Accurate prevention and control method for agricultural non-point source pollution based on dissolved oxygen hysteresis effect

Abstract

The invention discloses an accurate prevention and control method for agricultural non-point source pollution based on a dissolved oxygen hysteresis effect. The invention relates to the technical field of agricultural non-point source pollution prevention and control, and solves the problem that an existing agricultural non-point source pollution treatment method is difficult to meet accurate prevention and control requirements. The method comprises the following steps: constructing an SWAT model, and simulating monthly-scale nutritive salt load of each crop planting area; establishing a drainage basin-lake coupling model, inputting the nutritive salt load simulated by the SWAT model into a lake hydrodynamic water quality model, and performing calibration; adopting dissolved oxygen data output by the calibrated lake water power water quality model as a training set to train a dissolved oxygen concentration prediction model; determining dominant lag time of different crop areas by adopting an SHAP value analysis method; the optimal fertilization time window of each crop area is calculated by applying a backward model, and a differentiated fertilization reduction scheme is formulated based on the contribution degree weight of the nutritive salt, so that accurate regulation and control of agricultural non-point source pollution are realized, and the lake water quality is effectively improved while the use of chemical fertilizer is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural non-point source pollution prevention and control, and in particular to a method for precise prevention and control of agricultural non-point source pollution based on the dissolved oxygen hysteresis effect. Background Art

[0002] Dissolved oxygen (DO) is a core indicator of water ecological health, and its dynamic changes can comprehensively reflect the cumulative ecological effects of nutrient pollution. Compared with conventional indicators such as nitrogen and phosphorus, DO has three significant advantages: first, its concentration changes directly represent the self-purification capacity of the water body. When DO <3 mg / L, it will trigger an ecological crisis; second, DO's response to nutrient input has a significant lag of 3-12 months, providing a time window for prevention and control; third, DO changes can integrate the influence of multiple factors such as climate and hydrology to provide a more comprehensive ecological risk assessment. When lakes face seasonal hypoxia problems, with a high contribution rate from agricultural non-point source pollution, traditional management and control methods based on nitrogen and phosphorus concentrations cannot predict ecological risks, and there is an urgent need to establish a new DO-oriented regulatory system.

[0003] Despite the clear advantages of the DO-based regulation concept, its implementation faces three major technical bottlenecks: First, the DO lag effect is influenced by multiple factors, including watershed characteristics, climate conditions, and agricultural management practices, making it difficult to accurately quantify its temporal and spatial variability using traditional statistical methods. Second, existing water quality models inadequately capture physical-biochemical coupling, leading to significant deviations in DO predictions. Finally, the lack of standardized methods for translating DO response characteristics into specific fertilization plans restricts the technology's practical application. These bottlenecks have resulted in current agricultural non-point source pollution control remaining at an empirical, extensive stage, making it difficult to meet the demands of precise prevention and control. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for precise prevention and control of agricultural non-point source pollution based on the dissolved oxygen hysteresis effect, so as to solve the problem that existing agricultural non-point source pollution control methods are difficult to meet the needs of precise prevention and control.

[0005] The present invention provides a method for accurately preventing and controlling agricultural non-point source pollution based on the dissolved oxygen hysteresis effect, comprising:

[0006] Step 1: Build the SWAT model, input the digital elevation model, soil type, crop rotation and meteorological data, simulate the monthly nutrient load of each crop planting area and output it;

[0007] Step 2: Establish a watershed-lake coupling model, input the nutrient load simulated by the SWAT model into the lake hydrodynamic and water quality model, and calibrate the parameters of the lake hydrodynamic and water quality model using the measured dissolved oxygen data of the lake to ensure that the dissolved oxygen dynamics output by the lake hydrodynamic and water quality model are consistent with the actual situation;

[0008] Step 3: Using the dissolved oxygen data output by the calibrated lake hydrodynamic and water quality model as a training set, a long-short-term memory network model is constructed. This model is then fed with watershed nutrient load, meteorological data, and lake hydrological data lagged 1-12 months, to train a dissolved oxygen concentration prediction model.

