An ecological space risk intelligent evaluation method coupled with plant functional traits

Abstract

The application relates to the technical field of ecological space risk assessment and intelligent management and control, and discloses an ecological space risk intelligent assessment method coupled with plant functional traits, acquires multi-source heterogeneous data of a target ecological space, and constructs a unified space-time benchmark regional ecological risk database; the target ecological space is divided into multiple partitions according to predetermined partition conditions; the regional ecological risk database is input into a pre-constructed improved ecological risk index model, a weighted summation model or a graph neural network model is automatically selected according to the data characteristics of the target ecological space, and the ecological risk index of each partition is calculated; the ecological risk index of each partition is compared with a preset risk early warning threshold, a risk grade is divided, a hotspot area is identified through spatial superposition analysis, and an ecological risk space distribution map and a graded early warning report are generated. The application improves the precision and applicability of ecological risk assessment by adaptively selecting an assessment model and deeply coupling with plant functional traits.

Description

Technical Field

[0001] This application relates to the field of ecological space risk assessment and intelligent management technology, and in particular to an intelligent ecological space risk assessment method coupled with plant functional traits. Background Technology

[0002] Currently, with the increasing intensity of land development and the intensifying impact of climate change, the risks facing ecological spaces are becoming increasingly complex. Taking the construction of a highway next to an ecological park as an example, the human development stress during construction and operation will cause differentiated ecological damage to different areas within the park: the primary forest area has dense vegetation and deep root systems, making it highly resistant to disturbance; the wetland restoration area is in the middle of its recovery phase, with moderate resilience; while the artificial turf area has monoculture vegetation and exposed soil, resulting in poor resilience. Under the same external stresses (such as dust, soil erosion, and human disturbance), the actual ecological risks in the three areas differ greatly. However, traditional ecological risk assessment methods often overlook the differences in the inherent properties of vegetation, making it difficult to accurately quantify the moderating effect of this inherent vulnerability on risk outcomes. More broadly, the risks faced by ecological spaces are manifested as a complex superposition of multiple risk factors, such as ecological damage, environmental pollution, natural disasters, and human development stress. The response of ecosystems to risks exhibits significant spatial heterogeneity and nonlinearity. Vegetation communities in different regions show drastically different resistance and resilience under the same external stress due to differences in their intrinsic plant functional traits (such as specific leaf area, root depth, stress tolerance index, leaf nitrogen content, canopy closure, soil and water retention capacity, etc.).

[0003] Existing ecological risk assessment methods are mainly divided into two categories: one is a comprehensive evaluation method based on statistical indices or weighted summation, which selects indicators such as land use type, slope, soil texture, and vegetation cover, uses the analytic hierarchy process (AHP) or entropy weight method to determine weights, and then linearly superimposes them to obtain an ecological risk index; the other is an assessment method that uses machine learning or spatial analysis techniques, such as using remote sensing imagery to invert vegetation cover change, using geographic information systems for spatial overlay analysis, and introducing Bayesian networks or random forest models for risk probability prediction. Some studies have also begun to explore combining multi-source remote sensing data to invert vegetation physiological parameters to characterize ecosystem status.

[0004] However, all of the above methods have significant shortcomings. Weighted summation models assume independent contributions and linear superposition of risk factors, making it difficult to characterize nonlinear coupling effects such as synergistic amplification or antagonistic inhibition between factors. They also rarely consider the moderating role of intrinsic plant community resilience in risk transmission. While machine learning methods can fit complex relationships, they often rely on large training samples, and the models lack interpretability, making it difficult to directly trace risk assessment results back to specific ecological processes or vegetation characteristics. Furthermore, existing assessment methods typically employ fixed model architectures, failing to adaptively select the most suitable computational model based on the data completeness and spatial heterogeneity of the target ecological space, thus limiting the accuracy and applicability of assessments in different application scenarios. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides an intelligent ecological space risk assessment method coupled with plant functional traits to solve the technical problems of insufficient accuracy of ecological risk assessment, inability to quantify vulnerability correction mechanisms, lack of nonlinear coupling modeling, and poor interpretability of assessment results caused by ignoring the inherent stress resistance differences of plant communities in the prior art.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the main technical solutions adopted in this application include:

[0009] This application provides an intelligent ecological space risk assessment method coupled with plant functional traits, the specific steps of which include:

[0010] S1, the target ecological space is divided into multiple partitions according to the predetermined partitioning conditions; multi-source heterogeneous data of each partition are obtained and normalized to obtain multi-source heterogeneous data of a unified spatiotemporal reference for each partition, which serves as the basic data of each partition.

[0011] S2, Based on the basic data of each partition, the ecological risk index of each partition is calculated using a pre-constructed ecological risk index model;

[0012] The ecological risk index model is selected as an improved weighted summation model or a pre-trained improved graph neural network model based on the data characteristics in the basic data of each partition.

[0013] The improved weighted summation model introduces a vulnerability correction coefficient based on plant functional traits to multiplicatively adjust the weighted summation result;

[0014] The improved graph neural network model multiplicatively regulates the messages sent by nodes during message transmission using stress resistance regulation factors determined based on plant functional traits.

[0015] S3. Compare the ecological risk index of each partition with the preset risk warning threshold to obtain the risk level of each partition and generate a warning report.

[0016] Optionally, in some embodiments of this application, the ecological risk index model selects an improved weighted summation model or a pre-trained improved graph neural network model based on the data characteristics in the basic data of each partition, specifically including:

[0017] Each of the aforementioned partitions is divided into several grid units;

[0018] The effective data coverage C of the basic data of each partition is calculated using formula (1);

[0019] Formula (1) specifically includes: ;

[0020] Where, N total N represents the total number of grid cells within each of the aforementioned partitions. valid The number of raster cells in each of the aforementioned partitions that simultaneously possess plant functional trait data, meteorological element data, and human activity intensity data;

[0021] When the effective data coverage is lower than a preset threshold, an improved weighted summation model is used as the ecological risk index model; otherwise, a pre-trained improved graph neural network model is used as the ecological risk index model.

[0022] Optionally, in some embodiments of this application, when using an improved weighted summation model, the calculation steps for the ecological risk index include:

[0023] Based on the basic data of each partition, four types of ecological risk factors are extracted for each partition. The four types of ecological risk factors include ecological damage factors that characterize the degree of damage to the ecosystem structure, environmental pollution factors that characterize the degree of stress of pollutant input on the ecosystem, natural disaster factors that characterize the degree of impact of natural disasters on the ecosystem, and human development stress factors that characterize the degree of disturbance of human activities on the ecosystem.

[0024] The corresponding preset indicator data are extracted from the four types of ecological risk factors, and the combined weights of the four types of ecological risk factors are obtained by using the analytic hierarchy process coupled with the entropy weight method.

[0025] Plant functional trait parameters of each grid cell are extracted from the basic data of each partition and standardized to obtain standardized values ​​of each plant functional trait parameter in each grid cell; the stress resistance comprehensive score of each grid cell is calculated by weighted summation based on the standardized values.

[0026] The resilience level is determined based on the comprehensive resilience score, and the vulnerability correction coefficient is determined based on the preset level mapping relationship.

[0027] Extract the risk source intensity and ecosystem exposure of each raster unit corresponding to the four types of ecological risk factors from the basic data of each partition;

[0028] Calculate the ecological risk index ERI for each grid cell: ERI = Σ(W i ×S i ×E i ×V i ×V base );

[0029] Where i represents the i-th type of ecological risk factor; W i S represents the combined weight of the i-th type of ecological risk factor; i E represents the intensity of the risk source for the i-th type of ecological risk factor; i V represents the ecosystem exposure of the i-th type of ecological risk factor; i V is the vulnerability correction coefficient for the i-th type of ecological risk factor; base The vulnerability baseline coefficient is pre-set based on the land use type or soil type of each grid cell;

[0030] Based on the ecological risk index of each grid cell, the average ecological risk index of all grid cells within each partition is calculated as the ecological risk index of each partition.

