A forestry water conservation function evaluation method, system, device and medium

Abstract

The application relates to a forestry water source conservation function evaluation method, system, device and medium. The method comprises the following steps: acquiring and preprocessing multi-source data, generating initial attribute feature vectors, spatial topological relations and global context vectors of each evaluation unit, and constructing a basin space-time heterogeneous graph; inputting the basin space-time heterogeneous graph into a preset space-time heterogeneous graph neural network for feature fusion processing to obtain a high-order feature vector of each graph node in the basin space-time heterogeneous graph; and based on the high-order feature vector and the global context vector, performing dynamic decision-making through a preset adaptive multi-criteria decision model to obtain the final dynamic water source conservation function evaluation value of each evaluation unit. Through the fusion of multi-source data and nonlinear spatial correlation and in combination with an adaptive dynamic weight adjustment mechanism, the method improves the accuracy, spatial continuity and dynamic adaptability of water source conservation function evaluation in a complex ecological hydrological environment.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, system, equipment and medium for evaluating the water conservation function of forestry. Background Technology

[0002] Forestry water conservation is one of the core ecosystem services of forest ecosystems. Through the interception, absorption, storage, and regulation of atmospheric precipitation by the canopy, litter layer, and soil layer, it directly relates to regional water resource security, ecological balance maintenance, and climate change response. Therefore, scientific evaluation of forestry water conservation function can provide crucial support for sustainable forest management, the establishment of ecological compensation mechanisms, and water resource management decisions. Currently, the comprehensive evaluation method based on Geographic Information System (GIS), Remote Sensing (RS) technology, and Multi-Criterion Decision Analysis (MCDA) has become the mainstream technical solution in this field. This method selects evaluation factors such as vegetation cover, soil type, topographic features, and rainfall, determines the fixed weights of each factor using the Analytic Hierarchy Process (AHP), Delphi method, or entropy weight method, and then obtains the static evaluation value of the water conservation function of the evaluation unit using weighted linear summation or fuzzy comprehensive evaluation methods.

[0003] However, this comprehensive evaluation method still has significant inherent defects and is difficult to meet the dynamic evaluation needs under complex eco-hydrological environments. For example, this method simply regards water conservation function as a linear superposition of the independent contributions of each evaluation factor, ignoring the nonlinear interaction between factors such as forest stand structure and soil characteristics, topographic conditions and rainfall intensity, leading to distorted evaluation results in highly heterogeneous regions. Secondly, this type of method treats discrete and independent evaluation units as the object of processing, failing to consider the inherent spatial continuity of hydrological processes, severing the "source-sink" transmission relationship between upstream and downstream evaluation units, and failing to truly reflect the functional spatial pattern and network characteristics at the landscape scale. Furthermore, once the factor weights used in the evaluation process are determined, they remain fixed and cannot adapt to the dynamic characteristics of water conservation function that vary with seasonal changes and rainfall event types. It can only output averaged or typical static results, which are difficult to support emergency decision-making under extreme climate scenarios and refined seasonal water resource management. Summary of the Invention

[0004] Based on this, it is necessary to provide a method, system, equipment and medium for evaluating the water conservation function of forestry in response to the above-mentioned technical problems. The aim is to improve the accuracy and spatiotemporal resolution of the evaluation of water conservation function, enhance the ability to characterize the spatial correlation of watershed hydrological processes, and realize the transformation from static mean evaluation to dynamic adaptive evaluation.

[0005] Firstly, this application provides a method for evaluating the water conservation function of forestry, including:

[0006] Acquire multi-source data of the evaluation region including multiple evaluation units, perform standardized preprocessing on the multi-source data, and generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between evaluation units, and global context vectors.

[0007] Based on the initial attribute feature vector, spatial topological relationship and global context vector, a watershed spatiotemporal heterogeneous graph is constructed; the watershed spatiotemporal heterogeneous graph graph is input into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the watershed spatiotemporal heterogeneous graph graph.

[0008] Based on high-order feature vectors and global context vectors, dynamic decision-making is performed through a pre-set adaptive multi-criteria decision-making model to obtain the final dynamic water conservation function evaluation value of each evaluation unit.

[0009] In one embodiment, the multi-source data undergoes standardization preprocessing to generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between evaluation units, and a global context vector, including:

[0010] The process involves acquiring static attribute data and recent rainfall time-series data for each evaluation unit from multi-source data; performing vectorization encoding on the static attribute data to obtain static attribute vectors for each evaluation unit; extracting time-series features from the recent rainfall time-series data to obtain recent dynamic feature vectors for each evaluation unit; aligning the static attribute vectors and recent dynamic feature vectors dimensionally, and then performing cross-feature interactive encoding to obtain initial attribute feature vectors for each evaluation unit.

[0011] The digital elevation model of the evaluation area and the vector boundary data of the evaluation units are extracted from the multi-source data; spatial matching processing is performed on the digital elevation model and the vector boundary data of the evaluation units to obtain spatial fusion data; based on the spatial fusion data, the boundary adjacency relationship of each evaluation unit is traversed through the R-tree indexing algorithm to obtain the spatial adjacency relationship of each evaluation unit.

[0012] Water flow direction grids are extracted from the spatially fused data. Based on these grids, the inflow and outflow units of each evaluation unit are determined using a confluence path tracing algorithm. Slope and soil infiltration rate data for each evaluation unit are obtained from multi-source data. Combined with the positional relationship between the inflow and outflow units, the hydrological conduction efficiency is quantified to obtain the quantified hydrological conduction efficiency. Based on the quantified hydrological conduction efficiency, flow direction relationship analysis is performed to obtain the hydrological flow direction relationship for each evaluation unit.

[0013] The spatial adjacency relationship and the hydrological flow direction relationship are integrated and processed. A preset boundary interaction intensity weight is added to the spatial adjacency relationship and a preset transmission efficiency weight is added to the hydrological flow direction relationship to obtain the spatial topological relationship between each evaluation unit.

[0014] Climate event data and seasonal background data at the evaluation time point are extracted from multi-source data; feature extraction is performed on the climate event data and seasonal background data respectively to obtain climate event features and seasonal background features. Climate event features include peak intensity, duration proportion and spatiotemporal distribution concentration of events, while seasonal background features include matching phenological periods and hydrological influence coefficients.

[0015] Climate event features and seasonal background features are normalized to obtain first and second standardized feature data. The first and second standardized feature data are then input into a preset attention mechanism network for scenario priority calculation to obtain a scenario priority weight vector. Based on the scenario priority weight vector, the first and second standardized feature data are weighted and fused to obtain fused feature data. The fused feature data is then input into a fully connected network for feature integration and encoding to generate a global context vector.

[0016] In one embodiment, a spatiotemporal heterogeneous graph of the watershed is constructed based on initial attribute feature vectors, spatial topological relationships, and global context vectors, including:

[0017] Each evaluation unit is mapped to a graph node, and the initial attribute feature vector is compressed using a pre-set embedding network to obtain the initial node feature vector for each graph node.

[0018] Based on spatial adjacency relationships in spatial topology, bidirectional adjacency edges are established between graph nodes corresponding to spatial adjacency relationships. Bidirectional adjacency edges are used to characterize the lateral ecological interaction relationships between adjacent evaluation units.

[0019] Based on the hydrological flow direction relationship in the spatial topology, a one-way hydrological flow direction edge is established between the upstream evaluation unit graph node and the downstream evaluation unit graph node corresponding to the hydrological flow direction relationship. The one-way hydrological flow direction edge is used to characterize the confluence influence relationship between the upstream evaluation unit and the downstream evaluation unit.

[0020] By attaching the global context vector as a global attribute to the graph structure and integrating the initial node feature vectors, bidirectional adjacent edges, unidirectional hydrological flow edges, and global graph attributes, a spatiotemporal heterogeneous graph of the watershed is obtained.

[0021] In one embodiment, the preset spatiotemporal heterogeneous graph neural network includes a heterogeneous message generation module, a contextualized aggregation module, and a node state update module;

[0022] The spatiotemporal heterogeneous graph of the watershed is input into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing. The high-order feature vector of each graph node in the spatiotemporal heterogeneous graph of the watershed is obtained. The feature fusion processing corresponds to the following steps:

[0023] S1: Initialize the network parameters of the preset spatiotemporal heterogeneous graph neural network and set the preset iteration rounds; obtain the current feature vectors of each graph node in the spatiotemporal heterogeneous graph of the watershed; wherein, in the first iteration, the current feature vectors of each graph node are the initial node feature vectors;

[0024] S2: For each graph node, based on the adjacency message processing branch of the heterogeneous message generation module, the corresponding adjacent nodes are traversed through bidirectional adjacency edges, and the current feature vector of each adjacent node is linearly transformed through the first learnable parameter matrix to obtain the corresponding first transformed feature. The first transformed feature is aggregated through the adjacency message aggregation function to obtain the first aggregated message.

