A method and system for preventing and controlling risks of shallow furrows on sloping farmland

Abstract

This invention provides a method and system for risk prevention and control management of shallow ditches on sloping farmland, comprising: acquiring multidimensional shallow ditch data of sloping farmland; constructing a twin environment of shallow ditches on sloping farmland based on the multidimensional shallow ditch data; employing a cluster of shallow ditch risk prevention and control intelligent agents based on a neuroevolutionary architecture to perform risk simulation and strategy exploration; generating a candidate shallow ditch risk prevention and control strategy set; constructing a shallow ditch association map of sloping farmland; strengthening the candidate shallow ditch risk prevention and control strategy set according to the shallow ditch association map of sloping farmland; generating a strengthened shallow ditch risk prevention and control strategy; acquiring actual feedback data of the strengthened shallow ditch risk prevention and control strategy; and optimizing the strategy based on the actual feedback data, thereby improving the efficiency and reliability of shallow ditch risk prevention and control on sloping farmland.

Description

Technical Field

[0001] This invention relates to the field of agricultural risk prevention and control technology, and more specifically, to a method and system for risk prevention and control management of shallow ditches on sloping farmland. Background Technology

[0002] Controlling shallow gully erosion on sloping farmland is a complex systems engineering project involving the coupling of multiple processes such as hydrology and soil. However, traditional methods often have certain limitations in addressing this challenge.

[0003] On the one hand, traditional methods are usually based on historical experience or fixed scenario libraries for strategy design and evaluation, which often cannot make a rapid and targeted response to specific risk signals monitored in real time. Furthermore, when new or complex risks emerge, the strategy generation cycle is long, and it is often difficult to explore and generate comprehensive solutions to deal with unknown potential risks, resulting in insufficient reliability of the generated decisions.

[0004] On the other hand, traditional methods often treat various governance measures as independent variables and combine them, ignoring their synergistic, antagonistic, and resource competition relationships in spatiotemporal deployment. This can easily lead to the cancellation of the function of measures or resource conflicts in actual implementation, resulting in the generated strategy being fragile in real complex environments, or even completely ineffective. Summary of the Invention

[0005] To address the shortcomings of existing technologies, embodiments of the present invention provide a method and system for risk prevention and control management of shallow ditches on sloping farmland, in order to solve at least one of the aforementioned technical problems in existing technologies.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for risk prevention and control management of shallow ditches on sloping farmland, comprising:

[0008] Acquire multidimensional shallow ditch data of sloping farmland, and construct a ditch twin environment of sloping farmland based on the multidimensional shallow ditch data. The ditch twin environment of sloping farmland includes a ditch risk extrapolator and a soil carbon flow monitor.

[0009] A cluster of shallow ditch risk prevention and control intelligent agents based on a neuroevolutionary architecture is used to conduct risk simulation and strategy exploration, generating a candidate shallow ditch risk prevention and control strategy set;

[0010] Construct a shallow ditch association map of sloping farmland, and strengthen the candidate shallow ditch risk prevention and control strategy set based on the shallow ditch association map to generate a strengthened shallow ditch risk prevention and control strategy.

[0011] Obtain actual feedback data on strategies to strengthen shallow trench risk prevention and control, and optimize based on the actual feedback data.

[0012] Secondly, embodiments of the present invention also provide a risk prevention and control management system for shallow ditches on sloping farmland, comprising:

[0013] An environment construction module is used to acquire multidimensional sloping farmland shallow ditch data and construct a twin environment of sloping farmland shallow ditch based on the multidimensional sloping farmland shallow ditch data.

[0014] The risk simulation module is used to perform risk simulation and strategy exploration using a shallow trench risk prevention and control intelligent agent cluster based on a neuro-evolutionary architecture, and to generate a candidate shallow trench risk prevention and control strategy set.

[0015] The strategy generation module is used to construct a shallow ditch association map of sloping farmland and to strengthen the candidate shallow ditch risk prevention and control strategy set to generate a strengthened shallow ditch risk prevention and control strategy.

[0016] The feedback optimization module is used to obtain actual feedback data on the strategy for strengthening shallow trench risk prevention and control, and to perform feedback optimization based on the actual feedback data.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] Multidimensional data on shallow ditches in sloping farmland is obtained, and a twin environment of shallow ditches in sloping farmland is constructed based on the multidimensional data on shallow ditches in sloping farmland. This step constructs a high-fidelity twin environment by combining physical mechanisms. This twin environment provides a reliable simulation reasoning basis for the subsequent quantitative evaluation of strategies.

[0019] A cluster of intelligent agents for shallow ditch risk prevention and control based on a neuroevolutionary architecture is used to simulate risks and explore strategies, generating a set of candidate shallow ditch risk prevention and control strategies. This step combines the constructed twin environment of shallow ditches on sloping farmland for real-time anomaly monitoring, and simultaneously combines the generated strategy exploration task to evolve multiple strategy solutions to deal with the monitored anomalies, thereby improving the efficiency of identifying abnormal risks and providing a rich data foundation for subsequent strategy generation.

[0020] A shallow ditch association map of sloping farmland is constructed, and the candidate shallow ditch risk prevention and control strategy set is enhanced based on the map to generate an enhanced shallow ditch risk prevention and control strategy. This step optimizes the candidate shallow ditch risk prevention and control strategies by analyzing the synergistic and conflicting relationships among them and conducting robustness tests in both disturbed and undisturbed adversarial environments, thereby making the final enhanced shallow ditch risk prevention and control strategy more reliable.

[0021] By acquiring actual feedback data on the strategy to strengthen shallow trench risk prevention and control, and optimizing the strategy based on the actual feedback data, the accumulation of knowledge about shallow trench risk is achieved, thereby ensuring the reliability and stability of the generated shallow trench risk prevention and control strategy. Attached Figure Description

[0022] Figure 1 This is a flowchart of the steps in the present invention: a method for risk prevention and control management of shallow ditches on sloping farmland.

[0023] Figure 2 This is a flowchart of step S1 in the method for risk prevention and control management of shallow ditches on sloping farmland of the present invention;

[0024] Figure 3 This is a schematic diagram of a risk prevention and control management system for shallow ditches on sloping farmland according to the present invention. Detailed Implementation

[0025] To make the technical solution of the present invention clearer and its technical advantages more apparent, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of the present invention.

[0026] It should be noted that, in this document, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the present invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand, explicitly and implicitly, that the embodiments described herein can be combined with other embodiments.

[0027] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating the steps of a method for risk prevention and control management of shallow ditches on sloping farmland according to the present invention. Figure 2 This is a flowchart of step S1 in the method for risk prevention and control management of shallow ditches on sloping farmland of the present invention. The following is a detailed introduction to the method for risk prevention and control management of shallow ditches on sloping farmland.

[0028] Step S1: Obtain multidimensional shallow ditch data of sloping farmland, and construct a twin environment of shallow ditch of sloping farmland based on the multidimensional shallow ditch data of sloping farmland.

[0029] In this embodiment, step S1 includes:

[0030] Step S1-1: Construct a twin environment of shallow ditch on sloping farmland.

[0031] It should be noted that a training dataset is generated based on historical multidimensional sloping farmland shallow ditch data. The training dataset is used to train the shallow ditch risk inferrer and the soil carbon flow monitor. The historical multidimensional sloping farmland shallow ditch data includes at least historical sloping farmland shallow ditch risk case data and corresponding topographic geometric data such as digital elevation model (DEM) and surface roughness, soil property data such as texture, bulk density, organic carbon, and total nitrogen, vegetation property data such as NDVI and LAI, and environmental meteorological data such as precipitation, temperature, and humidity.

