Method for identifying hydrological effects of different vegetation restoration modes under dynamic model and parameters

Abstract

The application discloses a kind of dynamic model and parameter under different vegetation restoration mode hydrological effect identification method, for water resources scheduling control, correspondingly: obtain the multi-source basic data of research basin and establish spatial grid time series dataset;Dynamic vegetation parameters are generated based on remote sensing vegetation data;Trend separation processing is carried out by demarcating unchangeable and changeable area, and independent driving force of different restoration modes is extracted;Based on mapping relationship model, convert dynamic vegetation parameters into time-varying parameters of hydrological model under corresponding vegetation state;Reconstruct the input file of different vegetation change scene and carry out control variable simulation;Extract independent influence component and nonlinear synergistic effect component, and quantify output quantitative contribution rate.The application alleviates the problem of parameter static and driving force difficult to decouple, realizes the attribution identification of hydrological effect under different restoration modes, and improves the objectivity and accuracy of water resources adaptability evaluation under changing environment.

Description

Technical Field

[0001] This invention belongs to the field of hydrological modeling technology, and in particular, it is a method for identifying the hydrological effects of different vegetation restoration patterns under dynamic models and parameters. Background Technology

[0002] As an underlying surface element, vegetation not only exhibits dynamic characteristics driven by environmental factors but also exerts a reciprocal effect on these factors. Studying and determining the impact of different types of vegetation changes on water cycle processes helps to understand the changing characteristics of water resource systems under changing environments and ensure regional water security.

[0003] Regarding the hydrological effects of vegetation change, current research mainly focuses on the overall impact of vegetation greening on hydrological elements such as runoff, evapotranspiration, and soil moisture. For example, water balance models are used to analyze the impact of vegetation change on runoff in a given watershed, or to quantitatively assess the impact of vegetation greening on evapotranspiration and its components. Some studies have combined remote sensing leaf area index with hydrological models to assess the impact of interannual and intra-annual vegetation variations on runoff.

[0004] However, in complex and ever-changing eco-hydrological environments, existing hydrological effect assessment methods still face the problems of static parameters and difficulty in decoupling driving forces when dealing with the combined effects of vegetation and climate, resulting in errors in the assessment results. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying the hydrological effects of different vegetation restoration patterns under dynamic models and parameters, in view of the aforementioned technical problems.

[0006] Technical solution: A method for identifying the hydrological effects of different vegetation restoration patterns under dynamic models and parameters, including:

[0007] Acquire multi-source basic data for the research watershed and establish a spatial grid time series dataset;

[0008] Land use evolution characteristics are analyzed based on remote sensing vegetation data in spatial grid time series datasets to generate dynamic vegetation parameters that reflect the trajectory of vegetation change.

[0009] Based on dynamic vegetation parameters, we delineate areas with unchanged and changing land use types, perform trend separation processing, and extract the independent driving forces of different restoration modes that cause vegetation changes.

[0010] The mapping relationship model is invoked to convert dynamic vegetation parameters into time-varying parameters of the hydrological model under the corresponding vegetation conditions;

[0011] The vegetation input files for different vegetation change scenarios are reconstructed using independent driving forces. The hydrological model is driven by time-varying parameters to simulate controlled variables and obtain the simulated hydrological process corresponding to each vegetation change scenario.

[0012] Based on the extraction of independent influence components and nonlinear synergistic components from simulated hydrological processes, the quantitative contribution rate of different vegetation restoration models to hydrological effects is quantified and output.

[0013] Beneficial effects: This invention enables the attribution identification of hydrological effects under different recovery modes, improving the objectivity and accuracy of water resource adaptability assessment under changing environments. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0015] Figure 1 This is a flowchart illustrating a method for identifying the hydrological effects of different vegetation restoration patterns under a dynamic model and parameters, as described in this application.

[0016] Figure 2 This is a schematic diagram of a process in this application for delineating areas with unchanged and changing land use types based on dynamic vegetation parameters, performing trend separation processing, and extracting independent drivers of different restoration modes that cause vegetation changes.

[0017] Figure 3 This is a schematic diagram of a process for establishing a regression mapping relationship between dynamic vegetation parameters and corresponding spatiotemporal meteorological data, as described in this application.

[0018] Figure 4 This is a schematic diagram illustrating the process of obtaining a mapping relationship model in this application through pre-training via an offline model construction step. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a predetermined order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] In view of the above issues, the applicant conducted a search and analysis and found that:

[0022] Vegetation alters the spatiotemporal distribution of hydrological elements by influencing processes such as canopy interception, vegetation transpiration, and soil infiltration. Based on different restoration mechanisms, the hydrological effects of vegetation change can be categorized into two types: first, the hydrological effects resulting from natural vegetation restoration caused by climate change; and second, the hydrological effects resulting from human intervention in vegetation restoration.

[0023] On the one hand, traditional methods are difficult to separate and independently quantify the hydrological responses of natural vegetation succession and human intervention in the context of drastic changes in the underlying surface, and often ignore the interaction between the two on the spatiotemporal scales.

[0024] On the other hand, when hydrological models are applied to long-term simulations, the underlying physical parameters lack dynamic adaptive capabilities, and the model's accuracy and adaptability in depicting inconsistent evolution environments are still insufficient.

[0025] To solve these problems, combined with Figures 1 to 4 This application is elaborated through the following embodiments.

[0026] Accordingly, some nouns or terms appearing in the description of the embodiments of this application shall be interpreted as follows:

[0027] The predetermined evolution trend quantity can be an interannual linear evolution trend quantity in some embodiments; and a linear evolution trend quantity in other embodiments.

[0028] This application designs a method for identifying the hydrological effects of different vegetation restoration patterns under dynamic models and parameters, which is used for water resource scheduling and control. Specifically, it includes:

[0029] Step 101: Obtain multi-source basic data of the study watershed through ground observation stations and satellite remote sensing platforms, and establish a spatial grid time series dataset;

[0030] Furthermore, obtaining multi-source basic data for the research basin specifically refers to loading the hydrological, climatic, and current surface-related observational characteristics of the target research basin over consecutive historical years through ground meteorological observation stations, satellite remote sensing inversion platforms, or reanalysis datasets.

[0031] In this embodiment, by performing standard quality control procedures such as grid trimming, spatiotemporal resolution alignment, and outlier removal on the collected raw heterogeneous data, the dispersed data sources are uniformly mapped onto standard spatial grid nodes, thereby providing a benchmark for the hydrological simulation and dynamic change analysis of the underlying surface in the following text.

[0032] Among them, multi-source basic data includes at least pre-collected regional meteorological observation data.

[0033] Remote sensing vegetation data should include at least leaf area index, vegetation cover, and land use cover type.

[0034] Spatial grid time series datasets can be represented as a raster matrix structure containing three-dimensional coordinates of longitude, latitude, and time step.

[0035] In some scenarios, mapping relationship models, nonlinear mapping relationship networks, nonlinear time-varying mapping networks, and nonlinear time-varying mapping relationships are equivalent.

