Karst region ecological restoration effect prediction method and system based on machine learning

Abstract

The application discloses a karst region ecological restoration effect prediction method and system based on machine learning, comprising collecting multi-source heterogeneous data of a preset karst region, and preprocessing the multi-source heterogeneous data; the multi-source heterogeneous data comprises remote sensing time series data, geological data, restoration measure data, ecological monitoring data and meteorological data; a karst habitat comprehensive gradient index is constructed through the multi-source heterogeneous data, and a monitoring unit is obtained through karst habitat gradient self-adaptive partitioning according to the karst habitat comprehensive gradient index; geological features are obtained through double-layer hydrogeological prior feature extraction in the monitoring unit, and a karst pipeline network topology graph is constructed; a karst region ecological restoration effect prediction model is constructed according to the karst pipeline network topology graph and the geological features, and prediction data is input into the karst region ecological restoration effect prediction model to output a prediction result.

Description

Technical Field

[0001] This invention relates to the field of ecological restoration effect prediction technology, and in particular to a method and system for predicting the ecological restoration effect in karst areas based on machine learning. Background Technology

[0002] Karst is widely distributed in my country. These regions have unique geological structures, characterized by shallow soil, high rock exposure, complex hydrological processes, and fragile ecosystems, making them key areas for ecological restoration and management. Traditional assessments of karst ecological restoration effectiveness rely heavily on field monitoring and manual statistics, which suffer from drawbacks such as long cycles, high costs, uneven spatial coverage, and difficulty in dynamic tracking. These methods cannot quickly respond to the long-term effects and regional differences of restoration measures.

[0003] Existing ecological prediction models often focus on single dimensions such as surface vegetation and soil, failing to fully couple the dual-layer hydrogeological structure of karst areas and neglecting the crucial impact of underground spaces like conduits and fissures on water transport and solute transport. This leads to significant discrepancies between predicted results and actual restoration effects. Furthermore, karst ecological data exhibits multi-source heterogeneity, encompassing various types of information including remote sensing time series, geological, meteorological, restoration engineering, and ecological monitoring data. Traditional data processing methods struggle to efficiently integrate and uncover deeper patterns. Some machine learning-based ecological prediction models fail to adaptively partition karst habitats based on gradient differences, resulting in weak model generalization. Moreover, they lack prior hydrogeological knowledge and physical constraints, often relying purely on data-driven approaches and failing to reflect core physical processes such as water balance and solute transport in karst areas, thus lacking scientific interpretability and reliability in their predictions.

[0004] With the large-scale advancement of ecological restoration projects, there is an urgent need for an intelligent prediction method that can integrate multi-source data, couple surface-subsurface characteristics, and embed karst physical mechanisms to achieve accurate, rapid, and comprehensive quantitative assessment of restoration effects, and support the optimization and decision-making of ecological restoration plans in karst areas. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based method for predicting the effects of ecological restoration in karst areas.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Multi-source heterogeneous data of a pre-defined karst region were collected and preprocessed. The multi-source heterogeneous data included remote sensing time-series data, geological data, remediation measure data, ecological monitoring data, meteorological data, and soil moisture monitoring data. The geological data included digital elevation data, karst development intensity index, rock exposure rate, soil thickness, slope, aspect, surface layer thickness, conduit flow density, and fissure development degree. The remediation measure data included restoration measures, implementation period, engineering density, and species configuration patterns. The ecological monitoring data included vegetation cover, soil erosion modulus, biomass, groundwater quality indicators, and biodiversity index. The meteorological data included precipitation, evaporation, temperature, and relative humidity. A comprehensive gradient index for karst habitats is constructed using the multi-source heterogeneous data, and monitoring units are obtained by adaptively partitioning the karst habitat gradient based on the comprehensive gradient index. Geological features are obtained by performing double-layer hydrogeological prior feature extraction within the monitoring unit, and a karst pipeline network topology map is constructed. Based on the karst pipeline network topology and the geological features, a prediction model for the ecological restoration effect in karst areas is constructed. The data to be predicted is input into the prediction model for the ecological restoration effect in karst areas, and the prediction results are output.

[0007] Furthermore, the method for constructing a comprehensive gradient index of karst habitats using the aforementioned multi-source heterogeneous data includes: The pre-defined karst region is initially divided into multiple areas, and the comprehensive gradient index of karst habitat in each area is calculated. ; in Let z be the karst development intensity index of the z-th region. Let z be the normalized soil thickness of the z-th region. Let z be the rock exposure rate of the z-th region. Let z be the slope of the z-th region. Let z be the normalized surface water storage capacity index of the z-th region. Let z be the comprehensive gradient index of karst habitat in region z. As a weight for karst development, For fatness weight, Weight of exposed rock, For slope weight, Weighting based on drought and flood levels.

