A water-wetland-food collaborative optimization method, system, device and medium

Abstract

The application discloses a water-wetland-food collaborative optimization method, system, device and medium, relates to the technical field of multi-target collaborative optimization, and comprises the following steps: a multi-target optimization model is constructed with crop sowing area as a decision variable, wherein the optimization target of the objective function of the multi-target optimization model includes maximum food yield, minimum irrigation water demand and maximum wetland area, and the constraint conditions include crop yield constraint, crop irrigation water demand constraint and wetland area constraint; the multi-target optimization model is solved by using an NSGA-III algorithm to obtain a Pareto frontier solution set; a machine learning model is used to predict the dynamic change of the wetland area, and the prediction result is coupled with the NSGA-III algorithm to calculate the wetland area objective function. The application can ensure that the wetland ecosystem remains stable or develops well in the foreseeable future while meeting the yield and water-saving targets, and ensures the long-term ecological robustness and sustainability of the scheme.

Description

Technical Field

[0001] This invention relates to the field of multi-objective collaborative optimization technology, and in particular to a method, system, equipment and medium for collaborative optimization of water-wetland-food resources. Background Technology

[0002] In agricultural production, water and land resources are fundamental elements for ensuring sustainable food production. Optimizing planting structure and rationally allocating water and soil resources have become key measures for achieving sustainable agricultural development. Their necessity and urgency are mainly reflected in the following three aspects: First, optimizing planting structure is the core way to improve resource utilization efficiency. Traditional single-crop planting patterns often lead to low utilization rates of water and soil resources, making it difficult to adapt to the dynamic changes in climate change and market demand. By adjusting the crop planting structure, the resource utilization characteristics of different crops can be fully utilized to achieve efficient allocation of water and soil resources. Second, optimizing the allocation of water and soil resources is an inevitable choice to cope with resource shortages. As the problem of water shortage becomes increasingly prominent, how to achieve optimal utilization of land resources under strict water resource constraints has become the key to ensuring food security. Finally, the synergistic optimization of planting structure and water and soil resource allocation is an important measure to achieve multi-objective synergy.

[0003] Different agricultural ecosystems lead to differentiated land use patterns. In mid- to high-latitude regions, the rapid expansion of rice cultivation and the rapid decline of dryland crops have impacted national food security and overall land use planning. This research is crucial for understanding and guiding responses to food sustainability and environmental issues, and also provides a scientific basis for exploring planting structure optimization. The Northeast Black Soil Region, a "ballast stone" for my country's food security, suffers from an imbalanced planting structure of corn, rice, and soybeans. This imbalance not only hinders the stable development of grain production in Northeast China, but also poses a significant threat to the region's wetland ecosystem due to the continuously increasing agricultural water demand. Given the competition between agriculture and ecosystems for water, it is essential to adjust the planting structure to ensure the security of multi-system resources. Existing technologies employ a multi-objective collaborative optimization method, constructing a water-carbon-economy-ecology coupled model based on the NSGA-Ⅲ algorithm. This model achieves the comprehensive goals of alleviating water shortages, improving economic benefits, enhancing carbon sequestration capacity, and optimizing ecological benefits in Northeast China. The results show that the optimized rice planting area can still increase by 13.7%. Considering water-saving benefits, the study takes Heilongjiang Province as the research area and explores the optimization of planting structure based on the multi-objective model. The results indicate that in water-scarce areas, potatoes, which have a shorter growth cycle than corn, soybeans, and rice, can be used as substitutes. Although economic benefits decrease, ecological and water-saving benefits increase.

[0004] While some existing research has begun to focus on the interaction between water and soil resources, it usually only focuses on the optimal allocation of resource quantities. Wetland area is often treated as a fixed constraint or a static parameter independent of agricultural water use. The "optimal" planting scheme obtained in this way may gradually lead to wetland degradation or ecological risks after implementation because it does not take into account the cumulative and lagging effects on wetland hydrology. This makes it impossible to guarantee the long-term ecological robustness of the decision-making scheme. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a method, system, device, and medium for the synergistic optimization of water-wetland-grain resources, thereby solving the problems in the prior art.

[0006] The present invention specifically provides the following technical solution: A method for co-optimization of water-wetland-food resources includes the following steps: A multi-objective optimization model is constructed with crop planting area as the decision variable. The optimization objectives of the objective function in the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area. The constraints of the multi-objective optimization model include crop yield constraint, crop irrigation water demand constraint, and wetland area constraint. Based on the objective function and constraints, the multi-objective optimization model is solved using the NSGA-III algorithm to obtain the Pareto front solution set. Then, the optimal planting scheme is selected from the Pareto front solution set based on scenario analysis. The scenario analysis refers to the judgment of the ecological state of the planting structure under different scenarios. The precipitation, evapotranspiration, area of ​​different crops, irrigation water demand and elevation of the area to be predicted are collected as the data to be predicted. The dynamic changes of wetland area are predicted, and the dynamic changes of wetland area are coupled with the optimal planting scheme to obtain the optimal agricultural planting structure adjustment scheme.

[0007] Preferably, the objective objectives of the objective function in the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area, specifically: The objective function for maximizing grain output is expressed as: ; In the formula, G for K Within each prefecture-level city, J Within the year I Total yield of crops; for The first in the prefecture-level city Annual crops Yield per unit area; For the first k The first in the prefecture-level city Annual crops The sown area; The objective function for minimizing irrigation water demand is expressed as: ; In the formula, for K each prefecture-level city J Year I Total irrigation water requirement for crops For the first The first prefecture-level city Year The water deficit of crops is estimated by subtracting the effective precipitation from the water requirement during the crop's growth period. The objective function for maximizing wetland area is expressed as: ; In the formula, For wetland area, and The first prefecture-level cities Annual evapotranspiration and precipitation For the first prefecture-level cities Annual water demand for rice irrigation For the first prefecture-level cities Annual elevation.

[0008] Preferably, the constraints of the multi-objective optimization model include crop yield constraints, crop irrigation water requirement constraints, and wetland area constraints, specifically: The specific expression for the crop yield constraint is: ; The specific expression for crop irrigation water requirement constraints is as follows: ; The specific expression for the wetland area constraint is: ; In the formula, and The first k The first in the prefecture-level city Annual crops Minimum and maximum sowing area; for K each prefecture-level city J Year I Agricultural water consumption for crop cultivation; This represents the actual wetland area under the current scenario.

[0009] Preferably, the step of solving the multi-objective optimization model using the NSGA-III algorithm to obtain the Pareto front solution set, and then selecting the optimal planting scheme from the Pareto front solution set based on scenario analysis, specifically involves: Randomly generate an initial population containing various grain planting areas; Set crossover and mutation probabilities, and perform population evolution based on non-dominated sorting and crowding distance calculations; The Pareto front solution set is obtained through multiple iterations, and the optimal planting scheme is selected from the Pareto front solution set based on scenario analysis.

[0010] Preferably, the scenario analysis specifically includes: Scenario C1: Prioritize food supply, with the constraint that total food production should remain consistent with actual production; Scenario C2: Coordinated development of water, food and wetlands, with the irrigation water demand in the constraints remaining consistent with the actual irrigation water demand; Scenario C3: Prioritize wetland ecological restoration, and ensure that the wetland area in the constraints remains consistent with the actual wetland area.

[0011] This invention provides a water-wetland-food synergistic optimization system, comprising: The model building module is used to construct a multi-objective optimization model with crop planting area as the decision variable. The optimization objectives of the objective function in the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area. The constraints of the multi-objective optimization model include crop yield constraints, crop irrigation water demand constraints, and wetland area constraints. The multi-objective solution module is used to solve the multi-objective optimization model based on the objective function and constraints using the NSGA-Ⅲ algorithm, obtain the Pareto front solution set, and select the optimal planting scheme from the Pareto front solution set based on scenario analysis; the scenario analysis refers to the judgment of the ecological status of the planting structure under different scenarios. The collaborative optimization module is used to collect precipitation, evapotranspiration, different crop areas, irrigation water demand and elevation of the area to be predicted as the data to be predicted, predict the dynamic changes of wetland area, and couple the dynamic changes of wetland area with the optimal planting scheme to obtain the optimal agricultural planting structure adjustment scheme.

[0012] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described water-wetland-grain co-optimization method.

[0013] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described water-wetland-grain synergistic optimization method.

