Grassland degradation ecological risk assessment method based on multi-agent model

Abstract

This invention relates to the field of ecological environment assessment and spatial simulation technology, and discloses a method for grassland degradation ecological risk assessment based on a multi-agent model. The method includes a spatial risk assessment module that acquires multi-source basic data and divides it into discrete grid cells, calculating a landscape ecological risk index to generate spatial distribution results; a driving mechanism identification module that extracts driving factors to construct an improved regression tree model and outputs nonlinear response rules; and a multi-agent dynamic simulation module that constructs grassland environmental entities based on grid cells, injects risk indices and response rules, generates herder behavior entities, and inputs management policy parameters. Iterative calculations are performed using these two types of entities to simulate the dynamic process of herder relocation and resource consumption. After the iteration reaches the target year, the grid landscape attributes are reclassified according to the remaining grass cover, and the data is re-input into the assessment module for aggregation calculations, outputting a spatial distribution map to complete the evolution assessment. This invention can intuitively test the intervention effect of management policies on grassland degradation.

Description

Technical Field

[0001] This invention relates to the field of ecological environment assessment and spatial simulation technology, specifically to a method for assessing the ecological risk of grassland degradation based on a multi-agent model. Background Technology

[0002] Grassland ecosystems play a fundamental role in maintaining regional ecological balance, but grassland degradation is becoming increasingly prominent due to the dual impacts of natural succession and human activities. Accurately assessing the ecological risks of grassland degradation is a prerequisite for formulating ecological protection policies. Conventional landscape ecological risk assessments largely rely on remote sensing and basic geospatial data, extracting landscape pattern indices to construct spatial risk distribution models that reflect the macroscopic state of disturbance in the region.

[0003] However, grassland degradation is a complex process driven by climate, geographical conditions, and socio-economic activities such as grazing. Existing risk assessment methods, when quantifying these driving mechanisms, typically employ simple linear superposition or subjective weighting, which fails to effectively capture the nonlinear impacts and critical abrupt changes caused by natural factors and human activities on the ecological risk system, resulting in insufficient explanatory power of the models in understanding the risk evolution mechanism.

[0004] Meanwhile, current assessment methods are mostly limited to static profile analyses of historical or current time points, failing to establish a dynamic feedback process between human activities and the evolution of underlying natural resources. Due to the lack of characterization of micro-level individual behavior, existing models cannot simulate the spatial movement and resource consumption paths of herders under different grass conditions, environmental attractiveness, and policy constraints. This static research paradigm, detached from the underlying human-land interaction, cannot realistically predict the long-term evolutionary trends of grassland systems under continuous disturbance. Consequently, management departments find it difficult to conduct forward-looking spatial quantification and comparative simulations of the intervention effects of specific policies such as free grazing, grassland-livestock balance, or grazing bans and rotational grazing before implementing controls, limiting the guiding value of risk assessment results for practical work. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a grassland degradation ecological risk assessment method based on a multi-agent model. This method solves the technical problems of inaccurate risk spatial evolution prediction and policy control effect simulation caused by the difficulty in depicting the dynamic interaction between human grazing behavior and the underlying environment, as well as the neglect of the nonlinear characteristics of driving factors in existing grassland ecological risk assessment methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a method for assessing the ecological risk of grassland degradation based on a multi-agent model.

[0008] This method acquires multi-source basic data of the target area through a spatial risk assessment module and divides the target area into discrete grid cells. The multi-source basic data includes basic geographic vector data, landscape component raster data, elevation model data, meteorological raster data, and socioeconomic distribution data. Based on the landscape component raster data, the spatial risk assessment module extracts geometric distribution characteristic parameters of different landscape types within each discrete grid cell, calculates fragmentation index, separation index, and dominance index, and weights and fuses these indices to generate the landscape disturbance degree for each landscape type within the grid cell. Subsequently, a landscape vulnerability index is assigned based on the natural succession sensitivity of each landscape type, and the landscape loss index is obtained by multiplying the landscape vulnerability index by the landscape disturbance degree. Finally, the landscape ecological risk index is calculated by combining the area data of each landscape type within the discrete grid cell and the landscape loss index, outputting the spatial distribution results of the landscape ecological risk.

[0009] The driving mechanism of ecological risk is quantified. The driving mechanism identification module receives multi-source basic data and spatial distribution results of landscape ecological risk, extracting natural environmental characteristics, socio-economic characteristics, and spatial distance characteristics as driving factors. This module performs spatiotemporal alignment processing, mapping each feature element contained in the driving factors to discrete grid cells, and constructing an independent variable matrix by aggregating the mean of continuous variables within the discrete grid cells. A boosting regression tree model is constructed with the driving factors as independent variables and the landscape ecological risk index as the dependent variable. During the model training phase, the mean squared error is used as the loss function for node splitting and iterative training. A ten-fold cross-validation mechanism is used to perform grid search on hyperparameters such as tree complexity, learning rate, and maximum number of trees. The training stream is truncated when the validation set error no longer decreases for several consecutive iterations, and the optimal boosting regression tree model is output. This model quantifies the relative contribution of each driving factor to the change of the landscape ecological risk index and extracts nonlinear response rules reflecting the risk mutation boundary.

[0010] The multi-agent dynamic simulation module constructs an entity system using underlying data. This module assigns initial grassland biomass, elevation, slope, and water source distance to each discrete grid cell, and injects landscape ecological risk indices and nonlinear response rules to generate grassland environmental entities. Simultaneously, based on population and livestock statistics from socioeconomic distribution data, it generates herder behavior entities at the corresponding discrete grid cell settlement locations, assigns initial livestock numbers to these entities, and inputs management policy parameters including constraints on grazing ban periods, grazing ban areas, and carrying capacity.