[0008] Furthermore, a method for obtaining monitoring units by adaptive zoning of karst habitat gradients based on the comprehensive gradient index of karst habitats includes: The comprehensive gradient index of karst habitat is normalized, a fuzziness parameter is set, and a candidate range for the number of clusters is defined; AFCM clustering is performed for each candidate s value. Membership matrix randomly generated based on candidate cluster numbers ,satisfy and Update cluster centers Update membership The membership matrix is ​​updated based on the membership degrees. Iteration stops when the difference between the membership matrices of two iterations is less than a difference threshold. The center of cluster v is the center of cluster v in the (t+1)th iteration. Let r be the membership degree of sample r to cluster v in the t-th iteration. Let r be the membership degree of sample r to cluster v at the (t+1)th iteration. The normalized karst development intensity index for sample r; Calculate the Shebiny clustering effectiveness index The optimal number of clusters is selected by choosing the s-value with the smallest Shebiny clustering effectiveness index. ;in The total number of samples, This represents the number of clusters in the current test. Let r be the membership degree of sample r to cluster v. For fuzzy weighted index, The karst habitat gradient index for sample r. For the v-th cluster center, The squared distance from the sample to the cluster center; For each sample point, the index of the maximum membership degree is used as the hard partition label. Labels are assigned to corresponding spatial units based on geographic coordinates, using an 8-neighborhood majority voting mechanism. If the current label has a lower percentage of votes within its neighborhood, the majority vote is applied. If the label is incorrect, it should be corrected to a neighborhood-dominant label; among which... For the smoothing threshold, Let r be the hard partition label of sample r. Let r be the membership degree of sample r to cluster v under the optimal number of clusters. Comprehensive area weight Weights related to heterogeneity Calculate the number of monitoring units in each gradient domain Constructing spatial dispersion Habitat representativeness Using a joint objective function, a greedy iterative strategy is employed to progressively select the optimal subset, and the selected optimal subset is output as the monitoring unit. ;in The number of monitoring units for cluster v. This represents the total number of monitoring units in the target area. These are the weight coefficients for cluster v. It is a rounding function. The minimum number of monitoring units for cluster v. A unique number for the monitoring unit. The geographic coordinates of the monitoring unit For the hard partition label of the monitoring unit, The average karst habitat gradient index of the area where the monitoring unit is located. The key habitat indicator vector includes soil thickness, rock exposure rate, slope, and surface zone water storage capacity.

[0009] Furthermore, the method for obtaining geological features by extracting prior hydrogeological features from two layers within the monitoring unit includes: Obtain the surface zone storage capacity index of the i-th monitoring unit. Soil thickness and karst development intensity index Introducing the enhancement coefficient of karst development intensity on the surface water-holding capacity. The surface layer thickness is weighted and corrected to obtain the effective surface layer storage thickness. ;in It was determined by regional hydrogeological surveys; Obtain the region's average annual precipitation The surface zone storage coefficient is calculated based on the ratio of the effective surface zone storage thickness to the average annual precipitation. ; Obtain the slope, pipe flow density, and fracture development degree. Calculate the pipe flow response coefficient based on the sum of the pipe flow density and the equivalent fracture density. ;in This is a conversion factor between the water conductivity of the crack and the pipe. This is the slope attenuation coefficient. Let be the slope of the i-th monitoring unit. Let be the pipe flow density of the i-th monitoring unit. Let represent the crack development degree of the i-th monitoring unit; Based on hydrogeological surveys, tracer experiments, or geophysical exploration results, a binary connectivity indicator function is constructed. The function is 1 if there is pipe connectivity between monitoring units, and 0 otherwise. For areas where detailed exploration has not been conducted, the karst development intensity index and pipe flow density are used for joint inference, expressed as: ; in Let i be the karst development intensity index of the i-th monitoring unit. Let j be the karst development intensity index of the j-th spatial unit. Let J be the pipe flow density of the j-th monitoring unit. The connectivity threshold for karst development intensity. The connectivity threshold for the flow density in the pipe; Calculate the hydrological connectivity probability between spatial units: ; in Let be the Euclidean distance between the i-th spatial unit and the j-th spatial unit. For spatial attenuation parameters, A binary connectivity indicator function based on hydrogeological surveys; The surface zone regulation coefficient, pipeline flow response coefficient, and hydrological connectivity index are output as geological characteristics.

[0010] Furthermore, the method for constructing a karst conduit network topology includes: The monitoring unit is used as a node. An edge is established when two nodes are connected by a pipe or crack. Based on the results of hydrogeological surveys and tracing experiments, a karst hydrological connectivity map is constructed. ;in For a set of nodes, Let be the set of edges. It is an adjacency matrix; Constructing a temporal feature matrix of surface ecology A hybrid architecture combining convolutional neural networks and long short-term memory networks is used to extract temporal features of the land surface, obtaining the features of the land surface ecological channels, expressed as: The temporal characteristic matrix of land surface ecology includes remote sensing vegetation indices, meteorological drivers, soil physicochemical properties, and remediation engineering parameters; among which... For convolutional neural networks, extract spatial texture features. For long short-term memory networks, capturing seasonal to interannual dynamics; A hybrid architecture combining graph attention networks and Transformers is employed. The initial node features include physical embedding features and the hidden state of the surface branch output. The graph attention layer updates the node features. ;in Let i be the groundwater node feature updated at the graph attention layer. For activation function, Let be the set of neighboring nodes of node i. The weight matrix is ​​a learnable matrix. Let be the attention weight of node j to node i. Let be the initial feature vector of node j; Hydrological connectivity probability is used as the weight of the edges in the karst pipeline network topology graph. Adjacency matrix Elements, through hydrological connectivity Prior weighting is used to obtain attention weights, expressed as: ; in It is the ReLU activation function. For attention weight vectors, Let W be the feature vector of node i after transformation by the weight matrix W. This is a vector concatenation symbol; Introducing a temporal Transformer to capture the hysteresis of subsurface responses and obtain subsurface hydrological channels. ;in Let be the pipeline flow transmission delay time from node j to node i. The underground hydrological channel characteristics of node i.