[0025] S3: For each graph node, the hydrological flow direction message processing branch based on the heterogeneous message generation module traverses the corresponding upstream nodes through the unidirectional hydrological flow direction edge, performs a linear transformation on the current feature vector of each upstream node through the second learnable parameter matrix to obtain the corresponding second transformed feature vector, and aggregates each second transformed feature vector through the hydrological flow direction message aggregation function to obtain the second aggregated message; wherein, the second learnable parameter matrix and the first learnable parameter matrix are different parameter matrices;

[0026] S4: The weight calculation submodule based on the contextualized aggregation module concatenates the current feature vector of the current graph node with the global context vector to obtain the concatenated feature; after performing nonlinear mapping on the concatenated feature, the dynamic fusion weight is calculated through the activation function; through the message fusion submodule of the contextualized aggregation module, the first aggregated message and the second aggregated message are weighted and summed based on the dynamic fusion weight to obtain the total aggregated message of the current graph node;

[0027] S5: Through the node state update module, the current feature vector, total aggregate message and global context vector of the current graph node are input into the preset state update function to calculate the updated feature vector of the current graph node; the preset state update function is constructed based on the LSTM gating mechanism or the GRU gating mechanism.

[0028] S6: Use the updated feature vector as the current feature vector of the current graph node in the next iteration, and repeat steps S2 to S5 until the cumulative number of iterations reaches the preset number of iterations. Stop the iteration and output the updated feature vector obtained in the last iteration as the higher-order feature vector corresponding to each graph node.

[0029] In one embodiment, the adaptive multi-criteria decision-making model includes a criterion scoring network, a dynamic weight generation network, and a decision fusion module; the criterion scoring network includes a feature decoding submodule and a nonlinear mapping layer, the dynamic weight generation network includes a scenario interaction submodule and a normalization layer, and the decision fusion module includes a weighted calculation unit and a rationality verification unit.

[0030] Based on high-order feature vectors and global context vectors, dynamic decision-making is performed through a pre-set adaptive multi-criteria decision-making model to obtain the final dynamic water conservation function evaluation value of each evaluation unit, including:

[0031] K water conservation function criteria conforming to eco-hydrological mechanisms are preset, and an independent criterion scoring network is configured for each water conservation function criterion. The high-order feature vectors corresponding to each evaluation unit are input into the K criterion scoring networks respectively. The spatial correlation information and nonlinear interaction information in the high-order feature vectors are decoded by the feature decoding submodule to obtain the decoded features. The decoded features are then input into the nonlinear mapping layer for mapping to obtain the initial scores of K criteria for each evaluation unit under the K water conservation function criteria. Here, K is a preset positive integer. The water conservation function criteria include at least the canopy litter interception criterion, the soil infiltration water storage criterion, and the runoff regulation criterion.

[0032] Based on the dynamic weight generation network, the high-order feature vectors of each evaluation unit are interacted with the global context vector to generate scenario-attribute fusion features. The scenario-attribute fusion features are gated through the scenario interaction submodule to obtain gated features. The gated features are input into the normalization layer and normalized through the Softmax function to obtain K dynamic weights for each evaluation unit corresponding to K water conservation function criteria.

[0033] Through the weighted calculation unit of the decision fusion module, for each evaluation unit, K dynamic weights are multiplied element-wise with the corresponding K initial criterion scores to obtain K weighted scores; the K weighted scores are summed to obtain preliminary decision results; through the rationality verification unit of the decision fusion module, the validity of the preliminary decision results is verified based on the preset hydrological threshold range, and outliers exceeding the preset reasonable range are removed to obtain the final dynamic water conservation function evaluation value of the evaluation unit.

[0034] In one embodiment, the static attribute data is vectorized and encoded to obtain the static attribute vector of each evaluation unit; the recent rainfall time series data is subjected to time series feature extraction to obtain the recent dynamic feature vector of each evaluation unit, including:

[0035] Static attribute data is divided into static categorical feature data and static numerical feature data; one-hot encoding is performed on the static categorical feature data to obtain the categorical encoding vector; extreme value standardization is performed on the static numerical feature data to obtain the numerical standardization vector.

[0036] Based on the preset forestry water source rules, the classification coding vector and the numerical standardization vector are coupled and interactively coded to obtain the static attribute vector of each evaluation unit.

[0037] Preprocessing operations are performed on recent rainfall time-series data to obtain normalized time-series data; the preprocessing operations include outlier identification and interpolation repair.

[0038] Based on a preset time window, the normalized time series data is segmented to obtain multiple local time series segments; local dependency features are extracted from each local time series segment through a temporal convolutional network to obtain segment-level features.

[0039] By aggregating the temporal trends of features at each segment level using a long short-term memory network, the recent dynamic feature vectors of each evaluation unit are obtained.

[0040] In one embodiment, the second aggregated message is calculated using the following formula:

[0041]

[0042] in, For the first During the nth iteration The second aggregated message of each graph node; This is an aggregation function for hydrological flow direction messages; The second learnable parameter matrix for the hydrological flow direction message branch; For the first In the next iteration, the upstream node The current feature vector; upstream node To the node Normalized hydrological transport efficiency; For the first The set consisting of the upstream nodes of each graph node; Learnable vectors for scoring attention; for The transpose of ; A learnable weight matrix for scoring attention; It is the hyperbolic tangent activation function; It is a natural exponential function; For traversing the first Temporary variables for each upstream node of each graph node.

[0043] Secondly, this application also provides a forestry water conservation function evaluation system, including:

[0044] The data preprocessing and feature extraction module is used to acquire multi-source data of the evaluation area including multiple evaluation units, perform standardized preprocessing on the multi-source data, and generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between evaluation units, and global context vectors.

[0045] The graph neural network encoding module is used to construct a watershed spatiotemporal heterogeneous graph based on initial attribute feature vectors, spatial topological relationships, and global context vectors. The watershed spatiotemporal heterogeneous graph is then input into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the watershed spatiotemporal heterogeneous graph.

[0046] The dynamic multi-criteria evaluation module is used to make dynamic decisions based on high-order feature vectors and global context vectors through a preset adaptive multi-criteria decision model, and obtain the final dynamic water conservation function evaluation value of each evaluation unit.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0049] The aforementioned method, system, equipment, and medium for evaluating forestry water conservation function first acquires multi-source data of the evaluation area and performs standardized preprocessing to generate initial attribute feature vectors, spatial topological relationships, and global context vectors. This solves the problems of independent processing of traditional evaluation factors and fragmented spatiotemporal information, improving the correlation and completeness of input data. Second, a spatiotemporal heterogeneous watershed map is constructed based on the three types of data. High-order feature vectors are obtained through feature fusion using a pre-set spatiotemporal heterogeneous graphical neural network, realizing the modeling of nonlinear interaction of factors and spatial hydrological correlation, thus improving the fidelity and comprehensiveness of feature representation mechanisms. Finally, based on the high-order feature vectors and global context vectors, dynamic decision-making is performed through an adaptive multi-criteria decision model, improving the scenario adaptability and scientific validity of the evaluation results. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart of a method for evaluating the water conservation function of forestry is provided as an exemplary embodiment of the present invention;

[0052] Figure 2 A flowchart illustrating a method for dynamic decision-making using a preset adaptive multi-criteria decision-making model, provided as an exemplary embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of a forestry water conservation function evaluation system provided as an exemplary embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] In one embodiment, such as Figure 1 As shown, a method for evaluating the water conservation function of forestry is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0056] S101: Obtain multi-source data of the evaluation region including multiple evaluation units, perform standardized preprocessing on the multi-source data, and generate the initial attribute feature vector of each evaluation unit, the spatial topological relationship between each evaluation unit, and the global context vector.

[0057] Specifically, within the evaluation area, watershed sub-units, grid units, or forest compartment units can be used as basic evaluation units. Multi-source data covers key factors influencing water conservation function and can include static attribute data and dynamic time-series data, spatial geographic data and contextual background data. Static attribute data can be obtained through remote sensing image interpretation, ground plot surveys, and extraction from GIS vector databases. This includes inherent attribute data of the evaluation unit such as forest stand type, stand age, canopy closure, soil texture, soil thickness, topographic slope, and altitude. This data determines the basic water conservation potential of the evaluation unit. Dynamic time-series data mainly consists of recent rainfall time-series data, which can be obtained by aggregating real-time monitoring data from hydrological monitoring stations and meteorological observation points within the region. This includes time-series information such as rainfall intensity, rainfall duration, and rainfall intervals, reflecting the dynamic impact of short-term hydrological input on water conservation function. Spatial geographic data can include digital elevation models (DEMs), evaluation unit vector boundary data, etc., which are obtained through topographic mapping data and territorial spatial planning GIS databases. These data are used to capture the spatial location correlation and hydrological flow correlation of evaluation units. Scenario background data can include climate event data (such as the intensity, duration, and impact range of extreme climate events such as rainstorms and droughts) and seasonal background data (such as the phenological characteristics of spring, summer, autumn, and winter, and hydrological cycle patterns). These data are obtained through climate monitoring reports issued by meteorological departments and phenological observation data from ecological and environmental monitoring institutions. These data are used to adapt to different global evaluation scenarios.

[0058] Subsequently, the collected multi-source data can be standardized and preprocessed to eliminate dimensional differences between different data types, filter data noise, and unify data formats. Illustratively, the water conservation capacity of an evaluation unit depends on both inherent static attributes such as forest stands and soil, and is also affected by dynamic environmental changes such as recent rainfall. Therefore, the static attribute data in the preprocessed multi-source data can be integrated with recent rainfall time-series data to form a feature vector that comprehensively characterizes the basic characteristics and dynamic response capabilities of the evaluation unit. This vector provides the core input for subsequent graph node features. Since water conservation function has significant spatial correlations, there are lateral ecological effects such as forest stand interaction and water migration between adjacent evaluation units, and there are confluence and transmission effects between upstream and downstream evaluation units. Therefore, spatial data processing can be used to mine the adjacency relationships and hydrological flow directions between evaluation units, constructing a topological structure that reflects this spatial interaction mechanism. Furthermore, to adapt to the differences in different evaluation scenarios, and because climate events and seasonal backgrounds can affect the regional hydrological cycle process at a global level, thereby changing the evaluation logic of water conservation function, a global context vector can be constructed based on the scenario data in the preprocessed multi-source data to form contextual support that can guide subsequent evaluation decisions.