[0014] Compared with the prior art, the present invention has the following significant advantages: This invention combines machine learning algorithms with multi-objective optimization models to construct an integrated "prediction-optimization-decision" research framework. Various parameters are used as the objective functions of the multi-objective optimization model. From the top level of the optimization framework, wetland ecological status is elevated to an active optimization dimension of equal importance to production and water conservation, laying the structural foundation for subsequent dynamic coupling. The NSGA-III algorithm is used to solve the multi-objective optimization model, efficiently handling complex constraints and generating a uniformly distributed Pareto front solution set. This solution set intuitively reveals the quantitative trade-offs between food production, water conservation, and wetland protection, which cannot be simultaneously achieved. Instead of a single solution, this approach provides decision-makers with a set of candidate solutions covering different preferences (such as production-oriented or ecology-oriented). It effectively alleviates the unsustainable resource problem in the watershed while maximizing the benefits of the grain storage function and wetland system. It couples the dynamic changes in wetland area with the optimal planting scheme. If the prediction shows that the wetland area is evolving in an unsustainable direction, the model parameters need to be adjusted or the scenario selection needs to be re-examined until the obtained scheme can meet the production and water-saving targets while ensuring that the wetland ecosystem can maintain stable or healthy development in the foreseeable future, thus ensuring the long-term ecological robustness and sustainability of the scheme. Attached Figure Description

[0015] Figure 1 This is a flowchart of the NSGA-Ⅲ algorithm of the present invention; Figure 2 This is a flowchart of the multi-module joint simulation of the NSGA-Ⅲ algorithm and machine learning in this invention; Figure 3 This is a flowchart of the LSTM modeling process of the present invention; Figure 4 This is a diagram of the BP neural network topology of the present invention; Figure 5 This is a schematic diagram of the RF algorithm topology of the present invention; Figure 6 This is a comparative analysis chart showing the wetland area prediction results of different machine learning models of the present invention; wherein, Figure 6 (a) and (b) show the comparison of the prediction results and goodness of fit of the BP neural network model for wetland area, respectively. Figure 6 (c) and (d) show the comparison of the prediction results and goodness of fit of the LSTM model for wetland area, respectively. Figure 6 (e) and (f) represent the comparison of wetland area prediction results and goodness of fit of the RF model, respectively; Figure 7 This is a graph showing the interannual variation of crop planting structure optimization results for scenarios C1, C2, and C3 of this invention; Figure 7 (a), (b), and (c) show the interannual variation of wetland area under scenarios C1, C2, C3, and C0, respectively. Figure 7 (d), (e), and (f) represent the interannual variations in crop yields under scenarios C1, C2, and C3 compared to scenario C0, respectively. Figure 7 (g), (h), and (i) represent the interannual variations in irrigation water requirements for each crop under scenarios C1, C2, and C3 compared to scenario C0, respectively. Figure 7 (j), (k), and (l) represent the interannual variations in the sown area of ​​each crop in scenarios C1, C2, and C3 compared to scenario C0, respectively. Figure 8 The spatial variation diagram of crop yield, irrigation water requirement and wetland area after the optimization of crop planting structure in scenarios C1, C2 and C3 of the present invention; Figure 8 (a), (b), and (c) show the spatial variation of wetland area in each sub-region under scenarios C1, C2, and C3 compared to scenario C0, respectively. Figure 8 (d), (e), and (f) represent the spatial variation of the sown area of ​​each crop in scenarios C1, C2, and C3 compared to scenario C0, respectively. Figure 8 (g), (h), and (i) represent the spatial variation of crop yields in scenarios C1, C2, and C3 compared to scenario C0, respectively. Figure 8 (j), (k) and (l) represent the spatial variation of irrigation water requirements for each crop in scenarios C1, C2 and C3 compared to scenario C0, respectively. Figure 9 This is a diagram showing the adjustment ratio of crop planting structure under different hydrological year types according to the present invention; Figure 9 (a), (b), and (c) represent the adjustment ratios for scenario C1, respectively. Figure 9 (d), (e), and (f) represent the adjustment ratios for scenario C2, respectively. Figure 9 (g), (h), and (i) represent the adjustment ratios for scenario C3. Figure 10 The flowchart illustrates a water-wetland-grain synergistic optimization method provided by this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] In multi-objective optimization, the Genetic Algorithm (GA) consists of three parts: decision variables, objective function, and constraints. Decision variables are chromosomes with characteristic entities; these chromosomes generate a potential solution through random permutations and combinations. The objective function and constraints define a specific survival environment for these chromosomes, allowing them to compete for the optimal solution through crossover, mutation, and iterative updates, based on the principles of survival of the fittest. The GA's operational flow includes population initialization, fitness evaluation, selection, crossover, mutation, and iterative updates. Before population initialization, the potential solutions to the problem need to be converted into recognizable strings using binary encoding, floating-point encoding, etc., to generate the initial population. Fitness evaluation refers to the differences exhibited by each individual after being constrained by the multi-objective function; a higher value indicates a better individual. The selection operation selects individuals based on their fitness values ​​to generate the next generation. Two individuals serve as parents, exchanging some genetic information to generate new offspring individuals, which are then randomly mutated to increase population diversity. Finally, fitness evaluation, selection, crossover, and mutation operations are repeated until the maximum number of iterations is reached or fitness converges.

[0018] Genetic algorithms possess advantages such as strong global search capabilities and suitability for complex optimization problems, but they also have some limitations. For example, they struggle to simultaneously optimize multiple conflicting objectives in multi-objective optimization problems, and their computational complexity increases with the number of objectives. In the field of evolutionary algorithms, a significant improvement—NSGA-II—was proposed, improving both computational efficiency and overcoming the limitation of genetic algorithms in multi-objective computation. However, its computational performance significantly decreases with the number of objectives. Therefore, NSGA-III was reintroduced, employing the framework of NSGA-II and improving upon it to better handle optimization problems with multiple conflicting objectives. Furthermore, the introduction of reference points better maintains the diversity and uniformity of solutions, avoiding concentration in local regions, thereby improving the algorithm's performance in high-dimensional objective spaces. In NSGA-II, solutions are selected based on the crowding distance in the Ft optimal solution; the larger the crowding distance, the higher the probability of selection. Figure 1As shown, NSGA-III employs a reference point-based method. First, it finds the minimum value z (z = (z1, z2, ..., zm)) of the i-th objective in all solutions of the population St. Then, it normalizes the solutions and maps the ideal points to zero vectors. Finally, it partitions the objective space using reference lines (straight lines connecting the ideal points and reference points). This algorithm can partition all solutions by calculating the perpendicular distance from each reference line to the nearest reference point, thus achieving spatial partitioning management. This mechanism not only improves population diversity but also significantly enhances the algorithm's convergence performance and search efficiency. Therefore, this invention chooses NSGA-III to solve the multi-objective collaborative optimization problem of agricultural development and wetland ecosystems. Figure 1 ).

[0019] Traditional wetland area prediction methods mainly rely on statistical analysis and empirical models, which have limitations when dealing with complex nonlinear relationships and high-dimensional data. In recent years, intelligent algorithms, represented by deep learning, have demonstrated unique advantages. Their powerful feature extraction and pattern recognition capabilities provide a new technical path to solving the problem of predicting dynamic changes in wetlands. For example, by collecting soil, hydrological, topographic, and vegetation factors, and based on the Google Earth Engine platform, geographic big data, and machine learning algorithms, the distribution of potential wetlands in China was simulated, revealing that 39% of potential wetlands are located in Northeast China. Another approach also used a large wetland water level dataset and machine learning models such as random forests to predict flood dynamics and water seasons in wetlands of different sizes in the United States. While these technical solutions have significant advantages in solving single-system problems, there is currently no technology that couples machine learning with multi-objective optimization algorithms to address the challenge of water resource competition between agricultural development and wetland ecosystems. Therefore, this invention innovatively integrates a machine learning prediction model of the dynamic response of wetland area to changes in planting structure with the NSGA-III multi-objective optimization algorithm, constructing an integrated "prediction-optimization-decision" research framework. This framework, prioritizing grain production, ecological restoration, and the coordinated development of water, wetlands, and grain, has optimized the planting structure. It provides a feasible regulatory scheme for coordinating the improvement of grain production capacity with the protection of wetlands and water resources. This not only has significant practical guiding value but also advances research on the coordinated optimization of "water-agriculture-ecology" from three dimensions: theory, methodology, and application, offering new ideas and methodologies for similar research.

[0020] Given the contradiction between agricultural development and the stable development of wetland ecosystems in the Songhua River Basin, and based on the evaluation of the relationship between agricultural development and wetland ecosystems and the analysis of driving factors, it is evident that large-scale agricultural reclamation and the continuous expansion of paddy fields have led to a continuous decline in the stability and sustainability of the wetland ecosystem in the basin. This "water-for-grain" approach not only sacrifices water resources but also causes a sharp reduction in wetlands and degradation of the black soil layer. Although increased precipitation in recent years has helped alleviate the pressure on wetland ecosystems, the problem of water shortage will be further exacerbated during droughts, leading to severe agricultural yield reductions. Adjusting the planting structure can effectively alleviate this contradiction. However, current technologies for solving this problem mostly focus on simply adjusting the agricultural planting structure or using machine learning to predict key variables of water resources, agriculture, and wetland systems, failing to achieve synergistic optimization research on water-wetland-food systems.