[0011] Furthermore, the method for constructing a prediction model for the ecological restoration effect in the karst area includes: Obtain groundwater recharge modulus underground leakage rate and karst hydrodynamic regulation modulus Constructing physical information feature layers based on geological characteristics The expression is: ;in For transpose, The surface layer has a storage capacity coefficient. The pipe flow response coefficient; Based on the weak-form constraint coding of karst water balance and solute transport, the physical constraint layer of the karst conduit network topology is obtained, and its expression is: ; This includes soil water storage. For precipitation, Evaporation rate For pipeline flow rate, This refers to the fracture flow discharge rate. For the correction item of leakage caused by ecological restoration, This refers to the solute concentration. The hydrodynamic dispersion coefficient, The average pore flow velocity, For gradient operators, This represents the spatial gradient of solute concentration. For divergence operators, The rate of change of soil water storage over time. For adsorption, The weights for water balance constraints, The weights of the solute transport constraint term. Soil water storage capacity; A gating fusion mechanism is adopted to dynamically adjust the contribution weights of surface and underground channels based on the geological characteristics of each node, thereby obtaining fusion characteristics: ; ; in Let i represent the surface ecological corridor characteristics of node i. The underground hydrological channel characteristics of node i, Let i be the fusion feature of node i. Let i be the gating weight. The weight matrix is ​​a learnable matrix; The fused features are input into the multi-task output layer to simultaneously predict the multi-dimensional effects of ecological restoration, resulting in an ecological restoration prediction, expressed as: ; in For the ecological restoration prediction of node i, Let i be the vegetation cover restoration rate at node i. Let be the change in soil erosion modulus at node i. Let i be the comprehensive groundwater quality index. Let be the biodiversity recovery index for node i. The weight matrix of the output layer. This is the bias vector for the output layer; The total loss function is given based on the data-driven term, the physical consistency term, and the habitat partitioning regularization term. The expression is: in For data loss, For habitat regularization, For the actual observed value of node i, The total number of samples, The number of habitat gradient domains, Let be the parameter matrix of the a-th habitat-specific subnetwork. The basic parameter matrix is ​​shared globally. The square of the Hilbert-Schmidt norm. The weight of the habitat domain regularization term, The weight of the physical regularization term; Using Monte Carlo Dropout or deep ensemble methods to quantify prediction uncertainty, for K stochastic forward propagations or K independent sub-models, the prediction mean and variance are: in Let be the mean of node i. Let be the variance of node i. For the random number of forward propagations, For the k-th ecological restoration prediction, training continues until the variance is less than the variance threshold, at which point training stops.

[0012] Secondly, a machine learning-based system for predicting the effects of ecological restoration in karst areas includes: Data Acquisition and Processing Module: Used to collect multi-source heterogeneous data of a preset karst area and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes remote sensing time-series data, geological data, restoration measure data, ecological monitoring data, and meteorological data; the geological data includes digital elevation, karst development intensity index, rock exposure rate, soil thickness, slope, aspect, surface layer thickness, conduit flow density, and fissure development degree; the restoration measure data includes restoration measures, implementation period, engineering density, and species configuration pattern; the ecological monitoring data includes vegetation cover, soil erosion modulus, biomass, groundwater quality indicators, and biodiversity index; the meteorological data includes precipitation, evaporation, temperature, and relative humidity; Index construction and partitioning module: used to construct a comprehensive gradient index of karst habitat through the multi-source heterogeneous data, and to obtain monitoring units by adaptive partitioning of karst habitat gradient based on the comprehensive gradient index of karst habitat; Geological Feature Extraction and Topology Map Construction Module: Used to extract geological features by performing two-layer hydrogeological prior features within the monitoring unit, and to construct a karst pipeline network topology map; Model building and output module: This module is used to build a prediction model for the ecological restoration effect of karst areas based on the karst pipeline network topology map and the geological features. It inputs the data to be predicted into the prediction model for the ecological restoration effect of karst areas and outputs the prediction results.

[0013] The beneficial effects of this invention are: This invention relates to a machine learning-based method and system for predicting the effects of ecological restoration in karst areas. Compared with existing technologies, this invention has the following technical advantages: This invention integrates multi-source heterogeneous data and constructs a comprehensive gradient index for karst habitats to achieve adaptive zoning, improving the representativeness and spatial rationality of monitoring units. It extracts prior hydrogeological features from two layers to construct a karst conduit network topology map, accurately depicting the surface-to-ground hydrological connectivity mechanism and solving the problem of traditional models neglecting karst structures. It employs a hybrid neural network architecture with embedded physical constraints, balancing data-driven approaches with hydrological patterns, resulting in more accurate and interpretable predictions. It achieves simultaneous prediction of multi-dimensional restoration effects, outputting quantified uncertainty results to support scientific decision-making. The entire process is intelligent, reducing field monitoring costs and improving prediction efficiency and engineering applicability. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of the machine learning-based method for predicting the ecological restoration effect in karst areas according to the present invention. Detailed Implementation