[0009] Step 4: Using the SHAP value analysis method, the contribution of nutrient load to dissolved oxygen changes in each lag month was calculated to determine the dominant lag time in different crop areas;

[0010] Step 5: Based on the dissolved oxygen management objectives and the dominant lag time, the critical dissolved oxygen reverse calculation model is used to calculate the optimal fertilization time window for each crop area, and differentiated fertilization reduction plans are formulated based on the nutrient contribution weights. At the same time, real-time meteorological data is integrated to dynamically adjust management measures.

[0011] Furthermore, step one includes:

[0012] A distributed hydrological model was established using the SWAT model, with a spatial resolution of 100m × 100m grid cells to ensure that the spatial heterogeneity of different land use types is reflected. Input data included a 30-meter digital elevation model, a 1:50,000 scale soil type distribution map, monthly crop rotation data, and daily rainfall and temperature data from meteorological stations.

[0013] The runoff curve number, soil nitrate nitrogen content, and soil erosion control factor were calibrated, and the SUFI-2 algorithm was used for parameter sensitivity analysis and automatic optimization to ensure that the Nash efficiency coefficient of runoff and nutrient load simulation reached above 0.65.

[0014] The model outputs monthly total nitrogen and total phosphorus load data for each major crop planting area, providing basic input for subsequent dissolved oxygen response analysis.

[0015] Furthermore, step 2 includes:

[0016] The monthly total nitrogen and total phosphorus load data for each cropping area simulated by the SWAT model were input into the lake hydrodynamic and water quality model through a hydrological connection channel. A dissolved oxygen module was configured in the basin-lake coupling model to account for atmospheric reoxygenation, algal photosynthesis and respiration, and sediment oxygen consumption processes.

[0017] The model was calibrated using the measured dissolved oxygen data of the study lake area once a month for the past five years. The optimal parameter combination was determined through parameter sensitivity analysis, so that the Nash efficiency coefficient between the simulated values ​​and the measured values ​​reached above 0.65. The calibrated basin-lake coupling model was run to output the spatial distribution data of dissolved oxygen concentration in the lake area on a monthly basis for many years, which served as the training benchmark data set for the long-short-term memory network model.

[0018] Furthermore, step three includes:

[0019] The network structure was set as a two-layer long-short-term memory network model with 64 hidden units, using the ReLU activation function to capture nonlinear relationships in time series data. The input features included total nitrogen and total phosphorus load data for each crop region with a lag of 1-12 months, monthly average temperature, and monthly cumulative rainfall. The output was the predicted monthly average dissolved oxygen concentration in the lake area.

[0020] The five-fold cross validation method was used to evaluate the performance of the dissolved oxygen concentration prediction model, requiring the determination coefficient R of the test set to be 2 ≥0.75, the root mean square error of dissolved oxygen concentration prediction ≤0.5mg / L; early stopping and random dropout regularization were added during training to prevent overfitting.

[0021] Furthermore, step four includes:

[0022] Based on the trained dissolved oxygen concentration prediction model, the SHAP method is used to calculate the contribution of the nutrient load characteristics of each lagged month; for each sample x, the SHAP value of its feature i is calculated using the formula:

[0023]

[0024] Where φ_i represents the contribution of the historical nutrient load in month i to the current dissolved oxygen concentration; S represents any combination of months selected from the 1-12 month lag characteristics; M = 12, corresponding to a maximum lag time window of 12 months; f_x is a dissolved oxygen concentration prediction model, which takes the total nitrogen and phosphorus loads and meteorological and hydrological data of the past 12 months as input and outputs the dissolved oxygen concentration in the lake area; N{i} represents all historical months except month i; |S| represents the number of months included in the currently calculated month combination;

[0025] By analyzing the distribution characteristics of SHAP values, the dominant lag time τ of nutrient input in each crop area was determined.

