed as E=M*(τ / τ_c-1), where E is the suspension rate, M is the erosion rate constant, τ is the bed shear stress, and τ_c is the critical shear stress. The total suspended particles are then obtained by integrating the suspension rate over a period of time. This calculation is performed on all monitoring points, ultimately generating suspended particle distribution data covering the entire target wetland area.

[0074] By integrating saturation distribution data and suspended matter distribution data, the total re-release flux at each monitoring point under multiple water level fluctuations is calculated through flux overlay analysis.

[0075] To obtain the total nitrogen re-release flux and its fluctuation patterns, such as Figure 4As shown, this embodiment integrates flux overlay analysis. For dissolved nitrogen flux, based on the obtained spatial distribution of nitrogen saturation, the previously calculated nitrogen concentration (nitrogen solubility) is used, combined with the measured or empirical water nitrogen concentration, to calculate the diffusion flux driven by the concentration gradient using Fick's first law. For particulate nitrogen flux, based on the spatial distribution of suspended sediment particles and the pre-determined particulate nitrogen content C in the surface sediment, the flux J released due to particle resuspension is calculated using the formula J = R × C, where R is the suspended sediment particles. Subsequently, at each monitoring point and at each time step, the dissolved nitrogen flux and the particulate nitrogen flux are added to obtain the total re-release flux at that point. Finally, the entire multi-round historical water level fluctuation sequence is used as the model driver, and the above full-process calculation is repeated at each time step of the sequence to generate a time series of total re-release flux for each monitoring point. By analyzing the statistical characteristics such as peaks, troughs, and periods of these time series, the fluctuation pattern of surface sediment re-release flux reflecting the influence of seasons and hydrological events can be obtained.

[0076] S6: Construct a nitrogen loss flux model by combining fluctuation patterns, simulate the change in nitrogen loss ratio under multiple rounds of water level fluctuations, and perform spatial mapping to determine the net nitrogen loss flux.

[0077] In one embodiment, determining the net nitrogen loss flux includes the following steps: The surface oxide layer thickness data and sediment particle blockage data are integrated into the nitrogen loss flux model. Multiple rounds of water level fluctuation data are input to simulate and generate dynamic change data of nitrogen loss ratio.

[0078] The fluctuation pattern of surface sediment re-release flux obtained in the previous step, while revealing the temporal dynamics of dissolved and particulate nitrogen release from sediments to upper water layers, does not definitively answer whether and how much nitrogen permanently leaves the wetland ecosystem as nitrogen gas. This is because some of the re-released nitrogen is re-oxidized due to the presence of the surface oxide layer, and some particulate nitrogen may not be effectively released due to the blocking effect. These processes are crucial factors that must be considered when assessing wetland self-purification capacity and formulating nitrogen management strategies, and re-release flux alone cannot solve this problem. To address this issue, this invention combines the fluctuation pattern of surface sediment re-release flux with two key environmental constraints: the thickness of the surface oxide layer and the blocking effect of sediment particles, to construct a nitrogen loss flux model that reflects real environmental limitations. Subsequently, the model simulates the dynamic changes in the proportion of nitrogen loss under multiple rounds of water level fluctuations and finally performs spatial mapping to determine the fine distribution pattern of the portion of surface nitrogen lost in gaseous form.

[0079] An oxygen microelectrode was vertically inserted into the sediment from the overlying water in micrometer-level steps to record the dissolved oxygen concentration (DO) profile with depth in real time. The depth at which the DO concentration decreased to the detection limit was defined as the thickness of the oxide layer. Surface sediment samples were collected and subjected to laser particle size analysis in the laboratory to obtain the composition ratio and average particle size of sand, silt, and clay. The permeability coefficient of the sediment was estimated based on the Kozeny-Kaman equation or other relevant empirical models, combined with particle size and porosity data. The reciprocal of the permeability coefficient was used as the sediment particle retention effect data.