[0011] During the dynamic evolution phase, the multi-agent dynamic simulation module performs iterative calculations using grassland environmental entities and herder behavior entities. Under preset grassland-livestock balance management and grazing ban / rotational grazing protection scenarios, when the landscape ecological risk index of a certain grid or the livestock density generated by herder behavior entities exceeds the critical mutation point defined in the nonlinear response rule, the model adjusts the management policy parameters to issue a spatial grazing ban or limit the maximum carrying capacity to that grid, and enforces a mandatory blockade on herder behavior entities entering high-risk grids. Within a single time step, the model combines grassland biomass resource sufficiency score, mobility ease score, neighborhood resource agglomeration effect score, and policy constraint multiplier determined based on management policy parameters to calculate the comprehensive attraction utility value of herder behavior entities when moving to surrounding discrete grid cells, and uses a nonlinear mapping algorithm to convert it into a movement probability.

[0012] Based on movement probabilities, the model controls herder entities to perform spatial movement and trigger resource consumption logic, deducting the total daily grazing amount corresponding to the number of livestock from the grassland biomass reaching the grid. When the remaining grassland biomass reaching the grid is less than the total daily grazing amount, the grassland biomass of that grid is reset to zero, and the herder entities are forced to move outward in the next time step. Simultaneously, the model controls grassland environmental entities to execute resource regeneration logic based on meteorological data, calling the logistic growth equation constrained by meteorological data to calculate natural growth, using the intrinsic growth rate to reflect the effect of hydrothermal conditions on vegetation growth, and updating the grassland biomass status.

[0013] Once the iterative calculations reach the target year, the multi-agent dynamic simulation module extracts the remaining grassland biomass from all discrete grid cells and converts it into vegetation cover. Based on the decline in vegetation cover and remaining grassland biomass, the landscape attributes of the discrete grid cells are discretized and reclassified to generate updated landscape component data. Finally, the updated landscape component data is re-input into the spatial risk assessment module to perform spatial aggregation calculations, outputting a spatial distribution map of grassland degradation ecological risk for the target year as the system's evolutionary assessment result.

[0014] This invention provides a method for assessing the ecological risk of grassland degradation based on a multi-agent model. It has the following beneficial effects:

[0015] 1. This invention utilizes a driving mechanism identification module to construct an enhanced regression tree model, generate nonlinear response rules, and inject them into grassland environmental entities. Compared with the traditional linear superposition assessment method, it can accurately extract the nonlinear impact of natural and socio-economic factors on ecological risks. The identified risk critical mutation points are used as underlying constraints to directly participate in subsequent simulations, thereby improving the accuracy of grassland ecological risk assessment under complex driving factors.

[0016] 2. This invention introduces addressing and resource consumption logic based on attraction utility value into the multi-agent dynamic simulation module; the model deducts the biomass of the grid grassland in real time according to the number of livestock, and forcibly triggers the relocation addressing of herders according to the remaining grass status; this design establishes a direct interaction mechanism between grazing behavior and underlying grassland resources, and solves the problem that conventional static assessment cannot present the dynamic degradation process of grassland under continuous human interference.

[0017] 3. This invention presets different grazing control scenarios by inputting management policy parameters, and after simulating reaching the target year, it reclassifies landscape attributes using remaining grass and vegetation coverage, outputting an updated risk spatial distribution map. This closed-loop design transforms abstract management policies into quantitative constraints on grid and agent evolution, allowing direct comparison of grassland degradation under different policy interventions, and providing intuitive spatiotemporal evolution data support for ecological protection decisions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention;

[0019] Figure 2 This is a schematic diagram of the system architecture of the present invention;

[0020] Figure 3 This is a schematic diagram comparing the evolution trends of ecological risks under different grazing management policies according to the present invention.

[0021] Among them, 100 is the space risk assessment module; 200 is the driving mechanism identification module; and 300 is the multi-agent dynamic simulation module. Detailed Implementation

[0022] The technical solutions in 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 embodiments of the present invention, and not all 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 protection of the present invention.

[0023] Please see the appendix Figure 1 As the basic supporting architecture of the present invention, the present invention provides a grassland degradation ecological risk assessment and policy response simulation system, which may include: a spatial risk assessment module 100, a driving mechanism identification module 200 and a multi-agent dynamic simulation module 300.

[0024] Specifically, the spatial risk assessment module 100 is used to acquire multi-source basic data for the target area. This multi-source basic data includes basic geographic vector data, landscape component raster data, elevation model data, meteorological raster data, and socioeconomic distribution data. The spatial risk assessment module 100 discretizes continuous spatial data into grid cells containing attribute information by dividing the target area into equally spaced grids.

[0025] Furthermore, the spatial risk assessment module 100 is also used to calculate the landscape pattern index based on the area proportion of landscape components within each grid unit, thereby quantifying the vulnerability and disturbance degree of each landscape type. The spatial risk assessment module 100 generates the landscape ecological risk index for each grid unit through weighted calculation and outputs the spatial distribution results of landscape ecological risk in the target area.

[0026] In this embodiment, the driving mechanism identification module 200 is communicatively connected to the spatial risk assessment module 100. The driving mechanism identification module 200 receives the aforementioned multi-source basic data and the spatial distribution results of landscape ecological risks. The driving mechanism identification module 200 extracts natural environmental characteristics, socio-economic characteristics, and spatial distance characteristics from the multi-source basic data as driving factors.

[0027] To achieve accurate fitting of nonlinear relationships, the driving mechanism identification module 200 uses grid cells as a basis, extracts each driving factor as an independent variable, and uses the landscape ecological risk index as a dependent variable to construct an enhanced regression tree model. The driving mechanism identification module 200 trains the model and optimizes parameters through cross-validation, calculates the relative percentage contribution of each driving factor to the distribution of landscape ecological risk, and outputs the importance ranking of driving factors and nonlinear response rules.

[0028] The multi-agent dynamic simulation module 300 is communicatively connected to the spatial risk assessment module 100 and the driving mechanism identification module 200, respectively. The multi-agent dynamic simulation module 300 is used to construct the simulation environment. This simulation environment includes grassland environmental entities representing the surface state and herder behavior entities representing human activities.