[0015] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0016] The present invention provides a machine learning-based method and system for predicting the effects of ecological restoration in karst areas, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Multi-source heterogeneous data of a pre-defined karst region were collected and preprocessed. The multi-source heterogeneous data included remote sensing time-series data, geological data, remediation measure data, ecological monitoring data, meteorological data, and soil moisture monitoring data. The geological data included digital elevation data, karst development intensity index, rock exposure rate, soil thickness, slope, aspect, surface layer thickness, conduit flow density, and fissure development degree. The remediation measure data included restoration measures, implementation period, engineering density, and species configuration patterns. The ecological monitoring data included vegetation cover, soil erosion modulus, biomass, groundwater quality indicators, and biodiversity index. The meteorological data included precipitation, evaporation, temperature, and relative humidity. In the actual assessment, the ecological restoration area of ​​the karst peak cluster depression in H County was selected as the study area, which covers an area of ​​approximately 25 km². 2 It has a subtropical monsoon climate with an average annual precipitation of 1800 mm. The rock exposure rate is 45%~75%, the soil thickness is 5~30 cm, and karst channels and fissures are well developed. In 2020, a comprehensive ecological restoration project for rocky desertification control was launched. Six types of data were collected. After preprocessing including denoising, normalization, and spatial registration (30m resolution), some core data are as follows: Remote sensing time-series data includes NDVI and EVI vegetation indices, with quarterly averages ranging from 0.35 to 0.72 from 2019 to 2024; restoration measures data includes restoration measures of enclosure and afforestation, implemented for 5 years, with an engineering density of 0.12 and a species configuration of cypress and toon; ecological monitoring data includes a vegetation cover of 68% and a soil erosion modulus of 120 t / (km²). 2 •a) Biodiversity index: 2.8; A comprehensive gradient index for karst habitats is constructed using the multi-source heterogeneous data, and monitoring units are obtained by adaptively partitioning the karst habitat gradient based on the comprehensive gradient index. Geological features are obtained by performing double-layer hydrogeological prior feature extraction within the monitoring unit, and a karst pipeline network topology map is constructed. Based on the karst pipeline network topology and the geological features, a prediction model for the ecological restoration effect in karst areas is constructed. The data to be predicted is input into the prediction model for the ecological restoration effect in karst areas, and the prediction results are output.

[0017] In this embodiment, the method for constructing a comprehensive gradient index of karst habitats using the multi-source heterogeneous data includes: The pre-defined karst region is initially divided into multiple areas, and the comprehensive gradient index of karst habitat in each area is calculated. ; in Let z be the karst development intensity index of the z-th region. Let z be the normalized soil thickness of the z-th region. Let z be the rock exposure rate of the z-th region. Let z be the slope of the z-th region. Let z be the normalized surface water storage capacity index of the z-th region. Let z be the comprehensive gradient index of karst habitat in region z. As a weight for karst development, For fatness weight, Weight of exposed rock, For slope weight, Weighting based on drought and flood levels; In actual assessment, , , , , The sum is 1 when determined using the entropy weight method. The study area was divided into a 100m × 100m grid, comprising 625 units. It is 0.28. It is 0.22. It is 0.25. It is 0.15. The value is 0.1, for core repair area unit 1: It is 0.72. It is 0.65. It is 0.58. It is 22°. The value is 0.7, and the comprehensive gradient index of karst habitat is 0.562, indicating a medium habitat gradient.