[0080] Subsequently, these two environmental factors are integrated as adjustment coefficients into the nitrogen loss flux model. To accurately assess and optimize wetland nitrogen removal efficiency, it is necessary to quantify the regulatory role of key environmental factors on nitrogen loss. The nitrogen loss flux model constructed in this invention is a dynamic prediction model based on machine learning algorithms. This invention constructs a nitrogen loss flux model based on a gradient boosting decision tree, which can be implemented using mainstream industry frameworks. The model's input data includes water level fluctuation data, surface oxide layer thickness, sediment particle retardation effect number, water temperature, dissolved oxygen in overlying water, and sediment organic matter content. The output data is nitrogen loss flux. Before constructing the model, the training set is obtained through long-term in-situ monitoring in typical wetland areas. Techniques such as static chamber-gas chromatography or eddy covariance can be used to measure nitrogen loss flux under different hydrological and seasonal conditions, and the input variables required by the model are recorded simultaneously. Using a dataset containing input features and measured flux labels, the gradient boosting decision tree model is trained and validated through supervised learning. During the training phase, the measured nitrogen loss flux from the historical dataset is used as the output target for supervised learning, and key hyperparameters such as learning rate and maximum tree depth are optimized through cross-validation.

[0081] Using the parameter sensitivity analysis method, the surface oxide layer thickness data was set as a variable to calculate the nitrogen loss ratio corresponding to different surface oxide layer thickness data.

[0082] To quantify the regulatory effects of different environmental factors on nitrogen loss, such as Figure 3 As shown, this embodiment performs parameter sensitivity analysis on an established nitrogen loss flux model. Taking the analysis of the impact of surface oxide layer thickness as an example, it is set as a variable in the nitrogen loss flux model, while other parameters remain unchanged. In practice, incremental surface oxide layer thickness values ​​are set, and for each thickness value, the model is rerun under the same multiple rounds of water level fluctuations to simulate and generate the corresponding nitrogen loss ratio. By recording and comparing the nitrogen loss ratio curves at different thicknesses, the inhibitory effect of surface oxide layer thickness can be quantified.

[0083] Based on the data on the impact of sediment particles, the dynamic change data of nitrogen loss ratio is iteratively adjusted by a correction function to generate the adjusted dynamic change data of nitrogen loss ratio.

[0084] To capture explosive gas release events caused by particle trapping effects, this embodiment corrects the simulated dynamic change data of nitrogen loss ratio. Specifically, the time series of sediment particle trapping effect data is used as input, and the baseline nitrogen loss ratio curve is iteratively adjusted using a correction function. This correction function aims to simulate the physical process: when the sediment particle trapping effect continues to increase, the function suppresses the smooth growth of the nitrogen loss ratio, simulating the accumulation process of gas pressure; and once the trapping degree suddenly weakens due to factors such as water flow disturbance, the function simulates the explosive release of accumulated nitrogen. Through this step, the adjusted dynamic change characteristics of the nitrogen loss ratio can more realistically reflect the high-frequency fluctuation characteristics caused by particle trapping and sudden release.

[0085] This paper presents a method for implementing correction using a correction function. First, an internal state variable, called accumulated gas potential energy, is set to record the amount of nitrogen gas impeded. At each time step, based on the current particle impediment data, the baseline nitrogen loss ratio is dynamically reduced; the stronger the impediment, the greater the reduction, thus obtaining the suppressed loss ratio. The difference between the baseline ratio and this suppressed ratio is accumulated into the accumulated gas potential energy to simulate the accumulation process of gas pressure. Next, it is determined whether the reduction in the current particle impediment compared to the previous time step exceeds a preset burst threshold. If it exceeds this threshold, a burst release mechanism is triggered: the burst amount is calculated from the accumulated gas potential energy according to a preset burst release coefficient, and this burst amount is added to the current suppressed loss ratio to form the corrected loss ratio. Simultaneously, the released amount is subtracted from the accumulated gas potential energy. Conversely, if the reduction in the impediment does not reach the burst threshold, no additional release occurs, and the final loss ratio is the suppressed loss ratio. This complete iterative process enables the corrected data to accurately reflect the nonlinear physical phenomenon of gas bursting out instantaneously due to external disturbances after being blocked and accumulated for a long time.

