A method for estimating spatiotemporal dynamics of atmospheric nitrogen deposition in a gulf in land-sea synergy

Abstract

The present application belongs to the technical field of environmental data processing and pollution monitoring, and discloses a land-sea coordinated spatio-temporal dynamic estimation method for gulf atmospheric nitrogen deposition; the method comprises the following steps: S1, collecting and rasterizing social, economic and environmental data; S2, collecting atmospheric nitrogen data sets, integrating the three types of data and matching with the spatial position of atmospheric nitrogen deposition; S3, applying a CNN-LSTM model to estimate land atmospheric nitrogen deposition data; S4, applying a CNN-LSTM model to estimate marine anthropogenic source atmospheric nitrogen deposition data; S5, applying an LSTM model to estimate marine natural source atmospheric nitrogen deposition data; S6, integrating the results by means of inversion weighting method; S7, constructing a gulf atmospheric nitrogen deposition training data set; S8, carrying out three-dimensional verification based on the measured data of monitoring stations; S9, carrying out superimposition processing on the land and marine models, and outputting gulf atmospheric nitrogen deposition data after training the model; the present application improves the estimation accuracy of gulf atmospheric nitrogen deposition and provides a scientific basis for precise pollution control.

Description

Technical Field

[0001] This invention belongs to the field of environmental data processing and pollution monitoring technology, specifically relating to a land-sea coordinated method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in bays. Background Technology

[0002] Atmospheric nitrogen deposition research focuses on ecological issues such as eutrophication in bays. Currently, relying on ground stations, satellite remote sensing, and other numerical models, a technical foundation of "monitoring-simulation-source apportionment" has been established. However, it faces two major challenges: First, most studies are limited to small areas such as a single bay (e.g., Jiaozhou Bay), and large-scale cross-bay studies are difficult to advance due to inconsistent observation standards and insufficient model adaptability. Second, few researchers have quantitatively fitted socioeconomic indicators such as GDP and car ownership with nitrogen deposition flux, and the lack of multi-dimensional database support makes it difficult to quantify the driving mechanisms of human activities.

[0003] Therefore, how to effectively integrate socio-economic and meteorological data from multiple sources, thereby overcoming the bottlenecks of inconsistent observation standards and insufficient model adaptability in large-scale cross-bay studies, filling the gaps in multi-dimensional database support, achieving quantitative fitting between socio-economic indicators and nitrogen deposition flux, and ultimately accurately estimating atmospheric nitrogen deposition data to provide technical support for quantifying the driving mechanisms of human activities, has become an urgent problem to be solved in the field of environmental data processing. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a land-sea collaborative method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in bays. This method overcomes the bottlenecks of traditional atmospheric nitrogen deposition estimation methods (such as single-site interpolation and numerical models that do not integrate land-sea collaborative data), which suffer from low accuracy, insufficient spatiotemporal resolution, and low computational efficiency. While maintaining high accuracy and high spatiotemporal resolution, it significantly improves simulation speed and efficiently generates long-term, high-resolution hydrodynamic fields, providing reliable and efficient dynamic input for bay water quality models, ecological models, etc., and has promising prospects for widespread application.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A land-sea coordinated method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in bays includes the following steps:

[0007] S1. Collect social, economic, and environmental data from the Gulf region and perform rasterization processing;

[0008] S2. Collect atmospheric nitrogen datasets, integrate social, economic, and environmental data, and match them with the spatial location of atmospheric nitrogen deposition.

[0009] S3. Input 15-dimensional features, apply CNN-LSTM neural network model to estimate land atmospheric nitrogen deposition data, and output land nitrogen deposition flux;

[0010] S4. Input 8-dimensional features, apply CNN-LSTM neural network model to estimate marine anthropogenic atmospheric nitrogen deposition data, and output marine anthropogenic nitrogen deposition flux;

[0011] S5. Input 6-dimensional features, apply a long short-term memory network to estimate atmospheric nitrogen deposition data from natural marine sources, and output the nitrogen deposition flux from natural marine sources.

[0012] S6. Determine the weights of marine anthropogenic nitrogen deposition fluxes using the inversion weighting method, and calculate the total marine nitrogen deposition fluxes by weighted summation.

[0013] S7. Using the estimated terrestrial atmospheric nitrogen deposition data, marine anthropogenic atmospheric nitrogen deposition data, and marine natural atmospheric nitrogen deposition data, construct a training dataset for atmospheric nitrogen deposition in the bay based on the land-sea ratio in each grid.

[0014] S8. Using measured data from 10-20 monitoring stations along the coast and in the sea area of ​​the bay as the true values, the reliability of the training dataset for atmospheric nitrogen deposition in the bay is verified by accuracy, spatiality and stability.

[0015] S9. Construct a fusion estimation model for nitrogen deposition in the bay and train it using the atmospheric nitrogen deposition training dataset of the bay. Then, use the trained fusion estimation model for nitrogen deposition in the bay to estimate the spatiotemporal dynamics of atmospheric nitrogen deposition in the bay.