[0059] S102: Construct a watershed spatiotemporal heterogeneous graph based on initial attribute feature vectors, spatial topological relationships, and global context vectors; input the watershed spatiotemporal heterogeneous graph into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the watershed spatiotemporal heterogeneous graph.

[0060] Specifically, each evaluation unit can first be mapped as an independent node in the graph structure. The core information of the node is carried by the initial attribute feature vector generated by S101, ensuring that each node can initially represent the basic characteristics of the corresponding evaluation unit. Secondly, based on spatial topological relationships, two types of heterogeneous edge structures can be constructed: one type can be bidirectional edges based on spatial adjacency relationships, used to represent the lateral ecological interactions between adjacent evaluation units; the other type can be unidirectional edges based on hydrological flow direction relationships, used to represent the unidirectional confluence impact of upstream evaluation units on downstream evaluation units, such as the pressure of upstream rainfall runoff on downstream water storage capacity, and the regulating effect of upstream soil infiltration on downstream runoff, thus conforming to the unidirectional transmission characteristics of hydrological confluence. Finally, the global context vector can be added to the graph structure as a global attribute, so that the graph structure not only contains the association information of local nodes and edges, but also carries the constraint information of the global evaluation scenario, realizing the organic combination of local features and global scenarios.

[0061] Specifically, the pre-defined spatiotemporal heterogeneous graph neural network can handle heterogeneous messages transmitted by different types of edges based on a process of heterogeneous message generation, contextualized aggregation, and node state update, and achieve contextualized feature fusion by combining global context vectors. For example, the network parameters and iteration rounds can be initialized first, using the initial attribute feature vectors of the graph nodes as the input features for the first iteration. Subsequently, in each iteration, the network traverses its neighboring nodes and upstream nodes for each node through adjacency edges and hydrological flow direction edges, generating two types of heterogeneous messages. Adjacency messages represent the feature contribution of lateral ecological interactions, while hydrological flow direction messages represent the feature contribution of upstream confluence influence. Then, contextualized aggregation of the two types of heterogeneous messages is performed by combining global context vectors. The fusion weights of different types of messages can be dynamically adjusted according to the global evaluation context, avoiding feature bias caused by a single aggregation method. Finally, the current features of the node and the aggregated messages are fused and updated through a gating mechanism to generate the updated feature vector of the node, which is used as the input features for the next iteration. After multiple iterations, the network can output the high-order feature vector of each graph node. This high-order feature vector not only includes the static attributes and dynamic response features of the evaluation unit itself, but also deeply integrates the lateral interaction features of adjacent units, the confluence and transmission features of upstream units, and the global scenario constraint features. It can characterize the core influencing factors and mechanisms of the water conservation function of the evaluation unit from multiple dimensions and in depth.

[0062] S103: Based on high-order feature vectors and global context vectors, dynamic decision-making is performed through a preset adaptive multi-criteria decision-making model to obtain the final dynamic water conservation function evaluation value of each evaluation unit.

[0063] Specifically, water conservation is a complex eco-hydrological process with multiple dimensions. Its evaluation needs to cover multiple core criteria such as canopy litter interception, soil infiltration and water storage, and runoff regulation. However, the contribution weight of each criterion to the final evaluation result varies significantly depending on the attribute characteristics of different evaluation units and under different global evaluation scenarios. Therefore, an adaptive multi-criterion decision model can be adopted to achieve dynamic adaptation of criterion scoring and weight allocation.

[0064] For example, a predefined system of water conservation function criteria can be established first. This system, based on eco-hydrological mechanisms, includes multiple water conservation function criteria. Then, during the decision-making process of the adaptive multi-criteria decision model, an independent criterion scoring network can be configured for each water conservation function criterion. The high-order feature vectors of each evaluation unit are input into the corresponding network. This network can decode spatial correlation information and nonlinear interaction information related to specific criteria from the high-order features, and then obtain the initial score of each evaluation unit under each criterion through nonlinear mapping. Secondly, feature interaction can be performed based on the high-order feature vectors of the evaluation units and the global context vector to generate scenario-attribute fusion features that reflect the coupling relationship between the evaluation unit attributes and the global scenario. These features are then subjected to gating and normalization to obtain the dynamic weights corresponding to each criterion. These dynamic weights can adaptively adjust according to the characteristics of the evaluation unit and the global scenario, avoiding the evaluation bias caused by traditional fixed-weight decision-making. Finally, the initial scores of each criterion are multiplied element-wise with the corresponding dynamic weights to obtain weighted scores. Summing these weighted scores yields the preliminary decision results. This process achieves weighted integration of scores from various criteria, ensuring that the evaluation results reflect the dynamic contributions of each criterion. Furthermore, it allows for validity verification of preliminary decision-making results. For example, based on preset hydrological threshold ranges, outliers exceeding reasonable limits can be eliminated, ensuring the scientific validity and reliability of the final evaluation value. The hydrological threshold range can be determined based on regional historical hydrological data and ecological environment standards.

[0065] The aforementioned method first achieves an integrated representation of the evaluation unit's attribute features, spatial topology, and global context through multi-source data fusion and standardized preprocessing, overcoming the shortcomings of isolated data sources and fragmented spatial relationships in traditional methods, and enhancing the integrity of the data foundation. Second, high-order feature fusion based on spatiotemporal heterogeneous graphs and graph neural networks can capture the nonlinear interactions between factors such as forest stand structure, soil properties, topography, and rainfall, improving the depth and accuracy of feature representation. Finally, by dynamically adjusting weights through an adaptive multi-criteria decision model, the method achieves a dynamic response of water conservation function to seasonal changes and extreme events, effectively solving the problem that traditional static weights cannot adapt to the dynamic characteristics of hydrological processes, and enhancing the timeliness and decision support capability of the evaluation results.

[0066] In one embodiment, multi-source data undergoes standardized preprocessing to generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between evaluation units, and a global context vector, including:

[0067] The process involves acquiring static attribute data and recent rainfall time-series data for each evaluation unit from multi-source data; performing vectorization encoding on the static attribute data to obtain static attribute vectors for each evaluation unit; extracting time-series features from the recent rainfall time-series data to obtain recent dynamic feature vectors for each evaluation unit; aligning the static attribute vectors and recent dynamic feature vectors dimensionally, and then performing cross-feature interactive encoding to obtain initial attribute feature vectors for each evaluation unit.

[0068] The digital elevation model of the evaluation area and the vector boundary data of the evaluation units are extracted from the multi-source data; spatial matching processing is performed on the digital elevation model and the vector boundary data of the evaluation units to obtain spatial fusion data; based on the spatial fusion data, the boundary adjacency relationship of each evaluation unit is traversed through the R-tree indexing algorithm to obtain the spatial adjacency relationship of each evaluation unit.

[0069] Water flow direction grids are extracted from the spatially fused data. Based on these grids, the inflow and outflow units of each evaluation unit are determined using a confluence path tracing algorithm. Slope and soil infiltration rate data for each evaluation unit are obtained from multi-source data. Combined with the positional relationship between the inflow and outflow units, the hydrological conduction efficiency is quantified to obtain the quantified hydrological conduction efficiency. Based on the quantified hydrological conduction efficiency, flow direction relationship analysis is performed to obtain the hydrological flow direction relationship for each evaluation unit.

[0070] The spatial adjacency relationship and the hydrological flow direction relationship are integrated and processed. A preset boundary interaction intensity weight is added to the spatial adjacency relationship and a preset transmission efficiency weight is added to the hydrological flow direction relationship to obtain the spatial topological relationship between each evaluation unit.

[0071] Climate event data and seasonal background data at the evaluation time point are extracted from multi-source data; feature extraction is performed on the climate event data and seasonal background data respectively to obtain climate event features and seasonal background features. Climate event features include peak intensity, duration proportion and spatiotemporal distribution concentration of events, while seasonal background features include matching phenological periods and hydrological influence coefficients.

[0072] Climate event features and seasonal background features are normalized to obtain first and second standardized feature data. The first and second standardized feature data are then input into a preset attention mechanism network for scenario priority calculation to obtain a scenario priority weight vector. Based on the scenario priority weight vector, the first and second standardized feature data are weighted and fused to obtain fused feature data. The fused feature data is then input into a fully connected network for feature integration and encoding to generate a global context vector.