[0036] Step 102: Based on the remote sensing vegetation data in the spatial grid time series dataset, perform land use evolution characteristic analysis to generate dynamic vegetation parameters that reflect the vegetation change trajectory.

[0037] Accordingly, a land use evolution characteristic analysis is conducted, namely: statistically analyzing the area proportion of each land use type in each grid unit, and tracking the mutual conversion relationship between cultivated land, forest land, grassland and unused land in different years, so as to clarify whether the vegetation type in each grid unit has undergone substantial changes and the corresponding direction of change.

[0038] During the analysis, dynamic vegetation parameters reflecting the trajectory of vegetation change are generated. Specifically, the leaf area index and vegetation cover information extracted monthly are encapsulated with land use cover data year by year to generate monthly and yearly continuous time-series vector files that reflect the dynamic evolution characteristics of the underlying surface within a predetermined grid.

[0039] Dynamic vegetation parameters are used to drive distributed hydrological models or their improved physical models. They can dynamically capture the interannual and intra-annual phenological greening or degradation characteristics of the watershed underlying surface caused by projects such as returning farmland to forest and grassland or climate fluctuations, thereby avoiding the non-consistent systematic errors caused by the use of static constant parameters in traditional hydrological models.

[0040] Step 103: Based on the dynamic vegetation parameters, delineate areas with unchanged land use types and areas with changes, perform trend separation processing, and extract the independent driving forces of different restoration modes that cause vegetation changes.

[0041] To overcome the difficulty in separating natural and human-driven factors in existing technologies, this invention employs a spatial partitioning trend separation method in this step.

[0042] Among them, the land use type is divided into areas with unchanged land use type and areas with changes. Specifically, based on the preset land cover change detection threshold, two types of predetermined processing units are marked on the spatial grid of the study watershed. The first type is the land use type unchanged area, which has not changed throughout the entire time-series monitoring period. The second type is the land use type changed area, which has changed directionally due to human interventions such as afforestation, conversion of farmland to forest or grassland, or land reclamation.

[0043] Trend separation processing, namely: from the perspective of spatial heterogeneity, using areas with unchanged land use types as the background baseline driven by pure natural climate, extrapolating the response mechanism to areas with changes in land use types, and then using residual analysis to extract the pure human activity disturbance increment after deducting the natural climate background.

[0044] Based on this, the contributions of climate change drivers and human interventions are calculated and output, and each contribution is treated as an independent driving force.

[0045] At the whole watershed scale, quantitative decoupling of the driving factors of two vegetation restoration modes—natural restoration succession and anthropogenic restoration projects—was achieved.

[0046] Step 104: Based on the pre-built mapping relationship model, the dynamic vegetation parameters are converted into time-varying parameters of the preset hydrological model under the corresponding vegetation state.

[0047] Traditional hydrological simulations often use a fixed set of hydrological parameters that remain unchanged over time. However, this can lead to parameter incompatibility in ecological restoration zones where the underlying surface has changed. Therefore, this embodiment introduces a pre-constructed mapping model to solve the problem of static parameters.

[0048] Accordingly, dynamic vegetation parameters are used as feature inputs and substituted into the pre-trained nonlinear mapping network to transform and output the time-varying parameters of the preset hydrological model under the corresponding vegetation conditions.

[0049] Time-varying parameters are represented as one-dimensional feature vectors that are dynamically updated with time stamps. These vectors specifically include structural parameters used in the hydrological model to control surface runoff, soil water transport, groundwater generation, and evapotranspiration processes. Examples include infiltration parameters reflecting soil infiltration capacity and the maximum baseflow coefficient of the subsoil. The evolution of these model parameters reflects the transformation of watershed soil hydrodynamic mechanisms by vegetation construction.

[0050] Step 105: Reconstruct vegetation input files for different vegetation change scenarios using independent driving forces and dynamic vegetation parameters, and use time-varying parameters to drive the hydrological model to simulate controlled variables, thereby obtaining the simulated hydrological process corresponding to each vegetation change scenario.

[0051] In this step, a controlled variable mechanism is adopted, that is, the meteorological input conditions of the target watershed, such as precipitation, temperature, wind speed, etc., are kept consistent throughout the entire simulation time series, and the scenario assumptions are implemented by performing a predetermined combination of the underlying surface input parameters.

[0052] Accordingly, vegetation input files for different vegetation change scenarios are reconstructed. That is, by using the independent driving forces obtained from the stripping, four types of predetermined vegetation change scenarios are reconstructed to test the hydrological effects of different vegetation restoration models. These include the baseline scenario, the real dynamic change scenario, the vegetation anthropogenic restoration scenario, and the vegetation natural succession scenario.

[0053] Furthermore, the predetermined time series input files generated under each scenario, along with the derived time-varying parameters, are used as driving terms and substituted into the preset hydrological model for large-scale grid parallel evolution simulation.

[0054] This can lay the foundation for decoupling the water balance effect of vegetation under the dominant and synergistic effects of nature and human activities.

[0055] In the above text, the simulated hydrological process refers to the continuous spatiotemporal dataset that the model ultimately calculates and outputs, which includes the evapotranspiration time series field, surface runoff field, groundwater runoff field, and soil moisture content.

[0056] Step 106: Extract independent influence components and nonlinear synergistic effect components based on simulated hydrological processes, quantify and output the quantitative contribution rate of different vegetation restoration models to hydrological effects.

[0057] Since existing methods typically ignore the nonlinear interaction effect between natural recovery and human intervention, this invention, in this step,

[0058] Accordingly, independent influence components and nonlinear synergistic effect components are extracted, that is, a multi-scenario difference attribution identification framework is established. The simulated hydrological process corresponding to the real dynamic change scenario is subtracted from the simulated hydrological process corresponding to the baseline scenario to extract the overall influence component of vegetation change on the evolution of hydrological elements.

[0059] Accordingly, the simulation outputs of the vegetation restoration scenario and the vegetation succession scenario are subtracted from the baseline scenario to obtain the independent impact components of human intervention and the independent impact components of climate change.

[0060] Based on this, the sum of the two independent individual components, human and natural, is removed from the overall impact components. The remaining residual deviation is defined as the nonlinear synergistic effect component that characterizes the nonlinear coherence of the two components.

[0061] Furthermore, the proportion of each component in the overall impact component is calculated, and the quantitative contribution rate of different vegetation restoration models to hydrological effects is quantified and output; this can be used to provide quantitative decision support for watershed water security management and ecological restoration planning under changing environments.

[0062] In some embodiments, for dynamic vegetation parameters, regions with unchanged land use types and regions with changing land use types are delineated for trend separation processing, and independent driving forces causing different restoration modes of vegetation change are extracted, including:

[0063] Step 201: Within the area where land use type remains unchanged, establish a regression mapping relationship between dynamic vegetation parameters and the corresponding spatiotemporal meteorological data in the spatial grid time series dataset to obtain the climate impact baseline.

[0064] Correspondingly, areas with unchanged land use types represent spatial grid cell combinations where no land cover type flipping has occurred during the study period due to human activities such as land reclamation, reforestation, or urban expansion.