[0018] In this embodiment, the method for obtaining monitoring units by adaptive zoning of karst habitat gradient based on the comprehensive gradient index of karst habitat includes: The comprehensive gradient index of karst habitat is normalized, a fuzziness parameter is set, and a candidate range for the number of clusters is defined; AFCM clustering is performed for each candidate s value. Membership matrix randomly generated based on candidate cluster numbers ,satisfy and Update cluster centers Update membership The membership matrix is ​​updated based on the membership degrees. Iteration stops when the difference between the membership matrices of two iterations is less than a difference threshold. The center of cluster v is the center of cluster v in the (t+1)th iteration. Let r be the membership degree of sample r to cluster v in the t-th iteration. Let r be the membership degree of sample r to cluster v at the (t+1)th iteration. The normalized karst development intensity index for sample r; Calculate the Shebiny clustering effectiveness index The optimal number of clusters is selected by choosing the s-value with the smallest Shebiny clustering effectiveness index. ;in The total number of samples, This represents the number of clusters in the current test. Let r be the membership degree of sample r to cluster v. For fuzzy weighted index, The karst habitat gradient index for sample r. For the v-th cluster center, The squared distance from the sample to the cluster center; For each sample point, the index of the maximum membership degree is used as the hard partition label. Labels are assigned to corresponding spatial units based on geographic coordinates, using an 8-neighborhood majority voting mechanism. If the current label has a lower percentage of votes within its neighborhood, the majority vote is applied. If the label is incorrect, it should be corrected to a neighborhood-dominant label; among which... For the smoothing threshold, Let r be the hard partition label of sample r. Let r be the membership degree of sample r to cluster v under the optimal number of clusters. Comprehensive area weight Weights related to heterogeneity Calculate the number of monitoring units in each gradient domain Constructing spatial dispersion Habitat representativeness Using a joint objective function, a greedy iterative strategy is employed to progressively select the optimal subset, and the selected optimal subset is output as the monitoring unit. ;in The number of monitoring units for cluster v. This represents the total number of monitoring units in the target area. These are the weight coefficients for cluster v. It is a rounding function. The minimum number of monitoring units for cluster v. A unique number for the monitoring unit. The geographic coordinates of the monitoring unit For the hard partition label of the monitoring unit, The average karst habitat gradient index of the area where the monitoring unit is located. The key habitat indicator vector includes soil thickness, rock exposure rate, slope, and surface zone water storage capacity; In the actual assessment, considering the continuity of the karst habitat gradient, the initial value of the ambiguity parameter was set to 2.0. The greedy iteration algorithm maximized the spatial dispersion. Habitat representativeness Each time, the unit that maximizes the gain of the objective function is selected, until the preset total number of monitoring units is reached. , will be the largest and The corresponding monitoring unit output is the optimal subset; Candidate number of clusters: 3~6; minimum Shebini index: optimal number of clusters 4*, including light, moderate, heavy, and very heavy karst habitats; smoothing threshold is 0.3 to correct for fragmented patches; Based on area and heterogeneity weights, 45 monitoring units were finally determined. Unit 1 has the following attributes: =1, The coordinates are (110.45°E, 24.78°N). It is 2. It is 0.56. [0.18m, 0.58, 22°, 0.70]. In this embodiment, the method for obtaining geological features by extracting prior hydrogeological features from two layers within the monitoring unit includes: Obtain the surface zone storage capacity index of the i-th monitoring unit. Soil thickness and karst development intensity index Introducing the enhancement coefficient of karst development intensity on the surface water-holding capacity. The surface layer thickness is weighted and corrected to obtain the effective surface layer storage thickness. ;in It was determined by regional hydrogeological surveys; Obtain the region's average annual precipitation The surface zone storage coefficient is calculated based on the ratio of the effective surface zone storage thickness to the average annual precipitation. ; Obtain the slope, pipe flow density, and fracture development degree. Calculate the pipe flow response coefficient based on the sum of the pipe flow density and the equivalent fracture density. ;in This is a conversion factor between the water conductivity of the crack and the pipe. This is the slope attenuation coefficient. Let be the slope of the i-th monitoring unit. Let be the pipe flow density of the i-th monitoring unit. Let represent the crack development degree of the i-th monitoring unit; Based on hydrogeological surveys, tracer experiments, or geophysical exploration results, a binary connectivity indicator function is constructed. The function is 1 if there is pipe connectivity between monitoring units, and 0 otherwise. For areas where detailed exploration has not been conducted, the karst development intensity index and pipe flow density are used for joint inference, expressed as: ; in Let i be the karst development intensity index of the i-th monitoring unit. Let j be the karst development intensity index of the j-th spatial unit. Let J be the pipe flow density of the j-th monitoring unit. The connectivity threshold for karst development intensity. The connectivity threshold for the flow density in the pipe; Calculate the hydrological connectivity probability between spatial units: ; in Let be the Euclidean distance between the i-th spatial unit and the j-th spatial unit. For spatial attenuation parameters, A binary connectivity indicator function based on hydrogeological surveys; The surface zone storage coefficient, pipeline flow response coefficient, and hydrological connectivity index are output as geological characteristics; In actual assessment, , , Calibrated by regional hydrogeological tests, It is 0.1. It is 0.7. It is 0.18m. It is 0.72. It is 0.135m. It is 1.80m. It is 0.074. It is 0.35. It is 0.08. It is 22°. It is 0.8. It is 1.2. It is 0.68; With adjacent Unit 2: It is 120m. It is 80. The value is 1, the hydrological connectivity probability is 0.32, and the final geological features are [0.074, 0.68, 0.32].

[0019] In this embodiment, the method for constructing a karst conduit network topology includes: The monitoring unit is used as a node. An edge is established when two nodes are connected by a pipe or crack. Based on the results of hydrogeological surveys and tracing experiments, a karst hydrological connectivity map is constructed. ;in For a set of nodes, Let be the set of edges. It is an adjacency matrix; Constructing a temporal feature matrix of surface ecology A hybrid architecture combining convolutional neural networks and long short-term memory networks is used to extract temporal features of the land surface, obtaining the features of the land surface ecological channels, expressed as: The temporal characteristic matrix of land surface ecology includes remote sensing vegetation indices, meteorological drivers, soil physicochemical properties, and remediation engineering parameters; among which... For convolutional neural networks, extract spatial texture features. For long short-term memory networks, capturing seasonal to interannual dynamics; A hybrid architecture combining graph attention networks and Transformers is employed. The initial node features include physical embedding features and the hidden state of the surface branch output. The graph attention layer updates the node features. ;in Let i be the groundwater node feature updated at the graph attention layer. For activation function, Let be the set of neighboring nodes of node i. The weight matrix is ​​a learnable matrix. Let be the attention weight of node j to node i. Let be the initial feature vector of node j; Hydrological connectivity probability is used as the weight of the edges in the karst pipeline network topology graph. Adjacency matrix Elements, through hydrological connectivity Prior weighting is used to obtain attention weights, expressed as: ; in It is the ReLU activation function. For attention weight vectors, Let W be the feature vector of node i after transformation by the weight matrix W. This is a vector concatenation symbol; Introducing a temporal Transformer to capture the hysteresis of subsurface responses and obtain subsurface hydrological channels. ;in Let be the pipeline flow transmission delay time from node j to node i. The underground hydrological channel characteristics of node i; In the actual assessment, the neighborhood node set consists of nodes connected to node i in the karst conduit network, with 45 monitoring units as nodes and edges having a hydrological connectivity probability > 0.25, constructing an undirected weighted graph. Adjacency matrix A As weight.