[0073] Specifically, categorical data in static attribute data can be transformed into machine-processable numerical forms through vectorization encoding, ensuring the effective representation of discrete attributes such as stand type and soil texture. Numerical data, on the other hand, can undergo dimensional unification to eliminate differences in numerical ranges between different attributes. For recent rainfall time-series data, a sliding window method can be used to segment the data, extracting key features from each segment. This preserves the temporal variation information of rainfall intensity while avoiding redundancy in the original time-series data. Through this process, static attribute vectors and recent dynamic feature vectors can be obtained. These two vectors can then be dimensionally aligned to ensure consistency in feature dimensions. For example, since the core influencing factors of water conservation function are not independent but rather the result of the coupling of static attributes and dynamic rainfall conditions—for instance, the synergistic effect of stand canopy density and rainfall intensity directly affects interception efficiency—simple vector concatenation cannot capture such nonlinear relationships. Therefore, cross-feature interactive encoding can be performed on the aligned vectors. For example, the static attribute vector can be initially set as... The recent dynamic feature vector is ,in For the feature dimension, the interactively encoded feature vector satisfy , It represents the element-wise multiplication of a vector, used to characterize the point-to-point interaction between static attributes and dynamic features; It is a learnable weight matrix used to adjust the contribution of different interaction features; This is a bias term used to optimize the feature distribution; It is a non-linear activation function used to capture the complex non-linear relationships between interactive features, ultimately generating... This is the initial attribute feature vector of each evaluation unit, which simultaneously carries the coupled information of the inherent characteristics and dynamic response of the evaluation unit.

[0074] Specifically, after extracting the digital elevation model (DEM) of the evaluation area and the vector boundary data of the evaluation units from multi-source data, spatial matching is first performed to unify the two types of data in the spatial reference system. For example, coordinate transformation can be used to map the DEM and the vector boundary data to the same projected coordinate system, ensuring accurate spatial correspondence and thus obtaining spatially fused data. Based on this data, the R-tree index algorithm can be used to traverse the boundary adjacency relationships of each evaluation unit. The R-tree index algorithm can efficiently organize spatial data by constructing a hierarchical minimum boundary rectangle index structure to quickly locate other evaluation units that overlap or are adjacent to the target evaluation unit's boundary, avoiding the inefficiency of traditional brute-force search. Illustratively, a minimum boundary rectangle can be constructed for the vector boundary of each evaluation unit, and all minimum boundary rectangles can be organized into an R-tree index structure. Then, using the minimum boundary rectangle of the target evaluation unit as the query condition, the R-tree index quickly filters out potential adjacent evaluation units. Finally, the boundary geometric relationships of the filtered evaluation units are verified to determine the set of evaluation units with actual boundary adjacencies, thus obtaining the spatial adjacency relationships of each evaluation unit.

[0075] Furthermore, flow direction grids can be extracted from spatially fused data. This grid data is obtained through hydrological analysis of a digital elevation model, with each grid cell representing the direction of the flow within that cell. Based on these flow direction grids, a confluence path tracing algorithm can be used to trace the source of water flow in all grids within each evaluation unit, identifying the upstream evaluation unit (i.e., the inflow unit) supplying water to that unit. Simultaneously, by tracing the flow destination of all grids within the evaluation unit, the downstream evaluation unit (i.e., the outflow unit) receiving the water flow from that unit can be determined, thus achieving precise characterization of the confluence path. In addition, to quantify hydrological transmission efficiency, slope and soil infiltration rate data for each evaluation unit can be obtained from multi-source data. Slope data reflects the flow velocity within the evaluation unit, while soil infiltration rate reflects the unit's ability to retain water flow; both together determine the transmission efficiency between evaluation units. Considering the positional relationship between the inflow and outflow units and the transmission loss from the inflow unit to the target unit, the soil infiltration rate of the target evaluation unit can be initially set as [value missing]. If the slope is α, then the hydrological transport efficiency is... In the formula The slope influence coefficient characterizes the inhibitory effect of slope on water flow conduction. A steeper slope results in faster water flow velocity, shorter infiltration time, and greater water loss during conduction; therefore, conduction efficiency decreases with increasing slope. Based on the quantified hydrological conduction efficiency, flow direction relationships are analyzed and processed, retaining only those with conduction efficiencies greater than a preset threshold. This ensures that the hydrological flow direction relationships accurately reflect the effective confluence effects between evaluation units, ultimately yielding the hydrological flow direction relationships for each evaluation unit.

[0076] Specifically, spatial adjacency relationships and hydrological flow direction relationships can be uniformly organized to form structured topological data. Illustratively, this process can assign boundary interaction strength weights to spatial adjacency relationships. These weights can be determined based on the overlap length of the adjacent evaluation units' boundaries and the similarity of forest stand types. The longer the overlap length and the more similar the forest stand types, the stronger the lateral ecological interaction, and the larger the weight value. Similarly, a transmission efficiency weight can be assigned to hydrological flow direction relationships. This weight can directly adopt the quantified hydrological transmission efficiency; the higher the transmission efficiency, the more significant the confluence impact of upstream evaluation units on downstream evaluation units. Through these weight assignments, the integrated spatial topological relationships not only contain binary information of "whether they are related" but also carry quantitative information on the strength of the association, providing rich spatial association features for subsequent heterogeneous map construction.

[0077] Furthermore, after extracting climate event data and seasonal background data for the evaluation time points from multi-source data, feature extraction can be performed separately. For climate event data, key indicators reflecting event intensity, duration, and impact range can be extracted. Among these, peak event intensity characterizes the maximum impact of a climate event, duration ratio characterizes the proportion of the climate event's duration to the evaluation period, and spatiotemporal distribution concentration characterizes the spatial clustering of climate events within the evaluation area. For seasonal background data, features can be extracted focusing on the impact of seasons on the hydrological cycle. Matching phenological periods characterize the vegetation growth stage corresponding to the current season, and the hydrological influence coefficient characterizes the activity level of the regional hydrological cycle during that season.

[0078] Furthermore, the extracted climate event features and seasonal background features can be normalized separately, mapping the numerical ranges of the two types of features to the [0,1] interval to eliminate the influence of dimensional differences on subsequent fusion, resulting in first-standardized feature data and second-standardized feature data. Subsequently, the first-standardized feature data can be set as... The second standardized feature data is ,in For the context feature dimension, the two types of features are concatenated into a context feature matrix. The context priority weight vector can be calculated using the attention mechanism. softmax In the formula and These are learnable weight matrices, used for feature transformation and attention score calculation, respectively. and For bias terms; The activation function is used to capture the non-linear correlation between features, and the softmax function is used to normalize the attention scores into a probability distribution. Each element corresponds to the priority weight of a scenario feature. The larger the weight value, the more significant the impact of the feature on the evaluation scenario.

[0079] Specifically, based on the scenario priority weight vector, a vector dot product can be performed on the first and second standardized feature data to modulate the features with priority weights, resulting in fused feature data. This data concentrates key information about climate events and seasonal backgrounds and reflects the priority differences of features in different scenarios. The fused feature data is then input into a fully connected network for feature integration and encoding. The fully connected network can compress the high-dimensional fused features into a low-dimensional, compact global context vector through multi-layer linear transformations and non-linear activations. This vector can accurately represent the global scenario constraints of the current evaluation, providing scenario-oriented support for subsequent feature fusion and dynamic decision-making.

[0080] In one embodiment, static attribute data is vectorized and encoded to obtain static attribute vectors for each evaluation unit; recent rainfall time-series data is subjected to time-series feature extraction to obtain recent dynamic feature vectors for each evaluation unit, including:

[0081] Static attribute data is divided into static categorical feature data and static numerical feature data; one-hot encoding is performed on the static categorical feature data to obtain the categorical encoding vector; extreme value standardization is performed on the static numerical feature data to obtain the numerical standardization vector.

[0082] Based on the preset forestry water source rules, the classification coding vector and the numerical standardization vector are coupled and interactively coded to obtain the static attribute vector of each evaluation unit.

[0083] Preprocessing operations are performed on recent rainfall time-series data to obtain normalized time-series data; the preprocessing operations include outlier identification and interpolation repair.

[0084] Based on a preset time window, the normalized time series data is segmented to obtain multiple local time series segments; local dependency features are extracted from each local time series segment through a temporal convolutional network to obtain segment-level features; and temporal trend aggregation of each segment-level feature is performed through a long short-term memory network to obtain the recent dynamic feature vector of each evaluation unit.

[0085] Specifically, static categorical feature data refers to data representing discrete category attributes, such as forest stand type, soil texture, and land use type. Since these lack numerical comparability, they can be converted into machine-processable numerical forms through specific encoding methods. Static numerical feature data refers to attribute data that can be represented by continuous numerical quantification, such as canopy closure, soil thickness, altitude, and slope. These can be standardized first to eliminate the influence of data heterogeneity. For example, for static categorical feature data, an independent binary dimension can be assigned to each category. When the attribute of a certain evaluation unit belongs to a specific category, the corresponding dimension has a value of 1, and the other dimensions have a value of 0, to avoid assigning unreasonable numerical ordering relationships to categorical features. For example, first, count all unique categories for each categorical attribute in the static categorical feature data to construct a category dictionary. Then, for the categorical attribute of each evaluation unit, generate a corresponding binary vector based on the category dictionary. This vector is the categorical encoding vector, and its dimension is equal to the total number of categories for the corresponding categorical attribute, ensuring that the semantic information of each categorical attribute is completely preserved.

[0086] As an illustration, for static numerical feature data, extreme value standardization can be used to map numerical features of different dimensions and numerical ranges to a unified interval, resulting in a numerically standardized vector.