[0065] In this region, the natural growth and decline of vegetation are mainly controlled by the interannual and intra-annual fluctuations of meteorological factors such as precipitation and temperature.

[0066] In this embodiment, a regression mapping relationship is established, specifically by using a pre-built support vector machine regression model for baseline fitting. Monthly precipitation and temperature sequences are extracted as input independent variables for the support vector machine regression model within a land use type-invariant region, and the corresponding dynamic vegetation parameters are used as dependent variables. Supervised training is performed to fit a nonlinear response function of climate conditions to vegetation growth, and this response function is used as the climate impact baseline.

[0067] This baseline characterizes the natural evolution of vegetation parameters under ideal conditions free from human disturbance, as they fluctuate with weather patterns.

[0068] Alternatively, the regression mapping relationship can be replaced by statistical models such as the copula (connection function) multidimensional joint distribution model;

[0069] Alternatively, a multiple linear regression model or a random forest regression model can be used to iteratively obtain climate response parameters using the mean square error minimization criterion.

[0070] Step 202: Extrapolate the climate impact baseline to the land use type change area and calculate the theoretical vegetation evolution trend under the dominance of climate change.

[0071] Furthermore, the land use type change area includes spatial grids directly affected by human-led ecological engineering interventions.

[0072] The application of climate impact baseline extrapolation is as follows: the measured meteorological data of the time and space corresponding to the land use type change area are substituted into the support vector machine regression model that has been trained and converged in the land use type unchanged area;

[0073] The result of the model's forward propagation is the theoretical sequence of dynamic vegetation parameters assuming that the region is not disturbed by human activities and is only driven by meteorological factors.

[0074] Based on this, the theoretical sequence is calculated by sliding window difference or linear trend extraction along the time dimension to obtain the theoretical vegetation evolution trend; it represents the background baseline rate of vegetation development towards natural greening or degradation in the target watershed without artificial afforestation or farmland conversion intervention.

[0075] Step 203: Combine the actual vegetation evolution trend and the theoretical vegetation evolution trend in the land use type change area to perform residual analysis, separate the contribution of climate change and human activity intervention, and use each extracted contribution as an independent driving force; use it to separate the single physical driving factor from the mixed increment observed by remote sensing satellite.

[0076] Accordingly, the least squares method can be used to calculate the predetermined evolution trends of vegetation leaf area and vegetation cover under actual evolution, under the influence of climate change, and under human intervention.

[0077] This allows for the quantification of the ratio of independent driving forces that distinguish different driving sources.

[0078] In other words, the least squares method is used to perform time-series fitting on the dynamic vegetation parameter sequence of actual observations in the land use type change area, and the predetermined evolution trend quantity under the actual evolution is calculated.

[0079] Similarly, by using the least squares method to fit the theoretical sequence over time, the predetermined evolution trend under the influence of climate change can be calculated.

[0080] The residual sequence is obtained by subtracting the values ​​of the actual observed dynamic vegetation parameter sequence from the corresponding theoretical sequence at the same time. The least squares method is used to perform time series fitting on the residual sequence to calculate the predetermined evolution trend under human intervention.

[0081] The formula for calculating the leaf area index parameter is as follows:

[0082] r _LAI_C_y =ΔLAI _C_y / ΔLAI _obs_y ;

[0083] r _LAI_H_y =ΔLAI _H_y / ΔLAI _obs_y ;

[0084] Above, r _LAI_C_y ΔLAI represents the ratio of the driving force of climate change on the change in leaf area index in year y. _C_y ΔLAI represents the predicted evolution trend of the leaf area index (LAI) under the influence of climate change in year y, calculated using the least squares method. _obs_y r represents the predicted evolutionary trend of the leaf area index in year y, calculated using the least squares method based on the actual evolution. _LAI_H_y ΔLAI represents the ratio of human activity to the change in leaf area index in year y. _H_y This represents the predicted evolution trend of leaf area index under human activity intervention in year y, calculated using the least squares method.

[0085] Similarly, the calculation formula for the vegetation cover parameter is as follows:

[0086] r _FVC_C_y =ΔFVC _C_y / ΔFVC _obs_y ;

[0087] r _FVC_H_y =ΔFVC _H_y / ΔFVC _obs_y ;

[0088] In the formula, r _FVC_C_y ΔFVC represents the ratio of climate change to vegetation cover change in year y. _C_y ΔFVC represents the predicted evolution trend of vegetation cover (FVC) under the influence of climate change in year y, calculated using the least squares method. _obs_y r represents the predicted evolution trend of vegetation cover in year y, calculated using the least squares method. _FVC_H_y ΔFVC represents the ratio of human activity to changes in vegetation cover in year y. _H_yThis represents the predetermined evolution trend of vegetation cover under human intervention in year y, calculated using the least squares method.

[0089] Based on this, the driving ratios corresponding to the calculated and extracted parameters are uniformly merged as the contributions of human activity intervention and climate change driving forces for use in the simulation of control variables, i.e., independent driving forces.

[0090] Furthermore, the analysis of residual sequences can incorporate time lag smoothing operations, which weight the current residual with the residual of the previous time step to filter out noise fluctuations introduced by short-term extreme weather events, so that the extracted contribution of human intervention can reflect the long-term cumulative effect of sustainable ecological restoration projects.

[0091] In other embodiments, an optional implementation method is further described, which converts dynamic vegetation parameters into time-varying parameters of a preset hydrological model under the corresponding vegetation state based on a pre-built mapping relationship model.

[0092] Among them, the pre-built mapping relationship model is a pre-selected machine learning model;

[0093] Step 301: Use a pre-selected machine learning model to establish a nonlinear time-varying mapping relationship between dynamic vegetation parameters and multiple key parameters in the hydrological model, so as to realize the dynamic estimation of these time-varying parameters.

[0094] Alternatively, the above model can also be a statistical relationship model.

[0095] Optionally, a self-driven parameter estimation network can be used to complete the mapping transformation.

[0096] The pre-selected machine learning models are specifically chosen from any one of the following: Long Short-Term Memory Network Model, Random Forest Model, Multidimensional Joint Distribution Model, and Radial Basis Network Model.

[0097] Accordingly, a nonlinear time-varying mapping relationship is established, namely: leaf area index and vegetation cover sequence with time attributes are used as input layer feature data, and multi-dimensional feature mapping is performed through the hidden layer nodes of the machine learning model to output the dynamic parameter group of the bottom layer of the hydrological model used to control the physical process of water exchange.

[0098] Several key parameters, specifically including infiltration curve parameters, maximum baseflow parameters in the bottom layer, and soil thickness distribution parameters in each layer.

[0099] Taking the Long Short-Term Memory Network model as an example, the long-term lag response pattern of dynamic vegetation parameters in the time dimension is extracted by utilizing its internal gating unit structure, and the set of hydrological model parameters that conform to the physical characteristics of the underlying surface at the current time node is calculated and output.