[0016] Preferably, in step S1, the social data includes population, GDP, and car ownership; the economic data includes nitrogen fertilizer application, aquaculture scale, industrial wastewater discharge related to air pollution, industrial waste gas discharge related to air pollution, number of industries related to air pollution, ship voyage mileage, tourist arrivals, and oil and gas production; the environmental data includes wind speed, wind direction, precipitation, temperature, sea and land breeze frequency, altitude, slope, sea temperature, salinity, chlorophyll a concentration, solar radiation, sea-temperature difference, ocean current intensity, and distance from the coastline.

[0017] Preferably, in step S1, the specific process of rasterization includes:

[0018] S11. Group the data by indicator type-year-month, calculate the mean and standard deviation of each group, and then remove outlier data.

[0019] S12. Use ArcGIS software to convert meteorological station, land use and administrative region data into the WGS84 coordinate system.

[0020] S13. Use ArcGIS software's fishing net creation tool to generate a 1 km × 1 km regular grid, with the output format being GeoTIFF;

[0021] S14. Set the basic weights for land use types and store them in the land use vector data attribute table; use the clipping tool in ArcGIS software to clip the land use vector data according to the boundary of the study area, retaining the land use type patches within the target area; the basic weight setting rules for land use types are as follows: commercial land = 1.0, industrial land = 0.8, residential land = 0.6, agricultural land = 0.2 and forest land = 0.05;

[0022] S15. Using ArcGIS software's intersection tool, spatially overlay the clipped land use vector data with a 1 km × 1 km regular grid to obtain grid-land use type associated patches; use ArcGIS software's computational geometry tools to calculate the area of ​​each grid-land use type associated patch; for each grid, calculate the proportion of each type of land use area to the total area of ​​the selected grid, using the following formula: ,in, This represents the proportion of each type of land use area to the total area of ​​the selected grid. Associate the area of ​​patches for each grid-land use type; The total area of ​​each grid, and ;

[0023] S16. Initial Grid Weights: Calculate the overall weight of each grid, which represents the weighted sum of the proportion of various land uses and their corresponding weights. The calculation formula is as follows: ,in, The overall weight for each grid; The basic weight for land use type;

[0024] Total allocation by district / county: The sum of the comprehensive weights of all grids within the selected district / county is calculated according to the district / county boundaries, and the initial index value for each grid is calculated using the following formula: , ,in, This is the sum of the comprehensive weights of all grids within the selected district / county; Initial index values ​​for each grid; This represents the total statistical indicators for the selected districts and counties;

[0025] S17. Calculate the total deviation of the initial allocation values ​​for all grids within the selected district / county. The calculation formula is as follows: ,in, The total deviation of the initial allocation values ​​for all grids within the selected district / county;

[0026] Iterative correction: If Dev > 3%, repeat the iterative adjustment of the mesh values ​​until Dev ≤ 3%, to obtain the final mesh assignment value. The formula for iterative calculation is: ,in, Assign values ​​to the grid for the (t+1)th iteration; λ is the grid assignment value for the t-th iteration; λ is the adjustment coefficient, λ=0.08; The average value assigned to the grid within the selected district / county; To find the function with the maximum value; To find the minimum value function;

[0027] S18. Using the Kriging interpolation tool in ArcGIS software, interpolate the environmental data into a 1 km × 1 km grid; then convert the interpolation results into GeoTIFF format and associate them with the grid IDs of the social data grid and the economic data grid.

[0028] S19. Data Integration: Using Python Pandas software, merge the gridded social, economic, and environmental data according to the grid ID-year-month index to form socio-economic and environmental grid data with grid ID-indicator name-indicator value-spatial information.

[0029] Preferably, the specific process of step S2 includes:

[0030] S21. Unified grid benchmark and indexing rules: All data are constructed based on a 1 km × 1 km grid, and a grid ID is used to uniquely identify the spatial location; the time dimension is uniformly adopted using year-month timestamps, which ultimately form a three-dimensional index of year-month-grid ID, so that the spatiotemporal benchmark is consistent.

[0031] S22. Using the Add XY Data tool in ArcGIS software, load the socio-economic environment grid data and atmospheric nitrogen deposition grid data, both of which contain grid IDs and corresponding latitude and longitude coordinates, into the same spatial reference system. The spatial reference system adopts the WGS84 coordinate system.

[0032] S23. Using the spatial connection tool of ArcGIS software, with the atmospheric nitrogen deposition grid as the target element and the socio-economic environment grid as the connection element, select the complete containment matching rule to automatically associate the environmental, social and economic fields with the nitrogen deposition grid attribute table.

[0033] S24. For monthly time series data, use the Python geopandas library to write a batch processing script, and associate the data one-to-one through the grid ID field. The specific process is as follows: read the atmospheric nitrogen deposition grid and the socio-economic environment grid, execute the merge function based on the grid ID field, and output the associated data table.