[0021] Therefore, this study employs deep learning methods such as Long Short-Term Memory (LSTM), Random Forest (RF), and Backpropagation (BP) neural network to predict wetland area. The most accurate model is coupled with the crop yield and crop coefficient modules of NSGA-III. Considering the goals of food security, water security, and wetland security, and using the total crop yield, irrigation water consumption, crop planting area, and wetland area as constraints, three objective functions are set: "minimum irrigation water consumption," "no shrinkage of wetland area," and "no reduction in grain yield." The NSGA-III algorithm is used to explore the optimal agricultural planting structure adjustment scheme for coordinating agricultural development and wetland ecosystem in the Songhua River Basin under multiple objective constraints. This aims to effectively alleviate the unsustainable resource problem in the basin while maximizing the benefits of the grain storage function and the wetland system, and to provide a regulatory strategy for the coordinated optimization of grain production capacity improvement and wetland and water resource protection in the basin.

[0022] like Figure 10 This embodiment of a water-wetland-food synergistic optimization method includes the following steps: Step 1: Construct a multi-objective optimization model with crop planting area as the decision variable. The objective functions of the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area. The constraints of the multi-objective optimization model include crop yield constraints, crop irrigation water demand constraints, and wetland area constraints.

[0023] Multi-objective optimization problems typically include decision variables, objective functions, and constraints, and their formulas are as follows: (1); Subject to: (2); Among them, decision variables The independent variable of the problem refers to the planting area of ​​rice, corn, and soybeans in this invention (Table 1). Objective function Through decision variables The objective that the functional relationship aims to achieve is usually expressed in a minimization form, and we have... Three objective functions were established. Based on the current development status, existing problems, and functional positioning of the Songhua River Basin, optimization objectives for the three objective functions were set, along with corresponding constraints.

[0024] Table 1 Decision Variable Settings for NSGA-Ⅲ Algorithm In the multi-objective optimization model, the specific optimization objective of the objective function is as follows: (1) Largest grain output. The Songhua River Basin, as the main part of the black soil granary, produces 18% of my country's grain output using only 3.5% and 3.6% of the country's water and soil resources. Therefore, it is necessary to set an objective function to maintain grain output to ensure the granary's functional status, promote sustainable agricultural development, protect the ecological environment, enhance regional economic value, protect water resources, and even improve international competitiveness. The formula is as follows:

[0025] (3); In the formula, G represents the total output (in ten thousand tons) of crop I within K prefecture-level cities and year J. for The first in the prefecture-level city Annual crops Yield per unit area (kg / ha); The first in K prefecture-level city Annual crops The sown area (ten thousand ha).

[0026] (2) Minimize irrigation water demand. Water resources are a key factor in food production and the healthy development of wetland ecosystems; however, large-scale agricultural irrigation activities have caused many ecological problems, such as declining groundwater levels, shrinking wetland areas, and soil degradation. Therefore, taking effective measures to minimize irrigation water demand and improve water resource utilization efficiency is not only conducive to better protecting black soil resources and enhancing ecosystem service functions, achieving coordinated development of the ecological environment and food production, but also to improving agricultural water use efficiency through the promotion of water-saving irrigation technologies, ensuring the continuous growth and stability of the basin's food production capacity. The formula is as follows:

[0027] (4); In the formula, The total irrigation water demand (in billions of m³) for K prefecture-level cities and I crops in year J. 3 ), For the first The first prefecture-level city Year The water deficit (mm) of a crop is estimated by subtracting the effective precipitation from the crop's water requirement during its growth period.

[0028] (3) Largest wetland area. Wetlands, with their natural functions of both "water storage" and "water release," can buffer drought, mitigate floods, and replenish groundwater, possessing extremely high ecological, economic, and social value. However, large-scale farmland reclamation and irrigation, directly competing with wetlands for water and land, have led to severe wetland shrinkage, reducing the basin's ability to regulate extreme weather events. Therefore, maximizing wetland area not only helps promote the development of new and distinctive industries such as wetland agriculture and ecotourism, but also reduces the damage caused by extreme droughts and floods through the ecological regulation function of wetlands, protecting the stable development of agriculture. The formula is as follows:

[0029] (5); In the formula, The wetland area is measured in ten thousand ha. and The first prefecture-level cities Annual evapotranspiration and precipitation (mm) For the first prefecture-level cities Annual water demand for rice irrigation (100 million m³) 3 ), For the first prefecture-level cities Annual elevation (m). Establishing a relationship between wetland area and decision variables is a challenge. To address this, we first need to use machine learning to construct a wetland area prediction model by selecting factors such as crop planting area, precipitation, evaporation, irrigation crop water requirements, and DEM. Then, we combine the predicted and validated optimal model results with the NSGA-Ⅲ multi-module model to simulate and optimize the synergistic optimization of agricultural development and wetland ecosystem in the Songhua River Basin.

[0030] The constraints of the multi-objective optimization model are explained below: (1) Output constraint. That is, “actual minimum grain output” < “optimized grain output” < “expected maximum target grain output” (Formula 7); the crop area constraint for each sub-region is “historical minimum sown area” ≤ “optimized crop sown area” ≤ “historical maximum sown area” (Formula 6).

[0031] (2) Actual irrigation water demand constraint. That is, “optimized irrigation water demand” ≤ “actual irrigation water demand” (Formula 8); the crop area constraint for each sub-region is “the minimum sown area in previous years” ≤ optimized crop sown area ≤ “the maximum sown area in previous years”.

[0032] (3) Wetland area constraint. That is, “optimized wetland area” ≥ “actual wetland area” (Formula 9); the crop area constraint for each sub-region is “minimum sown area in previous years” ≤ optimized crop sown area ≤ “maximum sown area in previous years”.

[0033] (6); (7); (8); (9); In the formula, and They are respectively the kth prefecture-level city Annual crops Minimum and maximum sown area (ten thousand ha); and These represent the minimum and maximum expected yield (in ten thousand tons) of crop I within K prefecture-level cities and year J. Agricultural water consumption (in 100 million m³) for K prefecture-level cities and J years for I crops 3 ); This represents the actual wetland area (in ten thousand ha) under the current scenario.

[0034] Step 2: Based on the objective function and constraints, the NSGA-Ⅲ algorithm is used to solve the multi-objective optimization model to obtain the Pareto front solution set, and the optimal planting scheme is selected from the Pareto front solution set based on scenario analysis.

[0035] The multi-objective optimization model is solved using NSGA-III, and its computational process is similar to that of NSGA. However, this model is based on multi-module joint simulation and scenario optimization using the NSGA-III algorithm and machine learning. Figure 2 Therefore, the specific calculation steps are as follows: (1) Crop yield module: The total yield under different planting structures is calculated by multiplying the actual yield per unit area by the area.

[0036] (2) Crop coefficient method module: The crop coefficient method is used to obtain the water demand during the crop growth period and then subtract the effective precipitation to simulate the response of crop water consumption to changes in planting structure.

[0037] (3) Neural network module: Select the machine learning model with the highest accuracy in predicting wetland area to simulate the response of wetland area to changes in planting structure.

[0038] (4) Optimization Module: The NSGA-III algorithm is used, calling the three sub-modules mentioned above to continuously approach the goal of less water consumption, higher yield, and larger wetland area. First, 200 initial population individuals containing rice, corn, and soybean are randomly generated. Then, with a crossover probability of 0.8 and a mutation probability of 0.05, population evolution is performed based on non-dominated sorting and crowding distance calculation. By setting three iteration gradients of 100, 200, and 300 times and performing 1-2 rounds of independent computation, multiple Pareto front solutions are finally obtained. The optimal planting scheme is then selected based on scenario analysis.

[0039] In order to determine the potential impact of future agricultural development on watershed water resources and wetland ecosystems, and to explore how to effectively resolve this contradiction by optimizing the planting structure under different scenarios, this invention sets up three scenarios (Table 2): "Water-wetland optimization with food priority", "Water-food-wetland synergistic development", and "Water-food optimization with wetland ecology priority". To explore (1) the extent to which adjusting the planting area of ​​different crops in different years and different sub-regions can achieve synergistic optimization of water resources and wetlands under the condition that food production is maintained (Scenario C1); (2) the extent to which optimizing the planting area of ​​different crops in different years and different sub-regions can achieve synergistic optimization of agriculture and wetlands without increasing irrigation water demand (Scenario C2); and (3) the extent to which optimizing the planting area of ​​different crops in different years and different sub-regions can reduce crop water consumption and maximize the synergistic development of water resources and agricultural production under the condition that wetland area is not reduced (Scenario C3).