[0032] In this embodiment, step S1-1 includes:

[0033] Step S1-11: Construct a shallow trench risk extrapolator.

[0034] Specifically, a risk extrapolation kernel for a shallow gully risk extrapolator is constructed based on the WEPP model of the water erosion forecasting project. In the traditional WEPP model, key parameters such as gully erosion modulus and gully erodibility are often regarded as constants or simple empirical functions that are spatially related but time-invariant. In this embodiment, structural equation modeling (SEM) is used to deconstruct and reconstruct the traditional WEPP model, and the key parameters that were regarded as static, uniform or empirical are reconstructed into dynamic, spatially heterogeneous, and multi-factor-controlled stochastic process variables.

[0035] Taking the gully head advancement process as an example, this process is modeled as the net result of the following two sub-processes: gully bed scouring process, ,in, The distance the gully head advances due to direct scouring by water flow. For the shear force of the water flow at the bottom of the channel bed, The critical initiation shear force of the trench bed soil. This refers to the amount of sediment coming from upstream. For calibration coefficients, For time; the process of ditch wall collapse, ,in, The distance the trench head advances due to gravitational instability. Because of the depth of the trench, For the slope of the ditch wall, For the cohesion strength of the soil, It is a function of soil moisture content. To reflect the time-varying decay function of the soil-fixing effect of vegetation roots, For calibration coefficients, For time; ultimately, the net advance rate of the ditch head is modeled as .

[0036] It should be noted that the parameters of the above-mentioned stochastic process, such as the critical shear force... Coagulation strength The core soil mechanical parameters are defined as spatiotemporally heterogeneous Gaussian random fields. Specifically, the simulated region will be divided into multiple simulation grid cells. ,in These are represented as row and column indices of the mesh element, respectively. For each simulated mesh element, its parameters... It follows a time-varying Gaussian distribution, i.e. ,in, It is represented as a mean function determined by a set of observable covariates through a deterministic function, for example, by grid cells. The percentage of soil clay content by mass and soil organic carbon content, grid cell In time The weighted sum of soil volumetric water content and living root biomass density, with the weighting coefficient obtained by Bayesian regression on historical experimental data; The standard deviation is denoted as .

[0037] Furthermore, a spatiotemporal correction field is constructed using a Fourier neural operator. This field is used to perform residual repair on the risk simulation kernel. Specifically, a Fourier neural operator is introduced as a general residual learner. At each simulation time step, the Fourier neural operator receives the set of intermediate state field variables calculated by the risk simulation kernel, such as the runoff depth, water flow shear stress, sediment concentration, and current driving data such as rainfall intensity for each grid cell. It outputs a residual correction tensor with the same spatial dimension as the input space. This tensor is directly used to update the corresponding variables. For example, the risk simulation kernel calculates the shear stress of a certain cell as 5 Pa according to the hydrodynamic formula. However, historical data shows that under the current soil moisture content and previous drying cracking conditions, the actual critical threshold for erosion is lower. In this case, the Fourier neural operator performs convolution in the Fourier space and outputs a negative correction amount of -0.8 Pa, making the effective shear stress 4.2 Pa, thereby triggering the erosion simulation more accurately.

[0038] It should be noted that the training process is divided into the following stages: the hydrological basic learning stage, which trains on a subset of training data without severe erosion events, mainly training the hydrological module parameters based on the WEPP risk extrapolation kernel to ensure accurate runoff prediction; the erosion correction learning stage, which fixes the parameters of the risk extrapolation kernel and trains the spatiotemporal correction field based on Fourier neural operators on a subset of data containing severe erosion events, enabling it to learn to predict deviations in erosion occurrence; and the joint fine-tuning stage, which aims at overall simulation accuracy and performs small-scale joint fine-tuning of the risk extrapolation kernel and the spatiotemporal correction field.

[0039] The shallow ditch risk extrapolator is trained with the objective of minimizing a preset risk extrapolation loss function until a preset risk extrapolation qualification condition is met, at which point the training ends. The preset risk extrapolation loss function can be expressed as follows: ,in, The total loss term can be expressed as: , To simulate the number of events or time periods, and These represent the cumulative total sediment transport predicted by the model and observed values, respectively. This item is used to quantify the overall erosion magnitude. The spatial morphology loss term can be expressed as: , This represents the model's predicted probability of the true class. This is a focusing parameter, usually set to 2, used to reduce the weight of easily classified samples. These are balancing parameters used to give higher weights to rarer categories; The extreme event stability loss, used to enhance the model's predictive robustness under extreme weather conditions, can be expressed as: ,in, and These represent the predicted and actual gully head advance distances, respectively. This is a hyperparameter used to control the threshold for transitioning from squared loss to linear loss; The weighting coefficients for the corresponding terms can be determined through Pareto optimization based on the performance of the validation set; the preset risk simulation qualification conditions can be expressed as the average absolute error between the predicted gully head advance position and the measured gully head position being consistently lower than a preset threshold, and the Nash efficiency coefficients of the simulated and observed total sediment transport sequences reaching a preset threshold.

[0040] Furthermore, after the shallow ditch risk inferrer is trained, a spectral clustering algorithm is used to extract patterns from the spatiotemporal correction field, generating multiple shallow ditch risk pattern nodes. Simultaneously, counterfactual intervention reasoning is combined to perform causal mining on the shallow ditch risk pattern nodes, constructing a risk pattern graph.

[0041] Specifically, during the training process, dynamic mode decomposition is used to extract spatial modes that repeatedly appear during the simulation training and have specific oscillation frequencies or growth rates from the spatiotemporal correction field sequence output by the shallow ditch risk extrapolator. For each extracted mode, integral gradient backtracking analysis is used to identify which combinations of input conditions activated the mode, and the mode is associated with the corresponding combination of conditions. The above mode-condition association pairs are then clustered into different risk pattern types. For example, all correction patterns associated with "high erosion risk caused by the first heavy rainfall after the previous drought" are clustered into "cracking-infiltration mismatch pattern".

[0042] Each pattern is nodeified into a corresponding node, and its triggering condition, i.e., the combination of condition variables, is used as the node attribute. For example, soil moisture content exceeding 80% of field capacity, vegetation cover less than 20%, and previous tillage events less than 7 days can be used as a risk pattern triggering condition. The historical frequency and correction intensity of its occurrence are quantified into corresponding probability distributions. For example, the correction multiplier of the soil erodibility coefficient K value follows a Beta (α=2, β=5) distribution. Nodes are associated based on overlapping conditions or physical processes. For example, risk pattern A may exacerbate risk pattern B, thus establishing a directed edge between A and B, ultimately forming a Bayesian network, which is the risk pattern graph. This graph is used to store under which known scenarios the shallow ditch risk inferrer may be unreliable.

[0043] Steps S1-12: Construct a soil carbon flow monitor.

[0044] Specifically, a soil carbon flow monitor is constructed based on the Century model. Traditional Century models treat the turnover rate constants of different carbon pools, such as readily degradable and difficult-to-degrade carbon, as fixed values ​​weakly correlated with spatial location or only related to climate zones. This embodiment, however, parameterizes the turnover process between carbon pools. That is, for any carbon pool m, such as an active carbon pool or a slow-degradable carbon pool, in the grid cell... Turnover rate Instead of using constants, it is defined as a dynamic function driven by multiple physical environment covariates, and the corresponding carbon conversion rate prediction is generated based on the turnover rate of the dynamic function form.