[0034] S25. Filter related data based on year-month timestamps and remove invalid matches across timestamps; use the drop_duplicates function of the Python pandas library to remove duplicate matches and retain unique records with year-month-grid ID;

[0035] S26. Matching quality verification: Randomly select 10% of the atmospheric nitrogen deposition grid and the socio-economic environment grid, and manually check the consistency of latitude and longitude coordinates with grid ID; calculate the matching success rate, and use nearest neighbor interpolation to supplement the association for unmatched grids to ensure that no data breaks occur.

[0036] Preferably, the specific process of step S3 includes:

[0037] S31. Extract the terrestrial study area data from the data processed in step S2, and divide the training set and validation set into a 7:3 time series. Use a 12-month time window to capture seasonal features, and construct sample pairs of continuous 12-month features + nitrogen deposition flux in the 12th month to form the input feature tensor and label tensor.

[0038] S32. A three-level architecture of spatial feature extraction, temporal feature capture, and output mapping is adopted. In the input layer, a tensor of dimension (batch size × time window × number of features) is received. In the CNN spatial extraction layer, two convolutional layers are used to extract spatial correlation features, and ReLU is used as the activation function. In the LSTM temporal capture layer, a bidirectional LSTM captures the time series trend and outputs the bidirectional feature concatenation result. In the fully connected output layer, after suppressing overfitting, linear activation is used to output the nitrogen deposition flux.

[0039] S33. The optimizer is Adam, the initial learning rate is 0.001, and the batch size is set to 32; the loss function is the mean squared error adaptive regression task; an early stopping strategy is set, and L2 regularization is enabled.

[0040] S34: Build a data loader and iterate through 100 rounds of training; calculate the loss using the validation set after each training round and monitor the training curve in real time; save the current optimal model weights when early stopping is triggered.

[0041] S35. Load the optimal model to predict the validation set, and calculate the mean absolute error, root mean square error, and coefficient of determination. When the mean absolute error is ≤5, the root mean square error is ≤8, and the coefficient of determination is ≥0.8, the model is deemed qualified, and the monthly land nitrogen deposition flux at the 1 km×1 km grid scale is output. If the model does not meet the requirements, return to adjust the model architecture or parameters and retrain.

[0042] Preferably, in step S3, the 15-dimensional input features include population, GDP, car ownership, nitrogen fertilizer application, aquaculture scale, industrial wastewater discharge related to air pollution, industrial waste gas discharge related to air pollution, number of industries related to air pollution, wind speed, wind direction, precipitation, temperature, sea and land breeze frequency, altitude, and slope; in step S4, the 8-dimensional input features include ship voyage distance, tourist arrivals, oil and gas production, wind speed, wind direction, ocean current intensity, and distance from the coastline; in step S5, the 6-dimensional input features include seawater temperature, salinity, chlorophyll a concentration, solar radiation, sea-temperature difference, and wind speed.

[0043] Preferably, in step S6, the formula for calculating the total marine nitrogen deposition flux is: ,in, Total nitrogen deposition flux in the ocean; For anthropogenic nitrogen deposition flux in the ocean; For natural marine nitrogen deposition flux; Weighting of anthropogenic nitrogen deposition flux in the ocean.

[0044] Preferably, the specific process of step S7 includes:

[0045] S71. Based on the land-sea ratio, the weighted average atmospheric nitrogen deposition value for the Gulf grid is obtained using the following formula: ,in, This serves as the baseline value for atmospheric nitrogen deposition in the Gulf grid. For terrestrial nitrogen deposition flux; Total nitrogen deposition flux in the ocean; This is the ratio of the land area within the grid to the total area of ​​the grid.

[0046] S72. Multiply the baseline atmospheric nitrogen deposition value of the Gulf grid by the interaction adjustment factor to obtain the final atmospheric nitrogen deposition value of the Gulf grid. The calculation formula is as follows: ,in, This represents the final value of atmospheric nitrogen deposition in the Gulf grid. As an interactive modulatory factor, ∈[0.8,1.2];

[0047] S73. Organize the data in the format of year-month-grid ID-input feature set-training label, and output a land use vector data attribute table in CSV format and a raster in GeoTIFF format for training the stacked model.

[0048] Preferably, in step S9, the construction process of the Gulf nitrogen deposition fusion estimation model is as follows: a three-level structure of bottom-level embedding, middle-level fusion, and top-level output is adopted; in the bottom layer, the CNN-LSTM neural network model pre-trained in steps S3 and S4 is embedded, the core parameters of the CNN convolutional layer and LSTM layer are frozen, and only the 64-dimensional feature layer before the fully connected output is unfrozen; in the middle layer, the Transformer interaction layer has 8-head attention, a hidden layer dimension of 256, and the activation function is the GELU function, inputting multi-dimensional integrated features and capturing the land-sea interaction effect; in the top layer, the fully connected layer outputs the atmospheric nitrogen deposition flux in the Gulf.