[0031] Optionally, in some embodiments of this application, the method of coupling the analytic hierarchy process (AHP) with the entropy weight method to obtain the combined weights of the four types of ecological risk factors includes:

[0032] The subjective weights of the four types of ecological risk factors are obtained using the analytic hierarchy process (AHP), and the objective weights of the four types of ecological risk factors are obtained using the entropy weight method. An objective function is constructed with the goal of minimizing the information deviation between the subjective weights and the objective weights, and normalization constraints are introduced.

[0033] The objective function is solved using the Lagrange multiplier method to obtain the combined weights of the four types of ecological risk factors.

[0034] Optionally, in some embodiments of this application, when a pre-trained improved graph neural network model is used, the calculation steps of the ecological risk index include:

[0035] Multi-source spatial data layers are extracted from the basic data of each partition. The multi-source spatial data layers include a plant functional trait raster layer for characterizing the structure and functional characteristics of plant communities, a meteorological element raster layer for characterizing regional climate and hydrological conditions, a human activity intensity raster layer for characterizing the spatial distribution of human disturbance intensity, and a control partition vector map layer for defining the spatial range of each partition and the boundary of ecological management unit.

[0036] Using each grid cell as a node in the graph, the data values ​​corresponding to each grid cell are extracted from the plant functional trait grid layer, meteorological element grid layer, and human activity intensity grid layer, and then concatenated into the feature vector of each node.

[0037] A graph structure is constructed based on the control zone vector layer and the spatial adjacency relationships between each raster unit; the edges of the graph structure include spatial adjacency edges and ecological interaction edges;

[0038] The feature vectors of each node and the graph structure are input into a pre-trained graph neural network model for message passing, and the ecological risk index of each grid cell is output.

[0039] The ecological risk index of each grid cell is spatially aggregated according to the partition boundary in the control partition vector layer to obtain the ecological risk index of that partition.

[0040] Optionally, in some embodiments of this application, the message passing includes:

[0041] S11, extract the plant functional trait data corresponding to each node from the feature vector of each node;

[0042] S12, before sending a message to a neighboring node, each node calculates its stress resistance regulation factor based on the plant functional trait data;

[0043] S13, each node performs message passing; the message sent by each node to its neighboring node is equal to the result of linear transformation of the feature vector of each node multiplied by the corresponding resilience adjustment factor;

[0044] S14, each node aggregates the messages sent by its neighboring nodes through a graph attention mechanism and updates its own feature vector;

[0045] After repeating steps S11 to S14 a preset number of times, the ecological risk index of each grid cell is output.

[0046] Optionally, in some embodiments of this application, calculating the stress resistance regulating factor of each node based on the plant functional trait data includes:

[0047] Plant functional trait parameters are extracted from the feature vectors of each node;

[0048] The plant functional trait parameters are standardized, and the stress resistance comprehensive score of each node is calculated by weighted summation using preset weights.

[0049] When the overall stress resistance score is higher than the first threshold, the grid cell corresponding to the node is determined to be a high stress resistance cell, and the stress resistance adjustment factor is set to a value between 0.6 and 0.8, and the stress resistance adjustment factor is negatively correlated with the overall stress resistance score.

[0050] When the overall stress resistance score is lower than the second threshold, the grid cell corresponding to the node is determined to be a low stress resistance or degenerate cell, and the stress resistance adjustment factor is set to a value between 1.2 and 1.5, and the stress resistance adjustment factor is negatively correlated with the overall stress resistance score.

[0051] When the overall resilience score is between the first threshold and the second threshold, the resilience adjustment factor is set to 1.0; wherein the first threshold is greater than the second threshold.

[0052] Optionally, in some embodiments of this application, the training steps of the improved graph neural network model include:

[0053] S01, acquire training data from historical years, and extract historical multi-source spatial data layers from the training data. The historical multi-source spatial data layers include historical plant functional trait raster layers, historical meteorological element raster layers, historical human activity intensity raster layers, and historical control zone vector layers.

[0054] Using each of the grid cells as nodes, the data values ​​corresponding to each grid cell are extracted from the historical multi-source spatial data layer and concatenated to form the historical feature vector of each node.

[0055] S02, Construct a historical graph structure based on the historical control partition vector layer and the spatial adjacency relationship between each grid unit;

[0056] S03, Construct a heterogeneous graph neural network; the heterogeneous graph neural network includes a multilayer perceptron subnetwork, which is used to calculate stress resistance regulation factors based on the plant functional trait data of each node;

[0057] S04, input the historical feature vector and the historical graph structure into the heterogeneous graph neural network, and output the ecological risk index prediction value of each node;

[0058] S05, using the measured ecological loss data after historical disaster events as a monitoring signal, the error between the predicted value of the ecological risk index and the measured ecological loss data is calculated using a loss function. The parameters of the heterogeneous graph neural network are updated through backpropagation, and the iteration continues until convergence, resulting in a trained improved graph neural network model.

[0059] Optionally, in some embodiments of this application, the construction of the history graph structure further includes:

[0060] Each grid cell is expanded into four independent nodes, namely, a risk source node, an environment node, a plant functional trait node, and a receptor node;

[0061] Within the same grid cell, risk source nodes and environmental nodes, environmental nodes and plant functional trait nodes, and plant functional trait nodes and receptor nodes are connected through ecological interaction edges.

[0062] Different grid cells and nodes of the same type are connected by spatial adjacency edges; the spatial adjacency edges include a first type of spatial adjacency edge constructed based on geographical distance and a second type of spatial adjacency edge constructed based on hydrological flow direction.

[0063] Optionally, in some embodiments of this application, the loss function is a weighted sum of node regression loss, trait covariance constraint loss, hydrological connectivity constraint loss, and spatial smoothness constraint loss;

[0064] The node regression loss is the mean square error between the predicted ecological risk index of all nodes and the corresponding measured ecological loss data.

[0065] The trait covariation constraint loss is used to constrain the representation of plant functional trait nodes in the heterogeneous graph neural network, so that specific leaf area is positively correlated with leaf nitrogen content and stress tolerance index is positively correlated with root depth.

[0066] The hydrological connectivity constraint loss is used to constrain the representations of upstream and downstream nodes connected by second-type spatial adjacency edges, so as to minimize the difference between the representations of upstream and downstream nodes.

[0067] The spatial smoothing constraint loss is used to constrain the representations of adjacent nodes connected by first-type spatial adjacency edges, such that the difference in representations of adjacent nodes increases monotonically with increasing geographical distance.

[0068] (III) Beneficial Effects

[0069] This application presents an intelligent ecological space risk assessment method coupled with plant functional traits. Based on the data completeness and spatial heterogeneity of the target ecological space, it automatically selects a weighted summation model or a graph neural network model for calculation, avoiding the problem of insufficient applicability of a single model under different data conditions, and improving the adaptability and assessment accuracy of the assessment method to different ecological spaces.

[0070] By constructing a regional ecological risk database with a unified spatiotemporal benchmark and dividing the target ecological space into multiple zones, it is possible to achieve standardized processing of multi-source heterogeneous data and refined management of spatial units.

[0071] Based on the calculated regional ecological risk index, risk levels are classified through preset four-level early warning thresholds. Combined with spatial overlay analysis, hotspot areas with ecological risks significantly higher than those in neighboring areas are identified. The system can output intuitive risk spatial distribution maps and graded early warning reports, providing clear and operable decision-making basis for ecological and environmental management departments.

[0072] From data acquisition, zoning, adaptive model selection, risk index calculation to early warning report generation, a complete technical loop is formed. It does not depend on specific ecosystem types or regional scales, has good versatility and scalability, and can be widely applied to risk assessment and management of different types of ecological spaces. Attached Figure Description

[0073] Figure 1 This is a flowchart of an intelligent ecological space risk assessment method coupled with plant functional traits, as proposed in this application.

[0074] Figure 2 This is a flowchart of message passing in an improved graph neural network model in one embodiment of this application;

[0075] Figure 3 This is a flowchart illustrating the training process of an improved graph neural network model in one embodiment of this application. Detailed Implementation

[0076] To better explain and facilitate understanding of this application, a detailed description of its embodiments is provided below in conjunction with the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0077] Example 1:

[0078] A city plans to construct a two-way, four-lane highway approximately 12 kilometers long on the east side of an ecological park, with its closest point only 50 meters from the park boundary. The park comprises three typical ecological spaces: a primary forest area, a wetland restoration area, and an artificial turf area. The primary forest area is lush with vegetation, dominated by Masson pine and oak, and has a complete community structure. The wetland restoration area is primarily planted with reeds and cattails and is currently in the mid-recovery stage. The artificial turf area has a monoculture vegetation, mainly ryegrass, with significant soil exposure and poor resilience. The park management agency needs to assess the potential ecological risks that the highway construction and operation may pose to the park in order to develop differentiated management measures.