[0029] Based on this, the multi-agent dynamic simulation module 300 receives the spatial distribution results of landscape ecological risk output by the spatial risk assessment module 100 and the nonlinear response rules output by the driving mechanism identification module 200, and maps these results and rules as attribute parameters to each grassland environmental entity. The multi-agent dynamic simulation module 300 configures movement addressing rules and resource foraging rules for the herder behavior entities.

[0030] As a specific application scenario, the multi-agent dynamic simulation module 300 receives externally input management policy parameters. These parameters are configured as policy constraint variables that limit the activity boundaries and intensity of herder entities. The multi-agent dynamic simulation module 300 performs long-term iterative calculations according to a set time step, records data generated during entity interactions in real time, and outputs an evolutionary evaluation result of the grassland system state.

[0031] See appendix Figure 2 This method relies on the aforementioned grassland degradation ecological risk assessment and policy response simulation system. The grassland degradation ecological risk assessment and policy response simulation method provided by this invention may include the following steps:

[0032] In step S100, the spatial risk assessment module 100 acquires the spatial and attribute data of the target area and divides the target area into discrete grid units of a preset size. The spatial risk assessment module 100 calculates the landscape ecological risk index of each grid unit using preset landscape loss quantification rules, and generates and outputs a spatial distribution map of the landscape ecological risk level.

[0033] In step S201, the driving mechanism identification module 200 extracts topographic, climate, population, economic, and distance features at the corresponding grid cell scale, and constructs an independent variable matrix containing multidimensional driving factors. The driving mechanism identification module 200 uses the landscape ecological risk index generated in step S100 as the observation target and inputs it into the preset boosting regression tree algorithm framework.

[0034] In step S202, the driving mechanism identification module 200 iteratively trains the improvement regression tree algorithm framework by setting different combinations of decision tree depth and learning step size parameters. The driving mechanism identification module 200 determines the optimal prediction model under the condition of minimizing the fitting bias, and uses this optimal prediction model to extract the relative importance of driving factors, generating risk response rules for different environmental constraints.

[0035] In step S301, the multi-agent dynamic simulation module 300 constructs a grassland environment entity based on grid cells and injects the landscape ecological risk index output in step S100 and the risk response rules output in step S202 into the grassland environment entity. The multi-agent dynamic simulation module 300 generates corresponding herder behavior entities at the set starting grid positions based on preset population distribution data.

[0036] In step S302, the multi-agent dynamic simulation module 300 introduces time-dimensional constraints on grazing ban periods, spatial constraints on grazing ban areas, and subject-dimensional constraints on livestock carrying capacity. The multi-agent dynamic simulation module 300 applies these constraints to the operational logic of the herder's behavioral entity according to a preset scenario combination scheme.

[0037] In step S303, the multi-agent dynamic simulation module 300 triggers model execution with a daily time step. Within a single time step, the grassland environmental entity executes resource regeneration and degradation judgment logic based on climate input; the herder behavior entity executes spatial movement and foraging consumption logic based on terrain resistance and resource distribution. After a set simulation cycle, the multi-agent dynamic simulation module 300 extracts the final state data of the environmental entities and outputs policy response assessment results characterizing the total system resource quantity, degradation area, and spatial aggregation characteristics.

[0038] See appendix Figure 1 The spatial risk assessment module 100 provided by this invention quantifies the degree of disturbance and potential risks to the ecological environment of a target area. During the risk assessment and spatial mapping process, the spatial risk assessment module 100 specifically performs the following calculation and processing steps.

[0039] In step S101, the spatial risk assessment module 100 acquires basic geographic and landscape component data of the target area and performs scale division of the evaluation units. When conducting large-scale land use ecological risk assessment, the physical scale of the evaluation units directly determines the accuracy of the spatial pattern characteristics. If the grid division is too fine, it easily disrupts the physical integrity of individual landscape patches; if the grid is too large, it will obscure the ecological heterogeneity of local spaces. Based on the above general technical principles, to ensure the integrity of the map patches, the spatial risk assessment module 100 uses 2 to 5 times the average area of ​​the landscape patches in the study area as the benchmark range for the study scale. As a preferred specific implementation, this embodiment divides the entire study area into 4091 uniform grid units, with each grid unit having an area of ​​2.8 km by 2.8 km. This parameter configuration achieves a balance between spatial computational efficiency and the fidelity of local heterogeneity.

[0040] In step S102, after establishing the underlying computational grid system, the spatial risk assessment module 100 extracts geometric distribution characteristic parameters of different landscape types within each divided grid cell, and calculates the fragmentation index and separation index, which characterize the landscape pattern. The fragmentation index is used to characterize the severity of spatial discrete segmentation of a specific ecological landscape, reflecting the dissecting effect of external disturbances on continuous habitats. The formula for calculating this index is as follows:

[0041] ;

[0042] In the above formula, Represents the fragmentation index; Refers to the first The number of patches in the landscape; This represents the total area of ​​this type of landscape. To ensure the completeness of the underlying algorithm logic, when the extracted total landscape area... When the value approaches zero, the space risk assessment module 100 replaces the denominator with a preset minimum tolerance constant to avoid triggering a computer division-by-zero anomaly.

[0043] Besides the fragmentation of the landscape, the spatial barrier between patches also affects the energy flow of the system. Based on this, the spatial risk assessment module 100 calculates the separation index, and its formula is as follows:

[0044] ;

[0045] In the above formula, Represents the separation index; Represents the total area of ​​the entire study region; For the aforementioned extracted first Total area of ​​landscape types; Refers to the first The number of patches within a landscape class. A higher value for this indicator indicates greater spatial dispersion of patches within the same landscape class. Similarly, regarding the denominator in this formula... In some parts, the system synchronously applies the aforementioned minimum tolerance error prevention mechanism.