[0087] Specifically, the pre-defined forestry water source rules are attribute association rules constructed based on eco-hydrological mechanisms, such as the matching rules between tree species type and soil texture, the synergistic effect rules between canopy closure and litter thickness, and the association rules between slope and soil infiltration rate, etc., to capture the nonlinear coupling effect between different attributes on water conservation functions. For example, coupled interactive coding can be achieved by constructing associated feature terms. First, let the classification coding vector be... The numerically standardized vector is ,in For classification feature dimensions, For numerical feature dimensions, coupled with the interactively encoded feature vector satisfy:

[0088]

[0089] in, This represents the Hadamard product operation, used to achieve point-to-point interaction between categorical features and numerical features; This is a correlation weight matrix constructed based on forestry water source rules, with dimensions of... The values ​​of the elements in the matrix are determined based on the correlation strength between the corresponding categorical attributes and numerical attributes. The higher the correlation strength, the greater the weight value of the corresponding element, which is used to strengthen the expression of key correlation features. and These represent the transposes of the classification coding vector and the numerical standardization vector, respectively. Through this coupled interactive coding process, the final static attribute vector not only contains the original information of various attributes but also incorporates the coupling correlation information between attributes, enabling a more accurate representation of the basic water conservation potential of the evaluation unit.

[0090] Specifically, for recent rainfall time-series data, the reasonable fluctuation range of the data can be determined by calculating the mean and standard deviation of the time-series data. Values ​​exceeding this range are identified as outliers and removed. These outliers may be caused by factors such as monitoring equipment malfunction or extreme interference. Subsequently, linear interpolation or spline interpolation methods can be used to estimate the values ​​based on the valid data before and after the outliers or missing points, filling in the data gaps and ensuring the continuity and integrity of the time-series data, thus obtaining normalized time-series data. The size of the preset time window can be determined according to the temporal distribution characteristics of regional rainfall, such as using 24 hours, 12 hours, or 6 hours as a time window, ensuring that the time-series data within each window can represent a complete local rainfall process. Based on this window, the normalized time-series data can be traversed using a sliding time window approach, dividing it into multiple non-overlapping local time-series segments. Each local time-series segment corresponds to the rainfall data within a time window, reflecting the dynamic characteristics of rainfall within that time period.

[0091] Furthermore, temporal convolutional networks possess the characteristics of causal convolution and dilated convolution. Causal convolution ensures that the feature extraction process does not depend on data from future moments, conforming to the temporal irreversibility of rainfall time series. Dilated convolution, on the other hand, can capture the dependencies between data at different time steps within a local temporal segment by expanding the receptive field, such as the abrupt correlation of rainfall intensity within a short period or the correlation of intensity changes before and after the rainfall peak. Each local temporal segment is used as input to the temporal convolutional network. Features are extracted from the local temporal data through the network's convolutional layers. After processing by the activation function, segment-level features that characterize the rainfall dependencies within that local segment are output. These features are in low-dimensional vector form, condensing the dynamic information of local rainfall. Finally, the segment-level features can be sequentially input into a long short-term memory network (LSTM). The network's forget gate can selectively forget less important segment features, the input gate can fuse and update the current segment features with historical features from the network cell state, and the output gate can output an aggregated feature vector based on the current cell state. This vector is the recent dynamic feature vector of each evaluation unit. This vector not only contains dynamic information about each local rainfall process, but also captures the temporal trend and cumulative effect of rainfall in different time periods, such as the impact of early rainfall on later soil moisture and the intensity change trend of continuous rainfall processes.

[0092] In one embodiment, a spatiotemporal heterogeneous graph of the watershed is constructed based on initial attribute feature vectors, spatial topological relationships, and global context vectors, including:

[0093] Each evaluation unit is mapped to a graph node, and the initial attribute feature vector is compressed using a pre-set embedding network to obtain the initial node feature vector for each graph node.

[0094] Based on spatial adjacency relationships in spatial topology, bidirectional adjacency edges are established between graph nodes corresponding to spatial adjacency relationships. Bidirectional adjacency edges are used to characterize the lateral ecological interaction relationships between adjacent evaluation units.

[0095] Based on the hydrological flow direction relationship in the spatial topology, a one-way hydrological flow direction edge is established between the upstream evaluation unit graph node and the downstream evaluation unit graph node corresponding to the hydrological flow direction relationship. The one-way hydrological flow direction edge is used to characterize the confluence influence relationship between the upstream evaluation unit and the downstream evaluation unit.

[0096] By attaching the global context vector as a global attribute to the graph structure and integrating the initial node feature vectors, bidirectional adjacent edges, unidirectional hydrological flow edges, and global graph attributes, a spatiotemporal heterogeneous graph of the watershed is obtained.

[0097] Specifically, each evaluation unit is mapped to an independent graph node in the graph structure, with each graph node uniquely corresponding to one evaluation unit. This ensures that the spatial independence and attribute uniqueness of the evaluation units are accurately represented in the graph structure. Since the initial attribute feature vectors may have high dimensionality and feature redundancy, directly using them as graph node features would increase the computational complexity of subsequent networks and affect feature fusion efficiency. Therefore, a pre-defined embedding network can be used to compress the dimensionality of the vectors, reducing the vector dimension while retaining core feature information, thus generating initial node feature vectors. The pre-defined embedding network can employ a multi-layer perceptron structure, including an input layer, hidden layers, and an output layer. The input layer dimension is consistent with the initial attribute feature vector dimension, the output layer dimension is the pre-defined target feature dimension, and the hidden layers use non-linear activation functions to enhance the non-linear expressive power of the features. For example, the dimensionality compression process satisfies the formula... , This represents the generated initial node feature vector. This represents the initial attribute feature vector of the input. This is a learnable weight matrix embedded in the network, used to achieve feature dimension transformation and feature importance modulation. This is a bias term embedded in the network, used to optimize feature distribution and improve subsequent processing performance. As a non-linear activation function, the ReLU function can be used, which can effectively alleviate the gradient vanishing problem and enhance the sparsity representation of features. Through this embedded network processing, the initial node feature vectors not only retain the core correlation information between the static attributes and dynamic responses of the evaluation units, but also have the characteristics of low dimensionality and high discriminability, laying the foundation for feature interaction of graph nodes.

[0098] Specifically, the lateral ecological interactions between adjacent evaluation units are reciprocal. For example, forest stands in adjacent areas can exchange water and regulate microclimate through canopy interleaving, and soil moisture can migrate between adjacent units through lateral infiltration. Such interactions are not unidirectional but bidirectional and reciprocal. Therefore, bidirectional adjacency edges can be constructed to accurately represent these interactions. For instance, a list of spatial adjacency relationships for each evaluation unit corresponding to a graph node can be extracted from the spatial topology to clarify the set of adjacent graph nodes for each graph node. Then, for each graph node and its adjacent graph nodes, undirected bidirectional adjacency edges are established between them. The existence of the edge is directly determined by the spatial adjacency relationship; that is, if two evaluation units have a spatial adjacency relationship, then there must be a bidirectional adjacency edge between the corresponding graph nodes. In addition, to enhance the semantic information of edges, the weight of bidirectional adjacent edges can be determined based on the boundary overlap length of adjacent evaluation units and the similarity of forest stand types. The longer the boundary overlap length and the more similar the forest stand types, the higher the intensity of lateral ecological interaction and the larger the weight value of the corresponding edge. Furthermore, the weight value can be normalized and mapped to the [0,1] interval, so that the edge weight can quantitatively represent the strength of lateral ecological interaction.

[0099] Due to the unidirectional and irreversible nature of hydrological runoff processes, hydrological processes such as rainfall runoff generation and soil infiltration recharge in upstream evaluation units directly affect the water balance and water conservation load of downstream evaluation units. However, downstream units have no direct hydrological transmission effect on upstream units. Therefore, unidirectional hydrological flow edges can be constructed to align with the natural laws of hydrological processes. For example, the hydrological flow direction relationship corresponding to each graph node can be extracted from the spatial topology, clarifying the upstream and downstream graph node sets for each node. For each upstream graph node and its corresponding downstream graph node, a unidirectional hydrological flow edge is established between them, pointing from the upstream node to the downstream node. The direction of the edge follows the hydrological flow direction relationship, ensuring that the transmission direction of runoff influence is consistent with the actual hydrological process. The weight of a unidirectional hydrological flow edge is related to the corresponding hydrological transmission efficiency. The higher the hydrological transmission efficiency, the more significant the confluence effect of the upstream unit on the downstream unit, and the larger the edge weight value can be. Through this weight, the intensity difference of the confluence effect can be quantified, so that the edge not only has directional information, but also carries the quantitative characteristics of the influence intensity, providing an intensity basis for subsequent heterogeneous message transmission.

[0100] Specifically, the global context vector carries contextual information such as climate events and seasonal background at the evaluation time point. This information can influence the hydrological cycle process of the entire evaluation area at a global level, thereby changing the mechanism of water conservation function of each evaluation unit. Therefore, the global context vector can be embedded as a global attribute field of the graph into the metadata of the graph structure, adding a global contextual feature dimension to the graph structure in addition to node and edge features. Subsequently, the initial node feature vectors, bidirectional adjacent edges, unidirectional hydrological flow edges, and global graph attributes can be organized into an input form that can be recognized by the graph neural network according to a preset format, thus forming a complete spatiotemporal heterogeneous graph of the watershed. This graph simultaneously carries the attribute features of the evaluation units, two types of heterogeneous spatial correlation features, and global contextual constraint features, realizing the organic integration of attribute-space-time three-dimensional information, and providing structured input data support for the subsequent feature fusion of spatiotemporal heterogeneous graph neural networks.