[0079] In this embodiment, the intelligent ecological space risk assessment system is deployed on the government cloud platform of the city's ecological environment big data center, and achieves data interconnection with the city's land and space planning system, meteorological monitoring network, remote sensing data receiving system and ground ecological monitoring station network. Figure 1 This is a flowchart of an intelligent ecological space risk assessment method coupled with plant functional traits, as described in this application. The following is a summary of the process. Figure 1 This embodiment details the specific steps. When the park management agency enters the highway site selection plan (including parameters such as route alignment, construction scale, and construction period) into the system and submits an evaluation request, the system is triggered and automatically executes the following evaluation steps:

[0080] Step S1: Divide the target ecological space into multiple partitions according to the predetermined partitioning conditions; obtain multi-source heterogeneous data of each partition and perform normalization processing to obtain multi-source heterogeneous data of a unified spatiotemporal benchmark for each partition, which serves as the basic data for each partition.

[0081] Specifically, the assessment target is a 1km radius around the ecological park and the road on both sides. Based on the results of the "three zones and three lines" delineation, the target ecological space is divided into three zones: a primary forest protection zone, a wetland restoration and control zone, and a general artificial turf control zone.

[0082] Then, multi-source heterogeneous data for each zone are acquired. Specifically, basic ecological spatial data are obtained from the national land "one map" system; Sentinel-2 multispectral images are obtained from the remote sensing data receiving system, and combined with ground sampling data, a random forest algorithm is used to generate raster layers of six types of plant functional traits (specific leaf area, root depth, stress tolerance index, leaf nitrogen content, canopy closure, and soil and water conservation capacity); meteorological data such as rainfall, temperature, and evapotranspiration are obtained from the meteorological monitoring network; in a preferred embodiment, the acquired meteorological element time series data are smoothed using a 7-day sliding window to eliminate short-term random fluctuations; at the same time, the 3σ principle is used to detect outliers, and monitoring values ​​exceeding the mean ± 3 times the standard deviation are marked and manually verified. Human activity stress data such as population density and road density in highway site selection schemes are obtained from planning departments.

[0083] Missing value imputation, outlier removal, WGS-84 coordinate unification, and 10m pixel resampling are performed on the multi-source data for each partition. The Min-Max normalization method is then used to normalize the indicators for each of the above data types. In a preferred embodiment, a timeliness labeling system for the multi-source data is established. Data with an acquisition time deviation exceeding 15 days is adjusted using a phenological correction model, uniformly correcting the multi-source data acquired at different times to the evaluation benchmark period.

[0084] After the above processing, multi-source heterogeneous data with a unified spatiotemporal reference for each region are obtained, which serve as the basic data for each region.

[0085] In a preferred embodiment, before standardization using the Min-Max normalization method, the preprocessing steps further include: radiometric calibration and atmospheric correction of the acquired Sentinel-2 multispectral image, converting DN values ​​into true surface reflectance; declouding of cloud-covered areas in the image using a dark channel prior method or a multi-temporal composite method, and eliminating cloud scattering interference in thin cloud areas using a homomorphic filtering method; geometrical fine correction of the corrected image and spatial registration with GPS coordinates of ground-measured sample points, uniformly resampling to a 10m spatial resolution; median composite of multi-temporal cloudless images from the peak plant growth season to generate a reference image; filling in missing sporadic data values ​​remaining after composite using inverse distance weighted interpolation; and outlier detection and removal of the retrieved plant functional trait raster based on a preset normal physiological range threshold. After the above preprocessing, Min-Max normalization is then performed on the various index data.

[0086] In a preferred embodiment, for the sporadic missing values ​​still existing in the inverted plant functional trait raster, a neighborhood median interpolation method based on vegetation type is adopted: if there are valid raster units of the same vegetation type within a 200m radius of the missing raster unit, the median of each trait of that type of raster unit is used to fill the missing value; otherwise, an inverse distance weighted interpolation method is used for estimation, and the interpolation weight is inversely proportional to the square of the distance.

[0087] This step eliminates the interference of atmospheric scattering and absorption on surface reflectance through radiometric calibration and atmospheric correction, making reflectance data comparable across images from different time phases and regions. Cloud removal addresses the issue of missing or distorted characteristic band data required for plant functional trait inversion caused by cloud cover in optical remote sensing images. Geometric correction and image registration ensure accurate spatial alignment of multi-source data (remote sensing images, ground samples, and meteorological stations). Multi-temporal median synthesis and missing value interpolation reduce the uncertainty caused by phenological fluctuations and revisit cycles in single-temporal data. Outlier detection and removal prevent interference from abnormal data exceeding normal physiological ranges due to residual atmospheric correction errors, topographic shading, or mixed pixels. These preprocessing steps collectively provide a high-quality, highly consistent, and spatially comprehensive data foundation for accurate inversion of plant functional trait raster layers and subsequent risk assessment.

[0088] Step S2: Based on the basic data of each partition, calculate the ecological risk index of each partition using a pre-constructed ecological risk index model.

[0089] The ecological risk index model selects either an improved weighted summation model or a pre-trained improved graph neural network model based on the data characteristics of the basic data of each partition, including:

[0090] First, each partition is divided into several grid cells;

[0091] The effective data coverage C of the basic data of each partition is calculated using formula (1);

[0092] Formula (1) specifically includes: ;

[0093] Where, N total N represents the total number of grid cells within each of the aforementioned partitions. valid The number of raster cells in each of the aforementioned partitions that simultaneously possess plant functional trait data, meteorological element data, and human activity intensity data;

[0094] When the effective data coverage is lower than a preset threshold, a weighted summation model is selected as the improved ecological risk index model; otherwise, a pre-trained graph neural network model is selected as the improved ecological risk index model.

[0095] Specifically, each zone is divided into several 10m×10m raster units. The number of raster units in each zone that have complete data on plant functional traits (all six traits are non-empty), meteorological element data (rainfall, temperature, and evapotranspiration are all non-empty), and human activity intensity data (population density, road density, and construction land expansion rate are all non-empty) is counted. The proportion of the number of raster units that simultaneously meet all three conditions to the total number of raster units in the zone is the effective data coverage rate of the zone.

[0096] In this embodiment, the preset ratio threshold is set to 60%. The system executes the following determination logic partition by partition:

[0097] For the primary forest reserve: the effective data coverage rate is 93.7%, which is higher than the preset proportion threshold. Moreover, the community structure is complex and spatially dependent. Therefore, a pre-trained graph neural network model is selected to construct an independent graph structure using the grid within the partition for inference.

[0098] For the wetland restoration and control area: the effective data coverage rate is 71.7%, which is higher than the preset proportion threshold. There is a spatial recursive relationship between the upstream and downstream hydrological connectivity and pollutant diffusion. Therefore, a pre-trained graph neural network model is selected to construct an independent graph structure based on the grid inside the area for inference.

[0099] For the general control area of ​​artificial turf: the effective data coverage rate is 42.0%, which is lower than the preset proportion threshold, and the vegetation is monolithic with low spatial variability. Therefore, the weighted summation model is selected.

[0100] Although both the primary forest reserve and the wetland restoration and control area use GNN, they each use the grid within their respective zones to construct independent graph structures, share the same set of pre-trained model parameters, and complete forward propagation calculations separately without interfering with each other.

[0101] Each of the three zones completed its assessment independently, and the ecological risk index for each zone was output.