[0046] Step S103: To further identify the dominant position of specific landscapes in regional ecological succession, the spatial risk assessment module 100 calculates the dominance index, the formula of which is as follows:

[0047] ;

[0048] ;

[0049] In the above formula, The target of the operation on the left side of the equals sign is the first... The dominance index of landscape types; This represents the theoretical maximum value of the diversity index; in the summation operation of the numerator on the right side of the equation, the internal sign... It's a plaque. The percentage of the area to the total area of ​​the study region, indicated by the numbering here. Used as an internal summation index; This represents the total number of plaques; This represents the upper limit of the number of patches used for summation. Here, the principle of information entropy is introduced to calculate diversity, where the area ratio of a certain patch... When the value approaches zero, the system sets a very small positive lower limit threshold for participating in logarithmic operations to prevent the generation of meaningless singular solutions.

[0050] Step S104: A single pattern index often fails to fully reflect the combined effects of external stresses. Therefore, the spatial risk assessment module 100 uses a weighted fusion of the calculated fragmentation index, separation index, and dominance index to generate the landscape disturbance degree. The calculation formula is as follows:

[0051] ;

[0052] In the above formula, Refers to the degree of landscape disturbance; This represents the fragmentation index calculated above; Represents the separation index; Representative dominance index; , and This represents the weight of each indicator. In this embodiment, considering that the disruption of habitat physical connectivity has the most direct driving effect on ecological function loss, the system assigns the highest basic weight to fragmentation. Commonly assigned weights for each indicator are... It equals 0.5. It equals 0.3. It equals 0.2.

[0053] Step S105: Different types of underlying land cover have varying resistance to external environmental shocks due to differences in their internal structures. The spatial risk assessment module 100 assigns a landscape vulnerability index to various landscape types based on the sensitivity of different ecosystem types to natural succession. The system's preset assignment standard is as follows: types frequently subjected to human intervention or lacking vegetation protection are assigned higher values. According to different ecosystem types, grassland is assigned a value of 0.33, forest land 0.21, cultivated land 0.44, water area 0.56, construction land 0.1, and unused land 0.7.

[0054] The spatial risk assessment module 100 obtains the landscape loss index by multiplying the landscape vulnerability index by the landscape disturbance degree, in order to comprehensively characterize the potential degradation potential of this type of landscape. The calculation formula is as follows:

[0055] ;

[0056] In the above formula, The landscape loss index; This represents the aforementioned landscape vulnerability index; This refers to the landscape disturbance level calculated above.

[0057] Step S106: Based on the loss assessment of each individual landscape type, the spatial risk assessment module 100 integrates the area data of each landscape type within each evaluation unit to calculate the landscape ecological risk index of each grid. The calculation formula is as follows:

[0058] ;

[0059] In the above formula, The comprehensive landscape ecological risk index for the grid in which it is located; Representing the evaluation unit Interior landscape type The area; For unit The total area; The landscape loss index extracted above; This represents the total number of landscape types contained within the evaluation unit. For a specific grid located at an irregular boundary of the study area, its effective coverage area... If the value falls below the set geometric lower limit threshold, the system determines that the mesh is invalid and skips the weighted calculation to ensure the reliability of the global operation.

[0060] After completing the calculations for the entire grid unit, the spatial risk assessment module 100 assigns the calculated landscape ecological risk index value for each grid unit to the center position of each grid. The subsequent process of generating a spatially continuous distribution map and its classification based on this center point value can be implemented using conventional geographic information system interpolation techniques, which are well-known technologies in the field and will not be elaborated upon here.

[0061] See appendix Figure 1 After obtaining the spatial distribution characteristics of landscape ecological risks in the target area, in order to reveal the physical driving mechanism behind the evolution of the above spatial risk pattern and extract threshold rules that can be used for policy simulation, the driving mechanism identification module 200 provided by this invention specifically performs the following data processing and model calculation steps.

[0062] In step S201, the driving mechanism identification module 200 constructs an independent variable matrix containing multidimensional driving factors at a grid scale consistent with the aforementioned spatial risk assessment. The ecological risk pattern of a specific region is not only constrained by the carrying capacity of the underlying natural environment but also directly affected by the intensity of human socio-economic activities. Based on the above physical causal relationships, the driving mechanism identification module 200 extracts three types of driving factor features from multi-source heterogeneous data as input variables for the model. In terms of the selection of specific input parameters, natural environmental features include elevation and slope parameters reflecting topographic relief, as well as temperature and precipitation raster data reflecting hydrothermal conditions. These factors directly determine the natural resilience of the regional ecosystem. Socio-economic features include population density, GDP distribution, and livestock density data reflecting local carrying capacity pressure. These parameters are used to characterize the direct consumption of surface vegetation by human resource extraction behavior. Spatial distance features include distance from major water sources and distance from main roads to characterize the guiding effect of resource accessibility on the grazing trajectory of herders.

[0063] Considering the inherent differences in acquisition cycle and spatial accuracy among the aforementioned multi-source data, direct fusion would lead to singularities or severe biases in matrix operations. Therefore, the driving mechanism identification module 200 performs spatiotemporal alignment processing. In the time dimension, the system uses the remote sensing interpretation year of the landscape ecological risk index as the time base, extracting or interpolating to generate corresponding socioeconomic data slices for that year. In the spatial dimension, resampling and spatial interpolation techniques are used to uniformly map all feature elements to the aforementioned fixed-size discrete grid cells. As a preferred approach, the driving mechanism identification module 200 aggregates the mean of continuous variables within the grid, thereby constructing a dimension... The matrix of independent variables, where Represents the total number of grid cells in the entire domain. This represents the total number of extracted driving factor features. Simultaneously, the landscape ecological risk index value of this grid is used as the corresponding model observation label to complete the assembly of training sample pairs.