[0049] By adopting the above technical solution, the present invention has the following beneficial effects:

[0050] This invention constructs a land-sea collaborative estimation system, breaking through the limitations of traditional "land-sea" separate research. By integrating land, marine anthropogenic and natural nitrogen deposition data, and combining the land-sea ratio and interaction adjustment factors of the bay grid, it accurately captures the spatiotemporal dynamic characteristics of nitrogen deposition in the bay area, significantly improving the accuracy and applicability of large-scale estimation across bays.

[0051] This invention integrates heterogeneous data from multiple dimensions, including social, economic, and environmental data. Through standardized rasterization processing, outlier removal, iterative correction, and spatial matching, it eliminates data format barriers and quality risks, providing comprehensive and reliable data support for nitrogen deposition estimation and filling the gap in multi-dimensional database support.

[0052] This invention specifically employs CNN-LSTM models, LSTM models, and layered fusion models to adapt to the needs of land spatiotemporal feature capture, marine natural process time series simulation, and integration of land-sea interaction effects in the Gulf. While ensuring high-resolution output at a monthly scale of 1 km×1 km grid, it reduces the model's dependence on data and improves estimation efficiency and generalization ability.

[0053] This invention establishes a three-dimensional verification system encompassing accuracy, spatiality, and stability. Through multiple rounds of model training and optimization, along with calibration using measured data, the reliability of the estimation results is ensured. The output includes information such as flux values ​​and the proportion of land-based and marine source contributions, enabling intuitive location of high deposition areas and providing a scientific basis for eutrophication control and precise pollution management decisions in bay areas. Furthermore, the entire methodology is standardized, highly replicable, and compatible with existing common tools and data formats such as ArcGIS and Python. It does not rely on complex and expensive observation equipment and can be rapidly applied to atmospheric nitrogen deposition monitoring and assessment in different bay areas, demonstrating broad prospects for widespread application. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] like Figure 1 As shown, a land-sea coordinated method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in bays includes the following steps:

[0057] S1. Collect social, economic, and environmental data from the Gulf region and perform rasterization processing;

[0058] In step S1, the social data includes population, GDP, and vehicle ownership; the economic data includes nitrogen fertilizer application, aquaculture scale, industrial wastewater discharge related to air pollution, industrial waste gas discharge related to air pollution, number of industries related to air pollution, ship mileage, tourist arrivals, and oil and gas production; the environmental data includes wind speed, wind direction, precipitation, temperature, sea and land breeze frequency, altitude, slope, sea temperature, salinity, chlorophyll a concentration, solar radiation, sea-temperature difference, ocean current intensity, and distance from the coastline.

[0059] In step S1, the specific process of rasterization includes:

[0060] S11. Group the data by indicator type-year-month, calculate the mean and standard deviation of each group, and then remove outlier data.

[0061] S12. Use ArcGIS software to convert meteorological station, land use and administrative region data into the WGS84 coordinate system.

[0062] S13. Use ArcGIS software's fishing net creation tool to generate a 1 km × 1 km regular grid, with the output format being GeoTIFF;

[0063] S14. Set the basic weights for land use types and store them in the land use vector data attribute table; use the clipping tool in ArcGIS software to clip the land use vector data according to the boundary of the study area, retaining the land use type patches within the target area; the basic weight setting rules for land use types are as follows: commercial land = 1.0, industrial land = 0.8, residential land = 0.6, agricultural land = 0.2 and forest land = 0.05;

[0064] S15. Using ArcGIS software's intersection tool, spatially overlay the clipped land use vector data with a 1 km × 1 km regular grid to obtain grid-land use type associated patches; use ArcGIS software's computational geometry tools to calculate the area of ​​each grid-land use type associated patch; for each grid, calculate the proportion of each type of land use area to the total area of ​​the selected grid, using the following formula: ,in, This represents the proportion of each type of land use area to the total area of ​​the selected grid. Associate the area of ​​patches for each grid-land use type; The total area of ​​each grid, and ;

[0065] S16. Initial Grid Weights: Calculate the overall weight of each grid, which represents the weighted sum of the proportion of various land uses and their corresponding weights. The calculation formula is as follows: ,in, The overall weight for each grid; The basic weight for land use type;

[0066] Total allocation by district / county: The sum of the comprehensive weights of all grids within the selected district / county is calculated according to the district / county boundaries, and the initial index value for each grid is calculated using the following formula: , ,in, This is the sum of the comprehensive weights of all grids within the selected district / county; Initial index values ​​for each grid; This represents the total statistical indicators for the selected districts and counties;

[0067] S17. Calculate the total deviation of the initial allocation values ​​for all grids within the selected district / county. The calculation formula is as follows: ,in, The total deviation of the initial allocation values ​​for all grids within the selected district / county;

[0068] Iterative correction: If Dev > 3%, repeat the iterative adjustment of the mesh values ​​until Dev ≤ 3%, to obtain the final mesh assignment value. The formula for iterative calculation is: ,in, Assign values ​​to the grid for the (t+1)th iteration; λ is the grid assignment value for the t-th iteration; λ is the adjustment coefficient, λ=0.08; The average value assigned to the grid within the selected district / county; To find the function with the maximum value; To find the minimum value function;

[0069] S18. Using the Kriging interpolation tool in ArcGIS software, interpolate the environmental data into a 1 km × 1 km grid; then convert the interpolation results into GeoTIFF format and associate them with the grid IDs of the social data grid and the economic data grid.