[0100] In some scenarios, when the input feature data has high dimensionality or contains local data gaps, a random forest model can be used as an alternative. Multiple decision trees are constructed and computed in parallel, and their mean is calculated. This reduces the sensitivity of a single mapping channel to abnormal vegetation observation data, thereby improving the robustness of the time-varying parameter estimation process.

[0101] Taking the Long Short-Term Memory (LSTM) network model as an example, its internal structure includes an input layer, a hidden gating layer, and an output layer.

[0102] The number of nodes in the input layer matches the dimension of the dynamic vegetation parameter features input, and is used to receive the temporal features of leaf area index and vegetation cover of the current computational grid cell within a predetermined time step.

[0103] The hidden gating layer performs state memory and long-short-term dependency update on the temporal features of the input through the internally configured forget gate, input gate and output gate structure. The number of neurons in the hidden gating layer can be set by the technician.

[0104] The output layer consists of multiple linear neurons, with the number of nodes matching the number of key parameters involved in dynamic adjustment within the hydrological model. These parameters are used to map and transform the infiltration parameters, maximum daily baseflow, unsaturated zone runoff coefficient, soil moisture retention coefficient, intermediate soil layer thickness, and bottom soil layer thickness in the output hydrological model.

[0105] Accordingly, the feature vector received by the input layer is formed by concatenating the leaf area index sequence of the current month and the vegetation coverage sequence of the previous N months, forming a temporal input window with a dimension of 2×(N+1), where N is the time backtracking step size.

[0106] Step 302 introduces a comprehensive performance evaluation index composed of the Nash efficiency coefficient, the Kling-Gupta coefficient, and the relative error. Based on the deviation between the runoff simulation results output by the hydrological model and the measured runoff data in the spatial grid time series dataset, the overall accuracy of the runoff simulation results output by the hydrological model is evaluated and used as a basis for model selection.

[0107] Among them, the Nash efficiency coefficient is used to measure the degree of variance agreement between simulated runoff data and measured runoff data in the high flow peak range.

[0108] The Kling-Gupta coefficient is used to jointly test the consistency of time series based on three components: linear correlation, spatial variation ratio, and mean deviation ratio.

[0109] Relative error is a systematic deviation in water conservation used to calculate the total runoff volume over a long series.

[0110] By simultaneously calculating evaluation factors of three independent dimensions, a constraint matrix covering extreme value matching degree, process consistency and total volume deviation is formed, thereby determining whether the time-varying parameters generated by the selected machine learning model have practical applicability.

[0111] Optionally, the Nash efficiency coefficient, Kling-Gupta coefficient, and relative error are subjected to monotonic normalization mapping to prevent overflow, so as to unify the heterogeneous dimensions of each evaluation index.

[0112] Because the numerical boundaries and extreme convergence directions of the three evaluation indicators are different, directly performing algebraic weighting operations will cause the evaluation model to be biased towards a single variable whose value range is not limited.

[0113] The monotonic normalization mapping process to prevent overflow refers to using a nonlinear function to perform mathematical projection on the original measured index, forcing the statistical values ​​of all heterogeneous dimensions to converge to the standard constraint interval of 0-1.

[0114] For the Nash efficiency coefficient and the Kling-Gupta coefficient, the maximum value of the original calculation result and the boundary of the 0 value interval is extracted as the mapped data. A truncation operation is performed to filter out the negative interval values ​​generated under extremely poor fitting conditions, so as to prevent invalid features from diluting or distorting the overall evaluation distribution.

[0115] Optionally, for the relative error, a negative exponential mapping function containing the absolute value is used to convert it into a monotonically decaying scoring benchmark. Then, a weighted sum is made with the Nash efficiency coefficient and the Kling-Gupta coefficient, which have been processed by monotonically normalized mapping, to obtain the final comprehensive performance evaluation index.

[0116] The original result of the relative error has positive and negative algebraic signs representing overestimation or underestimation. If the sign difference is ignored and directly substituted into the exponential decay module, the negative bias term will cause the output value of the decay function to increase abnormally exponentially, violating the upper bound of the evaluation system. Further, a negative exponential mapping step, including absolute value extraction, is performed to achieve dimensionality reduction. Its mapping calculation expression is:

[0117] PBIAS _norm =exp(-|PBIAS| / P0);

[0118] Above, PBIAS _norm represents the normalized relative error value after mapping, exp represents the exponential function with the natural constant as the base, PBIAS represents the initial relative error value calculated based on the simulated hydrological process, and P0 represents the set reference deviation constant.

[0119] In some scenarios, the reference deviation constant can be set to 20%.

[0120] After obtaining the normalized basic variables for all dimensions, aggregation is performed using a weighted allocation mechanism. The specific weighted summation rule is executed according to the following formula:

[0121] α=w1×NSE _norm +w2×KGE _norm +w3×PBIAS _norm ;

[0122] In the formula, α represents the comprehensive performance evaluation index obtained, w1 represents the weight constant assigned to the characteristic dimension of the Nash efficiency coefficient, and NSE _norm KGE represents the Nash efficiency coefficient after truncated mapping, w2 represents the weight constants assigned to the feature dimensions of the Kling-Gupta coefficients, and KGE represents the efficiency coefficient. _norm w represents the Kling-Gupta coefficient after truncated mapping, and w3 represents the weight constant assigned to the relative error feature dimension.

[0123] The sum of the three weighting constants equals 1. When the comprehensive performance evaluation index of the calculated output approaches 1, it indicates that the physical accuracy of the current parameter mapping model has reached its optimum.

[0124] In other alternative implementations, the numerical proportions of the weighting constants can be dynamically reallocated according to predetermined application requirements. For example, for flood forecasting needs, the weighting constants allocated to the Nash efficiency coefficient characteristic dimension can be increased accordingly; for long-term water resource inventory accounting needs, the weighting constants allocated to the relative error characteristic dimension can be increased, thereby adapting the evaluation system to the predetermined engineering objectives.

[0125] On the one hand, this paper further explains how to reconstruct vegetation input files under different vegetation change scenarios using independent driving forces, and how to use time-varying parameters to drive a hydrological model for controlled variable simulation. Specifically, this includes:

[0126] Step 401: Keep the meteorological input conditions of the watershed studied in the spatial grid time series dataset unchanged, and set up four types of simulation scenarios, including the baseline scenario, the real dynamic change scenario, the vegetation human restoration scenario, and the vegetation natural succession scenario. Simulate the four types of simulation scenarios with controlled variables to obtain the simulated hydrological process corresponding to each vegetation change scenario.

[0127] Optionally, it is necessary to remove background interference caused by meteorological fluctuations.

[0128] In this paper, the meteorological input conditions are kept constant, that is, throughout the entire time series of the hydrological model, all scenarios use the same set of quality-controlled precipitation, temperature and wind speed sequences.

[0129] Based on this, the externally input dynamic vegetation parameter file is replaced to construct a model evolution matrix driven by a single variable.

[0130] The hydrological model is driven by time-varying parameters to perform calculations under four independently configured input matrices, and outputs four sets of simulated runoff and evapotranspiration data with equal time and spatial resolution.