[0064] Step S202: Traditional linear multivariate fitting methods struggle to capture the nonlinear response of ecosystems to environmental stresses, and collinearity often exists among multidimensional driving factors. To overcome these limitations, the driving mechanism identification module 200 introduces a boosting regression tree model. This type of model is based on the gradient boosting algorithm framework. By sequentially connecting multiple simple decision trees, each newly added decision tree specifically fits the residuals between the predictions of all previous trees and the true labels, thereby approximating complex nonlinear functions without increasing the complexity of a single tree. Based on the above general technical principles, the driving mechanism identification module 200 configures the model structure as an additive ensemble based on multiple classification and regression trees. Its core operational logic is limited by the following iterative approximation formula:

[0065] ;

[0066] In the above formula, This represents the predicted landscape ecological risk index output by the model after all iterations of training. This represents a multidimensional driving factor feature vector corresponding to a certain grid cell, and its feature dimension is the total number of driving factor features extracted in step S201. ; This represents the total number of base learners integrated within the model, which is the total number of decision trees. Index of iteration steps; This represents the learning rate parameter, whose value is set within the interval (0,1), and is used to control the learning rate. The step size of the model structure update in each iteration is used to prevent overfitting caused by excessive approximation in a single step. Representative at the The single classification and regression weak decision tree model constructed during the iteration.

[0067] In step S203, the driving mechanism identification module 200 divides the grid dataset containing multi-dimensional driving factor feature vectors and landscape ecological risk index labels into training and validation sets, and performs iterative training and hyperparameter optimization of the boosting regression tree model. During this training process, the system uses mean squared error as the specific loss function to measure the model's prediction bias. In each model iteration, the current decision tree module splits nodes with the optimization objective of reducing the cumulative residual gradient generated by all previous decision trees, thereby achieving gradual convergence of the overall prediction error.

[0068] To ensure the model's generalization ability and lock in the true ecological response threshold, the driving mechanism identification module 200 introduces a ten-fold cross-validation mechanism to perform grid search on three core hyperparameters: tree complexity, learning rate, and maximum number of trees. The tree complexity parameter determines the maximum depth of a single decision tree, its physical meaning corresponding to the number of levels at which interaction between factors is allowed. In this embodiment, considering that the interaction between ecological variables typically does not exceed five orders, the system limits the search range of tree complexity to 1 to 5; the learning rate is configured with a small step size parameter within the range of 0.001 to 0.1. During training, the driving mechanism identification module 200 monitors the loss function change trajectory on the validation set in real time. When the validation set error no longer decreases for several consecutive iterations, an early stopping mechanism is triggered to truncate the training flow, and the parameter configuration at this point is output as the optimal prediction model.

[0069] Step S204: Based on the trained optimal prediction model, the driving mechanism identification module 200 extracts and quantifies the relative contribution of each input feature to changes in ecological risk. The system extracts the sum of the squared error loss reduction brought about by each feature variable when it is a split node within all decision trees, and normalizes the sum of the reduction of each feature variable, outputting the relative importance percentage of each driving factor. This multi-dimensional weighted judgment result eliminates the interference of single statistical extrema, objectively reflects the dominance weight of different environmental and social factors on risk formation, and provides data support for the subsequent selection of key control factors.

[0070] In step S205, the driving mechanism identification module 200 uses the optimal prediction model to generate a biased dependency data column for the core driving factors with the highest relative importance. The logic of this biased dependency operation is that, under the constraint of keeping other feature variables at the mean of the entire sample, the system traverses the effective value range of a single target factor and inputs it into the model, recording the corresponding predicted change trajectory of the landscape ecological risk index. Based on this, the driving mechanism identification module 200 outputs nonlinear response rules reflecting the risk abrupt change boundary. For example, the system can identify the turning point where grassland degradation ecological risk exponentially increases when livestock density or slope exceeds a certain specific value. The extracted relative importance weights and the nonlinear response rules containing specific numerical boundaries are then encapsulated into structured data and transmitted to the downstream multi-agent dynamic simulation module 300 as core parameters for configuring management policies and restricting herders' grazing behavior.

[0071] See appendix Figure 1 After identifying the nonlinear mechanisms and key thresholds driving the evolution of grassland degradation ecological risks, in order to assess the future evolution trajectory of grassland resources and their degradation risks under different grazing management policy interventions, the multi-agent dynamic simulation module 300 provided by this invention specifically performs the following data processing and iterative deduction steps.

[0072] In step S301, the multi-agent dynamic simulation module 300 completes the initial configuration of grassland environmental entities and herder behavior agents within the underlying spatial grid system. The core technical principle of grassland ecological spatial simulation lies in adopting a bottom-up modeling approach, decomposing macro-grassland degradation into micro-level herder grazing behaviors under specific policy and resource constraints. The multi-agent dynamic simulation module 300 uses the 2.8 km orthogonal discrete grid constructed in the previous steps as the underlying grassland environmental cell, assigning each grid initial basic geographical attributes such as grassland biomass, elevation, slope, and distance to water sources. Based on population and livestock statistics for the target area, the system generates herder behavior entities representing human activities at the corresponding settlement grid locations and assigns an initial livestock number to each herder agent. Furthermore, the system defines a management agent responsible for implementing ecological baseline control, used to issue policy constraints to herders according to risk rules.

[0073] In step S302, the multi-agent dynamic simulation module 300 configures differentiated grazing policy simulation scenarios based on the nonlinear response rules output in step S200. To comprehensively evaluate the regulatory effectiveness of policy guidance on grassland degradation, the system presets three scenarios: unrestricted free grazing, grassland-livestock balance management, and grazing ban / rotational grazing protection. Under the unrestricted free grazing scenario, herders are driven solely by terrain and grass availability for grazing, without introducing additional policy intervention variables. Under the grassland-livestock balance and grazing ban / rotational grazing scenarios, the management agent directly invokes the risk mutation threshold identified by the aforementioned improved regression tree model. For example, when the ecological risk index or livestock density of a grid exceeds the critical mutation point output by the model, the management agent will be activated and issue a spatial grazing ban or limit on the maximum carrying capacity to that grid. This ban / restriction is manifested in the underlying algorithm as a forced blockade of herders entering specific high-risk grids; its physical purpose is to cut off the continuous grazing damage to the fragile grassland ecosystem by livestock.