[0070] S19. Data Integration: Using Python Pandas software, merge the gridded social, economic, and environmental data according to the grid ID-year-month index to form social, economic, and environmental grid data with grid ID-indicator name-indicator value-spatial information.

[0071] S2. Collect atmospheric nitrogen datasets, integrate social, economic, and environmental data, and match them with the spatial location of atmospheric nitrogen deposition.

[0072] The specific process of step S2 includes:

[0073] S21. Unified grid benchmark and indexing rules: All data are constructed based on a 1 km × 1 km grid, and a grid ID is used to uniquely identify the spatial location; the time dimension is uniformly adopted using year-month timestamps, which ultimately form a three-dimensional index of year-month-grid ID, so that the spatiotemporal benchmark is consistent.

[0074] S22. Using the Add XY Data tool in ArcGIS software, load the socio-economic environment grid data and atmospheric nitrogen deposition grid data, both of which contain grid IDs and corresponding latitude and longitude coordinates, into the same spatial reference system. The spatial reference system adopts the WGS84 coordinate system.

[0075] S23. Using the spatial connection tool of ArcGIS software, with the atmospheric nitrogen deposition grid as the target element and the socio-economic environment grid as the connection element, select the complete containment matching rule to automatically associate the environmental, social and economic fields with the nitrogen deposition grid attribute table.

[0076] S24. For monthly time series data, use the Python geopandas library to write a batch processing script, and associate the data one-to-one through the grid ID field. The specific process is as follows: read the atmospheric nitrogen deposition grid and the socio-economic environment grid, execute the merge function based on the grid ID field, and output the associated data table.

[0077] S25. Filter related data based on year-month timestamps and remove invalid matches across timestamps; use the drop_duplicates function of the Python pandas library to remove duplicate matches and retain unique records with year-month-grid ID;

[0078] S26. Matching quality verification: Randomly select 10% of the atmospheric nitrogen deposition grid and the socio-economic environment grid, and manually check the consistency of latitude and longitude coordinates with grid ID; calculate the matching success rate, and use nearest neighbor interpolation to supplement the association for unmatched grids to ensure that no data breaks occur.

[0079] S3. Input 15-dimensional features, apply CNN-LSTM neural network model to estimate land atmospheric nitrogen deposition data, and output land nitrogen deposition flux;

[0080] The specific process of step S3 includes:

[0081] S31. Extract the terrestrial study area data from the data processed in step S2, and divide the training set and validation set into a 7:3 time series. Use a 12-month time window to capture seasonal features, and construct sample pairs of continuous 12-month features + nitrogen deposition flux in the 12th month to form the input feature tensor and label tensor.

[0082] S32. A three-level architecture of spatial feature extraction, temporal feature capture, and output mapping is adopted. In the input layer, a tensor of dimension (batch size × time window × number of features) is received. In the CNN spatial extraction layer, two convolutional layers are used to extract spatial correlation features, and ReLU is used as the activation function. In the LSTM temporal capture layer, a bidirectional LSTM captures the time series trend and outputs the bidirectional feature concatenation result. In the fully connected output layer, after suppressing overfitting, linear activation is used to output the nitrogen deposition flux.

[0083] S33. The optimizer is Adam, the initial learning rate is 0.001, and the batch size is set to 32; the loss function is the mean squared error adaptive regression task; an early stopping strategy is set, and L2 regularization is enabled.

[0084] S34: Build a data loader and iterate through 100 rounds of training; calculate the loss using the validation set after each training round and monitor the training curve in real time; save the current optimal model weights when early stopping is triggered.

[0085] S35. Load the optimal model to predict the validation set, and calculate the mean absolute error, root mean square error, and coefficient of determination. When the mean absolute error is ≤5, the root mean square error is ≤8, and the coefficient of determination is ≥0.8, the model is deemed qualified, and the monthly land nitrogen deposition flux at the 1 km×1 km grid scale is output. If the model does not meet the requirements, return to adjust the model architecture or parameters and retrain.

[0086] In step S3, the 15-dimensional features input include population, GDP, car ownership, nitrogen fertilizer application, aquaculture scale, industrial wastewater discharge related to air pollution, industrial waste gas discharge related to air pollution, number of industries related to air pollution, wind speed, wind direction, precipitation, temperature, frequency of sea and land breezes, altitude, and slope.

[0087] S4. Input 8-dimensional features, apply CNN-LSTM neural network model to estimate marine anthropogenic atmospheric nitrogen deposition data, and output marine anthropogenic nitrogen deposition flux;

[0088] In step S4, the input 8-dimensional features include ship voyage distance, number of tourists, oil and gas production, wind speed, wind direction, ocean current intensity, and distance from the coastline.