[0102] The model selection process addresses the inherent differences in data completeness, monitoring network deployment years, and ecosystem structural complexity among different zones. It automatically adapts weighted summation models and graph neural network models using a quantitative criterion of effective data coverage. This allows graph neural networks to be used in areas with abundant data and complex spatial dependencies, such as primary forest reserves and wetland restoration management zones, to fully capture the cross-grid propagation and clustering effects of ecological risks. Conversely, in areas with sparse data or low spatial variability, such as general artificial turf management zones, the model automatically downgrades to a robust weighted summation model to avoid overfitting and result drift. This improves the accuracy and reliability of assessment results for each zone while avoiding the forced fitting and computational waste of graph neural networks on low-quality data. Furthermore, multiple zones share the same set of pre-trained model parameters and independently build and extrapolate models, further optimizing resource reuse efficiency in large-scale assessment scenarios. This enables the system to flexibly address the common challenge of uneven data accumulation across different zones in practical engineering, achieving adaptive model selection without manual intervention.

[0103] In a preferred embodiment, when a weighted summation model is selected, taking a general artificial turf control area as an example, the calculation steps for the ecological risk index of this area specifically include:

[0104] First, based on the basic data of each zone, four types of ecological risk factors and their corresponding specific indicator data are extracted, including:

[0105] Ecological damage factors: vegetation cover change rate (estimated difference before and after construction), soil erosion intensity (estimated soil loss after construction disturbance), and land degradation area ratio (the proportion of the area affected by construction to the total area of ​​the zone).

[0106] Environmental pollution factors: soil heavy metal content (Pb and Cd content, based on monitoring values ​​around the construction area), construction dust concentration (estimated daily average TSP concentration), and surface runoff suspended solids concentration (estimated SS concentration in runoff during the rainy season).

[0107] Natural disaster factors: frequency of rainstorms (average number of rainstorm days in the region per year), intensity of drought occurrence (average intensity index of drought events in the past 10 years), and number of days with strong winds (average number of days with wind speed ≥10m / s per year).

[0108] Human development stress factors: Disturbance area of ​​highway construction (area of ​​temporary land occupation and construction access roads), density of construction machinery (number of construction machines per unit area, units / km²) 2 Construction period (number of months) and frequency of transport vehicle traffic (average number of vehicles per day).

[0109] Then, using a coupled method of analytic hierarchy process (AHP) and entropy weighting, the combined weight W of each ecological risk factor was calculated. iExperts were invited to conduct pairwise comparisons of the relative importance of the four types of ecological risk factors, constructing a judgment matrix and calculating subjective weights. Simultaneously, monitoring values ​​of specific indicators were extracted from the basic data of the region, constructing an indicator data matrix and calculating objective weights using the entropy weight method. Finally, the Lagrange multiplier method was used to combine and optimize the subjective and objective weights to obtain the combined weights of each factor. The calculated values ​​were: ecological damage factor W1 = 0.27, environmental pollution factor W2 = 0.30, natural disaster factor W3 = 0.17, and human development stress factor W4 = 0.26.

[0110] Taking a typical grid unit in a general artificial turf control area adjacent to a highway construction area as an example (this area is actually divided into several grid units; this embodiment only uses one as an example, and the calculation process for the remaining grid units is the same), six types of plant functional trait parameters are extracted from the basic data of this partition: specific leaf area, root depth, stress tolerance index, leaf nitrogen content, canopy closure, and soil and water retention capacity. The six types of trait parameters are standardized using the Min-Max normalization method, and then weighted and summed according to preset weights. In this embodiment, the weights of the six traits are all the same, each taking 1 / 6. The calculated comprehensive stress resistance score of this grid unit is 0.33.

[0111] In this embodiment, the mapping relationship between the stress resistance level classification criteria and the vulnerability correction coefficient is pre-set based on a general method for evaluating plant stress resistance. Specifically, after obtaining a comprehensive stress resistance score by weighted summation of six types of plant functional trait parameters, and referring to the commonly used threshold range for classifying plant stress resistance levels, a comprehensive score ≥0.70 is set as a high stress resistance level, 0.40~0.70 is set as a medium stress resistance level, and ≤0.40 is set as a low stress resistance or degradation level. The vulnerability correction coefficient corresponding to each level is determined based on the ecological principle of high stress resistance community decline risk and low stress resistance community amplification risk. High stress resistance level corresponds to 0.6~0.8 (risk decline), medium stress resistance level corresponds to 1.0 (no correction), and low stress resistance or degradation level corresponds to 1.2~1.5 (risk amplification). In this embodiment, the comprehensive score of this typical grid unit is 0.33, which is judged as a low stress resistance / degradation level, and V is set to 1.45.

[0112] Extract the risk source intensity S of each ecological risk factor in the typical raster cell from the basic data of this partition. i and ecosystem exposure E i The risk source intensity of each factor is determined based on the actual spatial location and road distance of the grid unit, and the exposure level is uniformly set to 0.50 according to the control type of this zone. Ecological damage factor S1=0.75, environmental pollution factor S2=0.60, natural disaster factor S3=0.15, and human development stress factor S4=0.85.

[0113] Based on the land use type (artificial turf) of this typical grid unit, the vulnerability baseline coefficient V is determined. base In this embodiment, V base Based on pre-defined land use types, forest land has deep root systems and strong soil holding capacity, resulting in lower base vulnerability, so a base vulnerability coefficient of 0.3 is used; grassland has low vegetation cover and shallow root systems, resulting in higher base vulnerability, so a base vulnerability coefficient of 0.7 is used; the vulnerability base coefficient V for other land use types is... base Based on experience, the values ​​are set as follows: 0.5 for arable land, 0.9 for wetlands, 0.8 for water bodies, and 0.2 for construction land. The land use type of this typical grid cell is grassland, therefore V... base =0.7.

[0114] Calculate the ecological risk index ERI for each grid cell: ERI = Σ(W i ×S i ×E i ×V i ×V base ) = 0.319;

[0115] The ERI value of this grid cell is 0.319, which is in the medium-risk range.

[0116] After calculating the ERI value of all grid cells within the general control area of ​​the artificial turf, the arithmetic mean of the ERI values ​​of each grid cell was taken, resulting in an average ecological risk index of 0.69 for this zone. This value is in the high-risk range, indicating that the general control area of ​​the artificial turf faces a relatively high overall ecological risk due to highway construction disturbance.

[0117] This embodiment introduces a vulnerability correction mechanism based on plant functional traits into the weighted summation model. A comprehensive stress resistance score is obtained by weighted summation of six types of plant functional trait parameters, and this score is mapped to a vulnerability correction coefficient based on the stress resistance level. This upgrades vulnerability assessment from the traditional static assignment of land use type to a dynamic correction that integrates the actual stress resistance capacity of the plant community, effectively improving the accuracy of ecosystem vulnerability characterization. Simultaneously, the vulnerability correction coefficient is correlated with the baseline vulnerability coefficient V set based on land use type. base A two-layer vulnerability correction structure is formed, reflecting the differences in stress resistance at the plant community level and the differences in base vulnerability at the land cover level, respectively. The two work together to achieve a refined quantification of ecological vulnerability; combined weight W i The AHP-entropy weighting coupling method was adopted to determine the risk, taking into account both expert experience and objective data laws, and avoiding the bias of a single weighting method. Finally, the ERI calculation results at the grid scale were aggregated to the zoning scale, which not only preserved the spatial distribution details of the risk, but also formed an assessment result aligned with the boundary of the control zone.

[0118] In a preferred embodiment, when selecting a pre-trained improved graph neural network model, taking a primary forest reserve as an example, the specific steps for calculating the ecological risk index of this zone include:

[0119] Multi-source spatial data layers are extracted from the basic data of each partition. These multi-source spatial data layers include a plant functional trait raster layer, a meteorological element raster layer, a human activity intensity raster layer, and a control partition vector layer.

[0120] Among them, the plant functional trait raster layer, meteorological element raster layer and human activity intensity raster layer are used to construct the feature vector of each raster unit, and the control zone vector layer is used to construct the ecological interaction edge in the graph structure and aggregate the ecological risk index of each raster unit into the ecological risk index of each zone.

[0121] Each raster cell is used as a node in the graph. Data values ​​of each layer corresponding to the raster cell are extracted from the multi-source spatial data layer and concatenated to form the feature vector of the node.

[0122] A graph structure is constructed based on the control zone vector layer and the spatial location of the raster units; the edges of the graph structure include spatial adjacency edges and ecological interaction edges.