[0086] The calculated total re-release flux is combined with the adjusted dynamic change data of nitrogen loss ratio to calculate the net nitrogen loss flux.

[0087] For each monitoring point, data on the thickness of the corresponding surface oxide layer, the impact of sediment particle blockage, and the dynamic changes in the nitrogen loss ratio adjusted in the previous steps are integrated. The calculated total re-release flux value is multiplied by the adjusted nitrogen loss ratio obtained in this step to calculate the net nitrogen loss flux at that monitoring point at a specific time step. Subsequently, the average flux values ​​over the complete fluctuation period of all monitoring points are spatially mapped to generate the net nitrogen loss flux distribution.

[0088] Determining the net nitrogen loss flux also includes the following steps: Based on the dynamic changes in nitrogen loss ratio, combined with data on surface oxide layer thickness and sediment particle blockage, a distribution model for nitrogen re-release from the wetland surface is constructed.

[0089] To achieve a refined characterization of nitrogen loss distribution patterns, this embodiment introduces an adaptive analysis strategy based on the aforementioned nitrogen loss flux model. While the previous steps determined the overall spatial distribution pattern, they may not accurately capture the instantaneous flux peaks caused by extreme spatiotemporal variability in local regions. To address this issue, this invention further identifies these key regions by analyzing fluctuation characteristics and dynamically improves their spatial resolution for recalculation, thereby achieving a refined characterization of nitrogen loss distribution. The model used in this process is the aforementioned nitrogen loss flux model.

[0090] To identify key areas of nitrogen loss, this embodiment applies a pre-constructed nitrogen loss flux model to time series analysis. Specifically, the process involves first selecting sequence data containing multiple rounds of water level fluctuations as model input, then using the model to calculate and output the final nitrogen loss flux time series for each monitoring point throughout the entire sequence. Subsequently, for the time series data of each grid cell, its fluctuation characteristics are extracted, primarily including fluctuation amplitude and frequency. Finally, the fluctuation characteristics of all monitoring points are spatially integrated to form a spatial distribution map describing the intensity of the final nitrogen loss fluctuations across the entire wetland. Based on a preset amplitude threshold, high-amplitude fluctuation areas with severe fluctuations are identified on this map.

[0091] Multiple rounds of water level fluctuation data were input into the distribution model of nitrogen re-release from the wetland surface to simulate and generate time series data of nitrogen re-release flux at monitoring points. The time series data of nitrogen re-release flux were analyzed to extract amplitude and frequency, and to obtain the fluctuation characteristics data of nitrogen re-release.

[0092] Within these identified local regions, the resolution of the spatial grid is dynamically increased, for example, from the original 1m × 1m to 0.5m × 0.5m. Subsequently, using the established nitrogen loss flux model, the final nitrogen loss flux is recalculated on these refined grid cells with higher spatial precision input data. Because a smaller spatial scale captures greater environmental heterogeneity, the recalculated results more accurately reflect the impact of local microenvironments on nitrogen loss, thus generating a refined characterization dataset with variable spatial resolution directly correlated with fluctuation characteristics.