[0089] S5. Input 6-dimensional features, apply a long short-term memory network to estimate atmospheric nitrogen deposition data from natural marine sources, and output the nitrogen deposition flux from natural marine sources.

[0090] In step S5, the input 6-dimensional features include seawater temperature, salinity, chlorophyll a concentration, solar radiation, sea-air temperature difference, and wind speed.

[0091] S6. Determine the weights of marine anthropogenic nitrogen deposition fluxes using the inversion weighting method, and calculate the total marine nitrogen deposition fluxes by weighted summation.

[0092] In step S6, the formula for calculating the total marine nitrogen deposition flux is: ,in, Total nitrogen deposition flux in the ocean; For anthropogenic nitrogen deposition flux in the ocean; For natural marine nitrogen deposition flux; Weighting of marine anthropogenic nitrogen deposition flux;

[0093] S7. Using the estimated terrestrial atmospheric nitrogen deposition data, marine anthropogenic atmospheric nitrogen deposition data, and marine natural atmospheric nitrogen deposition data, construct a training dataset for atmospheric nitrogen deposition in the bay based on the land-sea ratio in each grid.

[0094] The specific process of step S7 includes:

[0095] S71. Based on the land-sea ratio, the weighted average atmospheric nitrogen deposition value for the Gulf grid is obtained using the following formula: ,in, This serves as the baseline value for atmospheric nitrogen deposition in the Gulf grid. For terrestrial nitrogen deposition flux; Total nitrogen deposition flux in the ocean; This is the ratio of the land area within the grid to the total area of ​​the grid.

[0096] S72. Multiply the baseline atmospheric nitrogen deposition value of the Gulf grid by the interaction adjustment factor to obtain the final atmospheric nitrogen deposition value of the Gulf grid. The calculation formula is as follows: ,in, This represents the final value of atmospheric nitrogen deposition in the Gulf grid. As an interactive modulatory factor, ∈[0.8,1.2];

[0097] S73. Organize the data in the format of year-month-grid ID-input feature set-training label, and output a land use vector data attribute table in CSV format and a raster in GeoTIFF format for training the stacked model.

[0098] S8. Using measured data from 10-20 monitoring stations along the coast and in the sea area of ​​the bay as the true values, the reliability of the training dataset for atmospheric nitrogen deposition in the bay is verified by accuracy, spatiality and stability.

[0099] S9. Construct a fusion estimation model for nitrogen deposition in the Gulf and train it using the Gulf atmospheric nitrogen deposition training dataset. Then, use the trained Gulf nitrogen deposition fusion estimation model to estimate the spatiotemporal dynamics of atmospheric nitrogen deposition in the Gulf.

[0100] In step S9, the construction process of the Gulf nitrogen deposition fusion estimation model is as follows: a three-level structure of bottom-level embedding, middle-level fusion, and top-level output is adopted; in the bottom layer, the CNN-LSTM neural network model pre-trained in steps S3 and S4 is embedded, the core parameters of the CNN convolutional layer and LSTM layer are frozen, and only the 64-dimensional feature layer before the fully connected output is unfrozen; in the middle layer, the Transformer interaction layer has 8-head attention, a hidden layer dimension of 256, and the activation function is the GELU function, inputting multi-dimensional integrated features and capturing the land-sea interaction effect; in the top layer, the fully connected layer outputs the atmospheric nitrogen deposition flux in the Gulf.

[0101] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A land-sea coordinated method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in bays, characterized in that, Includes the following steps: S1. Collect social, economic, and environmental data from the Gulf region and perform rasterization processing; S2. Collect atmospheric nitrogen datasets, integrate social, economic, and environmental data, and match them with the spatial location of atmospheric nitrogen deposition. S3. Input 15-dimensional features, apply CNN-LSTM neural network model to estimate land atmospheric nitrogen deposition data, and output land nitrogen deposition flux; S4. Input 8-dimensional features, apply CNN-LSTM neural network model to estimate marine anthropogenic atmospheric nitrogen deposition data, and output marine anthropogenic nitrogen deposition flux; S5. Input 6-dimensional features, apply a long short-term memory network to estimate atmospheric nitrogen deposition data from natural marine sources, and output the nitrogen deposition flux from natural marine sources. S6. Determine the weights of marine anthropogenic nitrogen deposition fluxes using the inversion weighting method, and calculate the total marine nitrogen deposition fluxes by weighted summation. In step S6, the formula for calculating the total marine nitrogen deposition flux is: ,in, Total nitrogen deposition flux in the ocean; For anthropogenic nitrogen deposition flux in the ocean; For natural nitrogen deposition flux in the ocean; Weighting of marine anthropogenic nitrogen deposition flux; S7. Using the estimated terrestrial atmospheric nitrogen deposition data, marine anthropogenic atmospheric nitrogen deposition data, and marine natural atmospheric nitrogen deposition data, construct a training dataset for atmospheric nitrogen deposition in the bay based on the land-sea ratio in each grid. The specific process of step S7 includes: S71. Based on the land-sea ratio, the weighted average atmospheric nitrogen deposition value for the Gulf grid is obtained using the following formula: ,in, This serves as the baseline value for atmospheric nitrogen deposition in the Gulf grid. For terrestrial nitrogen deposition flux; Total nitrogen deposition flux in the ocean; This represents the ratio of the land area within the grid to the total area of ​​the grid. S72. Multiply the baseline atmospheric nitrogen deposition value of the Gulf grid by the interaction adjustment factor to obtain the final atmospheric nitrogen deposition value of the Gulf grid. The calculation formula is as follows: ,in, This represents the final value of atmospheric nitrogen deposition in the Gulf grid. As an interactive modulatory factor, ∈[0.8,1.2]; S73. Organize the data in the format of year-month-grid ID-input feature set-training label, and output the land use vector data attribute table in CSV format and the raster in GeoTIFF format for training the stacked model; S8. Using measured data from 10-20 monitoring stations along the coast and in the sea area of ​​the bay as the true values, the reliability of the training dataset for atmospheric nitrogen deposition in the bay is verified by accuracy, spatiality and stability. S9. Construct a fusion estimation model for nitrogen deposition in the bay and train it using the atmospheric nitrogen deposition training dataset of the bay. Then, use the trained fusion estimation model for nitrogen deposition in the bay to estimate the spatiotemporal dynamics of atmospheric nitrogen deposition in the bay.