[0074] In step S303, the multi-agent dynamic simulation module 300 calculates the comprehensive attraction utility value for each herder's behavioral entity across the entire domain to move and locate in surrounding grids at each time step. The herder's decision to find grazing land is essentially a comprehensive consideration of grassland resource abundance, movement resistance, and policy restrictions. The system quantifies the current state of the [number]th step using the following mathematical expression. The herders of the first grid to the first The combined attraction utility value of moving each candidate grid:

[0075] ;

[0076] In the above formula, This indicates that the herders are from the current grid. Move to candidate grid The overall attraction utility value for grazing; Representing the The current grassland biomass resource sufficiency score of each candidate grid; Represents the current grid To candidate grid The mobility score is negatively correlated with the spatial distance between the two locations and the slope resistance, which is converted into a positive suitability score here; The score representing the neighborhood resource clustering effect of the candidate grid; , and These represent the weights of the three indicators in the herder's decision-making evaluation (usually each weight is set to be in the (0,1) range and the sum is 1). This represents the policy constraint multiplier issued by the management agent in step S302. If the candidate grid... If the area is a designated no-grazing zone or its livestock carrying capacity has reached its limit, triggering the red line, It is forcibly assigned a value of 0 to achieve absolute spatial prohibition; if the redline constraint is not triggered, then... The value is assigned to 1.

[0077] In the aforementioned comprehensive utility assessment, the neighborhood resource clustering effect reflects herders' preference for contiguous high-quality pastures. The multi-agent dynamic simulation module 300 calculates this neighborhood effect using the following local perception formula:

[0078] ;

[0079] In the above formula, Indicates the first Each candidate grid is a specific detection window radiating outward from the center (usually set to a 3x3 mole neighborhood). This represents the index of the surrounding neighborhood grid contained within the detection window; Indicates the first The actual grassland biomass of each neighborhood grid at the current moment; This is a conditional indicator function; it is invoked when the internal logical condition is true. Greater than or equal to the preset threshold for the lush grass period Output 1 if the condition is met, otherwise output 0. Represents the detection window The total number of valid grid cells contained within. When candidate grid cells... Located at the absolute boundary of the study area, resulting in a certain total number of effective grids within the detection window. When the value approaches zero, the system determines that the area lacks supporting neighboring resources and forcibly... Assign a value of 0 to avoid triggering a division-by-zero exception.

[0080] In step S304, after quantifying the herders' willingness to move to each surrounding grid, the multi-agent dynamic simulation module 300 executes the herders' final spatial movement and resource foraging actions through a probabilistic roulette mechanism. In complex grazing behavior, herders' decisions are not absolutely rational and are subject to uncertainty due to incomplete information. Therefore, the system uses a nonlinear mapping algorithm to transform the comprehensive attraction utility value into the herders' destination selection probability, the calculation formula of which is as follows:

[0081] ;

[0082] In the above formula, Represents the current grid The herders choose to move to the grid at the current time step. The probability of; Represents an exponential function with the natural constant as its base; This represents the overall attraction utility value output from the preceding steps. Represents the current grid herders to the first The combined attraction utility value when each candidate grid moves; To control the degree of rationality in herders' decision-making; This represents the total number of valid candidate grids within the maximum single movement radius of a herder. As a local summation index for traversing each candidate grid in the denominator; For a preset, extremely small positive real number (e.g., 10) 6 ), used for tolerance truncation, to prevent program crashes caused by the denominator approaching zero due to the minimal utility of all surrounding grids being de-grazing.

[0083] After obtaining the movement probability, the system determines the movement probability based on the probability of each candidate grid. The numerical values ​​are accumulated within the interval [0,1] to construct a continuous probability distribution line segment, generating pseudo-random numbers that follow a uniform distribution. The specific sub-interval into which the random number falls determines the actual destination of the herder's movement. After the herder reaches the target grid, the system triggers resource consumption logic, deducting the daily grazing allowance corresponding to the number of livestock from the grassland biomass of that grid.

[0084] Specifically, to standardize the consumption of different livestock species, the multi-agent dynamic simulation module 300 uses preset livestock conversion coefficients to uniformly convert all types of livestock owned by herders (such as cattle, sheep, and horses) into standard sheep units. Based on this, the total daily feed intake of a grid is configured as the product of the total number of standard sheep units held by the herder and the average daily roughage intake quota per standard sheep (e.g., 1.8 kg / day). If the remaining grassland biomass of the current grid is less than this total daily feed intake, the system will reset the grid's biomass to zero and forcibly trigger the herder's relocation action at the next time step.

[0085] In step S305, the multi-agent dynamic simulation module 300 executes the aforementioned state update logic with a daily feeding step size and an annual growth step size, iteratively advancing the simulation cycle until the preset target prediction year is reached. At natural growth cycle nodes, the grassland environmental entity executes resource regeneration logic based on climate data. The system introduces a logistic growth equation constrained by climate factors to calculate natural growth, and its state update formula is set as follows:

[0086] ;

[0087] In the above formula, and These represent the remaining grassland biomass of the grid before and after the update, respectively. Represents the intrinsic growth rate of grassland, which is obtained by dynamically mapping precipitation and temperature raster data within the current time step to reflect whether hydrothermal conditions promote or inhibit vegetation growth. This represents the maximum grassland biomass carrying capacity of the grid under ideal natural conditions; This represents the total cumulative feed intake of the herder's livestock during that natural growth cycle.

[0088] Once the simulation process reaches the target year, the multi-agent dynamic simulation module 300 extracts the remaining grass and vegetation cover of each grid across the entire domain, outputting a spatial distribution matrix of grassland degradation under various policy scenarios. To achieve data integration with the risk assessment module, the system introduces preset landscape type succession rules. Based on the decline in vegetation cover or remaining biomass, the system discretizes and reclassifies the landscape attributes of the grids (for example, the system sets 15% of the grassland cover baseline as the threshold for extremely severe irreversible degradation. When a grid's coverage calculated from its remaining biomass falls below this threshold due to continuous overgrazing, the system forcibly overwrites the landscape code of that grid from grassland to unused land / bare land in the underlying landscape attribute matrix). The system re-inputs the updated global landscape component raster data base into the aforementioned spatial risk assessment module 100, performs spatial aggregation calculations of landscape disturbance and vulnerability again, and finally outputs a spatial distribution map of grassland degradation ecological risk driven by different grazing control scenarios in the target year. By comparing the shrinkage of degraded area and the shift trajectory of risk center under multiple scenarios, the system can quantitatively analyze the actual marginal contribution of different livestock carrying capacity control and grazing ban policies to curb regional ecological degradation.