[0045] Decomposition rate of easily decomposable carbon pools For example, it can be represented as ,in, This represents the basic turnover rate of the carbon bank under standard reference conditions. and Temperature and volumetric water content The scaling functions, such as the Q10 model or the Arrhenius equation, are used to describe the response of biochemical reaction rates to temperature, and the water retention curve is used to describe the effect of water availability. For the reason The driving physical perturbation scaling function, the It is a soil structure state index used to reflect the stability of soil aggregates and pore connectivity, and can be expressed as: , For a moment The average weight diameter of the aggregates is used to represent stability. For a moment The saturated hydraulic conductivity is used to represent pore connectivity. Let be a decay function characterizing the recovery state after a disturbance, such as an exponential recovery function over time after tillage. and This refers to baseline reference values ​​for current soil locations, such as stable values ​​under long-term no-till conditions. This is a random field with a mean of 1, used to simulate and characterize the basic decomposition potential determined by soil background properties such as clay content and mineral surface activity; similarly, the parameters of the soil carbon flow monitor are defined as a spatiotemporally heterogeneous Gaussian random field, that is, for each grid cell... Its parameters The prior distribution is a Gaussian distribution. , and These represent the corresponding mean function and variance, respectively.

[0046] Furthermore, a multi-scale graph neural network is used to model shallow gullies in sloping farmland to generate corresponding soil carbon cycle variation maps. Specifically, the soil profile is discretized into supervoxel units in the vertical and horizontal directions, with each unit serving as a graph node. Node features include at least soil texture, bulk density, initial carbon storage, and carbon conversion rate predictions generated by a soil carbon flow monitor. Edge connections between nodes include physical transport edges and functional similarity edges based on process similarity, such as belonging to the same hydrological response unit. Physical transport edges are established based on spatial adjacency relationships to simulate the leaching of dissolved organic carbon and vertical carbon migration. Functional similarity edges are established using the K-nearest neighbor algorithm or attribute similarity thresholding method to connect nodes that are not spatially adjacent but have similar attributes. For example, all sandy soil nodes are connected to form a sandy soil functional cluster, thereby constructing a soil carbon flow map structure.

[0047] A message passing mechanism using a graph neural network is employed to propagate messages within the aforementioned soil carbon flow map structure. During training, the carbon flux between nodes is learned and predicted, and a residual correction vector is output for each node. Based on this residual correction vector, the predicted carbon conversion rate is residually corrected. For example, by learning how tillage disturbance affects the leaching rate of deep carbon by changing the pore network structure, or how carbon loss at specific locations on the slope affects the carbon input of downstream units through surface runoff, and other complex spatial relationships and nonlocal effects, a corresponding residual correction vector is generated to residually correct the predicted carbon conversion rate generated by the soil carbon flow monitor.

[0048] It should be noted that the soil carbon cycle anomaly map is represented as follows: for a trained graph neural network, diverse scenarios covering various meteorological, soil, and management measures are input into a trained soil carbon flow monitor. The residual correction vector output by the graph neural network in each scenario and the corresponding input data are recorded, i.e., parameterized data corresponding to environmental variables and management measures for each scenario. Graph neural network interpretability techniques such as GNNExplainer are used to identify and extract the subgraph structures that contribute the most to the final correction prediction in the soil carbon flow map structure under different scenarios, i.e., the set of key nodes and their connections. Combining subgraph isomorphism and clustering algorithms, the extracted subgraph structures are classified, and each subgraph represents the attention focus pattern of the graph neural network when dealing with a specific type of carbon cycle anomaly.

[0049] For each subgraph pattern, analyze the input environmental context at activation, such as soil moisture, temperature, and texture combination, as well as the resulting corrective behavior, such as the direction and magnitude of adjustment to the conversion rate of a certain carbon pool. Abstract this into a carbon cycle mutation node. For example, the asymmetric pulse release pattern of soil respiration during freeze-thaw cycles can be considered a carbon cycle mutation node. Its node attributes include the corresponding triggering conditions and influencing mechanisms. For instance, the triggering condition for a certain mutation pattern is that the soil temperature fluctuates within the range of 0℃±1℃ and the soil saturated hydraulic conductivity is lower than the corresponding preset threshold. Its influencing mechanism can be represented as: {Target process: decomposition rate of the active carbon pool, multiplication factor distribution during the heating phase:} The various carbon cycle variation nodes together form the soil carbon cycle variation map.

[0050] Furthermore, a training strategy combining multi-task learning and meta-learning is adopted, and the soil carbon flow monitor is trained simultaneously with a preset multi-task weighted loss function until the preset carbon flow monitoring qualification conditions are met.

[0051] Specifically, the main task of multi-task learning is to simulate the time series of soil respiration flux and the changes in organic carbon storage at different soil depths. The auxiliary tasks include at least predicting the leaching flux of dissolved organic carbon and the short-term changes in surface soil temperature and moisture content. Through shared representation learning, the auxiliary tasks provide additional physical process constraints such as water-carbon coupling and energy-carbon coupling to the main task, thereby improving the generalization ability and physical consistency of the soil carbon flow monitor. At the same time, in order to improve the rapid adaptation ability of the soil carbon flow monitor to new plots, model-independent meta-learning is adopted. Multiple slope datasets from different climate zones and soil types are constructed into a series of meta-tasks. The training objective of this process is to obtain a set of global initial parameters so that when facing any new task, it can achieve high-performance prediction based on only a small amount of observation data of the slope and through a small number of gradient update steps.

[0052] The loss function used in the training process can be expressed as: ,in, This is represented as a flux loss term, and the error between the predicted and observed soil respiration flux can be calculated using the Huber loss function. For the storage loss item, the error between the predicted and measured values ​​of carbon storage in each soil layer can be calculated using the mean square error. This is the auxiliary task loss term, which is the mean square error between the predicted and measured values ​​of the corresponding variables in the auxiliary task. The physical constraint loss term is a soft constraint used to penalize situations in the output that violate basic physical or biochemical laws. For example, a carbon-nitrogen ratio constraint loss can be introduced to penalize predictions of soil microbial biomass carbon-nitrogen ratios that exceed a reasonable range using the ReLU function, or a mass conservation loss can be introduced to check whether the inputs, outputs, and inventory changes of each carbon pool are balanced within the allowable error range. The weighting coefficients for the corresponding loss terms can be determined through Pareto front analysis on the validation set, such as... .

[0053] Training is performed with the goal of minimizing the above loss function until the following conditions are met: on a completely independent validation set, the mean absolute percentage error between the prediction and the observation is consistently lower than a preset threshold; the coefficient of determination and the efficiency coefficient between the soil respiration flux sequence generated by the soil carbon flow monitor and the observed data are both greater than the corresponding preset thresholds; or the preset maximum training period is reached, at which point training ends.

[0054] Steps S1-13: merging and generating twin environments of shallow ditches on sloping farmland.

[0055] Specifically, based on physical information neural networks and knowledge distillation, the shallow ditch risk inferrer and soil carbon flow monitor are distilled to generate a twin environment of shallow ditches on sloping farmland. In particular, the trained shallow ditch risk dynamic inferrer and soil carbon flow monitor are used as teacher models. Latin hypercube sampling is performed in the input parameter space of the teacher model to generate a scenario-outcome paired dataset. This scenario-outcome paired dataset is used as the training dataset for this stage. Each scenario includes a random initial environmental state, such as topography, soil, vegetation, and initial meteorological field, as well as a set of randomly generated governance strategy combinations, such as measure type, spatial layout, and implementation intensity. Each outcome is represented by the time series of core indicators simulated by the teacher model under that scenario, such as the cumulative erosion of each grid, gully erosion development status, changes in soil carbon pool storage, and net carbon flux.