2. The method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in a bay using a land-sea coordinated approach, as described in claim 1, is characterized in that... In step S1, the social data includes population, GDP, and car ownership; the economic data includes nitrogen fertilizer application, aquaculture scale, industrial wastewater discharge related to air pollution, industrial waste gas discharge related to air pollution, number of industries related to air pollution, ship voyage mileage, tourist arrivals, and oil and gas production; the environmental data includes wind speed, wind direction, precipitation, temperature, sea and land breeze frequency, altitude, slope, sea temperature, salinity, chlorophyll a concentration, solar radiation, sea-temperature difference, ocean current intensity, and distance from the coastline.

3. The method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in bays through land-sea collaboration as described in claim 1, characterized in that, In step S1, the specific process of rasterization includes: S11. Group the data by indicator type-year-month, calculate the mean and standard deviation of each group, and then remove outlier data. S12. Use ArcGIS software to convert meteorological station, land use and administrative region data into the WGS84 coordinate system. S13. Use ArcGIS software's fishing net creation tool to generate a 1 km × 1 km regular grid, with the output format being GeoTIFF; S14. Set the basic weights for land use types and store them in the land use vector data attribute table; use the clipping tool in ArcGIS software to clip the land use vector data according to the boundary of the study area, retaining the land use type patches within the target area; the basic weight setting rules for land use types are as follows: commercial land = 1.0, industrial land = 0.8, residential land = 0.6, agricultural land = 0.2 and forest land = 0.05; S15. Using ArcGIS software's intersection tool, spatially overlay the clipped land use vector data with a 1 km × 1 km regular grid to obtain grid-land use type associated patches; use ArcGIS software's computational geometry tools to calculate the area of ​​each grid-land use type associated patch; for each grid, calculate the proportion of each type of land use area to the total area of ​​the selected grid, using the following formula: ,in, This represents the proportion of each type of land use area to the total area of ​​the selected grid. Associate the area of ​​patches for each grid-land use type; The total area of ​​each grid, and ; S16. Initial Grid Weights: Calculate the overall weight of each grid, which represents the weighted sum of the proportion of various land uses and their corresponding weights. The calculation formula is as follows: ,in, The overall weight for each grid; The basic weight for land use type; Total allocation by district / county: The sum of the comprehensive weights of all grids within the selected district / county is calculated according to the district / county boundaries, and the initial index value for each grid is calculated using the following formula: , ,in, This is the sum of the comprehensive weights of all grids within the selected district / county; Initial index values ​​for each grid; This represents the total statistical indicators for the selected districts and counties; S17. Calculate the total deviation of the initial allocation values ​​for all grids within the selected district / county. The calculation formula is as follows: ,in, The total deviation of the initial allocation values ​​for all grids within the selected district / county; Iterative correction: If Dev > 3%, repeat the iterative adjustment of the mesh values ​​until Dev ≤ 3%, to obtain the final mesh assignment value. The iterative calculation formula is: ,in, Assign values ​​to the grid for the (t+1)th iteration; λ is the grid assignment value for the t-th iteration; λ is the adjustment coefficient, λ=0.08; The average value assigned to the grid within the selected district / county; To find the function with the maximum value; To find the minimum value function; S18. Using the Kriging interpolation tool in ArcGIS software, interpolate the environmental data into a 1 km × 1 km grid; then convert the interpolation results into GeoTIFF format and associate them with the grid IDs of the social data grid and the economic data grid. S19. Data Integration: Using Python Pandas software, merge the gridded social, economic, and environmental data according to the grid ID-year-month index to form socio-economic and environmental grid data with grid ID-indicator name-indicator value-spatial information.