[0026] Furthermore, step five includes:

[0027] Set target values ​​for dissolved oxygen management in the lake area and determine critical periods based on historical monitoring data;

[0028] The recommended fertilization time is calculated using the critical dissolved oxygen reverse calculation model:

[0029] t_fertilization = t_critical - τ - Δt_buffer (2)

[0030] Where t_fertilization is the recommended fertilization date for the crop area; t_critical is the date when the dissolved oxygen concentration in the lake area first falls below 5 mg / L; τ is the dominant lag time for the crop area; Δt_buffer = 15 days, which is the buffer period reserved for the fertilization measures to take effect;

[0031] Calculate the recommended fertilizer reduction ratio for each crop zone based on the SHAP value weight:

[0032] Reduction (%) = 100 × (SHAP_TN / ∑SHAP) × k (3)

[0033] Among them, SHAP_TN represents the contribution of the total nitrogen load in the current crop area to the decrease in dissolved oxygen; ∑SHAP is the sum of the SHAP values ​​of the total nitrogen and total phosphorus loads in all lagged months; and k is the safety factor set according to the fertilizer tolerance of the crop.

[0034] The present invention has the following beneficial effects: The present invention's precise control method for agricultural non-point source pollution based on the dissolved oxygen hysteresis effect establishes a method for analyzing the dissolved oxygen hysteresis effect based on a watershed-lake coupling model and an LSTM model, achieving precise quantification of the relationship between agricultural non-point source pollution and lake hypoxia. Furthermore, a comprehensive decision-making model for nutrient loading, dissolved oxygen response, and fertilization regulation is constructed. Through SHAP value analysis, the hysteretic effects of fertilization in different crop regions on lake dissolved oxygen are revealed. This method fully accounts for the spatiotemporal heterogeneity of meteorological and hydrological factors and the hysteresis effect of nutrient migration and transformation, more accurately reflecting the actual process of agricultural non-point source pollution's impact on dissolved oxygen in water bodies. Compared to traditional empirical fertilization or single water quality assessment methods, the proposed technical solution can accurately calculate the optimal fertilization time window and reduction ratio for each crop region based on historical monitoring data and model simulations. This effectively reduces the risk of lake hypoxia while ensuring crop yields, providing a quantifiable, actionable, and precise control method for agricultural non-point source pollution control. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0036] Figure 1 A flow chart of the method for precise prevention and control of agricultural non-point source pollution based on the dissolved oxygen hysteresis effect provided by the present invention;

[0037] Figure 2 This is the characteristic diagram of dissolved oxygen hysteresis effect;

[0038] Figure 3 This is a graph verifying the prediction effect of the LSTM model;

[0039] Figure 4 Schematic diagram of the precise fertilization control plan. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings.

[0041] See also Figure 1 The method for accurately preventing and controlling agricultural non-point source pollution based on the dissolved oxygen hysteresis effect provided by the embodiment of the present invention mainly includes the following steps:

[0042] (1) Construct the SWAT (Soil and Water Assessment Tool) model, input the Digital Elevation Model (DEM), soil type, crop rotation, and meteorological data, simulate the monthly nutrient load of each crop planting area, and output it. This step specifically includes:

[0043] (1-1) Construction of high-precision watershed hydrological model

[0044] A distributed hydrological model was developed using the SWAT model, with a spatial resolution of 100m × 100m grid cells to ensure that the spatial heterogeneity of different land use types is reflected. Input data included a 30-meter-precision DEM, a 1:50,000-scale soil type distribution map, monthly crop rotation data, and daily rainfall and temperature data from meteorological stations.

[0045] (1-2) Model parameter calibration and verification

[0046] Key parameters, such as the runoff curve number, soil nitrate nitrogen content, and soil erosion control factors, were calibrated. The Sequential Uncertainty Fitting Version 2 (SUFI-2) algorithm was used for parameter sensitivity analysis and automatic optimization. For one watershed, the calibrated Nash-Sutcliffe efficiency coefficient (NSE) for runoff simulation reached 0.85, the NSE for total nitrogen load simulation was 0.70, and the NSE for total phosphorus load simulation was 0.68, all meeting the model accuracy requirements.

[0047] (1-3) Model output and data preparation

[0048] The SWAT model outputs monthly total nitrogen and total phosphorus loading data for major crop-growing areas, such as rice paddies and vegetable fields. For one lake region, for example, the average monthly total nitrogen loading for rice paddies ranged from 2.5 to 8.7 kg / ha, and the average total phosphorus loading ranged from 0.3 to 1.2 kg / ha. These outputs provide the foundational input for subsequent dissolved oxygen response analysis. The SWAT model outputs serve as input data for the coupled model.