[0056] A physical information neural network is used as the student model, with its input and output consistent with the teacher model. Knowledge distillation technology is employed for synchronous integration. The model distillation is performed using the distillation loss function, where, The standard mean squared error loss is used to ensure the accuracy of the student model's fit to the teacher model's output data. For physical constraint loss, basic physical conservation laws such as sediment mass conservation equation and carbon mass balance equation are encoded as differentiable residual terms and incorporated into the loss function, forcing the student model to follow basic physical laws even in areas not covered by training data. The weighting coefficients for the corresponding loss terms can be obtained by analyzing historical data; the goal is to minimize the distillation loss function until the correlation coefficient between the student model output and the teacher model output reaches a preset threshold; at this point, the distilled student model is the core inference engine for the twin environment of sloping farmland and shallow ditch.

[0057] Furthermore, the risk pattern map and the soil carbon cycle variation map are encoded into a risk pattern memory for the shallow ditch twin environment of sloping farmland. The risk pattern memory contains multiple shallow ditch risk patterns, each of which consists of corresponding risk conditions from the risk pattern map and the soil carbon cycle variation map. Specifically, all pattern nodes are extracted from the risk pattern map and the soil carbon cycle variation map, and the nodes of both types of maps are uniformly encoded into shallow ditch risk pattern objects. Each shallow ditch risk pattern object contains at least a unique identifier, a trigger condition expression, a scope definition, an influence parameter, and an intensity distribution. The trigger condition expression represents a set of logical judgment rules based on environmental variables defined using a domain-specific language. The scope definition represents the spatial range to which the model applies, such as specific terrain location and soil type, and the temporal conditions, such as seasonality and specific post-management periods. The influence parameter represents the parameters affected by the model, such as infiltration rate, soil erodibility, and carbon decomposition rate. The intensity distribution is the probability distribution of the influence intensity.

[0058] Each ditch risk model object is instantiated as an effector module, which contains a condition discriminator, an intensity sampler, and a state modification function. The condition discriminator is represented as a lightweight decision tree or rule engine, used to determine in real-time whether the current simulated environment meets the triggering conditions for the ditch risk model. The intensity sampler is used to sample a specific intensity coefficient from the corresponding intensity distribution when the triggering conditions for a ditch risk model are met. The state modification function describes how to apply the intensity coefficient to the state variables of the ditch twin environment of sloping farmland. For example, for the erosion risk model corresponding to the risk model map, the modified effective soil shear strength = effective soil shear strength × (1 - intensity coefficient); for the carbon cycle anomaly model corresponding to the carbon cycle anomaly map, the modified active carbon pool decomposition rate = active carbon pool decomposition rate × (1 + intensity coefficient).

[0059] Step S1-2: Initialize the environmental state of the shallow ditch twin environment of sloping farmland.

[0060] Specifically, real-time multi-dimensional data on shallow ditches in sloping farmland is acquired, such as post-rain UAV orthophotos, lidar point clouds, and real-time soil temperature and humidity data from sensor networks. An integrated Kalman filter algorithm is then used to update the state of the twin environment of the shallow ditches in the sloping farmland.

[0061] Step S2: A cluster of shallow ditch risk prevention and control intelligent agents based on a neuroevolutionary architecture is used to conduct risk simulation and strategy exploration, generating a candidate shallow ditch risk prevention and control strategy set.

[0062] In this embodiment, step S2 includes:

[0063] Step S2-1: Conduct risk simulation based on the shallow ditch risk prevention and control intelligent agent cluster, and construct a set of strategy exploration tasks.

[0064] Specifically, risk monitoring is performed based on a preset risk monitoring mechanism to obtain an original risk anomaly dataset, and the original risk anomaly dataset is matched with the shallow ditch risk patterns stored in the risk pattern memory within the shallow ditch twin environment of sloping farmland.

[0065] If a match is found, it is determined to be a known risk pattern, and a calibration task is generated.

[0066] If a match fails, it is determined to be an unknown risk pattern. Spectral clustering and constraint-based causal discovery algorithms are used to explore pattern features and generate a set of verification tasks.

[0067] In one possible embodiment, the initialized twin environment of shallow ditches on sloping farmland undergoes forward simulation based on the current environmental state, outputting corresponding simulation prediction results. Specifically, the current environmental state vector of the twin environment is spatiotemporally aligned and spliced ​​with short-term deterministic meteorological data such as hourly rainfall and temperature sequences for the next 7 days provided by numerical weather prediction, generating a multidimensional state tensor. This multidimensional state tensor is then input into the core inference engine of the twin environment for inference. The core inference engine is a physical information neural network that takes the multidimensional state tensor as input and combines a series of hidden layers within the network for nonlinear transformation. The output is a predicted sequence of key state variables such as runoff depth, sediment concentration, changes in carbon pool storage, and soil respiration flux for each grid. The hidden layer has completed the encoding of basic physical laws such as mass conservation and energy balance, as well as their corresponding spatiotemporal correlations, during the distillation and training phases. For example, when features such as steep slope, high soil sand content, low vegetation cover, and forecast of heavy rainfall in a grid in the multidimensional state tensor are input into the physical information neural network, specific neurons in the middle layer of the network corresponding to physical processes such as "rapid runoff generation" and "easy soil particle stripping" will be strongly activated and used to perform nonlinear transformations on the corresponding features.

[0068] Based on the predicted sequences of the key state variables, corresponding risk indicators are generated. Taking the runoff depth, sediment concentration, changes in carbon pool storage, and soil respiration flux as examples, the sediment concentration corresponding to the same time stamp of each grid is multiplied by the runoff depth, and the runoff is accumulated along the slope to obtain the spatial distribution of sediment transport at the current moment. Based on the spatial distribution of sediment transport, an erosion risk heat map is generated for the corresponding moment. Each pixel value in the map represents the expected intensity of erosion at that location. The time dimension of the changes in carbon pool storage is differentiated, and the net carbon sink change rate of each grid is calculated in combination with the soil respiration flux. A positive value of this rate indicates carbon accumulation, and a negative value indicates carbon loss.

[0069] The generated erosion risk heat map and carbon sink change rate are compared with the corresponding historical baseline map or safety threshold grid by grid and time period. If one or more spatially continuous grid cells show statistically significant deviations, such as the predicted erosion modulus exceeding the 95th percentile of the same period in history and the spatial clustering is higher than the threshold, then all related contextual information such as soil properties, previous meteorological conditions, and topographic features, as well as the intermediate output features of the physical information neural network, are used to form a structured risk anomaly record. For example, according to the simulation of the shallow ditch twin environment of sloping farmland, the predicted soil loss suddenly increases to 85 tons / hectare in an area in the upper part of the slope where deep tillage was carried out last year, while the historical average baseline for this area is only 12 tons / hectare. Moreover, the high-risk area in the erosion risk heat map is spatially distributed in a strip shape. At this time, the geographical location polygon corresponding to the risk area, the abnormal indicators such as soil loss, the abnormal intensity, and the corresponding related environmental context are merged into a risk anomaly record.

[0070] The risk anomaly record is matched with the shallow ditch risk pattern stored in the risk pattern memory. That is, the feature vector of the risk anomaly record and the trigger condition corresponding to the shallow ditch risk pattern are calculated by cosine similarity. If the cosine similarity reaches a preset threshold, it is determined that the match is successful; otherwise, it is determined that the match is unsuccessful.

[0071] For successful matches, a calibration task is generated based on the matched shallow ditch risk pattern. Specifically, the intensity sampler within the effector module corresponding to the shallow ditch risk pattern is invoked to obtain one or more sets of intensity coefficients. The state vector of the current twin environment is then modified according to the corresponding state modification function to generate a calibration task environment state vector. The corresponding scenario trigger probability is generated by combining the historical occurrence frequency and cosine similarity of the matched shallow ditch risk pattern. For example, if the historical occurrence frequency of a certain shallow ditch risk pattern is three times a year, and the current cosine similarity is 0.8, then its scenario trigger probability is 3 / 12*0.8=0.32. This probability is the expected occurrence probability of the shallow ditch risk pattern in the future task cycle, such as the next growing season.