[0020] In this embodiment, the method for constructing a prediction model for the ecological restoration effect in the karst area includes: Obtain groundwater recharge modulus underground leakage rate and karst hydrodynamic regulation modulus Constructing physical information feature layers based on geological characteristics The expression is: ;in For transpose, The surface layer has a storage capacity coefficient. The pipe flow response coefficient; Based on the weak-form constraint coding of karst water balance and solute transport, the physical constraint layer of the karst conduit network topology is obtained, and its expression is: ; This includes soil water storage. For precipitation, Evaporation rate For pipeline flow rate, This refers to the fracture flow discharge rate. For the correction item of leakage caused by ecological restoration, This refers to the solute concentration. The hydrodynamic dispersion coefficient, The average pore flow velocity, For gradient operators, This represents the spatial gradient of solute concentration. For divergence operators, The rate of change of soil water storage over time. For adsorption, The weights for water balance constraints, The weights of the solute transport constraint term. Soil water storage capacity; A gating fusion mechanism is adopted to dynamically adjust the contribution weights of surface and underground channels based on the geological characteristics of each node, thereby obtaining fusion characteristics: ; ; in Let i represent the surface ecological corridor characteristics of node i. The underground hydrological channel characteristics of node i, Let i be the fusion feature of node i. Let i be the gating weight. The weight matrix is ​​a learnable matrix; The fused features are input into the multi-task output layer to simultaneously predict the multi-dimensional effects of ecological restoration, resulting in an ecological restoration prediction, expressed as: ; in For the ecological restoration prediction of node i, Let i be the vegetation cover restoration rate at node i. Let be the change in soil erosion modulus at node i. Let i be the comprehensive groundwater quality index. Let be the biodiversity recovery index for node i. The weight matrix of the output layer. This is the bias vector for the output layer; The total loss function is given based on the data-driven term, the physical consistency term, and the habitat partitioning regularization term. The expression is: in For data loss, For habitat regularization, For the actual observed value of node i, The total number of samples, The number of habitat gradient domains, Let be the parameter matrix of the a-th habitat-specific subnetwork. The basic parameter matrix is ​​shared globally. The square of the Hilbert-Schmidt norm. The weight of the habitat domain regularization term, The weight of the physical regularization term; Using Monte Carlo Dropout or deep ensemble methods to quantify prediction uncertainty, for K stochastic forward propagations or K independent sub-models, the prediction mean and variance are: in Let be the mean of node i. Let be the variance of node i. For the random number of forward propagations, For the k-th ecological restoration prediction, continue training until the variance is less than the variance threshold, then stop training. In actual assessment, , Optimized using grid search. It is 0.32. It is 0.18. The gating weight is 0.45; the gating weight is 0.55, integrating surface, underground, and physical features to output a 4-dimensional effect: It is 82%. -65t / (km) 2 ·a), Class I, It is 3.6; The total loss is 0.082 (converged), the variance threshold is 0.02, Monte Carlo Dropout (K=10): the prediction variance is 0.012 < 0.02, so training stops.

[0021] Secondly, a machine learning-based system for predicting the effects of ecological restoration in karst areas includes: Data Acquisition and Processing Module: Used to collect multi-source heterogeneous data of a preset karst area and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes remote sensing time-series data, geological data, restoration measure data, ecological monitoring data, and meteorological data; the geological data includes digital elevation, karst development intensity index, rock exposure rate, soil thickness, slope, aspect, surface layer thickness, conduit flow density, and fissure development degree; the restoration measure data includes restoration measures, implementation period, engineering density, and species configuration pattern; the ecological monitoring data includes vegetation cover, soil erosion modulus, biomass, groundwater quality indicators, and biodiversity index; the meteorological data includes precipitation, evaporation, temperature, and relative humidity; Index construction and partitioning module: used to construct a comprehensive gradient index of karst habitat through the multi-source heterogeneous data, and to obtain monitoring units by adaptive partitioning of karst habitat gradient based on the comprehensive gradient index of karst habitat; Geological Feature Extraction and Topology Map Construction Module: Used to extract geological features by performing two-layer hydrogeological prior features within the monitoring unit, and to construct a karst pipeline network topology map; Model building and output module: This module is used to build a prediction model for the ecological restoration effect of karst areas based on the karst pipeline network topology map and the geological features. It inputs the data to be predicted into the prediction model for the ecological restoration effect of karst areas and outputs the prediction results.

[0022] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for predicting the effects of ecological restoration in karst areas, characterized in that, Includes the following steps: Multi-source heterogeneous data of a pre-defined karst region were collected and preprocessed. The multi-source heterogeneous data included remote sensing time-series data, geological data, remediation measure data, ecological monitoring data, meteorological data, and soil moisture monitoring data. The geological data included digital elevation data, karst development intensity index, rock exposure rate, soil thickness, slope, aspect, surface layer thickness, conduit flow density, and fissure development degree. The remediation measure data included restoration measures, implementation period, engineering density, and species configuration patterns. The ecological monitoring data included vegetation cover, soil erosion modulus, biomass, groundwater quality indicators, and biodiversity index. The meteorological data included precipitation, evaporation, temperature, and relative humidity. A comprehensive gradient index for karst habitats is constructed using the multi-source heterogeneous data, and monitoring units are obtained by adaptively partitioning the karst habitat gradient based on the comprehensive gradient index. Geological features are obtained by performing double-layer hydrogeological prior feature extraction within the monitoring unit, and a karst pipeline network topology map is constructed. Based on the karst pipeline network topology and the geological features, a prediction model for the ecological restoration effect in karst areas is constructed. The data to be predicted is input into the prediction model for the ecological restoration effect in karst areas, and the prediction results are output.