[0049] (2) Establish a watershed-lake coupling model, input the nutrient load simulated by the SWAT model into lake hydrodynamic and water quality models, such as EFDC and MIKE, and calibrate the lake model parameters using a small amount of measured dissolved oxygen data from lakes to ensure that the dissolved oxygen dynamics output by the model are consistent with the actual situation. This step specifically includes:

[0050] (2-1) Basin-Lake Model Coupling

[0051] The monthly total nitrogen and phosphorus load data for each cropping area simulated by the SWAT model are input into the lake hydrodynamic-water quality model EFDC or MIKE through a hydrological link. A dissolved oxygen module is configured in the model to account for key processes such as atmospheric reoxygenation, algal photosynthesis / respiration, and sediment oxygen consumption.

[0052] (2-2) Model calibration and data generation

[0053] The model was calibrated using the measured dissolved oxygen data of the study lake area once a month for the past five years, and the optimal parameter combination was determined through parameter sensitivity analysis. Taking a certain lake as an example, the NSE of the dissolved oxygen simulation reached 0.72, and the correlation coefficient R 2 The coupled model after calibration is run to output the spatial distribution data of dissolved oxygen concentration in the lake area for many years and months, which is used as the training benchmark data set for the LSTM model. Figure 2 As shown, the data can clearly show the lagging response relationship between nutrient load and dissolved oxygen changes, with the lowest dissolved oxygen value being 3.2 mg / L from July to August and the highest value being 8.5 mg / L from December to January.

[0054] (3) Using the dissolved oxygen data output by the calibrated lake hydrodynamic and water quality model as the training set, a long short-term memory (LSTM) model is constructed. The nutrient load of the watershed, meteorological data, and lake hydrological data with a lag of 1-12 months are input to obtain a high-precision dissolved oxygen concentration prediction model. Meteorological data include temperature, rainfall, and wind speed, while lake hydrological data include water level and water temperature. This step specifically includes:

[0055] (3-1) LSTM model construction

[0056] The network structure was set as a two-layer LSTM model with 64 hidden units, using the ReLU activation function to capture nonlinear relationships in time series data. Input features included lagged total nitrogen and total phosphorus load data for each cropping region, ranging from 0.1 to 9.5 kg / ha for total nitrogen and 0.05 to 1.5 kg / ha for total phosphorus, as well as environmental factors such as average monthly temperature of 5°C to 32°C and monthly cumulative rainfall of 20 mm to 450 mm. The output was a predicted monthly average dissolved oxygen concentration of 3.0 mg / L to 9.0 mg / L for the lake region.

[0057] (3-2) Model training and verification

[0058] The five-fold cross-validation method is used to evaluate the model performance. Taking a lake application as an example, the coefficient of determination R of the test set is 2 The RMSE (Root Mean Square Error) and MAE (Mean Absolute Error) of the predicted dissolved oxygen concentration were 0.38 mg / L and 0.29 mg / L, respectively. Figure 3 As shown in the figure, the model predictions are well consistent with the measured values, verifying the reliability of the model. Early stopping and dropout regularization were added during training with a ratio of 0.2 to prevent overfitting.

[0059] (4) Using the SHAP (Shapley Additive Explanations) value analysis method, calculate the contribution of nutrient load to dissolved oxygen changes in each lag month and determine the dominant lag time in different crop areas. This step specifically includes:

[0060] (4-1) SHAP value contribution analysis

[0061] Based on the trained LSTM model, the SHAP method was used to calculate the contribution of nutrient load characteristics in each lagged month;

[0062] (4-2) Determination of the dominant lag time

[0063] By analyzing the distribution characteristics of SHAP values, the dominant lag time τ of nutrient salt input in each crop area was determined. For example, the total nitrogen load in rice fields lagged for 5.2±0.8 months, with a SHAP value of 0.32±0.05; the total nitrogen load in vegetable fields lagged for 3.5±0.6 months, with a SHAP value of 0.28±0.04, providing a scientific basis for the timing of precise fertilization. Figure 2 As shown in the figure, the lag time can be directly reflected by the phase difference between the nutrient loading and the dissolved oxygen change.