[0089] Specific application examples:

[0090] Overview and basic assessment of the study area:

[0091] The system selected a typical alpine meadow area with a total area of ​​10,000 square kilometers as the target region. The spatial risk assessment module 100 divided this region into 1275 grid units of 2.8 km × 2.8 km. By extracting landscape components (grassland, woodland, bare land, etc.) from 2023, calculating fragmentation, separation, and vulnerability, the system generated a landscape ecological risk index for the baseline year. The assessment results show that in the baseline year (2023), the area of ​​severe ecological risk in this region was 1000 square kilometers, mainly concentrated near water sources and along both sides of main roads. (See reference...) Figure 3 The vertical axis of the chart represents the total area (in square kilometers) of grassland within the target region that has been identified as being at severe ecological risk. The initial base value for this vertical axis is 1000 square kilometers, as determined by the assessment. The larger this value, the more severe the grassland degradation and the wider the area of ​​ecological damage; the smaller the value, the better the ecosystem is in a state of recovery.

[0092] Driver mechanism mining and rule extraction

[0093] The driving mechanism identification module 200 extracted independent variables such as elevation, slope, distance from water source, and livestock density, and input them into the landscape ecological risk index to train the improvement regression tree model. The model output the relative importance results: livestock density (45%) > slope (25%) > distance from water source (15%) > elevation (15%). Simultaneously, the model extracted a key nonlinear response rule (policy threshold): when the livestock density in a local grid exceeds 3.5 SU / hm²... ² When the standard sheep unit is reached per hectare, the landscape ecological risk index of the grid will undergo an exponential mutation, and the probability of degradation will increase sharply.

[0094] Multi-agent scenario simulation

[0095] The multi-agent dynamic simulation module 300 initialized 500 herder agents in a grid, with an initial total livestock quantity of 50,000 standard sheep units. The system was set to a simulation period of 10 years (2023-2033) with a step size of days. Combined with... Figure 3 As can be seen, the horizontal axis of the chart represents the time process of the dynamic evolution of the multi-agent system. The starting point is 2023, and the ending point is 2033. A time observation node is set every two years to illustrate the ecological evolution process over a long time series. The system simulates two comparative scenarios (the legend box is located in the upper left corner of the chart, used to correspond to the mapping relationship between the two scenarios and the line type):

[0096] Unrestricted free grazing scenario: Herders rely solely on the amount of grass and distance (overall attraction utility value) Free movement foraging, without policy restrictions (policy constraint multiplier) ).

[0097] Grassland-livestock balance management scenario: Introduction of a management intelligent agent. When herders attempt to enter areas where the vegetation density has reached 3.5 SU / hm²... ² When the nonlinear response rule threshold is in the grid, forced triggering of entry prohibition blocking (policy constraint multiplier) is performed. 0).

[0098] Comparison of simulation results and corroboration by graphs

[0099] After 10 years (3650 daily steps) of iterative computation of the underlying algorithm, the final evolution result output by the system is... Figure 3 A precise graphical representation was obtained:

[0100] The evolutionary outcome of unrestrained free grazing (corresponding to) Figure 3The solid black line marked with a solid square: In an unrestricted free-grazing scenario, excessive concentration of herders in prime pastures leads to the depletion of a large amount of grid biomass and triggers landscape reclassification (grassland degrades into bare land). By 2033, the area of ​​severely ecologically risky zones will expand dramatically to 1420 square kilometers (a 42% increase from the baseline year). Map interpretation: In this scenario, herders are not subject to any policy intervention (…). They pursue lush pastures entirely on their own will. This is reflected in... Figure 3 In the map, the solid black line marked with a solid square shows a steep and obvious upward trend (growing continuously from 1,000 square kilometers to 1,420 square kilometers). This feature visually reveals that a lack of scientific management leads to excessive livestock gathering, which in turn causes irreversible degradation of local grasslands, ultimately resulting in a sharp expansion of the overall area at severe ecological risk.

[0101] The evolution of grassland-livestock balance management (corresponding) Figure 3 (Black dashed line marked with a hollow circle): Under the grassland-livestock balance management scenario, due to the strict implementation of the 3.5 SU / hm extracted based on the lifting regression tree model... ² The threshold is set at which livestock are effectively moved to grids with excess carrying capacity. Grasslands recover naturally using the logistic growth equation. By 2033, the area of ​​severely ecologically risky zones will shrink to 850 square kilometers (a 15% reduction from the baseline year). Map interpretation: When local livestock density reaches the warning line identified by the model, the system forcibly prohibits new herders from entering. 0). Reflected in Figure 3 In the map, the black dotted line marked with hollow circles shows a gentle downward trend (gradually shrinking from 1,000 square kilometers to 850 square kilometers). This feature strongly demonstrates that the management rules effectively alleviated the pressure of grazing, giving degraded grasslands time and space to recover and regenerate according to natural laws, thereby effectively reducing the overall ecological risk of the region.