2. The method for predicting the ecological restoration effect in karst areas based on machine learning according to claim 1, characterized in that, The method for constructing a comprehensive gradient index of karst habitats using the aforementioned multi-source heterogeneous data includes: The pre-defined karst region is initially divided into multiple areas, and the comprehensive gradient index of karst habitat in each area is calculated. ； in Let z be the karst development intensity index of the z-th region. Let z be the normalized soil thickness of the z-th region. Let z be the rock exposure rate of the z-th region. Let z be the slope of the z-th region. Let z be the normalized surface water storage capacity index of the z-th region. Let z be the comprehensive gradient index of karst habitat in region z. As a weight for karst development, For fatness weight, Weight of exposed rock, For slope weight, Weighting based on drought and flood levels.

3. The method for predicting the ecological restoration effect in karst areas based on machine learning according to claim 1, characterized in that, Methods for obtaining monitoring units by adaptive zoning of karst habitat gradient based on the comprehensive gradient index of karst habitat include: The comprehensive gradient index of karst habitat is normalized, a fuzziness parameter is set, and a candidate range for the number of clusters is defined; AFCM clustering is performed for each candidate s value. Membership matrix randomly generated based on candidate cluster numbers ,satisfy and Update cluster centers Update membership The membership matrix is ​​updated based on the membership degrees. Iteration stops when the difference between the membership matrices of two iterations is less than a difference threshold. The center of cluster v is the center of cluster v in the (t+1)th iteration. Let r be the membership degree of sample r to cluster v in the t-th iteration. Let r be the membership degree of sample r to cluster v at the (t+1)th iteration. The normalized karst development intensity index for sample r; Calculate the Shebiny clustering effectiveness index The optimal number of clusters is selected by choosing the s-value with the smallest Shebiny clustering effectiveness index. ;in The total number of samples, This represents the number of clusters in the current test. Let r be the membership degree of sample r to cluster v. For fuzzy weighted index, The karst habitat gradient index for sample r. For the v-th cluster center, The squared distance from the sample to the cluster center; For each sample point, the index of the maximum membership degree is used as the hard partition label. Labels are assigned to corresponding spatial units based on geographic coordinates, using an 8-neighborhood majority voting mechanism. If the current label has a lower percentage of votes within its neighborhood, the majority vote is applied. If the label is incorrect, it should be corrected to a neighborhood-dominant label; among which... For the smoothing threshold, Let r be the hard partition label of sample r. Let r be the membership degree of sample r to cluster v under the optimal number of clusters. Comprehensive area weight Weights related to heterogeneity Calculate the number of monitoring units in each gradient domain Constructing spatial dispersion Habitat representativeness Using a joint objective function, a greedy iterative strategy is employed to progressively select the optimal subset, and the selected optimal subset is output as the monitoring unit. ;in The number of monitoring units for cluster v. This represents the total number of monitoring units in the target area. These are the weight coefficients for cluster v. It is a rounding function. The minimum number of monitoring units for cluster v. A unique number for the monitoring unit. The geographic coordinates of the monitoring unit For the hard partition label of the monitoring unit, The average karst habitat gradient index of the area where the monitoring unit is located. The key habitat indicator vector includes soil thickness, rock exposure rate, slope, and surface zone water storage capacity.

4. The method for predicting the ecological restoration effect in karst areas based on machine learning according to claim 1, characterized in that, A method for obtaining geological features by extracting prior hydrogeological features from two layers within the monitoring unit includes: Obtain the surface zone storage capacity index of the i-th monitoring unit. Soil thickness and karst development intensity index Introducing the enhancement coefficient of karst development intensity on the surface water-holding capacity. The surface layer thickness is weighted and corrected to obtain the effective surface layer storage thickness. ;in It was determined by regional hydrogeological surveys; Obtain the region's average annual precipitation The surface zone storage coefficient is calculated based on the ratio of the effective surface zone storage thickness to the average annual precipitation. ; Obtain the slope, pipe flow density, and fracture development degree. Calculate the pipe flow response coefficient based on the sum of the pipe flow density and the equivalent fracture density. ;in This is a conversion factor between the water conductivity of the crack and the pipe. This is the slope attenuation coefficient. Let be the slope of the i-th monitoring unit. Let be the pipe flow density of the i-th monitoring unit. Let represent the crack development degree of the i-th monitoring unit; Based on hydrogeological surveys, tracer experiments, or geophysical exploration results, a binary connectivity indicator function is constructed. The function is 1 if there is pipe connectivity between monitoring units, and 0 otherwise. For areas where detailed exploration has not been conducted, the karst development intensity index and pipe flow density are used for joint inference, expressed as: ； in Let i be the karst development intensity index of the i-th monitoring unit. Let j be the karst development intensity index of the j-th spatial unit. Let J be the pipe flow density of the j-th monitoring unit. The connectivity threshold for karst development intensity. The connectivity threshold for the flow density in the pipe; Calculate the hydrological connectivity probability between spatial units: ； in Let be the Euclidean distance between the i-th spatial unit and the j-th spatial unit. For spatial attenuation parameters, A binary connectivity indicator function based on hydrogeological surveys; The surface zone regulation coefficient, pipeline flow response coefficient, and hydrological connectivity index are output as geological characteristics.