[0040] Table 2. Scenario and Constraints for Optimizing Planting Structure in the Songhua River Basin Scenario C0 (Current Status and Development): The current state of planting structure, scale, total grain output, actual water consumption, and spatial and temporal distribution of wetland area, formed between 2000 and 2020 under the combined effects of human intervention and natural conditions. If policy regulation is excluded, and the current trend continues, the scale of crop planting will continue to expand, and the total output and irrigation volume will continue to increase, putting significant pressure on water resources and wetland systems.

[0041] Scenario C1 (Food Supply Priority): The core objective of this scenario is to optimize the spatial and temporal layout of planting structures to minimize irrigation water demand while increasing wetland area compared to Scenario S0, while ensuring that total food production remains unaffected. To this end, Formula 7 will... and Each crop was set as a sub-region within the study period. The total output is 99.99% and 100.01% of the total output, respectively, to ensure that the total grain output corresponding to the optimized planting structure is consistent with the total output under the actual planting structure.

[0042] Scenario C2 (Water-Food-Wetland Coordinated Development): Agricultural irrigation water use in the Songhua River Basin has increased by 7.5 billion cubic meters in the past 20 years. 3 The contribution of irrigation water use to water scarcity increased from 57% in 2000 to 79% in 2017, highlighting the urgent need to control irrigation water consumption and improve agricultural water efficiency to alleviate water shortages in the region. Therefore, the core objective of this scenario is to maximize total grain yield while restoring wetland area (this scenario may involve sacrificing some grain yield for increased wetland area) by optimizing the spatial and temporal layout of planting structures, ensuring that crop irrigation water demand does not increase. Furthermore, it aims to compensate for the decrease in total grain yield by improving the efficiency of grain production per unit of water, i.e., improving agricultural water resource utilization efficiency. In Formula 8... The values ​​are set to 99.99% and 100.01% of the actual irrigation water requirement to ensure that the crop irrigation water requirement corresponding to the optimized planting structure is basically consistent with the crop irrigation water requirement under the actual planting structure.

[0043] Scenario C3 (Prioritizing Wetland Ecological Restoration): The core objective is to maximize total grain yield while maintaining the same level of irrigation water demand by optimizing the spatial and temporal layout of planting structures, ensuring that the wetland area does not decrease across all sub-regions of the watershed at any given time. (In Formula 9) The optimal planting area is set to 99.99% and 100.01% of the actual wetland area to ensure that the wetland area corresponding to the optimized planting structure is basically consistent with the wetland area under the actual planting structure. This scenario, in order to maximize the wetland's ecological function, may sacrifice some crop planting scale, reducing irrigation water demand while also decreasing water yield efficiency per unit volume.

[0044] Step 3: Collect precipitation, evapotranspiration, different crop areas, irrigation water demand, and elevation of the area to be predicted as input variables for the machine learning model, predict the dynamic changes in wetland area, and couple the dynamic changes in wetland area with the optimal planting scheme to obtain the optimal agricultural planting structure adjustment scheme.

[0045] Traditional wetland area prediction methods mainly rely on statistical analysis and empirical models, but these methods have limitations when dealing with complex nonlinear relationships and high-dimensional data. In recent years, intelligent algorithms, represented by deep learning, have demonstrated unique advantages. Their powerful feature extraction and pattern recognition capabilities provide a new technical path for solving the problem of predicting dynamic changes in wetlands. This invention selects variables related to wetland area to construct a machine learning model for wetland area prediction. After verification, the results are coupled with the NSGA-Ⅲ algorithm to obtain a planting structure adjustment scheme. If the prediction shows that the wetland area is evolving in an unsustainable direction, the model parameters need to be adjusted or the scenario selection needs to be re-examined until the obtained scheme can meet the yield and water conservation targets while ensuring that the wetland ecosystem maintains stable or healthy development in the foreseeable future.

[0046] Model Input Variable Selection: In the variable selection stage, considering data availability and some dominant factors, seven variables were selected as input data from 2000 to 2020 for 20 prefecture-level cities in the Songhua River Basin: precipitation, evapotranspiration, rice, corn, and soybean planting area, rice irrigation water requirement, and DEM (Dual Earth Scale). Specifically: 1) Precipitation is one of the key natural factors affecting wetland area, directly influencing water level and volume, and thus wetland area. Existing studies have shown that increased precipitation typically leads to an expansion of wetland area, especially in inland wetlands and marsh wetlands. Studies in the Sanjiang Plain of China have also shown that topography and urbanization significantly affected wetland distribution between 1990 and 2000, while changes in precipitation played a dominant role between 2010 and 2020, and can indirectly affect wetland distribution by influencing topography and soil.

[0047] 2) Evapotranspiration is essentially a crucial component of the surface water cycle, encompassing two key mechanisms: first, the conversion of liquid water on the soil and vegetation surface into gaseous water; and second, physiological water transpiration by plants through stomata. The total amount of this complex process is evapotranspiration (ET), and its numerical characteristics reflect the intensity of regional water and heat exchange. Increased evapotranspiration leads to a decrease in wetland moisture, thus affecting wetland area, especially in arid and semi-arid regions where the impact of evapotranspiration on wetland area is particularly significant. Furthermore, the presence of wetland vegetation raises the question of whether its evapotranspiration increases or decreases wetland evapotranspiration, a matter of debate.

[0048] 3) The planting area of ​​rice, corn, and soybeans in the Songhua River Basin increased from 71.49% in 2000 to 97.28% in 2020. Besides improvements in agricultural policies and mechanization, meeting the increasing food demand of the population was a major contributing factor. In addition to the continuous reclamation of unused land resources, marsh wetlands with favorable hydrothermal conditions and rich soil organic matter have become the preferred sites for cultivation, leading to a decline in wetland ecological functions and a weakening of their flood control capacity, thus affecting grain yield. Therefore, changes in the planting area of ​​rice, corn, and soybeans directly affect the area and distribution of wetlands. Studies show that from 1990 to 2020, the area of ​​paddy fields in the Sanjiang Plain increased by 12,995.73 km², while the area of ​​marsh wetlands decreased by 1,031.9 km².

[0049] 4) Rice is the main irrigated crop in the Songhua River Basin, and its irrigation water demand directly affects the water resources of wetlands, significantly impacting the wetland system. Besides occupying wetland area, there is also the issue of competition with wetlands for water resources. Increased irrigation water demand may lead to a reduction in wetland area, especially accelerating wetland shrinkage when irrigation water is insufficient. This phenomenon is particularly pronounced in Northeast China. The construction of water conservancy projects such as the "Southern Diversion," "Central Diversion," and "Northern Diversion" in the Nenjiang River Basin for farmland irrigation has resulted in the interception or diversion of large amounts of water, leading to a continuous decrease in water inflow to downstream wetlands and lakes, causing a continuous shrinkage of wetland area.

[0050] 5) The DEM provides topographic information, including elevation, slope, and aspect. These topographic factors have a significant impact on the distribution and area of ​​wetlands. Low-lying areas are more prone to water accumulation and wetland formation, playing a major role in flood regulation, water purification, and fishery resources. Higher-elevation areas have fewer wetlands due to limited rainfall, but they are water conservation areas and play an important role in soil and water conservation and biodiversity protection. Existing research has also confirmed that topographic factors can be used to predict wetland distribution. Based on the statistics of wetland area and distribution in the Songhua River Basin in the previous chapters, 79% of the wetlands are located in the Songnen Plain area (elevation less than 200 m) including Baicheng City, Daqing City, Harbin City, Jilin City, Qiqihar City, Songyuan City, and Suihua City, as well as Jiamusi City in the Sanjiang Plain.

[0051] To improve the accuracy of wetland area prediction by the selected variables, the maximum and minimum wetland areas are introduced as constraints and run together with the loss function to continuously improve the prediction accuracy.

[0052] Model selection and training parameter tuning: 1) Long Short-Term Memory (LSTM) networks are a neural network method for predicting long-term data sequences. Unlike traditional RNN models, they employ a unique gating mechanism to address information loss as time progresses. The underlying logic involves three functional gating units: the forget gate, which is based on the hidden state from the previous time step. and current input The sigmoid activation function is used to determine which historical information needs to be discarded; the input gate also utilizes... and The update coefficients are calculated, and the tanh function is used to control which new information can be stored; the output gate synthesizes the current memory state. These three gating units, along with the input information, control the final output information. Through coordinated operation, the LSTM extracts key information in a timely manner during the learning, training, and prediction processes, thereby improving prediction accuracy. Its mathematical expression can be represented as:

[0053] Forgotten Gate: (10); Input Gate: (11); Candidate state: (12); Cell status update: (13); Output gate: (14); Hidden status update: (15); In the formula, σ is the sigmoid function, and tanh is the hyperbolic tangent function. and These are the forget gate weights and biases, respectively. For information that needs to be memorized; These are candidate memory units, used in the model to update memory units; and For input gate weights; and For input gate bias; For the output gate weights, The output gate is biased. The LSTM modeling process is as follows: Figure 3 As shown, the parameters are initialized and the input data is preprocessed before being input into the LSTM network. The error threshold and the number of iterations are then determined. If the threshold is not met, the number of iterations is increased. If the threshold is met, the training is considered complete. If the threshold is not met, Epoch is incremented by 1. Otherwise, the output is predicted directly.