[0072] The weights of a preset baseline fitness function are adjusted based on the matched shallow trench risk patterns. Specifically, the baseline fitness function can be expressed as follows: ,in, Represented as a comprehensive performance score, it measures the core return of the strategy under ideal conditions where no specific shallow trench risk pattern is activated. It can be expressed as... , To normalize the erosion risk reduction score, which is the erosion reduction rate relative to the baseline without measures, This is the normalized carbon sink gain fraction, which is the rate of increase in carbon storage relative to the baseline without measures. The corresponding preset weighting coefficient is 0.5 by default, indicating that erosion control and carbon sink gain are equally important. Represented as a normalized cost score, it measures the total cost of implementing a strategy and can be expressed as: ; Represented as a comprehensive robustness score, it measures the stability of a strategy's performance under general uncertainty, and can be expressed as: , For strategy Performance scores under a set of common random perturbation scenarios Standard deviation; These are the weighting coefficients for the corresponding items, and the sum of the three is 1. The default setting is... This is used to characterize tasks that prioritize performance while also considering cost and robustness. When generating calibration tasks for specific ditch risk patterns, the matching ditch risk pattern type is used to... Adjustments can be made; for example, assuming the matched shallow trench risk pattern corresponds to the erosion risk pattern in the risk pattern map, then... The value was increased from 0.5 to 0.8, thereby guiding the strategy exploration direction towards addressing erosion risks; and users can adjust the value based on their current risk control preferences. Adjustments can be made, such as those aimed at controlling costs. As the value of is increased, the other two weight coefficients are simultaneously decreased, thereby generating a fitness function that is adapted to the current calibration task.

[0073] Based on the characteristics of the matched shallow trench risk patterns, the most likely effective subset of measures is selected from a pre-set global measure library. For example, for soil structure fragile patterns, measures such as increasing organic matter input, planting deep-rooted green manure, and reducing tillage intensity are given priority for exploration. At the same time, the range of variation of key parameters of the selected measure subset will be guided according to the characteristics of the pattern. For example, to deal with structural fragility, the lower limit of the cover thickness for straw mulching may be increased to ensure a minimum protective effect.

[0074] Finally, the four types of task data—the calibration task environment state vector, scenario trigger probability, fitness function, list of allowed measures, and parameter range—are encapsulated into a calibration task.

[0075] For cases where a match fails—that is, a risk anomaly record cannot be effectively matched with any known shallow ditch risk pattern—the record is marked as a potential new risk pattern. First, all recent unmatched anomaly records are collected, and their multi-dimensional features, such as spatial coordinates, environmental condition combinations, and anomaly morphology indicators, are analyzed using a spectral clustering algorithm. This generates several anomaly clusters with internal consistency, each representing a possible prototype of a new shallow ditch risk pattern. For each anomaly cluster, a constraint-based causal discovery algorithm is used, under knowledge constraints such as "runoff is a necessary mediator of erosion," to infer the most probable causal structure graph from its data. Based on the generated causal structure graph, one or more verification tasks are formalized to form a set of verification tasks. For example, for a risk anomaly record with a certain carbon flux but no known pattern, spectral clustering is used to associate it with conditions such as "the 2nd-3rd rainfall after straw mulching" and "soil temperature 10-15℃". Then, a causal discovery algorithm is run to infer that "the mulch induces a local anaerobic environment under specific temperature and humidity conditions, generating a temporary CH4 pulse" is a possible causal path. At this point, a verification task is generated based on this causal path: {Hypothesis: There is an anaerobic carbon pulse under the straw mulch layer during low-temperature rainfall. Objective 1: Explore mulch management strategies that can minimize carbon sink loss under this hypothesis. Objective 2: Analyze the performance differences of different strategies under the two environments where the hypothesis is true and false, and simultaneously evaluate the reliability of the hypothesis}.

[0076] Step S2-2: Based on the set of strategy exploration tasks, conduct strategy exploration to generate a set of candidate shallow trench risk prevention and control strategies.

[0077] Specifically, the shallow ditch risk prevention and control intelligent agent cluster adopts an evolutionary algorithm, simultaneously combining the calibration task and the verification task set to perform forward simulation, obtain the corresponding simulation results, evaluate the fitness of the shallow ditch risk prevention and control intelligent agents based on the preset multi-objective evaluation function and the simulation results, and select the intelligent agents that reach the preset fitness conditions as parent intelligent agents, perform population evolution operation on the parent intelligent agents until the preset evolution termination condition is reached, and generate a candidate shallow ditch risk prevention and control strategy set.

[0078] It should be noted that the policy network of each shallow ditch risk prevention and control agent in the shallow ditch risk prevention and control agent cluster is represented as a spatiotemporal convolutional long short-term memory network. The policy network takes the state vector corresponding to the shallow ditch twin environment of sloping farmland as input and outputs the original candidate shallow ditch risk prevention and control policy.

[0079] Specifically, the shallow ditch risk prevention and control intelligent agent represents a complete and executable slope management solution. Its genotype consists of the following two parts: a spatiotemporal measure deployment sub-gene, which is represented as the weight parameter set of a spatiotemporal convolutional recurrent neural network. The spatiotemporal convolutional recurrent neural network takes the state space of the shallow ditch twin environment of sloping farmland as input and outputs a four-dimensional action tensor with dimensions of [measure type, spatial X coordinate, spatial Y coordinate, intensity level]. This network can generate differentiated and dynamically responsive measure configurations according to the spatiotemporal differences in environmental conditions. For example, it outputs the "high-intensity vegetation hedge" action in areas with steep slopes and low carbon storage, and the "low-intensity straw mulch" action in flat areas with low erosion risk. The adaptive logic rule sub-gene is represented as a set of "IF-THEN" rules written in a domain-specific language. Each rule defines the immediate adjustment of the output of the above spatiotemporal convolutional recurrent neural network when a specific monitoring indicator is triggered. For example, IF real-time rainfall intensity > 50 mm / h AND predicted soil moisture at the ditch head > level 4 THEN 50 meters upstream of the ditch head grid. The intensity level of the "straw mulching" measure in the motion tensor is temporarily increased to the highest level.

[0080] In one possible embodiment, a corresponding simulation environment copy is created for each task within the strategy exploration task set, and simulation is performed in conjunction with the corresponding task data. That is, for calibration tasks, simulation is performed directly based on the task data they contain, and the multi-objective evaluation function used in the simulation is the fitness function contained in the task data; for verification tasks, the corresponding relationships in the environment are adjusted according to the new hypothesis. For example, a low-temperature anaerobic methane generation module corresponding to the hypothesis is temporarily added, and simulation is performed based on the adjusted environmental state, and the multi-objective evaluation function used is the baseline fitness function.

[0081] An initial population of M shallow trench risk control agents is randomly generated. Under the specific environmental state specified by the strategy exploration task within the set of strategy exploration tasks, each shallow trench risk control agent in the population evaluates its comprehensive performance throughout the entire task cycle in combination with its corresponding multi-objective evaluation function, and calculates the corresponding fitness score.