[0093] like Figure 4 As shown, the present invention also provides a wetland nitrogen cycle optimization system based on machine learning, which is used to implement the above-mentioned method. The system includes: The data acquisition module collects multi-level environmental data of wetland soil profiles and generates a comprehensive environmental dataset. The first analysis module identifies lateral seepage paths and vertical priority flow paths based on a comprehensive environmental dataset through spatial overlay and temporal analysis, and uses a cellular automata model to simulate spatiotemporal evolution and construct nitrogen migration paths. The second analysis module analyzes the interaction between the nitrogen migration path analysis water flow-driven flux and the groundwater level on the path resistance to determine the degree of widening of the lateral seepage path. If the widening degree exceeds the preset threshold, it integrates water turbulence intensity data and flow velocity shear effect value to obtain the bubble escape rate and surface sediment disturbance frequency. The loss calculation module calculates the spatial distribution of nitrogen saturation and sediment particle suspension based on the bubble escape rate and surface sediment disturbance frequency. Through flux superposition analysis, the total re-release flux is obtained. Combined with the fluctuation law, a nitrogen loss flux model is constructed to simulate the change of nitrogen loss ratio under multiple rounds of water level fluctuations and perform spatial mapping to determine the net nitrogen loss flux.

[0094] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

[0097] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning-based method for optimizing wetland nitrogen cycle, characterized in that, include: Collect multi-level environmental data of wetland soil profiles to generate a comprehensive environmental dataset; Based on the comprehensive environmental dataset, we identified lateral seepage paths and vertical priority flow paths through spatial overlay and temporal analysis, and constructed nitrogen migration paths by using a cellular automata model to simulate spatiotemporal evolution. Based on the analysis of nitrogen migration pathways, the interaction between water flow-driven flux and groundwater level on path resistance is used to determine the degree of widening of lateral seepage pathways. If the widening exceeds the preset threshold, the water turbulence intensity data and flow velocity shear effect value are integrated to obtain the bubble escape rate and surface sediment disturbance frequency. Based on the bubble escape rate and the surface sediment disturbance frequency, the spatial distribution of nitrogen saturation and sediment particle suspension were calculated, and the total re-release flux was obtained through flux superposition analysis. A nitrogen loss flux model was constructed by combining the fluctuation pattern. The changes in the nitrogen loss ratio were simulated under multiple rounds of water level fluctuations, and spatial mapping was performed to determine the net nitrogen loss flux.

2. The method according to claim 1, characterized in that, Generate a comprehensive environment dataset, including the following steps: Monitoring points and sensors were set up in multiple soil profiles to form a sensor network. The redox potential sequence of each monitoring point in the target wetland soil profile was obtained through the sensor network, and the corresponding water level data and dissolved oxygen data were retrieved simultaneously. Time-aligned processing was performed on redox potential sequences, water level data, and dissolved oxygen data to generate multi-source time series data. Based on multi-source time series data, the potential gradient distribution data is obtained by calculating the ratio of the difference in redox potential values ​​between adjacent layers to the depth interval. Based on water level data and soil effective stress model, soil pore structure data is calculated. Based on potential gradient distribution data, the root penetration influence range data is extracted using a threshold segmentation algorithm. Potential gradient distribution data, soil pore structure data, and influence range data are used as a comprehensive environmental dataset.

3. The method according to claim 2, characterized in that, Constructing nitrogen migration pathways includes the following steps: Based on soil pore structure data and influence range data, lateral seepage paths are identified through spatial overlay. The comprehensive environmental dataset also includes real-time response sequences of seepage. By analyzing the time series, monitoring points with response time intervals less than a critical threshold are identified and connected to obtain the vertical priority flow path. The lateral seepage path is then fused with the vertical priority flow path to generate a preliminary migration path. Based on the initial migration path, the permeability adjustment coefficient is calculated and applied in conjunction with soil pore structure data to correct the initial migration path and obtain the adjusted migration path. Based on adjusting the migration path, a cellular automata model is used to simulate the spatiotemporal evolution and construct a dynamic nitrogen migration path.

4. The method according to claim 3, characterized in that, Determining the extent of widening of the lateral seepage path includes the following steps: The water level fluctuation sequence was obtained, and combined with the nitrogen migration path, an autoregressive moving average model including exogenous variables was used to simulate and generate the distribution data of water flow-driven flux. Analyze water level fluctuation sequences and soil pore structure data to generate data on the impact of path resistance; Based on distribution and impact data, the degree of widening of the lateral seepage path is determined by calculating the widening index and accumulating the widening increment.