[0054] The number of hidden layer neurons in an LSTM network directly affects the model's learning ability: too many neurons may cause the model to over-memorize detailed features of the training data and lose its generalization ability; too few neurons may prevent the model from accurately acquiring key features. Furthermore, optimizing parameters such as learning rate and batch size is equally crucial: an appropriate learning rate can balance convergence speed and stability, while a reasonable batch size affects the accuracy of gradient estimation. By systematically adjusting these parameter combinations, not only can model convergence be accelerated, but the final prediction accuracy can also be significantly improved. Therefore, this invention uses half the dimension of the input precipitation, evapotranspiration, rice, corn, and soybean planting area, rice irrigation water requirement, and DEM data as the number of hidden layer neurons, sets the gradient threshold to 2, and sets the learning rate to 0.1. After 1000 iterations, wetland area prediction is performed, and 75% of the input data is used as training samples, while the remaining 25% is used as real samples to verify the prediction results.

[0055] 2) Back propagation (BP) algorithm is widely used in pattern recognition, data classification, function approximation, prediction, and other fields. By simulating the information processing methods of the human nervous system, it can automatically learn patterns from large amounts of data, thereby accurately processing and predicting new data. The topology of a BP neural network typically includes an input layer, one or more hidden layers, and an output layer (…). Figure 4 During network operation, data is first passed from the input layer through the hidden layers, with each neuron performing a weighted summation of the input signal and transforming it using an activation function. Then, by comparing the network output with the expected value, the connection weights of each layer are adjusted in reverse. This iterative process of forward computation and error backpropagation enables the network to automatically extract data features and optimize prediction performance. Assuming a BP network has a set number of input nodes, hidden layer nodes, and output nodes, its learning process can be represented as follows:

[0056] (16); (17); in, This represents the output of the l-th layer. Indicates the first The weight matrix of the layer, Indicates the first The bias vector of layer l Indicates the activation function; This represents the actual output of the k-th node in the output layer. This indicates the expected output.

[0057] If the output deviates significantly from the expected result, backpropagation correction is initiated. Backpropagation first calculates the weight adjustment based on the error signal from the output layer, then propagates these error signals back to the hidden layers. Each neuron adjusts its connection weights based on the received error signal. This progressive error correction method ensures that the network can improve prediction accuracy by continuously adjusting its internal parameters. The formulas for calculating the error gradients of the output and hidden layers are as follows:

[0058] (18); (19); In the formula This represents the error gradient of the output layer. t represents the actual output, and t represents the expected output. This represents the derivative of the activation function of the output layer; Indicates the first The error gradient of the layer, Indicates the first The weight matrix of the layer, Indicates the first The error gradient of the layer, Indicates the first The derivative of the layer activation function.

[0059] The weights and biases are adjusted using the gradient descent method to minimize the error function, where... Indicates the learning rate: (20); (twenty one); Training and parameter tuning of the BP neural network model is a complex but systematic process. By scientifically configuring parameters such as learning rate, momentum term, weight decay, activation function, and batch size, the prediction accuracy of the network can be effectively improved. After multiple parameter optimization experiments, the network was finally determined to contain two fully connected layers with 64 and 32 nodes respectively. To prevent overfitting, a Dropout regularization layer was added to the network, with a neuron dropout rate of 30%. Model training adopted an iterative scheme of 500 epochs, with the training data randomly rearranged at the beginning of each epoch to enhance generalization ability. To monitor the training process in real time, the system dynamically plotted visualizations such as loss function change curves. The validation strategy was set to perform a validation test every 100 iterations. The learning rate adopted a dynamic adjustment mechanism: the initial value was set to 0.001, and a step decay strategy was used, reducing the learning rate by a decay coefficient of 0.01 after every 100 epochs.

[0060] 3) Random Forest (RF) improves predictive ability by constructing multi-layered binary trees to progressively filter sample features. Compared to a single decision tree, RF uses Bagging to continuously extract and replace subsets of samples from the original dataset to reduce model variance, filters subset features at the branching points of each tree to enhance model multidimensionality, and finally selects the optimal splitting feature using the Gini index to ensure that each tree captures the key features of the data. Figure 5 As shown, these decision trees trained with differences integrate prediction results through voting or averaging mechanisms, significantly improving prediction accuracy while maintaining good interpretability. The formula for calculating the Gini index is:

[0061] (twenty two); (twenty three); In the formula Represents a node. Indicates from The probability that a randomly selected sample belongs to class c. and They are respectively The physical meaning of the Gini index, a subset of features, lies in the fact that it quantifies feature-based... The degree of purity improvement of the two child node sets after partitioning is indicated by a smaller value, which means that the purity of the partitioned subsets is higher.

[0062] This invention establishes a predictive model (RF) based on parameters such as precipitation, evapotranspiration, rice, corn, and soybean planting area, rice irrigation water requirement, and DEM (Dense Earth Surface) and wetland area. During model training, the input features are first normalized to a range of [-1, 1]. Then, 70% of the samples are randomly selected as training data for the predictor variables, and the remaining 30% are used as test data to verify the accuracy of the prediction results. Finally, the RF model prediction results are evaluated using root mean square error and R². 2 Visualize it.

[0063] This invention employs three machine learning algorithms—LSTM, BP neural network, and random forest—to construct a prediction model. To ensure the model's prediction accuracy, a two-stage "training-validation" modeling strategy is adopted. The original dataset is divided into training and test sets in a 7:3 or 8:2 ratio. Key parameters that improve model accuracy are selected, and different threshold combinations are set to prevent overfitting. For example, the number of hidden layer nodes and dropout rate of LSTM, the learning rate and momentum coefficient of BP neural network, and the number of decision trees and feature subset size of random forest are selected. These parameters are iterated several times to obtain the optimal fitting result. All algorithms are implemented in MATLAB, and the results are evaluated using root mean square error (RMSE) and coefficient of determination (R²). RMSE reflects the absolute magnitude of the prediction error, and R² characterizes the model's ability to explain data variation. The smaller the RMSE and the closer R² is to 1, the better the model's prediction performance. The formulas for the specific metrics are as follows:

[0064] (twenty four); (25); In the formula, To input the total number of samples, The mean of all input samples. These are predicted values.

[0065] While traditional data partitioning methods can initially test a model's predictive ability, the fixed proportion of the test set is susceptible to the randomness of data distribution, potentially leading to significant fluctuations in evaluation results. This invention employs k-fold cross-validation to further enhance the applicability of model evaluation. k-fold cross-validation mainly includes the following steps: 1) First, the complete dataset is randomly divided into k equal-sized sub-sample groups (usually k=5 or 10), and then k rounds of iterative training and validation are performed. 2) In each iteration, a fixed number of sub-sample groups are randomly selected as the validation set, and the remaining k-1 groups are used as the training set. 3) Through this rotating validation mechanism, all data can participate in training and validation, and the average of the k validation results is taken as the overall performance index of the model. Compared to traditional single-partition methods, k-fold cross-validation has significant advantages. It utilizes all samples in the dataset through repeated cross-validation, significantly reducing the model evaluation distortion caused by single data partitioning, thus making the model performance more statistically robust. The selection of the fold number k requires a balance between computational efficiency and estimation accuracy. Usually, k=5 or k=10 is used as an empirical value. In this invention, the dataset is relatively small, so k=10 is selected to reduce the variance of the validation results.

[0066] 2. Comparative Study of Algorithms for Dynamic Simulation of Wetland Area: This invention evaluates the applicability and accuracy of different machine learning models—LSTM, BP neural network, and RF—in the field of wetland ecological prediction by comparing their predictive capabilities for wetland area. Figure 6 The study presents a comparison between the prediction results of each model and the actual values ​​under traditional data partitioning, as well as a goodness-of-fit analysis of the models.

[0067] The results show that the predicted and actual values ​​of the LSTM model are generally consistent in trend, demonstrating reliability in predicting the long-term trend of wetland area change. However, deviations exist at certain specific times. Combining the fitting results between the predicted and actual values, it can be concluded that R0 2 The RMSE was 287.37, which is 0.85. Although the model can explain about 85% of the data variability, the RMSE value is relatively high, indicating a large prediction error. Figure 6 (c) and (d)). Similarly, RF predictions also showed similar results to LSTM, but its RMSE was higher than that of the LSTM model, at 246.32 ( Figure 6 (e) and (f)). Compared to LSTM and RF, the BP neural network model outperforms LSTM and RF models in predicting wetland area, with an R2 value of 0.97 and the lowest RMSE of 205.56. Furthermore, the predicted and actual values ​​show a high degree of overlap, demonstrating its ability to effectively capture local changes in wetland area. Figure 6 (a) and (b)). K-fold cross-validation also yielded similar results (Table 3). Based on this advantage, the study integrated the optimized BP neural network with the NSGA-Ⅲ multi-objective optimization algorithm to construct a decision support system for sustainable development in the Songhua River Basin. This system aims to achieve coordinated, stable, and sustainable development across various sectors under scenarios involving adjustments to the planting structure, increased and stable grain production, no reduction in water resource supply, and no further shrinkage of wetlands and damage to ecological functions in the Songhua River Basin.