5. The method for predicting the ecological restoration effect in karst areas based on machine learning according to claim 1, characterized in that, Methods for constructing karst conduit network topology diagrams include: The monitoring unit is used as a node. An edge is established when two nodes are connected by a pipe or crack. Based on the results of hydrogeological surveys and tracing experiments, a karst hydrological connectivity map is constructed. ;in For a set of nodes, Let be the set of edges. It is an adjacency matrix; Constructing a temporal feature matrix of surface ecology A hybrid architecture combining convolutional neural networks and long short-term memory networks is used to extract temporal features of the land surface, obtaining the features of the land surface ecological channels, expressed as: The temporal characteristic matrix of land surface ecology includes remote sensing vegetation indices, meteorological drivers, soil physicochemical properties, and remediation engineering parameters; among which... For convolutional neural networks, extract spatial texture features. For long short-term memory networks, capturing seasonal to interannual dynamics; A hybrid architecture combining graph attention networks and Transformers is employed. The initial node features include physical embedding features and the hidden state of the surface branch output. The graph attention layer updates the node features. ;in Let i be the groundwater node feature updated at the graph attention layer. For activation function, Let be the set of neighboring nodes of node i. The weight matrix is ​​a learnable matrix. Let be the attention weight of node j to node i. Let be the initial feature vector of node j; Hydrological connectivity probability is used as the weight of the edges in the karst pipeline network topology graph. Adjacency matrix Elements, through hydrological connectivity Prior weighting is used to obtain attention weights, expressed as: ； in It is the ReLU activation function. For attention weight vectors, Let W be the feature vector of node i after transformation by the weight matrix W. This is a vector concatenation symbol; Introducing a temporal Transformer to capture the hysteresis of subsurface responses and obtain subsurface hydrological channels. ;in Let be the pipeline flow transmission delay time from node j to node i. The underground hydrological channel characteristics of node i.

6. The method for predicting the ecological restoration effect in karst areas based on machine learning according to claim 1, characterized in that, The method for constructing a prediction model for the ecological restoration effect in karst areas includes: Obtain groundwater recharge modulus underground leakage rate and karst hydrodynamic regulation modulus Constructing physical information feature layers based on geological characteristics The expression is: ;in For transpose, The surface layer has a storage capacity coefficient. The pipe flow response coefficient; Based on the weak-form constraint coding of karst water balance and solute transport, the physical constraint layer of the karst conduit network topology is obtained, and its expression is: ； This includes soil water storage. For precipitation, Evaporation rate For pipeline flow rate, This refers to the fracture flow discharge rate. For the correction item of leakage caused by ecological restoration, This refers to the solute concentration. The hydrodynamic dispersion coefficient, The average pore flow velocity, For gradient operators, This represents the spatial gradient of solute concentration. For divergence operators, The rate of change of soil water storage over time. For adsorption, The weights for water balance constraints, The weights of the solute transport constraint term. Soil water storage capacity; A gating fusion mechanism is adopted to dynamically adjust the contribution weights of surface and underground channels based on the geological characteristics of each node, thereby obtaining fusion characteristics: ； ； in Let i represent the surface ecological corridor characteristics of node i. The underground hydrological channel characteristics of node i, Let i be the fusion feature of node i. Let i be the gating weight. The weight matrix is ​​a learnable matrix; The fused features are input into the multi-task output layer to simultaneously predict the multi-dimensional effects of ecological restoration, resulting in an ecological restoration prediction, expressed as: ； in For the ecological restoration prediction of node i, Let i be the vegetation cover restoration rate at node i. Let be the change in soil erosion modulus at node i. Let i be the comprehensive groundwater quality index. Let be the biodiversity recovery index for node i. The weight matrix of the output layer. This is the bias vector for the output layer; The total loss function is given based on the data-driven term, the physical consistency term, and the habitat partitioning regularization term. The expression is: in For data loss, For habitat regularization, For the actual observed value of node i, The total number of samples, The number of habitat gradient domains, Let be the parameter matrix of the a-th habitat-specific subnetwork. The basic parameter matrix is ​​shared globally. The square of the Hilbert-Schmidt norm. The weight of the habitat regularization term, The weight of the physical regularization term; Using Monte Carlo Dropout or deep ensemble methods to quantify prediction uncertainty, for K stochastic forward propagations or K independent sub-models, the prediction mean and variance are: in Let be the mean of node i. Let be the variance of node i. For the random number of forward propagations, For the k-th ecological restoration prediction, training continues until the variance is less than the variance threshold, at which point training stops.

7. A machine learning-based system for predicting the effects of ecological restoration in karst areas, used to perform the method described in any one of claims 1-6, characterized in that, include: Data Acquisition and Processing Module: Used to collect multi-source heterogeneous data of a preset karst area and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes remote sensing time-series data, geological data, restoration measure data, ecological monitoring data, and meteorological data; the geological data includes digital elevation, karst development intensity index, rock exposure rate, soil thickness, slope, aspect, surface layer thickness, conduit flow density, and fissure development degree; the restoration measure data includes restoration measures, implementation period, engineering density, and species configuration pattern; the ecological monitoring data includes vegetation cover, soil erosion modulus, biomass, groundwater quality indicators, and biodiversity index; the meteorological data includes precipitation, evaporation, temperature, and relative humidity; Index construction and partitioning module: used to construct a comprehensive gradient index of karst habitat through the multi-source heterogeneous data, and to obtain monitoring units by adaptive partitioning of karst habitat gradient based on the comprehensive gradient index of karst habitat; Geological Feature Extraction and Topology Map Construction Module: Used to extract geological features by performing two-layer hydrogeological prior features within the monitoring unit, and to construct a karst pipeline network topology map; Model building and output module: This module is used to build a prediction model for the ecological restoration effect of karst areas based on the karst pipeline network topology map and the geological features. It inputs the data to be predicted into the prediction model for the ecological restoration effect of karst areas and outputs the prediction results.

Images