[0101] In one embodiment, the preset spatiotemporal heterogeneous graph neural network includes a heterogeneous message generation module, a contextualized aggregation module, and a node state update module. The watershed spatiotemporal heterogeneous graph is input into the preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the watershed spatiotemporal heterogeneous graph. The feature fusion processing corresponds to the following steps:

[0102] S111: Initialize the network parameters of the preset spatiotemporal heterogeneous graph neural network and set the preset iteration rounds; obtain the current feature vectors of each graph node in the spatiotemporal heterogeneous graph of the watershed; wherein, in the first iteration, the current feature vectors of each graph node are the initial node feature vectors;

[0103] S112: For each graph node, based on the adjacency message processing branch of the heterogeneous message generation module, the corresponding adjacent nodes are traversed through bidirectional adjacency edges, and the current feature vector of each adjacent node is linearly transformed through the first learnable parameter matrix to obtain the corresponding first transformed feature. The first transformed feature is aggregated through the adjacency message aggregation function to obtain the first aggregated message.

[0104] S113: For each graph node, the hydrological flow direction message processing branch based on the heterogeneous message generation module traverses the corresponding upstream nodes through the unidirectional hydrological flow direction edge, performs a linear transformation on the current feature vector of each upstream node through the second learnable parameter matrix to obtain the corresponding second transformed feature vector, and aggregates each second transformed feature vector through the hydrological flow direction message aggregation function to obtain the second aggregated message; wherein, the second learnable parameter matrix and the first learnable parameter matrix are different parameter matrices;

[0105] S114: The weight calculation submodule based on the contextualized aggregation module concatenates the current feature vector of the current graph node with the global context vector to obtain the concatenated feature; after performing nonlinear mapping on the concatenated feature, the dynamic fusion weight is calculated through the activation function; through the message fusion submodule of the contextualized aggregation module, the first aggregated message and the second aggregated message are weighted and summed based on the dynamic fusion weight to obtain the total aggregated message of the current graph node;

[0106] S115: Through the node state update module, the current feature vector, total aggregate message and global context vector of the current graph node are input into the preset state update function to calculate the updated feature vector of the current graph node; wherein the preset state update function is constructed based on the LSTM gating mechanism or the GRU gating mechanism;

[0107] S116: Use the updated feature vector as the current feature vector of the current graph node in the next iteration, and repeat steps S112 to S115 until the cumulative number of iterations reaches the preset number of iterations. Stop the iteration and output the updated feature vector obtained in the last iteration as the higher-order feature vector corresponding to each graph node.

[0108] Specifically, the Xavier initialization method can be used to initialize all learnable parameters of the pre-defined spatiotemporal heterogeneous graph neural network to ensure the stability of network training. A pre-defined number of iterations can be set, the value of which is determined based on the model's convergence characteristics and computational efficiency, typically by adaptively adjusting the loss trend on the validation set or using empirical values. Subsequently, the current feature vectors of each graph node in the spatiotemporal heterogeneous graph of the watershed are obtained. In the first iteration, the current feature vector is the initial node feature vector, which is obtained by dimensionality compression of the initial attribute feature vectors from the previous embedding network, possessing low dimensionality and high discriminative power.

[0109] Furthermore, for each graph node, its corresponding neighboring nodes can be traversed via bidirectional adjacency edges using the adjacency message processing branch of the heterogeneous message generation module. The first learnable parameter matrix can be an F×F dimension learnable weight matrix, used to linearly transform the current feature vectors of neighboring nodes, extracting information related to lateral ecological interactions from the neighboring node features to obtain the first transformed features. Subsequently, all the first transformed features can be aggregated using an adjacency message aggregation function, which can employ mean aggregation or attention aggregation to integrate the lateral interaction information of neighboring nodes into a unified feature vector, obtaining the first aggregated message. This message characterizes the comprehensive impact of the lateral ecological interactions between the graph node and its neighboring evaluation units on the water conservation function. Similarly, for each graph node, its corresponding upstream nodes can be traversed via unidirectional hydrological flow edges using the hydrological flow direction message processing branch of the heterogeneous message generation module. The second learnable parameter matrix can be an F×F dimension learnable weight matrix, but it is a different parameter matrix from the first learnable parameter matrix. The second learnable parameter matrix can be used to linearly transform the current feature vector of the upstream node, extracting feature information related to hydrological confluence, and obtaining the transformed feature vector. The hydrological flow direction message aggregation function is an attention-weighted summation aggregation, and the corresponding formula is:

[0110]

[0111] in, For the first During the nth iteration The second aggregated message of each graph node is an F-dimensional column vector, where F is the feature dimension hyperparameter of the pre-defined spatiotemporal heterogeneous graph neural network, which is pre-defined during the network structure design stage based on the data scale, feature complexity, and model training efficiency of the evaluation region. The second learnable parameter matrix for the hydrological flow direction message branch has dimensions of , used for linear transformation of upstream node features; For the first In the next iteration, the upstream node The current feature vector is an F-dimensional column vector; upstream node To the node The normalized hydrological transmission efficiency is obtained from the previous hydrological transmission efficiency quantification process, and its value range is [0,1]. It is used to characterize the hydrological transmission intensity from the upstream node to the target node. For the first The set consisting of all upstream nodes of a given graph node; The learnable vector for scoring attention, with dimension . It is obtained through model training; for The transpose of ; A learnable weight matrix for scoring attention, with dimensions of ; The hyperbolic tangent activation function is used to perform nonlinear mapping on the transformed features to generate intermediate features for attention scoring. This is the natural exponential function, used to calculate the exponential value of the attention score; For traversing the first Temporary variables for all upstream nodes of a given graph node.

[0112] Specifically, based on the above formula, we can first use... and The upstream node features are subjected to two linear transformations, followed by tanh activation to obtain the intermediate features for attention scoring, which are then further processed... The attention score for each upstream node is obtained by taking the dot product with the intermediate features. Then, the attention scores are expanded and normalized to obtain the attention weight for each upstream node. Finally, the attention weights are multiplied by the transformed features and the normalized hydrological transport efficiency, and the weighted sum is used to obtain the second aggregated message. This attention mechanism allows for dynamic focus on upstream nodes that have a greater impact on the water conservation function of the target node. For example, in heavy rainfall scenarios, nodes with steeper upstream slopes and lower soil infiltration rates have a more significant impact on downstream runoff, and their corresponding attention weights can be automatically increased. This allows the second aggregated message to more accurately represent the differentiated impact of upstream runoff.

[0113] Specifically, concatenating the current feature vector of the current graph node with the global context vector yields a concatenated feature. This concatenated feature contains both local attribute information and global contextual information of the graph node. After processing it through a nonlinear mapping such as a multilayer perceptron, it can be input into an activation function such as the Sigmoid function to calculate the dynamic fusion weight. This weight ranges from [0,1] and can be used to dynamically balance the contribution ratio of the first and second aggregated messages. For example, in a rainstorm scenario, the dynamic fusion weight can be tilted towards the second aggregated message to strengthen the representation of upstream confluence impact, while in a drought scenario, the dynamic fusion weight can be tilted towards the first aggregated message to strengthen the representation of lateral ecological interactions. Furthermore, the message fusion submodule of the contextualized aggregation module can perform a weighted summation of the first and second aggregated messages based on the dynamic fusion weight to obtain the total aggregated message of the current graph node. This message simultaneously carries comprehensive information on lateral ecological interactions and upstream confluence impact and is adapted to the current global context.

[0114] Specifically, the node state update module inputs the current feature vector, total aggregated message, and global context vector of the current graph node into a preset state update function built on an LSTM or GRU gating mechanism for node update. Illustratively, taking the GRU gating mechanism as an example, this function can include a reset gate and an update gate. The reset gate controls the degree of integration of the total aggregated message and global context vector, determining which information needs to be retained. The update gate controls the retention ratio of the current feature vector, determining the forgetting of old information and the integration of new information. Through the selective processing of this gating mechanism, the updated feature vector of the current graph node can be calculated. This vector retains the node's own historical features while incorporating new information from adjacency interactions, upstream confluence, and the global context, achieving dynamic feature updates.

[0115] Finally, multiple iterations and high-order feature outputs are performed. The updated feature vector can be used as the current feature vector of the current graph node in the next iteration. The above-mentioned adjacency message processing, hydrological flow direction message processing, contextual fusion, and node state update steps are repeated until the cumulative number of iterations reaches the preset number of iterations. At this point, the updated feature vector obtained in the last iteration is the high-order feature vector corresponding to each graph node. This vector deeply integrates the evaluation unit's own attributes, lateral interactions of adjacent units, the confluence influence of upstream units, and global scenario constraints. It can accurately characterize the core influencing factors of the evaluation unit's water conservation function from multiple dimensions, providing high-quality feature support for subsequent dynamic decision-making.