[0068] Table 3 Performance metrics for k-fold validation of machine learning models Figure 7This study presents the interannual variation results of simulations for the synergistic optimization of agricultural development and wetland ecosystems, revealing the trends in crop planting area, yield, irrigation water demand, and wetland area under different scenarios. Except for scenario C3, C1 and C2 generally show an increase in the planting proportion of water-intensive rice in wet years, and a decrease in the rice planting proportion while appropriately increasing the planting proportion of corn and soybeans in normal and dry years to maintain stable yields. The interannual variations in irrigation water demand and yield are largely consistent with the trends in planting area. Scenario C3, prioritizing wetland ecology, shows relatively small changes in crop planting structure, yield, and irrigation water demand compared to the current scenario C0. However, due to the doubling of rice planting proportion and yield after 2010, the model optimization process began to reduce the rice planting proportion to control its excessive water consumption and the resulting wetland area shrinkage.

[0069] Figure 7 (a), (b), and (c) show the differences in wetland area relative to scenario C0 under scenarios C1, C2, and C3, and the changes in wetland area are directly affected by crop planting area. Figure 7 The data shows that under scenario C1, the rice planting area increased by 42,600 hectares in 2015, leading to an increase in irrigation water demand of 72 million m³, while yield also increased significantly to 360,800 tons. This increase, to some extent, creates competition for water and land with wetlands. However, due to a decrease of 3,332,800 hectares in maize planting area, the demand for water resources decreased significantly, providing space for wetland recovery, resulting in an increase of 97,100 hectares in wetland area. In 2019, the rice planting area decreased the most, by -743,400 hectares, leading to a drop in yield to the lowest value between 2000 and 2020, -4,887,100 tons. However, the increase in the planting area of ​​maize and soybeans, which have relatively low water consumption, compensated for the loss in total yield, keeping the watershed yield relatively stable. This adjustment in planting structure contributed to the recovery of wetland area, with an increase of 118,700 hectares. By analogy, the interannual changes in irrigation water demand, yield, and wetland area corresponding to crop planting structure adjustments in other scenarios all exhibit the above patterns.

[0070] The optimization results for each scenario, adjusted for interannual crop planting area, reduced irrigation extraction to varying degrees, thereby improving water yield efficiency per unit volume and driving wetland area recovery. Figure 7 As shown in (a), (b), and (c), the average wetland area of ​​the watersheds under scenarios C1, C2, and C3 increased by 72,200 ha, 72,200 ha, and 75,400 ha, respectively, compared to scenario C0, at the end of the study period (2020). This is because the model, during its optimization process, increased the proportion of crop planting area in wet years and decreased the proportion in dry years. Figure 7(d), (e), and (f)) increased the proportion of crops with higher yield efficiency per unit area of ​​aquatic water, achieving the goals of wetland area restoration and grain yield increase to varying degrees. For example, compared with scenario C0, scenario C1 increased the average annual sown area of ​​maize by 268,700 ha / year in wet years, and increased the average annual sown area of ​​rice and soybean by 102,900 ha / year and 115,300 ha / year in normal years, respectively, while decreasing the average annual sown area of ​​rice and soybean by 64,100 ha / year and 80,500 ha / year in dry years, respectively; while scenario C2 decreased the average annual sown area of ​​crops by 542,800 ha / year, 1,176,500 ha / year, and 606,900 ha / year in wet, normal, and dry years, respectively; and scenario C3 decreased by 392,400 ha / year, 1,631,800 ha / year, and 597,300 ha / year under different hydrological year types, respectively.

[0071] Adjusting planting structures to achieve water conservation may result in a sacrifice of grain yield, such as... Figure 7 As shown in (g), (h), and (i), scenario C1 achieves an annual water saving of 680 million m³ at the cost of a 1% yield loss, while scenarios C2 and C3 achieve even greater water savings of 1.075 billion m³ and 418 million m³, respectively, despite yield losses of up to 19%. Taking scenario C1 as an example, the optimization results alleviate water resource pressure and restore wetland area by reducing the proportion of irrigated crops and increasing the proportion of rainfed crops. The proportion of rice cultivation decreased by 4%, but its demand for irrigation water also decreased by 3.6%. Furthermore, due to the influence of factors such as climate, crop unit price, and cost, even with the same adjustment ratio of crop planting area, the changes in yield, irrigation water demand, and wetland area in different years show significant differences. For example, in Scenario C2, the rice acreage was reduced by 152,600 ha and 150,100 ha in 2010 and 2016 (dry and wet years), respectively, but the yield decreased by 950,900 t and 621,700 t, respectively. Irrigation water demand was reduced by 393 million m³ and 184 million m³, respectively. Wetland area recovered by 68,500 ha and 101,700 ha, respectively, and the grain yield per unit area of ​​water increased by 2.4 kg / m³. 3 and 6.1 kg / m 3 .

[0072] By optimizing crop planting area, irrigation methods, and crop structure, the dual goals of agricultural production and ecological protection can be achieved under different scenarios. In each scenario, by rationally adjusting the planting area of ​​different crops, grain yields were increased while effectively reducing dependence on water resources and promoting wetland restoration. This optimization scheme achieved dual benefits: on the one hand, it significantly improved the sustainable operation capacity of the agricultural production system; on the other hand, it established an effective guarantee mechanism for regional ecological security. In subsequent promotion, a differentiated implementation strategy is recommended: dynamically adjust the crop spatial allocation scheme according to the water resource endowment characteristics and ecological sensitivity of each region, and maximize the synergistic benefits of the three major benefits (water resources-wetland-food) through precise regulation.

[0073] Spatial Differentiation Characteristics of Crop Layout Optimization under Multiple Scenario Simulations: Due to significant differences in geographical location and resource conditions across regions, adjustments to planting structures need to be made according to local conditions. The following section presents the spatial variation results of scenario simulations for the synergistic optimization of agricultural development and wetland ecosystems, revealing the spatial differences in crop planting area, yield, irrigation water requirement, and wetland area under different scenarios compared to the current scenario. Figure 8 It can be seen that the optimized model spatially adjusts the crop irrigation water demand corresponding to prefecture-level cities with significant wetland area reductions in historical data, making the spatial distribution of wetland shrinkage more balanced. Taking scenario C1 as an example, the model reduces the proportion of rice cultivation in severely wetland-shrinking areas of the Songhua River basin plain in Baicheng, Daqing, Qiqihar, Harbin, Suihua, and Jiamusi cities (reducing by an average of 12,000 ha, 1,900 ha, 16,300 ha, 9,400 ha, 5,800 ha, and 45,400 ha respectively over many years) and irrigation water demand (reducing by an average of 27 million m³ respectively over many years). 3 0.13 billion m 3 0.73 billion m 3 0.40 billion m 3 0.18 billion m 3 and 0.16 billion m 3 The corresponding wetland areas restored were 37,800 ha, 24,100 ha, 14,500 ha, 13,200 ha, 11,400 ha and 6,000 ha, respectively. Figure 8 (a), (d), and (j)); while Changchun, Mudanjiang, and Heihe maintain a balance between agricultural development and ecological and social benefits by increasing the proportion of rice cultivation and reducing the proportion of corn and soybean cultivation. This adjustment plan helps to achieve synergistic optimization between agricultural development and wetland ecosystems, as well as more rational allocation and utilization of water resources in the basin, ensuring the sustainability of agricultural production while protecting the ecological environment.

[0074] Since there was no data on the actual sown area of ​​rice, corn, and soybeans in Hailar and Ulanhot, these two prefecture-level cities were not considered in the optimization process. The spatial variation trend of rice irrigation area in scenarios C2 and C3 is consistent with that in C1, but the adjustment range differs due to different objective function constraints. In scenarios C2 and C3, to control irrigation amount and wetland area to be less than C0, the model optimization results reduced the planting ratio of irrigated crops in all prefecture-level cities, and further reduced it on the basis of the reduction in planting ratio in scenario C1. This resulted in a decrease in yield of 88,600 tons, 25,300 tons, 227,800 tons, 207,700 tons, 120,400 tons, and 340,600 tons in Baicheng, Daqing, Qiqihar, Harbin, Suihua, and Jiamusi in the two scenarios, respectively. Except for the increased planting ratio of corn in Heihe City and soybeans in Tonghua City in scenario C2, the planting ratio of these two crops decreased to varying degrees in other prefecture-level cities and in scenario C3. Figure 8 (g), (h), (i), (j), (k), and (l)).