[0123] Specifically, the spatial adjacency edges include a first type of spatial adjacency edges constructed based on geographical distance and a second type of spatial adjacency edges constructed based on hydrological flow direction. The first type of spatial adjacency edges characterizes the mutual influence of geographical proximity on ecological processes; the edge weight is inversely proportional to the geographical distance between grid cells, meaning that the closer the grid cells are, the stronger the ecological association. In a preferred embodiment, the weight of the first type of spatial adjacency edges is assigned a distance attenuation coefficient γ based on the ecosystem type: γ=0.05 for forest ecosystems, γ=0.08 for wetland ecosystems, and γ=0.12 for grassland ecosystems. The edge weight calculation formula is w=exp(-γ×d), used to determine the influence strength of neighboring nodes on the current node during message passing, where d is the geographical distance between grid cells.

[0124] The second type of spatial adjacency edge is used to characterize the upstream-downstream ecological stress transmission effect brought about by hydrological connectivity. The edge direction is from upstream to downstream, and the edge weight is positively correlated with the intensity of the upstream influence on the downstream. That is, the ecological stress influence of the upstream grid cell on the downstream grid cell increases with the enhancement of hydrological connectivity.

[0125] The ecological interaction edge is constructed based on the vector layer of the control zone and is used to represent the ecological associations (such as material cycling, energy flow, species migration, etc.) between different grid units within the same control zone due to belonging to the same ecological management unit. Specifically, an ecological interaction edge is constructed between any two grid units belonging to the boundary of the same control zone, with the edge weight being a preset fixed value (1.0 in this embodiment). This ensures that each grid unit node within the same control zone receives equal message interaction weight, guaranteeing that information within the same zone can be fully propagated during the message transmission process of the graph neural network.

[0126] The merging of the three types of edges results in the final graph structure: the first type of spatial adjacency edge connects geographically adjacent raster cells; the second type of spatial adjacency edge connects upstream and downstream hydrological raster cells unidirectionally along the water flow direction; and the ecological interaction edge connects all raster cells within the same control zone. These three types of edges coexist and overlap during message passing, with each node simultaneously receiving messages from all three types of neighboring nodes and performing aggregation updates.

[0127] The feature vectors of each node and the graph structure are input into a pre-trained improved graph neural network model. Figure 2 This is a message passing flowchart of the improved graph neural network model in this embodiment, as shown below. Figure 2 As shown, the message passing process of the improved graph neural network model specifically includes the following steps:

[0128] Step S11 involves extracting the plant functional trait data corresponding to each node from its feature vector. Specifically, the feature vector of each node is composed of data values ​​from the multi-source spatial data layer corresponding to that raster unit, including plant functional trait data (including six types of traits: specific leaf area, root depth, stress tolerance index, leaf nitrogen content, canopy closure, and soil and water retention capacity). This step independently extracts these six types of trait parameters from the feature vector for subsequent calculation of stress resistance regulating factors.

[0129] Step S12: Before sending messages to neighboring nodes, each node calculates its stress resistance regulation factor based on the plant functional trait data. The specific calculation method is as follows:

[0130] First, six types of plant functional trait parameters are extracted from the feature vector of this node, including specific leaf area, root depth, stress tolerance index, leaf nitrogen content, canopy closure, and soil and water retention capacity. Then, the six types of plant functional trait parameters are standardized, and a comprehensive stress resistance score is calculated by weighted summation using preset weights.

[0131] Based on the calculated comprehensive stress resistance score, the value of the stress resistance adjustment factor is determined according to the following rules:

[0132] (1) When the comprehensive stress resistance score is higher than the first threshold, it is judged as a high stress resistance level. The stress resistance adjustment factor is in the range of 0.6 to 0.8, and the stress resistance adjustment factor is negatively correlated with the comprehensive stress resistance score.

[0133] (2) When the comprehensive stress resistance score is lower than the second threshold, it is judged as low stress resistance or degradation level. The stress resistance adjustment factor is in the range of 1.2 to 1.5, and the stress resistance adjustment factor is negatively correlated with the comprehensive stress resistance score.

[0134] (3) When the comprehensive stress resistance score is between the first threshold and the second threshold, the stress resistance adjustment factor is 1.0.

[0135] The first threshold is greater than the second threshold.

[0136] Taking a typical grid unit within a primary forest reserve as an example, this unit is dominated by Masson pine and Quercus glauca, with a first threshold of 0.70 and a second threshold of 0.40. After the above standardization process, a weighted summation is performed, resulting in a comprehensive stress resistance score of 0.82 for this unit, which is higher than the first threshold of 0.70, thus classifying it as having a high stress resistance level. According to a preset level mapping relationship, the stress resistance adjustment factor α corresponding to a high stress resistance level ranges from 0.6 to 0.8. In this embodiment, based on the relative position of the comprehensive stress resistance score within the interval, a linear interpolation method is used to calculate α = 0.68, meaning that when this node sends a message to its neighboring nodes, the message strength is attenuated to 68%.

[0137] Step S13: Each node performs message passing; the message sent by each node to its neighbor is equal to the result of a linear transformation of its feature vector multiplied by the corresponding resilience adjustment factor. It should be noted that in a standard graph neural network, the message sent by a node to its neighbor is the result of a linear transformation of its feature vector (i.e.,... The message strength is entirely data-driven. In this step, the message sent by the node is multiplied by a resilience adjustment factor (i.e., This results in the risk signal of nodes in highly stress-resistant plant communities being attenuated (α<1), while the risk signal of nodes in low-stress-resistant or degraded communities being amplified (α>1).

[0138] This design embeds the ecosystem resilience characterized by plant functional traits into the message passing process of a graph neural network in the form of physical priors.

[0139] In step S14, each node aggregates messages sent by its neighboring nodes using a graph attention mechanism and updates its own feature vector. The graph attention mechanism can adaptively learn the importance weights of different neighboring nodes, enabling the model to focus on neighboring nodes with greater influence when aggregating messages.

[0140] Loop Check: Determine if the current number of message passing rounds has reached the preset number K. If it has not reached K, return to step S11 and proceed to the next round of message passing and node update; if it has reached K, terminate the loop and output the ecological risk index of each grid cell.

[0141] The preset number of iterations K represents the number of message passing layers in the improved graph neural network model, which is set to 2 in this embodiment. This value is not fixed and can be adaptively adjusted according to the spatial dependence range and data scale of the evaluation area in practical applications.

[0142] In summary, in this application, the messages sent by the nodes are: α is determined by the functional traits of the plant, which realizes the physical constraint of ecological prior knowledge on the message transmission intensity, so that the model output conforms to both the statistical laws of data and the common sense of ecological physics.

[0143] After calculating the ecological risk index of each raster unit, spatial aggregation is performed based on the control zone vector layer according to the zone boundary. That is, the ERI values ​​of all raster units falling within the same zone boundary are arithmetically averaged to obtain the ecological risk index of each zone.

[0144] Figure 3 This is a flowchart of the training process for the improved graph neural network model in this embodiment. The specific training steps of the improved graph neural network model include:

[0145] Step S01: Obtain training data for historical years, and extract a historical multi-source spatial data layer from the training data. The historical multi-source spatial data layer includes a historical plant functional trait raster layer, a historical meteorological element raster layer, a historical human activity intensity raster layer, and a historical control zone vector layer. Using each raster unit as a node, extract the data value corresponding to each raster unit from the historical multi-source spatial data layer and concatenate them into a historical feature vector for each node.

[0146] Step S02: Construct a historical map structure based on the historical control zone vector layer and the spatial adjacency relationships between each raster unit. The edges of the historical map structure include a first type of spatial adjacency edge constructed based on the geographical distance between raster units, a second type of spatial adjacency edge constructed based on the hydrological flow direction between raster units, and ecological interaction edges constructed based on land use type or ecosystem type.

[0147] Step S03: Construct a heterogeneous graph neural network; the heterogeneous graph neural network includes a multilayer perceptron subnetwork, which is used to calculate stress resistance regulation factors based on the plant functional trait data of each node.

[0148] Step S04: Input the historical feature vector and the historical graph structure into the heterogeneous graph neural network, and output the predicted value of the ecological risk index for each node; the message passing process is the same as in steps S11-S14.