[0116] In one embodiment, the adaptive multi-criteria decision-making model includes a criterion scoring network, a dynamic weight generation network, and a decision fusion module; the criterion scoring network includes a feature decoding submodule and a nonlinear mapping layer, the dynamic weight generation network includes a context interaction submodule and a normalization layer, and the decision fusion module includes a weighted calculation unit and a rationality verification unit. Figure 2 As shown, based on high-order feature vectors and global context vectors, a pre-set adaptive multi-criteria decision-making model is used for dynamic decision-making to obtain the final dynamic water conservation function evaluation value of each evaluation unit, including:

[0117] S201: K water conservation function criteria conforming to eco-hydrological mechanisms are preset, and an independent criterion scoring network is configured for each water conservation function criterion; the high-order feature vectors corresponding to each evaluation unit are input into the K criterion scoring networks respectively, and the spatial correlation information and nonlinear interaction information in the high-order feature vectors are decoded by the feature decoding submodule to obtain the decoded features, and the decoded features are input into the nonlinear mapping layer for mapping to obtain the initial scores of K criteria for each evaluation unit under the K water conservation function criteria; where K is a preset positive integer; the water conservation function criteria include at least the canopy litter interception criterion, the soil infiltration water storage criterion, and the runoff regulation criterion;

[0118] S202: Based on the dynamic weight generation network, the high-order feature vectors of each evaluation unit are interacted with the global context vector to generate scenario-attribute fusion features; the scenario interaction submodule is used to gating the scenario-attribute fusion features to obtain gating features; the gating features are input into the normalization layer and normalized by the Softmax function to obtain K dynamic weights for each evaluation unit corresponding to K water conservation function criteria;

[0119] S203: Through the weighted calculation unit of the decision fusion module, for each evaluation unit, the K dynamic weights are multiplied element-wise with the corresponding K initial criterion scores to obtain K weighted scores; the K weighted scores are summed to obtain the preliminary decision result; through the rationality verification unit of the decision fusion module, the validity of the preliminary decision result is verified based on the preset hydrological threshold range, and outliers exceeding the preset reasonable range are removed to obtain the final dynamic water conservation function evaluation value of the evaluation unit.

[0120] Specifically, water conservation function criteria can be constructed by following the core eco-hydrological processes of forest water conservation. These criteria can include canopy litter interception criteria, soil infiltration and water storage criteria, and runoff regulation criteria. The canopy litter interception criteria characterize the interception capacity of the canopy and litter layers for rainfall; the soil infiltration and water storage criteria characterize the soil layer's capacity for water absorption and storage; and the runoff regulation criteria characterize the evaluation unit's ability to delay and regulate surface runoff. Furthermore, an independent criterion scoring network can be configured for each water conservation function criterion. Each scoring network includes a feature decoding submodule and a nonlinear mapping layer. The feature decoding submodule can adopt a multilayer perceptron structure, with its input dimension consistent with the higher-order feature vector dimension and its output dimension being a preset decoding feature dimension. This is used to specifically decode spatial correlation information and nonlinear interaction information related to the current criterion from the higher-order feature vector. For example, regarding the canopy litter interception criterion, the feature decoding submodule can extract the correlation features between stand canopy closure, litter thickness, and recent rainfall intensity from higher-order features, as well as the lateral influence features of adjacent evaluation unit stand structure on interception efficiency. For the soil infiltration and water storage criterion, it can extract the nonlinear interaction features of soil texture, slope, and the cumulative effect of previous rainfall. Illustratively, after inputting the higher-order feature vectors corresponding to each evaluation unit into the K-criterion scoring network, the decoded features are output. These features focus on the core influencing factors of the current criterion, eliminating irrelevant information interference. Subsequently, the decoded features can be input into a nonlinear mapping layer, using the Sigmoid activation function to map the decoded features to the [0,1] interval, obtaining the initial scores of each evaluation unit under the K water conservation function criteria.

[0121] Furthermore, the high-order feature vectors of each evaluation unit can be concatenated with the global context vector to obtain the scenario-attribute fusion feature. This feature simultaneously carries the spatial heterogeneity and nonlinear interaction information of the evaluation unit, as well as the climate scenario and seasonal background information at the current evaluation time. The scenario-attribute fusion feature is then input into the scenario interaction submodule, which is built based on a gating mechanism and includes a reset gate and an update gate. The reset gate controls the degree of integration of global context information; for example, in a rainstorm scenario, the reset gate can strengthen scenario information related to runoff regulation criteria. The update gate controls the retention ratio of spatial correlation information in the high-order features; for example, in a drought scenario, the update gate can strengthen attribute information related to soil infiltration and water storage criteria. Through gating adjustment, the gated feature can be output, which achieves dynamic adaptation between attribute information and scenario information, avoiding the rigidity problem of fixed weights. By inputting the gated feature into the normalization layer, the Softmax function can be used for normalization processing to obtain K dynamic weights for the K water conservation function criteria corresponding to each evaluation unit, and the sum of the dynamic weights of all criteria is 1. Furthermore, this weight value can be adaptively adjusted according to changes in scenario and attribute. For example, in a rainstorm scenario, the dynamic weight of the runoff regulation criterion can be significantly increased, while in a light rain scenario, the dynamic weight of the canopy litter interception criterion can be increased accordingly.

[0122] Finally, through the decision fusion module, for each evaluation unit, the K dynamic weights are first multiplied element-wise with the corresponding K initial criterion scores to obtain K weighted scores. These weighted scores represent the actual contribution of each criterion under the current scenario and attributes. Summing the K weighted scores yields the preliminary decision result, which is the sum of the contributions of each criterion and reflects the comprehensive water conservation capacity of the evaluation unit under the current spatiotemporal scenario. Furthermore, this result can be verified through a rationality verification unit. For example, based on regional historical hydrological monitoring data, ecological environment standards, and the theoretical upper limit of water conservation function, the hydrological threshold range can be determined first. For instance, the reasonable range for the water conservation function evaluation value is [0,1]. Preliminary decision results exceeding this range are considered outliers, which may originate from model noise or extreme data interference. For outliers, interpolation calibration or threshold truncation can be used for correction. For example, preliminary results exceeding 1 can be truncated to 1, and preliminary results less than 0 can be truncated to 0. Finally, the final dynamic water conservation function evaluation value of each evaluation unit can be obtained. This evaluation value has both mechanistic rationality and dynamic adaptability, and can provide accurate decision-making basis for forest management and water resource management.

[0123] Based on the same inventive concept, this application also provides a forestry water conservation function evaluation system for implementing the above-mentioned forestry water conservation function evaluation method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the forestry water conservation function evaluation system provided below can be found in the limitations of the forestry water conservation function evaluation method described above, and will not be repeated here.

[0124] In one exemplary embodiment, such as Figure 3 As shown, a forestry water conservation function evaluation system 300 is provided, comprising:

[0125] The data preprocessing and feature extraction module 301 is used to acquire multi-source data of the evaluation area including multiple evaluation units, perform standardized preprocessing on the multi-source data, and generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between each evaluation unit, and global context vectors.

[0126] The graph neural network encoding module 302 is used to construct a watershed spatiotemporal heterogeneous graph based on the initial attribute feature vector, spatial topological relationship and global context vector; the watershed spatiotemporal heterogeneous graph is input into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the watershed spatiotemporal heterogeneous graph.

[0127] The dynamic multi-criteria evaluation module 303 is used to make dynamic decisions based on high-order feature vectors and global context vectors through a preset adaptive multi-criteria decision model, so as to obtain the final dynamic water conservation function evaluation value of each evaluation unit.

[0128] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the forestry water conservation function evaluation method of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0129] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the forestry water conservation function evaluation method of this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0130] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for evaluating the water conservation function of forestry, characterized in that, The method includes: Acquire multi-source data of an evaluation region including multiple evaluation units, perform standardized preprocessing on the multi-source data, and generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between each evaluation unit, and global context vectors. Based on the initial attribute feature vector, the spatial topology relationship, and the global context vector, a watershed spatiotemporal heterogeneous graph is constructed; the watershed spatiotemporal heterogeneous graph is input into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the watershed spatiotemporal heterogeneous graph. Based on the higher-order feature vector and the global context vector, dynamic decision-making is performed through a preset adaptive multi-criteria decision-making model to obtain the final dynamic water conservation function evaluation value of each evaluation unit.

2. The method according to claim 1, characterized in that, The standardization preprocessing of the multi-source data to generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between the evaluation units, and global context vectors includes: The static attribute data and recent rainfall time-series data of each evaluation unit in the multi-source data are obtained; the static attribute data is vectorized and encoded to obtain the static attribute vector of each evaluation unit; the recent rainfall time-series data is extracted and processed to obtain the recent dynamic feature vector of each evaluation unit; the static attribute vector and the recent dynamic feature vector are dimensionally aligned and then cross-feature interactive encoding is performed to obtain the initial attribute feature vector of each evaluation unit. The digital elevation model of the evaluation area and the vector boundary data of the evaluation units are extracted from the multi-source data; spatial matching processing is performed on the digital elevation model and the vector boundary data of the evaluation units to obtain spatial fusion data; based on the spatial fusion data, the boundary adjacency relationship of each evaluation unit is traversed through the R-tree indexing algorithm to obtain the spatial adjacency relationship of each evaluation unit. The water flow direction grid is extracted from the spatial fusion data. Based on the water flow direction grid, the inflow unit and outflow unit of each evaluation unit are determined by the confluence path tracing algorithm. The slope data and soil infiltration rate data of each evaluation unit are obtained from the multi-source data. Combined with the positional relationship between the inflow unit and the outflow unit, the hydrological conduction efficiency is quantified to obtain the quantified hydrological conduction efficiency. Based on the quantified hydrological conduction efficiency, the flow direction relationship is analyzed to obtain the hydrological flow direction relationship of each evaluation unit. The spatial adjacency relationship and the hydrological flow direction relationship are integrated and processed. A preset boundary interaction intensity weight is added to the spatial adjacency relationship, and a preset transmission efficiency weight is added to the hydrological flow direction relationship to obtain the spatial topological relationship between each evaluation unit. Climate event data and seasonal background data at the evaluation time point are extracted from the multi-source data; feature extraction is performed on the climate event data and the seasonal background data respectively to obtain climate event features and seasonal background features. The climate event features include event peak intensity, duration proportion and spatiotemporal distribution concentration. The seasonal background features include matching phenological period and hydrological influence coefficient. The climate event features and the seasonal background features are normalized to obtain first standardized feature data and second standardized feature data. The first standardized feature data and the second standardized feature data are input into a preset attention mechanism network to perform scenario priority calculation to obtain a scenario priority weight vector. Based on the scenario priority weight vector, the first standardized feature data and the second standardized feature data are weighted and fused to obtain fused feature data. The fused feature data is input into a fully connected network to perform feature integration encoding to generate the global context vector.