[0082] After fitness evaluation, an iterative optimization loop begins. First, based on the fitness scores, individuals in the population are classified according to Pareto levels using fast non-dominated sorting. Combined with crowding distance calculation, individuals with high frontier levels and sparse distribution in the target space are preferentially selected as parents. Subsequently, task-guided crossover and mutation are performed on the selected parents to generate new offspring individuals. Their genotypes are recombinated through methods such as simulated binary crossover, and mutation operations incorporate domain knowledge. For example, in tasks dealing with fragile soil structures, the mutation probability of gene segments encoding rules such as "increasing organic fertilizer application" and "reducing tillage" is increased. This allows for targeted exploration of the relevant strategy space. Then, newly generated offspring individuals, along with some elite parents, form a new generation population. For example, 50% of the parent individuals are randomly selected to form a new strategy exploration population with all offspring individuals. Based on this strategy exploration population, a cyclical process of simulation, fitness score calculation, elite selection, task-guided crossover and mutation, and population renewal is executed until the hypervolume index of the population's Pareto front changes below a preset threshold for multiple consecutive generations, or reaches a preset maximum number of generations. This outputs a Pareto optimal solution set for a candidate shallow ditch risk control strategy, which is the candidate shallow ditch risk control strategy set.

[0083] Step S3: Construct a shallow ditch association map of sloping farmland, and strengthen the candidate shallow ditch risk prevention and control strategy set based on the shallow ditch association map to generate a strengthened shallow ditch risk prevention and control strategy.

[0084] In this embodiment, step S3 includes:

[0085] Step S3-1: Based on the candidate shallow ditch risk prevention and control strategy set, perform hypergraph modeling and combine hypergraph convolutional neural network for information propagation to construct shallow ditch association map of sloping farmland.

[0086] Specifically, the candidate shallow ditch risk prevention and control strategy set is analyzed and the pattern is extracted to generate a corresponding candidate strategy atom set. The atomic attributes of each candidate strategy atom in the candidate strategy atom set include at least spatial unit information such as "geographic grid ID", management measure type such as "constructing contour hedges" or "implementing straw mulching", implementation intensity information such as "double row planting, density of 5 plants / meter" or "mulching thickness of 5 cm" and activation context condition information such as "only implemented when the slope is greater than 15 degrees and the soil organic matter is less than 2%".

[0087] Furthermore, based on the candidate strategy atom set and corresponding atom attributes, an original sloping farmland shallow ditch association map is constructed. A hypergraph convolutional neural network is used to propagate and aggregate information on the attributes of each node in the original sloping farmland shallow ditch association map, thus constructing the sloping farmland shallow ditch association map.

[0088] Understandably, the shallow ditch association map of sloping farmland contains at least candidate strategy nodes, land unit nodes, environmental status nodes, and resource constraint nodes. Specifically, each candidate strategy atom is a candidate strategy node; land unit nodes are used to store topography and soil attributes; environmental status nodes are used to store environmental status information, such as soil moisture exceeding 80% of field capacity; and resource constraint nodes are used to store available resource information, such as available tractor shifts and total straw supply.

[0089] The edge connections between nodes in the shallow ditch association graph of sloping farmland include at least implementation edges, conditional edges, competitive edges, and cooperative edges. Implementation edges connect candidate strategy nodes to their corresponding land unit nodes, thus representing the implementation location of the candidate shallow ditch risk control strategy. Conditional edges connect candidate strategy nodes to the environmental state nodes that activate them, thus representing the effective conditions of the candidate shallow ditch risk control strategy. Competitive edges connect two or more candidate strategy nodes that share the same resource constraint node, representing their competition for limited resources. For example, candidate strategy node 1 representing "deep cultivation" and candidate strategy node 2 representing "drainage ditch excavation" compete for the resource constraint node 1 representing "large machinery hours." Cooperative edges connect strategy atomic nodes that mutually promote each other in physical processes or ecological functions. For example, a cooperative hyperedge connects candidate strategy node 3 representing "vegetation hedges," candidate strategy node 4 representing "straw mulch," and environmental state node 1 representing "runoff reduction," thus representing the effect of runoff reduction achieved by constructing vegetation hedges to slow runoff and applying straw mulch to increase infiltration.

[0090] In one possible embodiment, firstly, neural network attribution analysis and pattern extraction techniques are used to deconstruct each candidate shallow ditch risk prevention and control strategy within the candidate shallow ditch risk prevention and control strategy set into a set of candidate strategy atoms. Each atom represents a basic decision-making unit that executes a specific type and intensity of governance measures under specific spatiotemporal conditions. For example, a candidate strategy atom can be expressed as {Location: Global, Measures: Reduce fertilizer application, Intensity: Reduce by 20%, Condition: Soil nitrogen, phosphorus, and potassium detection values ​​are higher than the threshold}. Next, an original sloping farmland shallow ditch association map is constructed to store the association relationships between all candidate strategy atoms, geographic grid units, environmental status categories, and resource elements. Then, a hypergraph convolutional neural network is used to learn the constructed original sloping farmland shallow ditch association map. Through multi-layer message passing, the information stored in different nodes and hyperedges is aggregated to generate an embedding vector for each node in the original sloping farmland shallow ditch association map. The original sloping farmland shallow ditch association map, which completes information transmission and aggregation, serves as the final sloping farmland shallow ditch association map.

[0091] Step S3-2: Based on the shallow ditch association map of sloping farmland, the candidate shallow ditch risk prevention and control strategy set is enhanced to generate enhanced shallow ditch risk prevention and control strategy.

[0092] Specifically, cognitive uncertainty perturbations are injected into the twin environment of shallow ditches on sloping farmland based on the perturbation injection mechanism to generate a perturbed twin environment of shallow ditches on sloping farmland. In both the undisturbed and perturbed twin environments of shallow ditches on sloping farmland, a set of candidate shallow ditch risk prevention and control strategies is simulated and executed.

[0093] Furthermore, based on the simulation results and the preset robustness assessment mechanism, the top N candidate shallow ditch risk prevention and control strategies in terms of robustness are selected as primary shallow ditch risk prevention and control strategies. Based on the primary shallow ditch risk prevention and control strategies, multi-objective strategy optimization is carried out simultaneously by combining the shallow ditch correlation map of sloping farmland to generate enhanced shallow ditch risk prevention and control strategies.

[0094] In one possible embodiment, firstly, based on the nodes associated with each candidate strategy node in the shallow ditch association map of sloping farmland and their corresponding connecting edges, an adaptive undisturbed environment vector is constructed for each candidate shallow ditch risk prevention strategy. For example, if a candidate shallow ditch risk prevention strategy specifically addresses the risk of "fragile soil structure after tillage leading to increased erosion," then its adaptive undisturbed environment vector is the environment vector simulating "fragile soil structure after tillage," used to reproduce the environment vector corresponding to the occurrence of this risk. Subsequently, from the risk pattern memory of the sloping farmland shallow ditch twin environment, according to the history of each pattern... Based on the frequency of occurrence, seasonal characteristics, and matching degree with the current scenario, one or more effector modules are randomly selected. The triggering conditions of each effector module are forcibly met to generate a corresponding intensity coefficient. Combined with its state modification function, the undisturbed environment vector is perturbed to generate a set of perturbed environment vectors containing multiple perturbed environment vectors. For example, the "dry crack-infiltration mismatch mode" and the "freeze-thaw carbon pulse release mode" are activated at the same time, so that the originally undisturbed environment vector shows a decrease in soil infiltration rate and a short-term increase in carbon mineralization, thereby simulating a complex and unfavorable composite perturbation scenario.