[0064] (5) Based on the dissolved oxygen management target and the dominant lag time, the critical dissolved oxygen reverse calculation model is used to calculate the optimal fertilization time window for each crop area, and a differentiated fertilization reduction plan is formulated based on the nutrient contribution weight. At the same time, real-time meteorological data is integrated to dynamically adjust management measures. This step specifically includes:

[0065] (5-1) Dissolved oxygen management target setting

[0066] The target value for dissolved oxygen management in the lake area is set according to the Surface Water Environmental Quality Standard (GB3838-2002), with a monthly average DO ≥ 5 mg / L. The critical period is determined based on historical monitoring data. For example, during the high temperature period from July to August, the lowest measured DO value is 3.2 mg / L.

[0067] (5-2) Calculation of the optimal fertilization time window

[0068] The critical dissolved oxygen reverse calculation model is used to calculate the recommended fertilization time, such as Figure 4 As shown, for rice fields with a lag time of 5.2 months, to avoid the DO trough in August, the recommended fertilization time should be February 10 ± 10 days, which is calculated as August 1 - 5.2 months - 15 days;

[0069] (5-3) Development of differentiated fertilization plans

[0070] The recommended fertilization reduction ratio for each crop area is calculated based on the SHAP value weight, and specific plans are formulated in combination with farming practices, such as reducing base fertilizer in rice fields by 42.5%±3.2% and reducing topdressing in vegetable fields by 18.6%±2.8%. These plans are then verified through field trials. Figure 4 As shown in the middle color block, different crop types and fertilization periods correspond to differentiated reduction ratios, 35%-45% for rice fields and 15%-25% for vegetable fields, achieving precise control.

[0071] Technologies not described in detail in this invention are existing technologies, such as the SWAT model, lake hydrodynamic and water quality model, LSTM model, etc., which will not be described in detail here.

[0072] The above-described embodiments of the present invention do not limit the protection scope of the present invention.

Claims

1. A method for precise prevention and control of agricultural non-point source pollution based on the dissolved oxygen hysteresis effect, characterized in that: include: Step 1: Build the SWAT model, input the digital elevation model, soil type, crop rotation and meteorological data, simulate the monthly nutrient load of each crop planting area and output it; Step 2: Establish a watershed-lake coupling model, input the nutrient load simulated by the SWAT model into the lake hydrodynamic and water quality model, and calibrate the parameters of the lake hydrodynamic and water quality model using the measured dissolved oxygen data of the lake to ensure that the dissolved oxygen dynamics output by the lake hydrodynamic and water quality model are consistent with the actual situation; Step 3: Using the dissolved oxygen data output by the calibrated lake hydrodynamic and water quality model as a training set, a long-short-term memory network model is constructed. This model is then fed with watershed nutrient load, meteorological data, and lake hydrological data lagged 1-12 months, to train a dissolved oxygen concentration prediction model. Step 4: Using the SHAP value analysis method, the contribution of nutrient load to dissolved oxygen changes in each lag month was calculated to determine the dominant lag time in different crop areas; Step 5: Based on the dissolved oxygen management objectives and the dominant lag time, the critical dissolved oxygen reverse calculation model is used to calculate the optimal fertilization time window for each crop area, and differentiated fertilization reduction plans are formulated based on the nutrient contribution weights. At the same time, real-time meteorological data is integrated to dynamically adjust management measures.

2. The method for accurately preventing and controlling agricultural non-point source pollution based on the dissolved oxygen hysteresis effect according to claim 1, wherein: Step one includes: A distributed hydrological model was established using the SWAT model, with a spatial resolution of 100m × 100m grid cells to ensure that the spatial heterogeneity of different land use types is reflected. Input data included a 30-meter digital elevation model, a 1:50,000 scale soil type distribution map, monthly crop rotation data, and daily rainfall and temperature data from meteorological stations. The runoff curve number, soil nitrate nitrogen content, and soil erosion control factor were calibrated, and the SUFI-2 algorithm was used for parameter sensitivity analysis and automatic optimization to ensure that the Nash efficiency coefficient of runoff and nutrient load simulation reached above 0.