5. The method according to claim 4, characterized in that, Obtaining the bubble escape rate and the surface sediment disturbance frequency includes the following steps: Three-dimensional velocity data of monitoring points are collected by acoustic Doppler current meter. The turbulence intensity data of water body is obtained by the velocity fluctuation component in the three-dimensional velocity data. The velocity shear effect value is obtained by analyzing the vertical velocity gradient in the three-dimensional velocity data. A weighted average method was used to integrate the water turbulence intensity data and velocity shear effect values ​​from the monitoring points to generate a comprehensive value. Based on the comprehensive values ​​from the monitoring points, the bubble escape rate and the surface sediment disturbance frequency are calculated.

6. The method according to claim 5, characterized in that, The total re-release flux is obtained through flux superposition analysis, including the following steps: Calculate the nitrogen partial pressure and saturation at each monitoring point to generate saturation distribution data; Based on surface sediment disturbance frequency data, the particulate nitrogen release model was applied to calculate the suspended nitrogen at each monitoring point and obtain suspended nitrogen distribution data. By integrating saturation distribution data and suspended matter distribution data, the total re-release flux at each monitoring point under multiple water level fluctuations is calculated through flux overlay analysis.

7. The method according to claim 6, characterized in that, Determining the net nitrogen loss flux also includes the following steps: The surface oxide layer thickness data and sediment particle blockage data are integrated into the nitrogen loss flux model. Multiple rounds of water level fluctuation data are input to simulate and generate dynamic change data of nitrogen loss ratio. Using the parameter sensitivity analysis method, the surface oxide layer thickness data was set as a variable to calculate the nitrogen loss ratio corresponding to different surface oxide layer thickness data; Based on the data on the impact of sediment particle blockage, the dynamic change data of nitrogen loss ratio is iteratively adjusted by a correction function to generate the adjusted dynamic change data of nitrogen loss ratio. The calculated total re-release flux is combined with the adjusted dynamic change data of nitrogen loss ratio to calculate the net nitrogen loss flux.

8. The method according to claim 7, characterized in that, Determining the net nitrogen loss flux also includes the following steps: Based on the dynamic change data of nitrogen loss ratio, combined with the surface oxide layer thickness data and sediment particle blockage data, a distribution model of wetland surface nitrogen re-release is constructed. Multiple rounds of water level fluctuation data were input into the distribution model of nitrogen re-release from the wetland surface to simulate and generate time series data of nitrogen re-release flux at monitoring points. The time series data of nitrogen re-release flux were analyzed to extract amplitude and frequency, and to obtain the fluctuation characteristics data of nitrogen re-release.

9. A machine learning-based wetland nitrogen cycle optimization system, used to implement the method as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition module collects multi-level environmental data of wetland soil profiles and generates a comprehensive environmental dataset. The first analysis module identifies lateral seepage paths and vertical priority flow paths based on a comprehensive environmental dataset through spatial overlay and temporal analysis, and uses a cellular automata model to simulate spatiotemporal evolution and construct nitrogen migration paths. The second analysis module analyzes the interaction between the nitrogen migration path analysis water flow-driven flux and the groundwater level on the path resistance to determine the degree of widening of the lateral seepage path. If the widening degree exceeds the preset threshold, it integrates water turbulence intensity data and flow velocity shear effect value to obtain the bubble escape rate and surface sediment disturbance frequency. The loss calculation module calculates the spatial distribution of nitrogen saturation and sediment particle suspension based on the bubble escape rate and surface sediment disturbance frequency. Through flux superposition analysis, the total re-release flux is obtained. Combined with the fluctuation law, a nitrogen loss flux model is constructed to simulate the change of nitrogen loss ratio under multiple rounds of water level fluctuations and perform spatial mapping to determine the net nitrogen loss flux.

Images