[0131] Among them, the baseline scenario is used to output the hydrological reference state that is not disturbed by ecological evolution;

[0132] Real-world, dynamic scenarios are used to reflect the overall mapping of hydrological processes under current comprehensive real-world conditions;

[0133] The scenarios of human-induced vegetation restoration and natural vegetation succession serve as decoupling scenarios, specifically designed to isolate the hydrological response corresponding to a single driving force extracted from the test.

[0134] Step 402: For the baseline scenario, extract the monthly vegetation parameter sequence of the year before the land use change occurred from the dynamic vegetation parameters as the corresponding vegetation input file.

[0135] For real-world dynamic change scenarios, the annual dynamic vegetation parameter sequence actually observed after land use change occurs is used as the corresponding vegetation input file.

[0136] Correspondingly, extracting the monthly vegetation parameter sequence for the year preceding the occurrence of land use change refers to using the time node of the launch of large-scale ecological engineering in the target watershed as a time segmentation barrier, and calculating the monthly average value of remote sensing vegetation data from several years before this barrier node.

[0137] For example, using the year when the policy of returning farmland to forest and grassland was first implemented as the dividing point, the monthly leaf area index and the mean vegetation cover of previous historical years were extracted. Combined with the spatial grid of early land use types, a static or mean-stationary parameter matrix was generated and provided as input conditions to the hydrological model to calculate and generate a theoretical water cycle process without vegetation greening or degradation disturbance.

[0138] For real-world dynamic scenarios, directly extract the measured remote sensing time series after segmentation nodes;

[0139] It retains the absolute dynamic change characteristics of the underlying surface on a monthly and yearly basis, and uses it as the input of actual observation status into the hydrological model to verify the overall accuracy of the model in reproducing the evolution of real hydrological elements, and as the overall control term for subsequent difference calculations.

[0140] Step 403: For the vegetation restoration scenario and the vegetation succession scenario, the corresponding independent driving force is converted into a contribution rate ratio and applied to the actual observed vegetation parameter increment, and then superimposed on the vegetation parameter sequence of the baseline state, so as to reconstruct the predetermined time series input under the independent driving force of human intervention and the independent driving force of natural succession, respectively, as the corresponding vegetation input file.

[0141] In this step, the predetermined timing input file is reconstructed, that is, the original mixed greening increment is physically disassembled and reconstructed according to the driving source.

[0142] Accordingly, for the scenario of human restoration of vegetation, the incremental part of the actual observed value compared with the baseline mean is extracted by using the contribution ratio of human activity intervention. The two are multiplied to form the pure parameter increment caused by human intervention, which is then superimposed on the baseline sequence before the land was converted back to forest.

[0143] The reconstruction calculation of the leaf area index parameter can be performed according to the following formula:

[0144] LAI _S2 (m,y)=LAI _0 (m)+r _LAI_H_y ×[LAI _obs (m,y)-LAI _0 (m)];

[0145] Among them, LAI _S2 (m,y) represents the leaf area index (LAI) in the m-th month of year y under the reconstructed vegetation restoration scenario. _0 (m) represents the baseline sequence of monthly average leaf area index values ​​for month m under the baseline scenario, r _LAI_H_y To express the ratio of human activity driving changes in leaf area index (LAI) in year y, converted into a contribution rate ratio. _obs (m,y) represents the leaf area index under the actual evolution state in the m-th month of the y-th year, obtained through actual observation.

[0146] The same principle applies to the reconstruction calculation of vegetation cover parameters under human restoration scenarios;

[0147] For vegetation succession scenarios, the contribution ratio of independent climate drivers needs to be multiplied by the actual increment and superimposed on the baseline sequence to output the vegetation time series regulated by pure climate natural state.

[0148] The reconstructed calculation for leaf area index is shown in the following formula:

[0149] LAI _S3 (m,y)=LAI _0 (m)+r _LAI_C_y ×[LAI _obs (m,y)-LAI _0 (m)];

[0150] Among them, LAI _S3 (m,y) represents the leaf area index in the m-th month of year y under the reconstructed vegetation succession scenario, r _LAI_C_yThe ratio of the contribution rate to the change in leaf area index in year y is the ratio of the contribution rate to the contribution rate.

[0151] The same principle applies to the reconstruction of vegetation cover under natural succession scenarios.

[0152] In some cases, if the remote sensing observation errors within the study basin meet the pre-configured convergence boundary conditions, the ratio mapping strategy can be abandoned, and the absolute difference sequence output by the residual analysis can be used directly as an alternative input source.

[0153] The specific operation involves directly using the extracted absolute residual sequences of human activity interventions and climate change-driven absolute residual sequences as correction operators, performing linear additive budgeting at the corresponding time steps, and compensating them onto the baseline monthly smoothing curve to generate the corresponding predetermined time series input file. This alternative solution can directly inherit the dimensions of the residual sequences by eliminating the scaling operation in the ratio transformation.

[0154] On the other hand, establishing a hydrological effect attribution identification framework, based on the simulation output results of various scenarios, quantifies and outputs the quantitative contribution rate of different vegetation restoration models to hydrological effects, which can be:

[0155] Step 501: Based on the differences between the simulated hydrological processes corresponding to each vegetation change scenario, the overall impact of vegetation change on hydrological elements is decomposed into independent impact components corresponding to natural recovery and independent impact components corresponding to human intervention.

[0156] Furthermore, it is necessary to perform difference analysis on the different time-series results of the control variable simulation output.

[0157] The overall impact is calculated by subtracting the simulated hydrological process corresponding to the baseline scenario from the simulated hydrological process corresponding to the actual dynamic change scenario.

[0158] The independent impact components corresponding to natural recovery are calculated by subtracting the simulated hydrological process corresponding to the baseline scenario from the simulated hydrological process corresponding to the natural vegetation succession scenario.

[0159] Similarly, by subtracting the simulated hydrological process corresponding to the baseline scenario from the simulated hydrological process corresponding to the vegetation restoration scenario, the independent impact components corresponding to human intervention are calculated. The decomposition calculation process is as follows:

[0160] W _VEG =f _veg -f _control ;

[0161] W _VH =f _human -f _control ;

[0162] W _VC=f _natural -f _control ;

[0163] Above, W _VEG To represent the overall impact of vegetation change on hydrological elements, f _veg f represents the simulated hydrological process results corresponding to real dynamic changing scenarios. _control The results represent the simulated hydrological processes corresponding to the baseline scenario;

[0164] W _VH f represents the independent impact component corresponding to human intervention. _human The results of simulated hydrological processes corresponding to human-induced vegetation restoration scenarios;

[0165] W _VC f represents the independent influence component corresponding to natural recovery. _natural The results represent the simulated hydrological processes corresponding to the natural succession scenario of vegetation.

[0166] Hydrological elements can be runoff, actual evapotranspiration, or soil moisture content;

[0167] Overall impact refers to the deviation of target hydrological elements caused by comprehensive changes in the underlying surface of the watershed;

[0168] Independent influence component refers to the hydrological response bias when it is assumed that only a single vegetation change driver exists.