4. The method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in a bay using a land-sea coordinated approach, as described in claim 1, is characterized in that... The specific process of step S2 includes: S21. Unified grid benchmark and indexing rules: All data are constructed based on a 1 km × 1 km grid, and a grid ID is used to uniquely identify the spatial location; the time dimension is uniformly adopted using year-month timestamps, which ultimately form a three-dimensional index of year-month-grid ID, so that the spatiotemporal benchmark is consistent. S22. Using the Add XY Data tool in ArcGIS software, load the socio-economic environment grid data and atmospheric nitrogen deposition grid data, both of which contain grid IDs and corresponding latitude and longitude coordinates, into the same spatial reference system. The spatial reference system adopts the WGS84 coordinate system. S23. Using the spatial connection tool of ArcGIS software, with the atmospheric nitrogen deposition grid as the target element and the socio-economic environment grid as the connection element, select the complete containment matching rule to automatically associate the environmental, social and economic fields with the nitrogen deposition grid attribute table. S24. For monthly time series data, use the Python geopandas library to write a batch processing script, and associate the data one-to-one through the grid ID field. The specific process is as follows: read the atmospheric nitrogen deposition grid and the socio-economic environment grid, execute the merge function based on the grid ID field, and output the associated data table. S25. Filter related data based on year-month timestamps and remove invalid matches across timestamps; use the drop_duplicates function of the Python pandas library to remove duplicate matches and retain unique records with year-month-grid ID; S26. Matching quality verification: Randomly select 10% of the atmospheric nitrogen deposition grid and the socio-economic environment grid, and manually check the consistency of latitude and longitude coordinates with grid ID; calculate the matching success rate, and use nearest neighbor interpolation to supplement the association for unmatched grids to ensure that no data breaks occur.

5. The method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in a bay using a land-sea coordinated approach, as described in claim 1, is characterized in that... The specific process of step S3 includes: S31. Extract the terrestrial study area data from the data processed in step S2, and divide the training set and validation set into a 7:3 time series. Use a 12-month time window to capture seasonal features, and construct sample pairs of continuous 12-month features + nitrogen deposition flux in the 12th month to form the input feature tensor and label tensor. S32. A three-level architecture of spatial feature extraction, temporal feature capture, and output mapping is adopted. In the input layer, a tensor of dimension (batch size × time window × number of features) is received. In the CNN spatial extraction layer, two convolutional layers are used to extract spatial correlation features, and ReLU is used as the activation function. In the LSTM temporal capture layer, a bidirectional LSTM captures the time series trend and outputs the bidirectional feature concatenation result. In the fully connected output layer, after suppressing overfitting, linear activation is used to output the nitrogen deposition flux. S33. The optimizer is Adam, the initial learning rate is 0.001, and the batch size is set to 32; the loss function is the mean squared error adaptive regression task; an early stopping strategy is set, and L2 regularization is enabled. S34: Build a data loader and iterate through 100 rounds of training; calculate the loss using the validation set after each training round and monitor the training curve in real time; save the current optimal model weights when early stopping is triggered. S35. Load the optimal model to predict the validation set, and calculate the mean absolute error, root mean square error, and coefficient of determination. When the mean absolute error is ≤5, the root mean square error is ≤8, and the coefficient of determination is ≥0.8, the model is deemed qualified, and the monthly land nitrogen deposition flux at the 1 km×1 km grid scale is output. If the model does not meet the requirements, return to adjust the model architecture or parameters and retrain.

6. The method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in a bay using a land-sea coordinated approach, as described in claim 1, is characterized in that: In step S3, the 15-dimensional input features include population, GDP, car ownership, nitrogen fertilizer application, aquaculture scale, industrial wastewater discharge related to air pollution, industrial waste gas discharge related to air pollution, number of industries related to air pollution, wind speed, wind direction, precipitation, temperature, sea and land breeze frequency, altitude, and slope. In step S4, the 8-dimensional input features include ship voyage distance, tourist arrivals, oil and gas production, wind speed, wind direction, ocean current intensity, and distance from the coastline. In step S5, the 6-dimensional input features include seawater temperature, salinity, chlorophyll a concentration, solar radiation, sea-temperature difference, and wind speed.

7. The method for estimating the spatiotemporal dynamics of atmospheric nitrogen deposition in a bay using a land-sea coordinated approach, as described in claim 1, is characterized in that... In step S9, the construction process of the Gulf nitrogen deposition fusion estimation model is as follows: a three-level structure of bottom-level embedding, middle-level fusion, and top-level output is adopted; in the bottom layer, the CNN-LSTM neural network model pre-trained in steps S3 and S4 is embedded, the core parameters of the CNN convolutional layer and LSTM layer are frozen, and only the 64-dimensional feature layer before the fully connected output is unfrozen; in the middle layer, the Transformer interaction layer has 8-head attention, a hidden layer dimension of 256, and the activation function is the GELU function, inputting multi-dimensional integrated features and capturing the land-sea interaction effect; In the top layer, the fully connected layer outputs atmospheric nitrogen deposition flux into the bay.

Images