[0095] Next, the state space of the twin environment of shallow ditches on sloping farmland is updated based on the vectors in the sets of undisturbed and disturbed environments, thereby simulating the corresponding undisturbed and disturbed environments. Corresponding candidate shallow ditch risk control strategies are executed in these environments. The comprehensive performance score in the undisturbed environment is obtained by combining the comprehensive performance score calculation function in the baseline fitness function, and this score is used as the expected performance value. Simultaneously, the worst 5% comprehensive performance score in the disturbed environment is obtained and used as the risk performance value. For example, if a candidate shallow ditch risk control strategy has 100 versions of disturbed environments (i.e., 100 different disturbed environments are generated from 100 different disturbed environment vectors), the comprehensive performance scores corresponding to the candidate shallow ditch risk control strategies in these 100 disturbed environments are obtained, and the average of the five lowest comprehensive performance scores is used as the risk performance value. The average comprehensive performance score of all disturbed environments is obtained and used as the stability score. The risk performance value and stability score are weighted and averaged to obtain the comprehensive robustness score corresponding to each candidate shallow ditch risk control strategy. The top strategies with the highest comprehensive robustness scores are selected. The candidate shallow ditch risk prevention and control strategies for N are selected, and the selected candidate shallow ditch risk prevention and control strategies are used as the primary shallow ditch risk prevention and control strategies.

[0096] Next, using neural network attribution analysis and pattern extraction techniques, the N primary shallow ditch risk control strategies are further deconstructed into candidate strategy atoms to obtain a robust candidate strategy atom set. Based on this robust candidate strategy atom set, a reinforcement strategy exploration space is constructed, and a multi-objective Bayesian optimization algorithm is run to maximize expected effectiveness and risk effectiveness while minimizing total cost as the exploration objective. In the reinforcement strategy exploration space, corresponding robust candidate strategy atoms are selected for each spatial unit on the ditch slope. At each key time window, such as before spring plowing or before the rainy season, a set of robust candidate strategy atoms is selected and their specific parameters are set to form a candidate reinforcement shallow ditch risk control strategy. For example, for a certain spatial unit, its decision variable can be represented as [robust candidate strategy atom combination type: {constructing a plant hedge and performing strip straw mulching}, implementation intensity: {row spacing 1.2 meters, coverage 80%}]. The decision variables of all spatial units together constitute a complete candidate reinforcement shallow ditch risk control strategy. By continuously exploring strategies in the reinforcement strategy exploration space, a reinforcement shallow ditch risk control strategy set is finally output.

[0097] It should be noted that during the strategy exploration process using the multi-objective Bayesian optimization algorithm, the domain knowledge and relationships contained in the sloping farmland ditch association graph need to be used to constrain the strategy exploration process. Specifically, the competitive edges in the sloping farmland ditch association graph are transformed into hard constraints of the optimization problem. For example, if "deep plowing with large machinery" and "manual digging of fish-scale pits" share "resource pool 1", and the daily availability of resource pool 1 is limited, then when generating candidate solutions, the optimizer must ensure that the total demand for R1 of these two types of measures deployed within the same time period does not exceed the upper limit. The cooperative edges in the sloping farmland ditch association graph are transformed into reward terms in the objective function, thereby guiding the optimizer to give higher utility evaluations to solutions where robust candidate strategy atomic pairs with strong cooperative relationships co-occur. Similarly, the conditional edges in the sloping farmland ditch association graph are transformed into conditional relationship-guided soft constraints, so that the optimizer learns exploration experience such as "deploying measures that rely on high humidity is inefficient under drought conditions", thereby avoiding searching invalid regions.

[0098] Step S4: Obtain actual feedback data on the strategy to strengthen the prevention and control of shallow trench risks, and optimize the feedback based on the actual feedback data.

[0099] Specifically, the enhanced shallow ditch risk prevention and control strategy is implemented, and corresponding actual feedback data is obtained. The actual feedback data includes at least the actual environmental change data after the implementation of the strategy, such as surface micro-topography change data, vegetation coverage data, soil volumetric water content, temperature, electrical conductivity and groundwater level, as well as carbon flux data such as soil CO2 and CH4 flux.

[0100] Furthermore, based on the actual feedback data, the corresponding parameters of the twin environment of shallow ditch on sloping farmland and the intelligent agent cluster for shallow ditch risk prevention and control are optimized.

[0101] In one possible embodiment, an ensemble Kalman filter algorithm or assimilation algorithm is used to fuse the actual observation data in the feedback data with the predicted results of the sloping farmland ditch twin environment. By adjusting the state variables and key process parameters of the sloping farmland ditch twin environment, the generated simulation trajectory is made to approximate the observed facts as closely as possible. For example, if the observation data shows that the simulated erosion is significantly higher than the measured value in a specific soil texture area, the assimilation algorithm is used to backpropagate the error, inferring that the "aggregate water stability parameter" of the soil in this area may be underestimated, and the posterior distribution of this parameter in the model is automatically calibrated. At the same time, the measured carbon flux pulse data is used to calibrate the parameters related to soil disturbance in the carbon cycle process.

[0102] The strategy, environment, and outcome triplet corresponding to the actual implementation of the shallow ditch risk prevention and control strategy will be used as a new high-quality training sample to fine-tune the parameters of the fitness function adopted by the shallow ditch risk prevention and control agent cluster, so that its prediction of future strategy performance is closer to the actual benefits. Furthermore, the triggering conditions and effect intensity of each shallow ditch risk mode in the risk mode memory of the shallow ditch twin environment of sloping farmland will be corrected based on actual feedback data.

[0103] Figure 3 This is a schematic diagram of a risk prevention and control management system for shallow ditches on sloping farmland according to the present invention.

[0104] Specifically, a risk prevention and control management system for shallow ditches on sloping farmland includes:

[0105] An environment construction module is used to acquire multidimensional sloping farmland shallow ditch data and construct a twin environment of sloping farmland shallow ditch based on the multidimensional sloping farmland shallow ditch data.

[0106] The risk simulation module is used to perform risk simulation and strategy exploration using a shallow trench risk prevention and control intelligent agent cluster based on a neuroevolutionary architecture, and to generate a candidate shallow trench risk prevention and control strategy set.

[0107] The strategy generation module is used to construct a correlation map of shallow ditches on sloping farmland and to strengthen the candidate shallow ditch risk prevention and control strategy set to generate a strengthened shallow ditch risk prevention and control strategy.

[0108] The feedback optimization module is used to obtain actual feedback data on the strategy for strengthening shallow trench risk prevention and control, and to perform feedback optimization based on the actual feedback data.

[0109] It should be noted that the above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0110] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0111] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0112] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for risk prevention and control management of shallow ditches on sloping farmland, characterized in that, It includes the following steps: Acquire multidimensional shallow ditch data of sloping farmland, and construct a ditch twin environment of sloping farmland based on the multidimensional shallow ditch data. The ditch twin environment of sloping farmland includes a ditch risk extrapolator and a soil carbon flow monitor. A cluster of shallow ditch risk prevention and control intelligent agents based on a neuroevolutionary architecture is used to conduct risk simulation and strategy exploration, generating a candidate shallow ditch risk prevention and control strategy set; Construct a shallow ditch association map of sloping farmland, and strengthen the candidate shallow ditch risk prevention and control strategy set based on the shallow ditch association map to generate a strengthened shallow ditch risk prevention and control strategy. Obtain actual feedback data on strategies to strengthen shallow trench risk prevention and control, and optimize based on the actual feedback data.

2. The method for risk prevention and control management of shallow ditches on sloping farmland according to claim 1, characterized in that, Acquire multidimensional shallow ditch data of sloping farmland, and construct a twin environment of shallow ditches of sloping farmland based on the multidimensional shallow ditch data, including: The multidimensional sloping farmland shallow ditch data includes at least topographic geometric data, soil attribute data, vegetation attribute data, environmental meteorological data, and historical sloping farmland shallow ditch risk case data. A training dataset is generated based on historical multidimensional sloping farmland shallow ditch data. The training dataset is used to train the shallow ditch risk extrapolator and the soil carbon flow monitor. The risk extrapolation kernel of the shallow trench risk extrapolator is constructed based on the WEPP model of the water erosion prediction project, and the spatiotemporal correction field is constructed using Fourier neural operators. The spatiotemporal correction field is used to perform residual repair on the risk extrapolation kernel. A phased learning strategy is adopted to train the shallow ditch risk extrapolator with the goal of minimizing the preset risk extrapolation loss function until the preset risk extrapolation qualification conditions are met and the training ends. A soil carbon flow monitor was built based on the Century model, and a multi-scale graph neural network was used to model shallow gullies on sloping farmland to generate corresponding soil carbon cycle variation maps. A training strategy combining multi-task learning and meta-learning is adopted, and the soil carbon flow monitor is trained simultaneously with a preset multi-task weighted loss function until the preset carbon flow monitoring qualification conditions are met.