[0149] In a preferred embodiment, when the number of grid cells in the historical training data exceeds 100,000, a small-batch training strategy using graph sampling is adopted: 10,000 nodes are randomly sampled in each iteration, and their K-hop neighborhood subgraphs are extracted for training. During the inference phase, the entire graph is used for one forward propagation.

[0150] Step S05: Using the measured ecological loss data after historical disaster events as a monitoring signal, the error between the predicted value of the ecological risk index and the measured ecological loss data is calculated using a loss function. The parameters of the heterogeneous graph neural network are updated through backpropagation and iterated until convergence to obtain a trained improved graph neural network model.

[0151] Specifically, such as Figure 2 As shown, steps S04 and S05 constitute the loop in the training process: in each iteration, forward propagation is first performed to obtain the predicted value (S04), then the loss is calculated and the parameters are updated through backpropagation (S05); after completing this iteration, the process returns to step S04 to enter the next round of forward propagation. This loop is repeated until the model converges and training ends.

[0152] In a preferred embodiment, a hierarchical training strategy is used to train the heterogeneous graph neural network: in the first stage, the parameters of the graph attention layer and output layer are fixed, and only the multilayer perceptron subnetwork is trained; in the second stage, all modules are trained end-to-end, and the overall network parameters are fine-tuned.

[0153] In step S05, the loss function is used to calculate the error between the predicted ecological risk index and the measured ecological loss data. It is a weighted sum of nodal regression loss, trait covariance constraint loss, hydrological connectivity constraint loss, and spatial smoothness constraint loss. The physical meanings of each loss are as follows:

[0154] (1) Node regression loss; The node regression loss is the mean square error between the predicted ecological risk index of all nodes and the corresponding measured ecological loss data. It is the main supervision signal for model training and is used to constrain the predicted ecological risk index output by the model to approximate the measured loss data. In this embodiment, the measured ecological loss data comes from the field survey and remote sensing monitoring results after historical disaster events (drought, rainstorm, construction disturbance, etc.), including three indicators: vegetation coverage decline rate, soil erosion and biomass reduction rate. After normalization, these are used as training labels.

[0155] (2) Trait covariation constraint loss; The trait covariation constraint loss is used to constrain the representation of plant functional trait nodes in the heterogeneous graph neural network, making specific leaf area positively correlated with leaf nitrogen content and stress tolerance index positively correlated with root depth. This loss term is based on the leaf economic spectrum theory in plant ecology—there is a synergistic relationship between specific leaf area and leaf nitrogen content, and a positive correlation between stress tolerance and root depth. This constraint ensures that the plant functional trait representations learned by the model conform to ecological common sense and avoids abnormal embeddings in the model output that violate plant physiological laws.

[0156] (3) Hydrological connectivity constraint loss; The hydrological connectivity constraint loss is used to constrain the representations of upstream and downstream nodes connected by second-type spatial adjacency edges, so as to minimize the difference between the representations of upstream and downstream nodes. This loss term is based on the principle of watershed hydrological connectivity - changes in the upstream ecological state will be transmitted downstream along the direction of water flow. Therefore, the representations of upstream and downstream nodes should maintain a continuous and smooth trend of change in the embedded space, and there should be no abrupt changes.

[0157] (4) Spatial smoothing constraint loss; The spatial smoothing constraint loss is used to constrain the representations of adjacent nodes connected by first-type spatial adjacency edges, so that the difference in representations of adjacent nodes increases monotonically with increasing geographical distance. This loss term is based on the first law of geography—the closer the raster cells are, the more similar their ecological characteristics are, and the farther apart the raster cells are, the greater the difference is. This constraint ensures that the risk distribution output by the model maintains reasonable smoothness in space, avoiding random oscillations.

[0158] The total loss function is expressed as:

[0159] ;

[0160] Among them, L reg For node regression loss, L trait For trait covariance constraint loss, L hydro For hydrological connectivity constraint loss, L smooth λ1, λ2, λ3, and λ4 are the corresponding weight coefficients, representing the spatial smoothing constraint loss.

[0161] In a preferred embodiment, the four weights of the loss function are adaptively adjusted during training: Initially, the forward propagation of each loss term is run individually to calculate its initial magnitude. Then, the initial weights are set according to the reciprocal of the magnitude, ensuring that each loss term contributes equally in the early stages of training. Subsequently, the magnitude of each loss term is recalculated every 10 epochs, and the weights are dynamically adjusted to maintain a relative balance in the contributions of each loss term. In this embodiment, after initial magnitude calculation and adjustment, the weight λ1 of the node regression loss is 0.6, the weight λ2 of the trait covariance constraint loss is 0.15, the weight λ3 of the hydrological connectivity constraint loss is 0.15, and the weight λ4 of the spatial smoothness constraint loss is 0.10.

[0162] This step introduces multi-dimensional physical constraints into model training by constructing a weighted sum loss function: Node regression loss uses measured ecological loss data as a supervision signal to ensure that the model's predicted values ​​closely approximate the measured results; trait covariance constraint loss embeds the ecological covariance laws among plant functional traits (specific leaf area is positively correlated with leaf nitrogen content, and stress tolerance index is positively correlated with root depth) into the model training process, ensuring that the trait representations learned by the model conform to ecological common sense and avoiding abnormal embeddings that violate plant physiological laws; hydrological connectivity constraint loss uses second-type spatial adjacency edges (hydrological flow direction) to minimize the representation differences between upstream and downstream nodes, ensuring that the risk distribution output by the model maintains a continuous and smooth trend along the flow direction, conforming to the principle of watershed hydrological connectivity; spatial smoothness constraint loss uses first-type spatial adjacency edges (geographical distance) to constrain the representation differences between adjacent nodes to increase monotonically with increasing geographical distance, ensuring that the risk distribution conforms to the first law of geography; simultaneously, the four weights are adaptively adjusted during training, avoiding training imbalance caused by differences in the magnitude of different loss terms, ensuring that each physical constraint can effectively play its role in training. The aforementioned multi-constraint collaborative mechanism enables the model output to simultaneously satisfy four physical constraints: data statistical regularity, ecological common sense, hydrological connectivity, and spatial smoothness, significantly improving the physical interpretability and ecological rationality of the assessment results.

[0163] In a preferred embodiment, the construction of the history graph structure further includes:

[0164] Each grid cell is expanded into four independent nodes: a risk source node, an environment node, a plant functional trait node, and a receptor node. The input data for the risk source node is the intensity of the risk source corresponding to that grid cell (including the intensity values ​​of four risk factors: ecological damage, environmental pollution, natural disasters, and human development stress). The input data for the environment node is the meteorological elements (rainfall, temperature, evapotranspiration) and topographic factors (elevation, slope) corresponding to that grid cell. The input data for the plant functional trait node is the six types of plant functional trait parameters corresponding to that grid cell. The input data for the receptor node is the ecosystem exposure and land use type corresponding to that grid cell.

[0165] Within the same grid cell, risk source nodes are sequentially connected to environmental nodes, environmental nodes to plant functional trait nodes, and plant functional trait nodes to receptor nodes via ecological interaction edges. This represents the complete causal chain of ecological stress generated by the risk source, which, through environmental transmission and plant response, ultimately acts on the ecological receptor. Between different grid cells, nodes of the same type are connected via spatial adjacency edges. These spatial adjacency edges include a first type based on geographical distance and a second type based on hydrological flow direction.

[0166] In this step, a heterogeneous graph structure is constructed by extending a single grid unit into four independent nodes: risk source, environment, plant functional trait, and receptor, so that the four types of information perform message passing along their respective independent propagation paths, avoiding the mixing of different types of information in a single node; at the same time, the four ecological interaction edges within the unit completely retain the causal chain from the occurrence of stress to the exposure of the receptor, making the information propagation process of the graph neural network correspond to the real ecological process, which effectively improves the physical interpretability and ecological rationality of the assessment results.