[0075] In optimization scenarios C1, C2, and C3, the spatial distribution trends of irrigation volume, crop yield, and wetland area in each sub-region generally correspond to the changes in irrigated area. However, due to significant differences in irrigation water demand, total grain yield, and wetland area among the different scenarios, the magnitude of these changes varies considerably. Particularly in scenario C3, the efficiency of grain production per unit area of ​​water production continues to decline as irrigated area and irrigation water consumption decrease significantly. This phenomenon indicates that without advanced irrigation technologies, the restoration of wetland area will inevitably lead to a decrease in grain yield, thereby threatening the strategic position of the Songhua River Basin as a major grain-producing area in my country and posing challenges to sustainable agricultural development. Overall, scenarios C1, C2, and C3, by adjusting crop planting structure and scale, have alleviated the agricultural and ecological contradictions in areas with severe wetland shrinkage to some extent. In areas with relatively abundant water and wetland resources, a relative balance in yield is achieved by increasing rice planting area and reducing the proportion of corn and soybean planting.

[0076] Optimization and Regulation Strategies for the Water-Wetland-Grain System under Different Hydrological Year Types: Due to geographical differences, the precipitation distribution in various prefecture-level cities in the Songhua River Basin exhibits significant spatial heterogeneity. The Songnen Plain in the western part of the basin, as the main grain crop planting area, suffers from relatively scarce precipitation, while the southeastern part, although rich in precipitation, is limited by insufficient arable land. To address the spatial mismatch of water and soil resources in the study area, this invention constructs an adaptive planting optimization system based on differences in hydrological year types. Based on the spatiotemporal variation characteristics of regional precipitation, crop configuration schemes for three typical hydrological year types (high-water year, normal-water year, and low-water year) were designed. By setting a baseline scenario (C0) and three optimized scenarios (C1-C3), the system quantifies the changes in irrigation area of ​​major crops in each administrative region under different scenarios, calculates the average irrigation area for each scenario under the three hydrological year types, and further derives the adjustment ratios of scenarios C1, C2, and C3 relative to scenario C0. The impact of implementing the optimization scheme on total yield, irrigation water demand, and wetland area is shown in Table 6.

[0077] Studies show that, except for the C1 scenario (a normal water year) which increases the irrigated rice area by 103,000 hectares (6% of the irrigated rice area in the C0 scenario), resulting in a yield increase of 77,000 tons and an increase in irrigation water demand of 10 million m³, the irrigated rice area under other hydrological year types all show varying degrees of reduction. To alleviate the yield decline caused by the reduction in irrigated rice area and soybean planting area, the C1 scenario increases the maize planting area by 478,000 hectares in wet years and 69,000 hectares in dry years (3% and 1% of the maize planting area in the C0 scenario, respectively). The optimized scheme can effectively increase the maize planting area and yield in wet and normal water years, but in dry years, the crop planting area and yield generally decrease, and the wetland restoration area is the lowest among the three scenarios. Analysis shows that the optimized scheme has significant differences in its impact on planting structure, yield, and water resource utilization efficiency under different hydrological year types. Further optimization of the planting structure in various prefecture-level cities is needed, taking into account regional water resource distribution characteristics and ecological environment needs, to achieve sustainable agricultural development.

[0078] Table 6. Adjustment plans for crop sown area, yield, irrigation water requirement, and wetland area under different hydrological year types. Note: Crop planting area (10,000 hectares), crop yield (10,000 tons), crop irrigation water requirement (100 million cubic meters) 3 ), wetland area (ten thousand ha).

[0079] Figure 9The study showcases optimized planting structures under different hydrological year types. Under high-water conditions (annual precipitation greater than 577 mm), in scenario C1, 44% of prefecture-level cities showed an increasing trend in irrigated rice area compared to scenario C0. Among these, Baishan, Harbin, Heihe, Jilin, and Changchun saw increases exceeding 2%, while Baicheng, Hegang, Jiamusi, and Qiqihar experienced decreases exceeding 10%. In optimizing maize planting structures, 50% of prefecture-level cities showed an increasing trend, with Heihe and Suihua showing the highest increases at 10%. Meanwhile, Jilin, Qiqihar, Songyuan, and Changchun, located in the core area of ​​the Songnen Plain, all achieved significant increases exceeding 5%. Although the proportion of soybean planting is relatively low, the adjusted area is significant. The increases in Harbin, Liaoyuan, Qitaihe, and Jiamusi were 19.16%, 15.01%, 10.08%, and 7.83%, respectively, while the decreases in Daqing, Liaoyuan, and Qitaihe were 19.16%, 15.01%, and 10.08%, respectively. Under scenario C2, due to strict control over irrigation water demand, except for a few prefecture-level cities where the planting area of ​​rice, corn, and soybeans increased, over 67% of prefecture-level cities saw reductions to varying degrees, with the rate of reduction decreasing from the northwest to the southeast of the basin. Soybeans saw the largest reduction, followed by corn, while rice saw the smallest reduction. Scenario C3 is similar to C2, but the overall reduction is smaller. The above analysis shows that the optimization schemes for planting structure under different scenarios exhibit significant differences in spatial distribution. Particularly under wet year conditions, scenario C1 significantly adjusts the planting area of ​​rice and maize, while scenarios C2 and C3 focus more on controlling irrigation water demand, leading to a general reduction in planting area. These results provide important references for regional agricultural water resource management and planting structure optimization.

[0080] Figure 9Figures (d), (e), and (f) illustrate the spatial distribution characteristics of planting structure optimization under scenarios C1, C2, and C3 under normal water year conditions (annual precipitation between 527 mm and 577 mm). In scenario C1, the adjustment ratios for rice, maize, and soybeans increased several times compared to wet year conditions. Over 61% of prefecture-level cities saw a reduction in rice irrigated area, with the most significant reductions occurring in Jiamusi, Qitaihe, Yichun, Harbin, Baicheng, and Songyuan in the Northeast-Southwest region, ranging from 20.32% to 48.55%. In contrast, Heihe, Qiqihar, and Daqing in the Nenjiang River basin, and Baishan and Tonghua in the Second Songhua River basin, showed slight increases. In maize planting structure optimization, except for slight increases in Suihua and Yichun, all other prefecture-level cities experienced varying degrees of decrease. To achieve a balanced planting structure, soybean adjustments followed the opposite trend to rice and maize, with over 61% of prefecture-level cities showing increases. In scenarios C2 and C3, the adjustment proportions for rice, corn, and soybeans were relatively small. The small adjustment proportions for rice, corn, and soybeans under scenarios C2 and C3 indicate that these two scenarios result in relatively mild adjustments to the planting structure under normal water conditions. Overall, scenario C1 shows a larger adjustment to the planting structure under normal water conditions, particularly in the reduction of rice and corn planting areas, while scenarios C2 and C3 tend to maintain a relatively stable planting structure. Figure 9 (g), (h), and (i) represent the spatial distribution characteristics of crop planting structure optimization under scenarios C1, C2, and C3 at the dry year level (precipitation below 527 mm). Under dry year conditions, the crop planting ratio in all prefecture-level cities generally shows a decreasing trend. However, because Hegang, Liaoyuan, Mudanjiang, Yichun, and Qitaihe are located in mountainous areas with high altitudes and belong to water conservation areas, the model prioritizes yield maximization in the optimization process of scenarios C1 and C2, without fully incorporating the constraints of arable land resources, resulting in a slight increase in the crop planting ratio in these prefecture-level cities. In contrast, the crop planting ratio in all prefecture-level cities remains unchanged or shows a decreasing trend in scenario C3. Overall, scenarios C1 and C2 have certain limitations in adjusting the crop planting ratio in water conservation areas under dry year conditions. Therefore, it is extremely important to comprehensively consider yield targets and arable land resource constraints under different hydrological year types.

[0081] A dynamic response prediction model for wetland area changes in planting structure was constructed based on three machine learning models: LSTM, RF, and BP neural networks. Through comparative analysis, the model with the highest prediction accuracy was selected and integrated with the NSGA-Ⅲ multi-objective optimization algorithm, achieving systematic optimization of agricultural development and wetland ecosystem in the Songhua River Basin. Specifically, the trained BP neural network model was encapsulated as a callable module and embedded into the NSGA-Ⅲ optimization framework along with crop coefficient and crop yield modules, constructing a closed-loop system based on "prediction-optimization-decision". Based on this, by setting up multi-scenario simulation schemes, the crop planting structure of each sub-region under different hydrological year types was optimized and adjusted in different years. The main research conclusions are as follows:

[0082] (1) Based on the k-fold cross-validation results, LSTM, RF, and BP neural networks all showed good reliability in predicting the long-term trend of wetland area change. However, the LSTM model had a determination coefficient of 0.74 and a root mean square error of 283.04, while the random forest model performed slightly better. 2 The R² value was 0.88, and the RMSE was 194.21, both inferior to the BP neural network (R²=0.91, RMSE=163.21). This indicates that the BP neural network has a significant advantage in capturing local changes in wetland area, with more outstanding accuracy and stability in its prediction results. Therefore, this invention, by coupling the BP neural network and the NSGA-Ⅲ algorithm, provides reliable technical support for adjusting the planting structure, stabilizing and increasing grain production, promoting sustainable water resource utilization, and restoring wetland ecosystems in the Songhua River Basin.