[0169] Step 502: Simultaneously extract all independent influence components from the overall influence, and define the remaining deviation as the cooperative residual caused by the nonlinear superposition effect;

[0170] Accordingly, the difference between the total impact and the independent impact components corresponding to natural recovery and human intervention is the residual that cannot be explained by the independent components.

[0171] The residual is defined as the cooperative residual caused by the nonlinear superposition effect. This cooperative residual characterizes the hydrological interaction and intervention features between natural climate evolution and human activities on a spatiotemporal scale. Its stripping calculation process is shown below:

[0172] W _η =W _VEG -W _VH -W _VC ;

[0173] In the formula, W _η This is the cooperative residual caused by the nonlinear superposition effect.

[0174] Step 503: Calculate the corresponding quantitative contribution rate by using the proportion of each independent influence component and the collaborative residual in the overall influence.

[0175] Furthermore, the influence components with specific unit pairs are converted into percentage relative values.

[0176] The independent impact component corresponding to the calculated natural recovery is divided by the overall impact to obtain the quantitative contribution rate of natural recovery.

[0177] Divide the calculated independent impact component corresponding to human intervention by the total impact to obtain the quantitative contribution rate of human restoration.

[0178] Dividing the calculated synergistic residual by the overall impact yields the quantitative contribution rate of the synergistic effect.

[0179] For example, η _VC =W _VC / W _VEG ×100%;

[0180] η _VH =W _VH / W _VEG ×100%;

[0181] η _η =W _η / W _VEG ×100%;

[0182] Where, η _VC Quantitative contribution rate η to natural recovery _VH W is the quantitative contribution rate of human restoration. _VH η represents the independent influence component corresponding to human intervention. _η Quantify the contribution rate of synergistic effect.

[0183] In other embodiments, for some ecologically extremely fragile watersheds, if the calculated quantitative contribution rate of synergistic effect is negative and the absolute value is large, it indicates a strong mutual weakening effect and water resource competition between natural restoration and human restoration. In this case, an early warning mechanism can be triggered based on the feedback of this negative contribution rate to guide subsequent ecological engineering to dynamically adjust the vegetation planting density in the area, thereby providing targeted ecological water resource stock allocation decision parameters.

[0184] Furthermore, offline modeling uses measured runoff as the calibration basis, divides the time periods when vegetation status is stable, and calibrates the hydrological model in different time periods to obtain the optimal parameters;

[0185] A training dataset for a mapping model is constructed using contemporaneous vegetation features as input and corresponding calibration parameters as labels.

[0186] An end-to-end training mode is adopted, and the mapping model parameters are optimized by relying on the deviation between simulated and measured runoff.

[0187] The pre-built mapping model is obtained through the following offline model building steps:

[0188] Accordingly, the hydrological simulation error loss for the target watershed is obtained, and the physical constraint penalty term is extracted based on the physical feasible interval boundaries of each time-varying parameter in the hydrological model.

[0189] Among them, the hydrological simulation error loss for the target watershed can be calculated based on the deviation between the simulated runoff sequence of the hydrological model and the measured runoff sequence in the spatial grid time series dataset.

[0190] Blindly fitting hydrological parameters to machine learning models can easily lead to output parameters that violate basic hydrodynamic laws. Therefore, physical boundary constraints must be introduced during the offline model building phase.

[0191] That is, to obtain the hydrological simulation error loss for the target watershed, specifically, to calculate the mean square error between the simulated runoff sequence output by the hydrological model forward propagation and the corresponding measured runoff sequence, and use this as the first-level data fitting fidelity assessment standard.

[0192] The physical feasible interval boundary represents the reasonable range of parameter values ​​pre-calibrated based on the predetermined hydrological model structure.

[0193] Extracting the physical constraint penalty term involves extracting the parameter estimates from the machine learning network output in real time during each iteration of model training, comparing them with the corresponding upper and lower bounds, and applying a quadratic penalty to any values ​​exceeding the bounds. This extraction process can be performed according to the following formula:

[0194] L _range =sum(max(0,θ _j -θ _j_max ) 2 +max(0,θ _j_min -θ _j ) 2 );

[0195] Above, L _range The extracted physical constraint penalty term, sum is the summation operator over all output dimensions, and θ is the summation operator. _j θ is the estimated value of the j-th time-varying parameter output by the machine learning model in the current iteration. _j_max θ is the upper limit of the physically feasible interval of the set j-th time-varying parameter. _j_min is the lower limit of the physical feasible interval of the j-th time-varying parameter, and max is the function to take the maximum value.

[0196] Before calculating the physical constraint penalty term, the out-of-bounds deviation of each time-varying parameter is normalized according to the span of the corresponding physical feasible interval boundary, so that the penalty contribution of each parameter dimension is on the same order of magnitude, avoiding the parameter with a large interval span from dominating the optimization direction of the penalty term.

[0197] Accordingly, the physical constraint penalty term is superimposed on the hydrological simulation error loss to form the overall physical constraint loss function;

[0198] Optionally, a unified optimization objective matrix can be established.

[0199] The overall physical constraint loss function is constructed by introducing predetermined penalty weight parameters. The physical constraint penalty term, representing the degree of violation of the model boundary, is used as a regularization term and merged into the initial hydrological simulation error loss according to a linear addition rule. The superposition process is as follows:

[0200] L _total =L _hydro +λ _1 ×L _range ;

[0201] In the formula, L _total For the overall physical constraint loss function, L _hydro To obtain the hydrological simulation error loss, λ _1 L is the penalty weight parameter assigned to the physical constraint penalty term. _range This refers to the extracted physical constraint penalty term.

[0202] Accordingly, iterative optimization training is performed with the goal of minimizing the overall physical constraint loss function to obtain a pre-built mapping relationship model that has been trained.

[0203] Furthermore, because the hydrophysical model itself contains non-differentiable operations, the gradient flow in the deep learning error backpropagation algorithm breaks down when traversing the hydrophysical model, making it impossible to directly pass the backpropagation gradient of the overall physical constraint loss function located at the end of the network to the mapping relationship model at the front end. Therefore, this embodiment provides two parallel optimization strategies to achieve iterative optimization training.

[0204] First, a heuristic global optimization algorithm is used to implement the update. That is, a complex evolutionary algorithm or genetic algorithm is introduced to treat the mapping relationship model as a black box mapper that does not depend on gradient information.

[0205] In each round of population reproduction iteration, the parameter vector output by the black-box mapper is substituted into the hydrological model to calculate the overall physical constraint loss function. The reciprocal of this loss function is used as the fitness evaluation function of each individual parameter matrix. Crossover, mutation and global optimization operations are performed to gradually guide the network hyperparameters of the mapping relationship model to converge to the better solution region.

[0206] Secondly, a surrogate model technique is used to facilitate gradient propagation. That is, a differentiable hydrological surrogate model is constructed by pre-fitting the input-output response surface of the hydrological model using a feedforward neural network with continuous differentiability.