3. The method for risk prevention and control management of shallow ditches on sloping farmland according to claim 2, characterized in that, The method further includes: The spatiotemporal correction field is subjected to pattern extraction using a spectral clustering algorithm to generate multiple shallow trench risk pattern nodes; Counterfactual intervention reasoning is used to perform causal mining on the shallow trench risk pattern nodes to construct a risk pattern map; Based on physical information neural network and knowledge distillation, shallow ditch risk extrapolator and soil carbon flow monitor are distilled to generate shallow ditch twin environment of sloping farmland. The risk pattern map and the soil carbon cycle variation map are encoded into a risk pattern memory for the shallow ditch twin environment of sloping farmland. The risk pattern memory contains multiple shallow ditch risk patterns, each of which consists of a risk pattern map and the corresponding risk conditions in the soil carbon cycle variation map. The state initialization of the twin environment of the shallow ditch in the sloping farmland is performed based on real-time multidimensional data of shallow ditch in the sloping farmland.

4. The method for risk prevention and control management of shallow ditches on sloping farmland according to claim 1, characterized in that, A shallow ditch risk prevention and control intelligent agent cluster based on a neuroevolutionary architecture is used for risk simulation and strategy exploration, generating a candidate shallow ditch risk prevention and control strategy set, including: Risk monitoring is conducted based on a pre-defined risk monitoring mechanism to obtain the original risk anomaly dataset; The original risk anomaly dataset is matched with the shallow ditch risk patterns stored in the risk pattern memory within the shallow ditch twin environment of sloping farmland. If a match is found, it is determined to be a known risk pattern, and a calibration task is generated. If the matching fails, it is determined to be an unknown risk pattern. Spectral clustering and constraint-based causal discovery algorithms are used to explore pattern features and generate a set of verification tasks. The shallow ditch risk prevention and control intelligent agent cluster explores candidate strategies based on the set of calibration and verification tasks, and generates a candidate shallow ditch risk prevention and control strategy set.

5. A method for risk prevention and control management of shallow ditches on sloping farmland according to claim 4, characterized in that, The shallow ditch risk prevention and control intelligent agent cluster explores candidate strategies based on the calibration and verification task set, generating a candidate shallow ditch risk prevention and control strategy set, including: The policy network of each shallow ditch risk prevention agent in the shallow ditch risk prevention agent cluster is represented as a spatiotemporal convolutional long short-term memory network. The strategy network takes the state vector corresponding to the twin environment of shallow ditch on sloping farmland as input and outputs the original candidate shallow ditch risk prevention and control strategy. The shallow ditch risk prevention and control intelligent agent cluster uses an evolutionary algorithm to simultaneously combine the calibration task and the verification task set to perform forward simulation and obtain the corresponding simulation results. The fitness of the shallow ditch risk prevention agent is evaluated based on the preset multi-objective evaluation function and the simulation results, and the agent that meets the preset fitness conditions is selected as the parent agent. The parent agent is subjected to population evolution operation until a preset evolution termination condition is reached, generating a candidate shallow trench risk prevention and control strategy set.

6. The method for risk prevention and control management of shallow ditches on sloping farmland according to claim 1, characterized in that, A shallow ditch association map of sloping farmland is constructed, and the candidate shallow ditch risk prevention and control strategy set is enhanced based on the shallow ditch association map to generate enhanced shallow ditch risk prevention and control strategies, including: Hypergraph modeling is performed based on a candidate shallow ditch risk prevention and control strategy set, and information propagation is performed by combining hypergraph convolutional neural networks to construct a shallow ditch association map of sloping farmland. Based on the perturbation injection mechanism, cognitive uncertainty perturbation is injected into the twin environment of shallow ditch in sloping farmland to generate a perturbed twin environment of shallow ditch in sloping farmland. Simulate the execution of candidate shallow ditch risk control strategy sets in both undisturbed and disturbed sloping farmland shallow ditch twin environments; Based on the simulation results and the pre-set robustness assessment mechanism, the top N candidate shallow ditch risk prevention and control strategies in terms of robustness were selected as primary shallow ditch risk prevention and control strategies. Based on the aforementioned primary shallow ditch risk prevention and control strategy, multi-objective strategy optimization is carried out simultaneously by combining the shallow ditch correlation map of sloping farmland to generate an enhanced shallow ditch risk prevention and control strategy.

7. A method for risk prevention and control management of shallow ditches on sloping farmland according to claim 6, characterized in that, Hypergraph modeling is performed based on a candidate shallow ditch risk prevention and control strategy set, and information propagation is performed using a hypergraph convolutional neural network to construct a shallow ditch association map of sloping farmland, including: The candidate shallow trench risk prevention and control strategy set is analyzed and the pattern is extracted to generate the corresponding candidate strategy atom set; Among them, the atomic attributes of each candidate strategy atom in the candidate strategy atom set include at least spatial unit information, governance measure type, implementation intensity information, and activation context condition information; Based on the candidate strategy atom set and the corresponding atom attributes, a correlation map of shallow ditches in sloping farmland is constructed. A hypergraph convolutional neural network was used to propagate and aggregate the attribute information of each node in the original sloping farmland shallow ditch association map to construct the sloping farmland shallow ditch association map. Among them, the shallow ditch association map of sloping farmland contains at least candidate strategy nodes, land unit nodes, environmental status nodes, and resource constraint nodes. The edge connections between nodes in the shallow ditch association graph of sloping farmland include at least implementation relationship edges, conditional relationship edges, competitive relationship edges, and collaborative relationship edges.

8. The method for risk prevention and control management of shallow ditches on sloping farmland according to claim 1, characterized in that, Obtain actual feedback data on strategies to strengthen shallow trench risk prevention and control, and optimize these strategies based on the actual feedback data, including: Implement the enhanced shallow ditch risk prevention and control strategy and obtain corresponding actual feedback data; The actual feedback data includes at least the actual environmental change data and carbon flux data after the strategy was implemented; Based on the actual feedback data, the corresponding parameters of the twin environment of shallow ditches on sloping farmland and the intelligent agent cluster for shallow ditch risk prevention and control are optimized.

9. A risk prevention and control management system for shallow ditches on sloping farmland, used to implement the method described in any one of claims 1 to 8, characterized in that, include: An environment construction module is used to acquire multidimensional sloping farmland shallow ditch data and construct a twin environment of sloping farmland shallow ditch based on the multidimensional sloping farmland shallow ditch data. The risk simulation module is used to perform risk simulation and strategy exploration using a shallow trench risk prevention and control intelligent agent cluster based on a neuro-evolutionary architecture, and to generate a candidate shallow trench risk prevention and control strategy set. The strategy generation module is used to construct a shallow ditch association map of sloping farmland and to strengthen the candidate shallow ditch risk prevention and control strategy set to generate a strengthened shallow ditch risk prevention and control strategy. The feedback optimization module is used to obtain actual feedback data on the strategy for strengthening shallow trench risk prevention and control, and to perform feedback optimization based on the actual feedback data.

Images