[0083] (2) Compared with the current scenario (C0), the food supply priority scenario, under the constraint of maintaining the total grain output basically unchanged, reduces the annual irrigation water demand of the basin by 680 million m³ by optimizing the planting ratio of rice, corn and soybeans. 3 The efficiency of grain production per unit volume of aquatic organisms increased by 0.41 kg / m³. 3 This effectively alleviated the contradiction between water supply and demand. At the same time, the wetland area increased from 694,400 ha / year in the baseline scenario to 791,600 ha / year, a net increase of 97,200 ha / year, or 14%, effectively restoring the function of the wetland ecosystem.

[0084] (3) Under the constraint of not increasing irrigation water demand, the water-grain-wetland coordinated development scenario achieves a 0.32 kg / m² increase in grain yield per unit area by reducing the planting area of ​​rice, corn, and soybeans, sacrificing the total yield of rice, corn, and soybeans (a decrease of 1.19, 1.56, and 410,000 tons per year, respectively). 3 The result showed that wetland area recovery reached 96,800 hectares per year. This scenario indicates that, under water-saving conditions, a certain amount of food production must be sacrificed in exchange for stable ecological benefits.

[0085] (4) Wetland Ecological Priority Scenario: Under the constraint of maintaining the wetland area without reduction, the dual goals of water conservation and ecological protection are achieved through the optimization and adjustment of planting structure. Under this scenario, the annual yields of rice, corn, and soybeans decrease by 940,000 tons, 2.61 million tons, and 400,000 tons, respectively, and the total water consumption of the basin decreases by 418 million cubic meters. 3 Although the yield per unit of aquatic grain decreased by 0.12 kg / m³ 3 However, the wetland ecosystem function was maintained.

[0086] (5) In view of the uneven distribution of water resources and arable land resources in the basin, this invention proposes a planting structure optimization scheme adapted to different hydrological year types based on the spatiotemporal distribution patterns of precipitation in different prefecture-level cities. By quantitatively analyzing the dynamic changes of crop sowing area in each zone under three hydrological scenarios of abundant water year, normal water year, and dry water year, the adjustment ratios and spatial differentiation characteristics of yield, irrigation water demand, and wetland area are clarified, providing a scientific basis for formulating differentiated planting structure adjustment strategies.

[0087] This invention combines machine learning algorithms with multi-objective optimization models to construct an integrated "prediction-optimization-decision" research framework, providing new methodological support for coordinating agricultural production and ecological protection in watersheds. The research results not only enrich the theoretical system of optimal allocation of agricultural water resources but also provide a practical path for the sustainable development of similar watersheds.

[0088] Based on the above description, this invention proposes a water-wetland-food synergistic optimization system, comprising: The model building module constructs a multi-objective optimization model with crop planting area as the decision variable. The objective functions of the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area. The constraints of the multi-objective optimization model include crop yield constraints, crop irrigation water demand constraints, and wetland area constraints. The multi-objective solution module solves the multi-objective optimization model using the NSGA-III algorithm based on the objective function and constraints to obtain the Pareto front solution set. Based on scenario analysis, the optimal planting scheme is selected from the Pareto front solution set. Scenario analysis represents the judgment of the ecological state of the planting structure under different scenarios. The collaborative optimization module collects precipitation, evapotranspiration, different crop areas, irrigation water demand, and elevation of the area to be predicted as the data to be predicted, predicts the dynamic changes of wetland area, and couples the dynamic changes of wetland area with the optimal planting scheme to obtain the optimal agricultural planting structure adjustment scheme.

[0089] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of a water-wetland-grain co-optimization method.

[0090] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0091] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a water-wetland-grain co-optimization method.

[0092] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0093] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for synergistic optimization of water-wetland-food resources, characterized in that, Includes the following steps: A multi-objective optimization model is constructed with crop planting area as the decision variable. The optimization objectives of the objective function in the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area. The constraints of the multi-objective optimization model include crop yield constraint, crop irrigation water demand constraint, and wetland area constraint. Based on the objective function and constraints, the multi-objective optimization model is solved using the NSGA-Ⅲ algorithm to obtain the Pareto front solution set, and the optimal planting scheme is selected from the Pareto front solution set based on scenario analysis. The scenario analysis refers to the judgment of the ecological status of the planting structure under different scenarios; The precipitation, evapotranspiration, area of ​​different crops, irrigation water demand and elevation of the area to be predicted are collected as the data to be predicted. The dynamic changes of wetland area are predicted, and the dynamic changes of wetland area are coupled with the optimal planting scheme to obtain the optimal agricultural planting structure adjustment scheme.

2. The water-wetland-food synergistic optimization method as described in claim 1, characterized in that, The objective functions in the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area, specifically: The objective function for maximizing grain output is expressed as: ； In the formula, G for K Within each prefecture-level city, J Within the year I Total yield of crops; for The first in the prefecture-level city Annual crops Yield per unit area; For the first k The first in the prefecture-level city Annual crops The sown area; The objective function for minimizing irrigation water demand is expressed as: ； In the formula, for K each prefecture-level city J Year I Total irrigation water requirement for crops For the first The first prefecture-level city Year The water deficit of crops is estimated by subtracting the effective precipitation from the water requirement during the crop's growth period. The objective function for maximizing wetland area is expressed as: ； In the formula, For wetland area, and The first prefecture-level cities Annual evapotranspiration and precipitation For the first prefecture-level cities Annual water demand for rice irrigation For the first prefecture-level cities Annual elevation.

3. The water-wetland-food synergistic optimization method as described in claim 1, characterized in that, The constraints of the multi-objective optimization model include crop yield constraints, crop irrigation water requirement constraints, and wetland area constraints, specifically: The specific expression for the crop yield constraint is: ； The specific expression for crop irrigation water requirement constraints is as follows: ； The specific expression for the wetland area constraint is: ； In the formula, and The first k The first in the prefecture-level city Annual crops Minimum and maximum sowing area; for K each prefecture-level city J Year I Agricultural water consumption for crop cultivation; This represents the actual wetland area under the current scenario.

4. The water-wetland-food synergistic optimization method as described in claim 1, characterized in that, The process involves using the NSGA-III algorithm to solve the multi-objective optimization model, obtaining the Pareto front solution set, and then selecting the optimal planting scheme from the Pareto front solution set based on scenario analysis. Specifically: Randomly generate an initial population containing various grain planting areas; Set crossover and mutation probabilities, and perform population evolution based on non-dominated sorting and crowding distance calculations; The Pareto front solution set is obtained through multiple iterations, and the optimal planting scheme is selected from the Pareto front solution set based on scenario analysis.

5. The water-wetland-food synergistic optimization method as described in claim 4, characterized in that, The scenario analysis specifically refers to: Scenario C1: Prioritize food supply, with the constraint that total food production should remain consistent with actual production; Scenario C2: Coordinated development of water, food and wetlands, with the irrigation water demand in the constraints remaining consistent with the actual irrigation water demand; Scenario C3: Prioritize wetland ecological restoration, and ensure that the wetland area in the constraints remains consistent with the actual wetland area.

6. A water-wetland-food synergistic optimization system, characterized in that, include: The model building module is used to construct a multi-objective optimization model with crop planting area as the decision variable. The optimization objectives of the objective function in the multi-objective optimization model include maximizing grain yield, minimizing irrigation water demand, and maximizing wetland area. The constraints of the multi-objective optimization model include crop yield constraints, crop irrigation water demand constraints, and wetland area constraints. The multi-objective solution module is used to solve the multi-objective optimization model based on the objective function and constraints using the NSGA-Ⅲ algorithm, obtain the Pareto front solution set, and select the optimal planting scheme from the Pareto front solution set based on scenario analysis. The scenario analysis refers to the judgment of the ecological status of the planting structure under different scenarios; The collaborative optimization module is used to collect precipitation, evapotranspiration, different crop areas, irrigation water demand and elevation of the area to be predicted as the data to be predicted, predict the dynamic changes of wetland area, and couple the dynamic changes of wetland area with the optimal planting scheme to obtain the optimal agricultural planting structure adjustment scheme.

7. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of the water-wetland-food co-optimization method as described in any one of claims 1 to 5.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the water-wetland-grain synergistic optimization method according to any one of claims 1 to 5.

Images