[0207] After the hydrological proxy model is connected in series to the mapping relationship model, the partial derivatives generated by the overall physical constraint loss function can be smoothly backpropagated to each layer node of the mapping relationship model through the standard chain rule via the hydrological proxy model. The weights of the mapping network are continuously updated using the gradient descent method until the loss function converges, resulting in a pre-built mapping relationship model that has been trained.

[0208] In some scenarios, the offline build mechanism is configured with a range of basic runtime hyperparameters.

[0209] The basic operating hyperparameters involved in the iterative optimization training process, such as learning rate, batch size, and weight decay coefficient, as well as the control parameters in the evolutionary algorithm, such as population size, crossover probability, and maximum number of iterations, can be determined through conventional hyperparameter tuning methods such as cross-validation or grid search.

[0210] The above ensures that the mapping relationship model converges smoothly to the globally optimal computational space.

[0211] Based on the offline model building mechanism, the iterative optimization training process aimed at minimizing the overall physical constraint loss function also includes a closed-loop adjustment mechanism for parameter boundaries to strengthen the rigid constraints of the model. Specifically, it includes the following steps:

[0212] Optionally, the physical constraint satisfaction rate is calculated by counting the number of parameters in the total output of the statistical model that fall within the corresponding physical feasible interval. The calculation of the physical constraint satisfaction rate means that after each training cycle or a fixed number of iteration batches, a comprehensive scan and comparison is performed on all target parameters currently output by the mapping relationship model.

[0213] For example, the calculation formula is as follows:

[0214] CR=N _valid / N _total ×100%;

[0215] In the formula, CR represents the calculated physical constraint satisfaction rate, and N... _valid N represents the total number of parameters whose values ​​fall within the preset physically feasible range. _total This represents the total number of time-varying parameters output by the model in this study.

[0216] Optionally, it can be determined whether the physical constraint satisfaction rate of the current round has reached the pre-configured satisfaction rate threshold. If it has not reached the satisfaction rate threshold, the penalty weight parameter corresponding to the physical constraint penalty term is increased and the model is retrained until the satisfaction rate threshold limit is finally met.

[0217] In some scenarios, the pre-configured satisfaction rate threshold can be set to 95%; however, this is not the only limitation and can be adjusted according to the actual situation.

[0218] If the physical constraint satisfaction rate calculated in the current round is found to be lower than the satisfaction rate threshold, it indicates that although the current model may have achieved local superiority in terms of hydrological simulation error loss, it has come at the cost of sacrificing hydrological regularity.

[0219] At this point, the closed-loop feedback control mechanism is triggered, which multiplies or amplifies the penalty weight parameter by a preset step size factor, clears the gradient accumulation cache of the current batch, and forces the job scheduling module to load the updated penalty weight parameter and re-execute iterative optimization training.

[0220] As the penalty weight parameters continue to increase, regions in the model optimization plane that violate physical boundaries will be constructed as penalty barriers, forcing the mapping relationship model to adaptively adjust the weights of internal nodes and compress the output parameter cluster back into the physical compliance domain.

[0221] This cyclical verification and penalty amplification process will continue to be triggered until the physical constraint satisfaction rate exceeds the satisfaction rate threshold limit. This not only makes the numerical accuracy of the offline mapping relationship model higher, but also ensures that each set of vegetation time-varying parameters output by the model has hydrological interpretation significance.

[0222] In some other embodiments, the hydrological model is a distributed hydrological model;

[0223] The transformation and derivation of time-varying parameters and the simulation process of control variables are all run one by one on each independent computational grid unit divided by the spatial grid time series dataset, so as to reflect the spatial heterogeneity of vegetation and hydrological element evolution within the watershed.

[0224] In the above text, the distributed hydrological model can be specifically constructed using a variable infiltration capacity model.

[0225] That is, obtain the geographic coordinate boundaries in the spatial grid time series dataset and divide the target watershed into multiple independent computational grid units arranged in an array.

[0226] During computation, a parallel computing framework or loop structure is invoked, and mapping instructions are issued one by one to each independent computation grid cell, using the two-dimensional spatial coordinates (x, y) as the index.

[0227] At this point, each independent computational grid cell independently triggers the nonlinear time-varying mapping network to update local physical parameters based on its own extracted local dynamic vegetation parameters and microclimate measured sequences, and performs control variable simulations respectively.

[0228] Since the terrain slope, soil texture and degree of human engineering disturbance vary in each grid, the mechanism of independent calculation for each grid can preserve the spatial heterogeneity of vegetation and hydrological element evolution within the watershed.

[0229] In some scenarios, the hydraulic connection characteristics between each independent computational grid unit can also be obtained according to the water system topology, which can be used to improve the computational efficiency of the whole watershed grid set.

[0230] Based on the water flow confluence hierarchy, a directed acyclic graph is used to allocate parallel computing tasks, scheduling grid cells without upstream and downstream physical dependencies to the same computing batch for independent processing.

[0231] Based on this, after quantifying and outputting the quantitative contribution rate of different vegetation restoration models to hydrological effects, the following is also included:

[0232] Based on the quantitative contribution rate of each spatial grid output in the spatial grid time series dataset, a spatial distribution map of the watershed hydrological effect attribution is generated and rendered to visualize the spatial differences in the impact of climate change and human intervention on regional water resources.

[0233] In this step, the grid-level data is converted into a two-dimensional continuous surface field with geographic constraints.

[0234] Accordingly, after obtaining the quantitative contribution rates of natural restoration, human restoration, and synergistic effects from the independent computational outputs of all independent computational grid cells, a mapping table between quantitative data and spatial coordinates is established.

[0235] The mapping table is processed using an inverse distance weighted interpolation algorithm to output a continuous and smooth data surface covering the entire watershed.

[0236] Furthermore, a mapping dictionary between numerical intervals and color gradients is established, and the extracted color information is used to perform raster coloring on the continuous smooth data surface to render a spatial distribution map of the watershed hydrological effect attribution.

[0237] For example, geographical areas where the quantitative contribution rate of human intervention is greater than a predetermined proportion are mapped to red-based feature values, while geographical areas where the quantitative contribution rate of natural climate-driven factors is greater than a predetermined proportion are mapped to blue-based feature values.

[0238] This output format can provide evidence solidification means for commercial software systems or environmental monitoring platforms, clearly define the driving sources of water resource loss or increase in various geographical areas, and complete the technical transformation from pure numerical results to a quantitative decision-making interface.

[0239] In other embodiments, the study watershed includes areas where ecological restoration projects such as converting farmland back to forest or grassland have been implemented;

[0240] The independent driving forces of different restoration modes causing vegetation changes are extracted and combined with quantitative contribution rates to calculate and output the assessment results of water resources and vegetation carrying capacity under the dual driving forces of natural succession and human restoration.

[0241] Correspondingly, taking some ecologically fragile areas as an example, these areas have undergone long-term ecological restoration projects of returning farmland to forests and grasslands, resulting in increased vegetation coverage on the underlying surface, which in turn leads to the risk of hydrological imbalance, such as increased ecological water consumption and reduced available surface runoff.