Claims

1. A method for grassland degradation ecological risk assessment based on multi-agent model, characterized in that, include: The spatial risk assessment module (100) acquires multi-source basic data of the target area, divides the target area into discrete grid units, calculates the landscape ecological risk index of the discrete grid units based on the multi-source basic data, and generates and outputs the spatial distribution results of landscape ecological risk. The driving mechanism identification module (200) receives the multi-source basic data and the spatial distribution results of the landscape ecological risk, extracts driving factors from the multi-source basic data, constructs an enhanced regression tree model with the driving factors as independent variables and the landscape ecological risk index as dependent variables, and outputs nonlinear response rules through the enhanced regression tree model. The multi-agent dynamic simulation module (300) constructs a grassland environment entity based on the discrete grid cell, injects the landscape ecological risk index and the nonlinear response rule into the grassland environment entity, generates a herder behavior entity in the discrete grid cell, obtains management policy parameters and inputs the management policy parameters into the herder behavior entity, performs iterative calculations using the grassland environment entity and the herder behavior entity, and outputs the evolution evaluation results of the grassland system state. The multi-agent dynamic simulation module (300) assigns initial grassland biomass, elevation, slope and water source distance to each discrete grid cell based on the multi-source basic data, and combines the injected landscape ecological risk index and the nonlinear response rule to generate the grassland environmental entity. The multi-agent dynamic simulation module (300) generates the herder behavior entity at the settlement location of the corresponding discrete grid cell based on the population and livestock statistics in the socio-economic distribution data in the multi-source basic data, and assigns the initial livestock stock quantity and the management policy parameters to the herder behavior entity. Within a single time step of the iterative operation, the multi-agent dynamic simulation module (300) calculates the comprehensive attraction utility value of the herder behavior entity when it moves to the surrounding discrete grid cells for addressing. The comprehensive attraction utility value is obtained by comprehensively quantifying the grassland biomass resource sufficiency score, mobility convenience score, neighborhood resource agglomeration effect score, and policy constraint multiplier determined based on the management policy parameters. The multi-agent dynamic simulation module (300) uses a nonlinear mapping algorithm to convert the comprehensive attraction utility value into the movement probability of the herder behavior entity; The multi-agent dynamic simulation module (300) controls the herder behavior entity to perform spatial movement according to the movement probability and triggers resource consumption logic. It deducts the total daily feed intake of the herder behavior entity corresponding to the number of livestock from the grassland biomass of the grassland environment entity corresponding to the arrived discrete grid cell as the remaining grassland biomass. When the remaining grassland biomass of the arrived discrete grid cell is less than the total daily foraging amount, the grassland biomass of the arrived discrete grid cell is reset to zero, and the herder behavior entity is forced to perform an addressing action to move outward in the next time step.

2. The method according to claim 1, wherein, The multi-source basic data includes basic geographic vector data, landscape component raster data, elevation model data, meteorological raster data, and socioeconomic distribution data; The spatial risk assessment module (100) extracts geometric distribution characteristic parameters of different landscape types within each discrete grid cell based on the landscape component raster data, calculates the fragmentation index and separation index, and calculates the dominance index. The spatial risk assessment module (100) uses the fragmentation index, the separation index and the dominance index to generate the landscape disturbance degree of each landscape type within the grid unit by weighted fusion; The spatial risk assessment module (100) assigns a landscape vulnerability index to each landscape type based on the natural succession sensitivity of each landscape type, obtains a landscape loss index by multiplying the landscape vulnerability index by the landscape disturbance degree, and calculates the landscape ecological risk index by combining the area data of each landscape type in the discrete grid cell and the landscape loss index.

3. The method according to claim 1, wherein, The driving factors include natural environmental characteristics, socio-economic characteristics, and spatial distance characteristics; The driving mechanism identification module (200) performs spatiotemporal alignment processing, maps each feature element contained in the driving factor to the discrete grid cell, and aggregates the continuous variables in the discrete grid cell by mean to construct the independent variable matrix; The driving mechanism identification module (200) uses the mean squared error as a loss function to measure the model prediction bias, performs node splitting, and executes iterative training and hyperparameter optimization of the boosting regression tree model.

4. The method according to claim 3, wherein, The hyperparameters include tree complexity, learning rate, and maximum number of trees; The driving mechanism identification module (200) introduces a ten-fold cross-validation mechanism to perform grid search on the hyperparameters. When the validation set error no longer decreases for several consecutive iterations, the training stream is truncated, and the optimal boosting regression tree model is output. The driving mechanism identification module (200) uses the optimal boosting regression tree model to extract and quantify the relative contribution of each driving factor to the change of the landscape ecological risk index, and generates the nonlinear response rule that reflects the risk mutation boundary.

5. The method according to claim 1, wherein, The management policy parameters include grazing ban period constraints, grazing ban area constraints, and livestock carrying capacity constraints. The multi-agent dynamic simulation module (300) presets unrestrained free grazing scenarios, grass-livestock balance management scenarios, and grazing ban and rotational grazing protection scenarios. During the execution of the iterative calculation, under the grassland-livestock balance management scenario and the grazing ban and rotational grazing protection scenario, when the landscape ecological risk index of a certain grid or the livestock density generated by the herder's behavior entity exceeds the critical mutation point defined in the nonlinear response rule, the management policy parameters are adjusted to issue a spatial grazing ban or limit the maximum carrying capacity to the certain grid, and the underlying algorithm performs forced blocking on the herder's behavior entity entering the high-risk grid.

6. The method according to claim 1, wherein, During the execution of the iterative operation, the multi-agent dynamic simulation module (300) controls the grassland environment entity to execute resource regeneration logic based on the meteorological data in the multi-source basic data; The resource regeneration logic calls the logistic growth equation constrained by the meteorological data to calculate the natural growth rate, so as to update the grassland biomass status of the grassland environmental entity. The logistic growth equation uses the intrinsic growth rate to reflect the promoting and inhibiting effects of hydrothermal conditions on vegetation growth.

7. The method according to claim 1, wherein, After the iterative calculation reaches the target year, the multi-agent dynamic simulation module (300) extracts the remaining grassland biomass of all grassland environmental entities in the discrete grid cells, calculates the vegetation cover based on the remaining grassland biomass, and discretizes and reclassifies the landscape attributes of the discrete grid cells based on the decrease in vegetation cover and the remaining grassland biomass to generate updated landscape component data. The multi-agent dynamic simulation module (300) re-inputs the updated landscape component data into the spatial risk assessment module (100) to perform spatial aggregation calculations and outputs the spatial distribution map of grassland degradation ecological risk for the target year as the evolution assessment result of the grassland system state.

Images