[0167] In summary, this step adopts an improved graph neural network model. By constructing a three-layer graph structure including spatial adjacency edges of geographic distance type, spatial adjacency edges of hydrological flow direction type, and ecological interaction edges of management control zones, the graph neural network can simultaneously capture three spatial dependencies in message passing: geographic proximity, hydrological upstream-downstream correlation, and ecological management unit consistency, which effectively breaks through the limitation of existing methods that only construct a single type of edge based on geographic distance; in the message passing process, a stress resistance regulatory factor determined by plant functional traits is introduced to perform multiplicative modulation on the messages sent by nodes, so that the risk signal of communities with high stress resistance is attenuated, and the risk signal of communities with low stress resistance is amplified, embedding ecological prior knowledge into the forward propagation process of the graph neural network in the form of physical constraints, which overcomes the defect that the message intensity of the standard graph neural network is purely data-driven and lacks physical interpretability; at the same time, by constructing a weighted sum loss function including node regression loss, trait covariation constraint loss, hydrological connectivity constraint loss, and spatial smoothness constraint loss, multiple physical constraints such as ecological covariation law, watershed hydrological connectivity, and the first law of geography are further embedded in the training stage, so that the model output satisfies both the statistical law of data and the physical common sense of ecology, and improves the physical interpretability and ecological rationality of the assessment results.

[0168] Step S3: Compare the ecological risk index of each of the zones with a preset risk early warning threshold to obtain the risk level of each of the zones, and generate an early warning report.

[0169] Specifically, this embodiment adopts the natural breakpoint method to classify the ecological risk index, and divides the risk levels into four levels from low to high: low risk zone, medium risk zone, high risk zone, and extremely high risk zone, which correspond to blue early warning, yellow early warning, orange early warning, and red early warning respectively. The natural breakpoint method minimizes the variation within each level and maximizes the variation between levels through iterative calculation, so as to determine the optimal division threshold for each level. Based on the distribution of ecological risk indices of all grid units in the whole domain, calculated by the natural breakpoint method, the thresholds for the low risk zone, medium risk zone, high risk zone, and extremely high risk zone are determined as ERI≤0.25, 0.25<ERI≤0.50, 0.50<ERI≤0.75, and ERI>0.75 respectively.

[0170] In a preferred embodiment, the risk warning threshold adopts a dynamic update mechanism: the threshold is determined by the natural breakpoint method in the initial assessment; in subsequent assessments, the historical ERI distribution and the current ERI distribution are subjected to the KS test. If there is no significant difference in the distribution, the historical threshold is used; otherwise, it is recalculated.

[0171] The ecological risk indices of the three zones were compared by substituting them into the aforementioned threshold ranges:

[0172] The average ERI (Effective Risk Index) of the primary forest reserve was 0.29, falling between 0.25 and 0.50, classifying it as a medium-risk area and corresponding to a yellow alert. Although the primary forest community structure is intact and highly resilient, the edge effect of highway construction still poses a certain threat to the forest edge area, requiring continuous monitoring of changes in forest edge vegetation.

[0173] The average ERI (Effective Relationship Index) of the wetland restoration and control zone is 0.58, falling between 0.50 and 0.75, classifying it as a high-risk area and corresponding to an orange alert. The wetland vegetation is still in the mid-stage of recovery and has weak resilience. Runoff pollution from highway construction and disruption of hydrological connectivity pose a significant threat to this area, requiring strengthened monitoring and the development of specific protection plans.

[0174] The average ERI (Effective Relationship) of the general control area for artificial turf is 0.69, falling within the range of 0.50 to 0.75, classifying it as a high-risk area and corresponding to an orange alert. Artificial turf has a single vegetation type, shallow root system, and poor resilience, resulting in an extremely high risk of soil erosion after construction disturbance, requiring focused prevention and control.

[0175] Based on the risk level assessment results, the system automatically generates an ecological risk early warning report. The early warning report includes a spatial distribution map of ecological risks for the entire region and each sub-region (output in raster layer format, with different colors representing different risk levels), the risk level and early warning color markings for each sub-region, an analysis of the dominant risk factors for each sub-region (the dominant risk factor for primary forest reserves is the edge effect, the dominant risk factors for wetland restoration and control areas are hydrological connectivity disruption and runoff pollution, and the dominant risk factors for general artificial turf control areas are soil erosion and vegetation degradation), and differentiated early warning recommendations.

[0176] The process of generating the spatial distribution map of ecological risk is as follows: Based on the ecological risk index of each grid cell output in step S14, a spatial overlay analysis method is used to compare the ecological risk index of each grid cell with its neighboring grid cells, identifying hotspot areas in each partition where the ecological risk index is significantly higher than that of its neighbors. Specifically, Getis-Ord is used. Local spatial autocorrelation statistics are used to calculate the degree of spatial clustering of each grid cell. When the ERI value of a grid cell is higher than the global average and the ERI value of its neighboring grid cells is also significantly higher than the global average, the grid cell is determined to belong to a risk hotspot area.

[0177] In a preferred embodiment, when conducting multi-period assessments, a time-series tracking file is established for the identified hotspot areas: the spatial overlap and ERI change rate of hotspot areas in two adjacent periods are compared. If the ERI increases by more than 20% compared to the previous period, the warning level is automatically upgraded; if it decreases by more than 20%, the warning level is automatically downgraded.

[0178] The hotspot area identification results are highlighted with bright color blocks on the risk spatial distribution map to identify clusters of high-risk grids, providing spatial positioning basis for precise control. In this embodiment, the identified hotspot areas are mainly distributed in the grid strips adjacent to the highway construction area in the general control area of ​​artificial turf, and in the downstream area of ​​the wetland restoration control area near the road drainage outlet.

[0179] In this embodiment, both the wetland restoration control area and the artificial turf general control area are identified as high-risk areas (orange alert). The park management agency must develop a specific ecological protection plan before road construction and submit the alert report to the Municipal Ecological Environment Bureau for record-keeping, while also sending copies to the Transportation Bureau and the construction unit to facilitate subsequent cross-departmental collaborative management. The primary forest reserve is identified as a medium-risk area (yellow alert), requiring routine management measures and continuous monitoring of vegetation changes in the forest edge area to prevent further escalation of the risk.

[0180] This step iteratively calculates the ERI value distribution of the entire raster unit using the natural breakpoint method, minimizing the differences within each level and maximizing the differences between levels. This avoids the subjectivity and arbitrariness of manually setting thresholds, ensuring the objectivity and scientific nature of risk classification. Based on this, independent risk assessments are performed for each zone, resulting in differentiated risk levels and warning indicators for primary forest reserves, wetland restoration and control zones, and general artificial turf control zones. This provides a clear basis for matching differentiated control measures to different ecological spaces. Simultaneously, Getis-Ord is employed. Local spatial autocorrelation statistics identify hotspots with significantly higher risk than their neighboring areas, accurately locating high-risk clusters on spatial distribution maps. This provides quantifiable spatial guidance for management departments to implement precise policies, effectively improving the precision and operability of risk warnings.

[0181] Example 2:

[0182] Finally, this application also proposes a computing device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the instructions stored in the memory so that the computer device performs an intelligent assessment method for ecological space risk coupled with plant functional traits as described in the above embodiments.

[0183] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligent assessment of ecological spatial risk coupled with plant functional traits, characterized in that, include: S1, the target ecological space is divided into multiple partitions according to the predetermined partitioning conditions; multi-source heterogeneous data of each partition are obtained and normalized to obtain multi-source heterogeneous data of a unified spatiotemporal reference for each partition, which serves as the basic data of each partition. S2, Based on the basic data of each partition, the ecological risk index of each partition is calculated using a pre-constructed ecological risk index model; The ecological risk index model is selected as an improved weighted summation model or a pre-trained improved graph neural network model based on the data characteristics in the basic data of each partition. The improved weighted summation model introduces a vulnerability correction coefficient based on plant functional traits to multiplicatively adjust the weighted summation result; The improved graph neural network model multiplicatively regulates the messages sent by nodes during message transmission using stress resistance regulation factors determined based on plant functional traits. S3. Compare the ecological risk index of each partition with the preset risk warning threshold to obtain the risk level of each partition and generate a warning report.

2. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 1, characterized in that, The ecological risk index model selects either an improved weighted summation model or a pre-trained improved graph neural network model based on the data characteristics in the basic data of each partition, specifically including: Each of the aforementioned partitions is divided into several grid units; The effective data coverage C of the basic data of each partition is calculated using formula (1); Formula (1) specifically includes: ; Where, N total N represents the total number of grid cells within each of the aforementioned partitions. valid The number of raster cells in each of the aforementioned partitions that simultaneously possess plant functional trait data, meteorological element data, and human activity intensity data; When the effective data coverage is lower than a preset threshold, an improved weighted summation model is used as the ecological risk index model. Otherwise, a pre-trained improved graph neural network model is used as the ecological risk index model.

3. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 2, characterized in that, When using the improved weighted summation model, the calculation steps for the ecological risk index include: Based on the basic data of each partition, four types of ecological risk factors are extracted for each partition. The four types of ecological risk factors include ecological damage factors that characterize the degree of damage to the ecosystem structure, environmental pollution factors that characterize the degree of stress of pollutant input on the ecosystem, natural disaster factors that characterize the degree of impact of natural disasters on the ecosystem, and human development stress factors that characterize the degree of disturbance of human activities on the ecosystem. The corresponding preset indicator data are extracted from the four types of ecological risk factors, and the combined weights of the four types of ecological risk factors are obtained by using the analytic hierarchy process coupled with the entropy weight method. Plant functional trait parameters of each grid cell are extracted from the basic data of each partition and standardized to obtain standardized values ​​of each plant functional trait parameter in each grid cell; the stress resistance comprehensive score of each grid cell is calculated by weighted summation based on the standardized values. The resilience level is determined based on the comprehensive resilience score, and the vulnerability correction coefficient is determined based on the preset level mapping relationship. Extract the risk source intensity and ecosystem exposure of each raster unit corresponding to the four types of ecological risk factors from the basic data of each partition; Calculate the ecological risk index ERI for each grid cell: ERI = Σ(W i ×S i ×E i ×V i ×V base ); Where i represents the i-th type of ecological risk factor; W i S represents the combined weight of the i-th type of ecological risk factor; i E represents the intensity of the risk source for the i-th type of ecological risk factor; i V represents the ecosystem exposure of the i-th type of ecological risk factor; i V is the vulnerability correction coefficient for the i-th type of ecological risk factor; base The vulnerability baseline coefficient is pre-set based on the land use type or soil type of each grid cell; Based on the ecological risk index of each grid cell, the average ecological risk index of all grid cells within each partition is calculated as the ecological risk index of each partition.

4. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 3, characterized in that, The method employing the coupling of the analytic hierarchy process (AHP) and the entropy weight method yields the combined weights of the four types of ecological risk factors, including: The subjective weights of the four types of ecological risk factors are obtained using the analytic hierarchy process (AHP), and the objective weights of the four types of ecological risk factors are obtained using the entropy weight method. An objective function is constructed with the goal of minimizing the information deviation between the subjective weights and the objective weights, and normalization constraints are introduced. The objective function is solved using the Lagrange multiplier method to obtain the combined weights of the four types of ecological risk factors.

5. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 4, characterized in that, When using a pre-trained improved graph neural network model, the steps for calculating the ecological risk index include: Multi-source spatial data layers are extracted from the basic data of each partition. The multi-source spatial data layers include a plant functional trait raster layer for characterizing the structure and functional characteristics of plant communities, a meteorological element raster layer for characterizing regional climate and hydrological conditions, a human activity intensity raster layer for characterizing the spatial distribution of human disturbance intensity, and a control partition vector map layer for defining the spatial range of each partition and the boundary of ecological management unit. Using each grid cell as a node in the graph, the data values ​​corresponding to each grid cell are extracted from the plant functional trait grid layer, meteorological element grid layer, and human activity intensity grid layer, and then concatenated into the feature vector of each node. A graph structure is constructed based on the control zone vector layer and the spatial adjacency relationships between each raster unit; the edges of the graph structure include spatial adjacency edges and ecological interaction edges; The feature vectors of each node and the graph structure are input into a pre-trained graph neural network model for message passing, and the ecological risk index of each grid cell is output. The ecological risk index of each grid cell is spatially aggregated according to the partition boundary in the control partition vector layer to obtain the ecological risk index of that partition.

6. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 5, characterized in that, The message passing includes: S11, extract the plant functional trait data corresponding to each node from the feature vector of each node; S12, before sending a message to a neighboring node, each node calculates its stress resistance regulation factor based on the plant functional trait data; S13, each node performs message passing; the message sent by each node to its neighboring node is equal to the result of linear transformation of the feature vector of each node multiplied by the corresponding resilience adjustment factor; S14, each node aggregates the messages sent by its neighboring nodes through a graph attention mechanism and updates its own feature vector; After repeating steps S11 to S14 a preset number of times, the ecological risk index of each grid cell is output.

7. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 6, characterized in that, The calculation of stress resistance regulating factors for each node based on the plant functional trait data includes: Plant functional trait parameters are extracted from the feature vectors of each node; The plant functional trait parameters are standardized, and the stress resistance comprehensive score of each node is calculated by weighted summation using preset weights. When the overall stress resistance score is higher than the first threshold, the grid cell corresponding to the node is determined to be a high stress resistance cell, and the stress resistance adjustment factor is set to a value between 0.6 and 0.8, and the stress resistance adjustment factor is negatively correlated with the overall stress resistance score. When the overall stress resistance score is lower than the second threshold, the grid cell corresponding to the node is determined to be a low stress resistance or degenerate cell, and the stress resistance adjustment factor is set to a value between 1.2 and 1.5, and the stress resistance adjustment factor is negatively correlated with the overall stress resistance score. When the overall resilience score is between the first threshold and the second threshold, the resilience adjustment factor is set to 1.0; Wherein, the first threshold is greater than the second threshold.

8. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 7, characterized in that, The training steps of the improved graph neural network model include: S01, acquire training data from historical years, and extract historical multi-source spatial data layers from the training data. The historical multi-source spatial data layers include historical plant functional trait raster layers, historical meteorological element raster layers, historical human activity intensity raster layers, and historical control zone vector layers. Using each of the grid cells as nodes, the data values ​​corresponding to each grid cell are extracted from the historical multi-source spatial data layer and concatenated to form the historical feature vector of each node. S02, Construct a historical graph structure based on the historical control partition vector layer and the spatial adjacency relationship between each grid unit; S03, Construct a heterogeneous graph neural network; the heterogeneous graph neural network includes a multilayer perceptron subnetwork, which is used to calculate stress resistance regulation factors based on the plant functional trait data of each node; S04, input the historical feature vector and the historical graph structure into the heterogeneous graph neural network, and output the ecological risk index prediction value of each node; S05, using the measured ecological loss data after historical disaster events as a monitoring signal, the error between the predicted value of the ecological risk index and the measured ecological loss data is calculated using a loss function. The parameters of the heterogeneous graph neural network are updated through backpropagation, and the iteration continues until convergence, resulting in a trained improved graph neural network model.

9. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 8, characterized in that, The construction of the history graph structure also includes: Each grid cell is expanded into four independent nodes, namely, a risk source node, an environment node, a plant functional trait node, and a receptor node; Within the same grid cell, risk source nodes and environmental nodes, environmental nodes and plant functional trait nodes, and plant functional trait nodes and receptor nodes are connected through ecological interaction edges. Different grid cells and nodes of the same type are connected by spatial adjacency edges; the spatial adjacency edges include a first type of spatial adjacency edge constructed based on geographical distance and a second type of spatial adjacency edge constructed based on hydrological flow direction.

10. The intelligent ecological spatial risk assessment method coupled with plant functional traits according to claim 9, characterized in that, The loss function is a weighted sum of node regression loss, trait covariance constraint loss, hydrological connectivity constraint loss, and spatial smoothness constraint loss. The node regression loss is the mean square error between the predicted ecological risk index of all nodes and the corresponding measured ecological loss data. The trait covariation constraint loss is used to constrain the representation of plant functional trait nodes in the heterogeneous graph neural network, so that specific leaf area is positively correlated with leaf nitrogen content and stress tolerance index is positively correlated with root depth. The hydrological connectivity constraint loss is used to constrain the representations of upstream and downstream nodes connected by second-type spatial adjacency edges, so as to minimize the difference between the representations of upstream and downstream nodes. The spatial smoothing constraint loss is used to constrain the representations of adjacent nodes connected by first-type spatial adjacency edges, such that the difference in representations of adjacent nodes increases monotonically with increasing geographical distance.

Images