3. The method according to claim 1, characterized in that, The construction of a watershed spatiotemporal heterogeneous graph based on the initial attribute feature vector, the spatial topological relationship, and the global context vector includes: Each evaluation unit is mapped to a graph node, and the initial attribute feature vector is compressed using a preset embedding network to obtain the initial node feature vector of each graph node. Based on the spatial adjacency relationship in the spatial topology, bidirectional adjacency edges are established between the graph nodes corresponding to the spatial adjacency relationship. The bidirectional adjacency edges are used to characterize the lateral ecological interaction relationship between adjacent evaluation units. Based on the hydrological flow direction relationship in the spatial topology, a one-way hydrological flow direction edge is established between the upstream evaluation unit graph node and the downstream evaluation unit graph node corresponding to the hydrological flow direction relationship. The one-way hydrological flow direction edge is used to characterize the confluence influence relationship between the upstream evaluation unit and the downstream evaluation unit. The global context vector is attached to the graph structure as a global graph attribute, and the initial node feature vector, the bidirectional adjacent edges, the unidirectional hydrological flow edges, and the global graph attribute are integrated to obtain the watershed spatiotemporal heterogeneous graph.

4. The method according to claim 3, characterized in that, The preset spatiotemporal heterogeneous graph neural network includes a heterogeneous message generation module, a contextualized aggregation module, and a node state update module; The step of inputting the spatiotemporal heterogeneous graph of the watershed into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the spatiotemporal heterogeneous graph of the watershed, the feature fusion processing corresponds to the following steps: S1: Initialize the network parameters of the preset spatiotemporal heterogeneous graph neural network and set the preset iteration rounds; obtain the current feature vector of each graph node in the spatiotemporal heterogeneous graph of the watershed; wherein, in the first iteration, the current feature vector of each graph node is the initial node feature vector; S2: For each graph node, based on the adjacency message processing branch of the heterogeneous message generation module, the corresponding adjacent nodes are traversed through the bidirectional adjacency edge, and the current feature vector of each adjacent node is linearly transformed through the first learnable parameter matrix to obtain the corresponding first transformed feature. The first transformed feature is aggregated through the adjacency message aggregation function to obtain the first aggregated message. S3: For each graph node, based on the hydrological flow direction message processing branch of the heterogeneous message generation module, the corresponding upstream nodes are traversed through the unidirectional hydrological flow direction edge. The current feature vector of each upstream node is linearly transformed using the second learnable parameter matrix to obtain the corresponding second transformed feature vector. The second transformed feature vector is then aggregated using the hydrological flow direction message aggregation function to obtain the second aggregated message. The second learnable parameter matrix and the first learnable parameter matrix are different parameter matrices. S4: Based on the weight calculation submodule of the contextualized aggregation module, the current feature vector of the current graph node is concatenated with the global context vector to obtain the concatenated feature; after performing nonlinear mapping processing on the concatenated feature, the dynamic fusion weight is calculated through the activation function; through the message fusion submodule of the contextualized aggregation module, the first aggregated message and the second aggregated message are weighted and summed based on the dynamic fusion weight to obtain the total aggregated message of the current graph node; S5: The node state update module inputs the current feature vector of the current graph node, the total aggregated message, and the global context vector into a preset state update function to calculate the updated feature vector of the current graph node; wherein the preset state update function is constructed based on an LSTM gating mechanism or a GRU gating mechanism. S6: Use the updated feature vector as the current feature vector of the current graph node in the next iteration, and repeat steps S2 to S5 until the cumulative number of iterations reaches the preset number of iterations. Stop the iteration and output the updated feature vector obtained in the last iteration as the higher-order feature vector corresponding to each graph node.

5. The method according to claim 3, characterized in that, The adaptive multi-criteria decision model includes a criterion scoring network, a dynamic weight generation network, and a decision fusion module; the criterion scoring network includes a feature decoding submodule and a nonlinear mapping layer, the dynamic weight generation network includes a scenario interaction submodule and a normalization layer, and the decision fusion module includes a weighted calculation unit and a rationality verification unit. The process of making dynamic decisions based on the higher-order feature vector and the global context vector, using a preset adaptive multi-criteria decision model, to obtain the final dynamic water conservation function evaluation value for each evaluation unit includes: K water conservation function criteria conforming to eco-hydrological mechanisms are preset, and an independent criterion scoring network is configured for each water conservation function criterion. The high-order feature vectors corresponding to each evaluation unit are input into the K criterion scoring networks respectively. The spatial correlation information and nonlinear interaction information in the high-order feature vectors are decoded by the feature decoding submodule to obtain decoded features. The decoded features are then input into the nonlinear mapping layer for mapping to obtain the K initial scores of each evaluation unit under the K water conservation function criteria. Here, K is a preset positive integer. The water conservation function criteria include at least the canopy litter interception criterion, the soil infiltration water storage criterion, and the runoff regulation criterion. According to the dynamic weight generation network, the high-order feature vectors of each evaluation unit are interacted with the global context vector to generate scenario-attribute fusion features; the scenario interaction submodule is used to gating the scenario-attribute fusion features to obtain gating features; the gating features are input into the normalization layer and normalized using the Softmax function to obtain K dynamic weights for each evaluation unit corresponding to K water conservation function criteria; Through the weighted calculation unit of the decision fusion module, for each evaluation unit, the K dynamic weights are multiplied element-wise with the corresponding K initial criterion scores to obtain K weighted scores; the K weighted scores are summed to obtain a preliminary decision result; through the rationality verification unit of the decision fusion module, the validity of the preliminary decision result is verified based on a preset hydrological threshold range, and outliers exceeding the preset reasonable range are removed to obtain the final dynamic water conservation function evaluation value of the evaluation unit.

6. The method according to claim 2, characterized in that, The static attribute data is vectorized and encoded to obtain the static attribute vector of each evaluation unit; the recent rainfall time series data is subjected to time series feature extraction to obtain the recent dynamic feature vector of each evaluation unit, including: The static attribute data is divided into static categorical feature data and static numerical feature data; the static categorical feature data is one-hot encoded to obtain a categorical encoding vector; the static numerical feature data is subjected to extreme value standardization to obtain a numerical standardization vector. Based on the preset forestry water source rules, the classification coding vector and the numerical standardization vector are coupled and interactively coded to obtain the static attribute vector of each evaluation unit. Preprocessing operations are performed on the recent rainfall time-series data to obtain normalized time-series data; the preprocessing operations include outlier identification and interpolation repair. Based on a preset time window, the normalized time-series data is segmented to obtain multiple local time-series segments; local dependency features are extracted from each local time-series segment through a temporal convolutional network to obtain segment-level features. By aggregating the temporal trends of each segment-level feature using a long short-term memory network, the recent dynamic feature vector of each evaluation unit is obtained.

7. The method according to claim 4, characterized in that, The second aggregated message is calculated using the following formula: in, For the first During the nth iteration The second aggregated message of each graph node; This is the aggregation function for the hydrological flow direction message; This is the second learnable parameter matrix for the hydrological flow direction message branch; For the first In the next iteration, the upstream node The current feature vector; upstream node To the node Normalized hydrological transport efficiency; For the first The set consisting of the upstream nodes of each of the graph nodes; Learnable vectors for scoring attention; for The transpose of ; A learnable weight matrix for scoring attention; It is the hyperbolic tangent activation function; It is a natural exponential function; For traversing the first Temporary variables of each upstream node of each graph node.

8. A forestry water conservation function evaluation system, characterized in that, The system includes: The data preprocessing and feature extraction module is used to acquire multi-source data of the evaluation region including multiple evaluation units, perform standardized preprocessing on the multi-source data, and generate initial attribute feature vectors for each evaluation unit, spatial topological relationships between each evaluation unit, and global context vectors. The graph neural network encoding module is used to construct a watershed spatiotemporal heterogeneous graph based on the initial attribute feature vector, the spatial topology relationship and the global context vector; and input the watershed spatiotemporal heterogeneous graph graph into a preset spatiotemporal heterogeneous graph neural network for feature fusion processing to obtain the high-order feature vector of each graph node in the watershed spatiotemporal heterogeneous graph graph. The dynamic multi-criteria evaluation module is used to make dynamic decisions based on the higher-order feature vector and the global context vector through a preset adaptive multi-criteria decision model, so as to obtain the final dynamic water conservation function evaluation value of each evaluation unit.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images