65. The model outputs monthly total nitrogen and total phosphorus load data for each major crop planting area, providing basic input for subsequent dissolved oxygen response analysis.

3. The method for accurately preventing and controlling agricultural non-point source pollution based on the dissolved oxygen hysteresis effect according to claim 1, wherein: Step 2 includes: The monthly total nitrogen and total phosphorus load data for each cropping area simulated by the SWAT model were input into the lake hydrodynamic and water quality model through a hydrological connection channel. A dissolved oxygen module was configured in the basin-lake coupling model to account for atmospheric reoxygenation, algal photosynthesis and respiration, and sediment oxygen consumption processes. The model was calibrated using the measured dissolved oxygen data of the study lake area once a month for the past five years. The optimal parameter combination was determined through parameter sensitivity analysis, so that the Nash efficiency coefficient between the simulated values ​​and the measured values ​​reached above 0.

65. The calibrated basin-lake coupling model was run to output the spatial distribution data of dissolved oxygen concentration in the lake area on a monthly basis for many years, which served as the training benchmark data set for the long-short-term memory network model.

4. The method for accurately preventing and controlling agricultural non-point source pollution based on the dissolved oxygen hysteresis effect according to claim 1, wherein: Step three includes: The network structure was set as a two-layer long-short-term memory network model with 64 hidden units, using the ReLU activation function to capture nonlinear relationships in time series data. The input features included total nitrogen and total phosphorus load data for each crop region with a lag of 1-12 months, monthly average temperature, and monthly cumulative rainfall. The output was the predicted monthly average dissolved oxygen concentration in the lake area. The five-fold cross validation method was used to evaluate the performance of the dissolved oxygen concentration prediction model, requiring the determination coefficient R of the test set to be 2 ≥0.75, the root mean square error of dissolved oxygen concentration prediction ≤0.5mg / L; early stopping and random dropout regularization were added during training to prevent overfitting.

5. The method for accurately preventing and controlling agricultural non-point source pollution based on the dissolved oxygen hysteresis effect according to claim 1, wherein: Step 4 includes: Based on the trained dissolved oxygen concentration prediction model, the SHAP method is used to calculate the contribution of the nutrient load characteristics of each lagged month; for each sample x, the SHAP value of its feature i is calculated using the formula: Where φ_i represents the contribution of the historical nutrient load in month i to the current dissolved oxygen concentration; S represents any combination of months selected from the 1-12 month lag characteristics; M = 12, corresponding to a maximum lag time window of 12 months; f_x is a dissolved oxygen concentration prediction model, which takes the total nitrogen and phosphorus loads and meteorological and hydrological data of the past 12 months as input and outputs the dissolved oxygen concentration in the lake area; N{i} represents all historical months except month i; |S| represents the number of months included in the currently calculated month combination; By analyzing the distribution characteristics of SHAP values, the dominant lag time τ of nutrient input in each crop area was determined.

6. The method for accurately controlling agricultural non-point source pollution based on the dissolved oxygen hysteresis effect according to claim 1, characterized in that: Step five includes: Set target values ​​for dissolved oxygen management in the lake area and determine critical periods based on historical monitoring data; The recommended fertilization time is calculated using the critical dissolved oxygen reverse calculation model: t_fertilization = t_critical - τ - Δt_buffer (2) Where t_fertilization is the recommended fertilization date for the crop area; t_critical is the date when the dissolved oxygen concentration in the lake area first falls below 5 mg / L; τ is the dominant lag time for the crop area; Δt_buffer = 15 days, which is the buffer period reserved for the fertilization measures to take effect; Calculate the recommended fertilizer reduction ratio for each crop zone based on the SHAP value weight: Reduction (%) = 100 × (SHAP_TN / ∑SHAP) × k (3) Among them, SHAP_TN represents the contribution of the total nitrogen load in the current crop area to the decrease in dissolved oxygen; ∑SHAP is the sum of the SHAP values ​​of the total nitrogen and total phosphorus loads in all lagged months; and k is the safety factor set according to the fertilizer tolerance of the crop.

Images