[0242] Assessing the water and vegetation carrying capacity under the dual drivers of natural succession and anthropogenic restoration involves extracting the base amount of allocable water resources under the baseline scenario and subtracting the actual evapotranspiration consumption increases brought about by the anthropogenic restoration and natural succession scenarios. The carrying capacity calculation can be performed using the following formula:

[0243] Q _avail =Q _base -ΔE _H -ΔE _C -ΔE _η ;

[0244] In the formula, Q _avail Q represents the remaining available water resources calculated from the assessment. _base ΔE represents the baseline runoff stock output under the baseline scenario. _H This represents the actual increase in evapotranspiration consumption due to human-induced repairs, derived from the aforementioned independent driving force analysis, ΔE. _C ΔE represents the actual increase in evapotranspiration consumption due to natural succession. _η This represents the deviation in evapotranspiration consumption resulting from the nonlinear synergistic effect of the two.

[0245] The above formula is based on the principle of long-term water balance, that is, under the control condition that the precipitation input remains unchanged, the increase in evapotranspiration is approximately equal to the decrease in available runoff resources.

[0246] The calculation formula is used to determine whether the remaining available water resources are lower than the pre-configured ecological security warning threshold.

[0247] When the comparison result is determined to be below the threshold, the vegetation density reduction ratio is reversed by using the extracted incremental feature parameters of each part, and the environmental intervention strategy parameters containing the reduction target are output.

[0248] The evaluation process extracted the evapotranspiration loss component caused by human responsibility, providing a quantitative engineering constraint basis for formulating differentiated watershed ecological red lines and restricting the expansion of certain water-consuming vegetation.

[0249] In other specific embodiments, the remaining available water resources calculated above can be coupled and compared with the time series of agricultural irrigation demand stored in the study watershed.

[0250] If the comparison results show that a time-series water shortage conflict has occurred, a trigger control command is generated, the scheduling limit parameters are output and written into the water resource allocation execution module, and the system-level automated decision-making closed loop of ecological restoration monitoring and available water resource scheduling is completed.

[0251] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the hydrological effects of different vegetation restoration patterns under a dynamic model and parameters, characterized in that, include: Acquire multi-source basic data for the research watershed and establish a spatial grid time series dataset; Land use evolution characteristics are analyzed based on remote sensing vegetation data in spatial grid time series datasets to generate dynamic vegetation parameters that reflect the trajectory of vegetation change. Based on dynamic vegetation parameters, we delineate areas with unchanged and changing land use types, perform trend separation processing, and extract the independent driving forces of different restoration modes that cause vegetation changes. The mapping relationship model is invoked to convert dynamic vegetation parameters into time-varying parameters of the hydrological model under the corresponding vegetation conditions; The vegetation input files for different vegetation change scenarios are reconstructed using independent driving forces. The hydrological model is driven by time-varying parameters to simulate controlled variables and obtain the simulated hydrological process corresponding to each vegetation change scenario. Based on the extraction of independent influence components and nonlinear synergistic effect components from simulated hydrological processes, this study quantifies and outputs the quantitative contribution rate of different vegetation restoration models to hydrological effects. Specifically, this includes: based on the differences between simulated hydrological processes corresponding to various vegetation change scenarios, decomposing the overall impact of vegetation change on hydrological elements into independent influence components corresponding to natural restoration and human intervention; simultaneously extracting all independent influence components from the overall impact, defining the remaining deviation as the synergistic residual caused by nonlinear superposition effect; and calculating the corresponding quantitative contribution rate by using the proportion of each independent influence component and the synergistic residual in the overall impact. Among them, the mapping relationship model is a machine learning model, which specifically uses a machine learning model to establish a time-varying mapping relationship between dynamic vegetation parameters and parameters in the hydrological model; the vegetation input files of different vegetation change scenarios are reconstructed using independent driving forces, and the hydrological model is driven by time-varying parameters to perform controlled variable simulation. Specifically, it includes: keeping the meteorological input conditions of the study watershed unchanged, setting four types of simulation scenarios, including the baseline scenario, the real dynamic change scenario, the vegetation human restoration scenario, and the vegetation natural succession scenario, and performing controlled variable simulation for each of the four types of simulation scenarios. For both human-induced vegetation restoration and natural vegetation succession scenarios, the corresponding independent driving forces are converted into contribution rate ratios, which are then applied to the actual observed vegetation parameter increments and superimposed on the vegetation parameter sequence of the baseline state. This process reconstructs the predetermined time series inputs under both human intervention-driven and natural succession-driven independent driving scenarios, which serve as the corresponding vegetation input files.

2. The method according to claim 1, characterized in that, Multi-source basic data should include at least regional meteorological observation data; Remote sensing vegetation data includes at least leaf area index, vegetation cover, and land use cover type.

3. The method according to claim 1, characterized in that, For dynamic vegetation parameters, areas with unchanged and changing land use types were delineated for trend separation processing, and independent driving forces causing different restoration modes of vegetation change were extracted, including: Within areas where land use types remain unchanged, a regression mapping relationship is established between dynamic vegetation parameters and corresponding spatiotemporal meteorological data to obtain a climate impact baseline. By extrapolating the climate impact baseline to areas of land use change, we can calculate the theoretical vegetation evolution trend under the dominance of climate change. By combining the actual and theoretical vegetation evolution trends in areas of land use change, residual analysis was conducted to separate the contributions of climate change and human intervention, and each extracted contribution was taken as an independent driving force.

4. The method according to claim 3, characterized in that, Establish a regression mapping relationship between dynamic vegetation parameters and corresponding spatiotemporal meteorological data, specifically by using a pre-constructed support vector machine regression model for baseline fitting; Residual analysis was conducted by combining the actual vegetation evolution trend with the theoretical vegetation evolution trend in the land use type change area. Specifically, the least squares method was used to calculate the predetermined evolution trend of vegetation leaf area and vegetation cover under actual evolution, climate change influence, and human activity intervention, and to quantify the independent driving ratio that distinguishes different driving sources.

5. The method according to claim 1, characterized in that, The mapping relationship model is pre-trained through the following offline model building steps: Obtain the hydrological simulation error loss for the target watershed, and extract physical constraint penalty terms based on the physical feasible interval boundaries of each time-varying parameter in the hydrological model; The physical constraint penalty term is superimposed on the hydrological simulation error loss to form the overall physical constraint loss function; The mapping relationship model is trained iteratively with the goal of minimizing the overall physical constraint loss function. The hydrological proxy model is then connected to the mapping relationship model in series. The partial derivatives of the overall physical constraint loss function are smoothly backpropagated to each layer node of the mapping relationship model through the hydrological proxy model using the standard chain rule. The weights of the mapping network are continuously updated using the gradient descent method until the loss function converges, thus obtaining the pre-built mapping relationship model after training. The iterative optimization training process aimed at minimizing the overall physical constraint loss function also includes a closed-loop adjustment mechanism: the number of parameters in the model training output that fall within the corresponding physical feasible interval is counted to calculate the physical constraint satisfaction rate; it is determined whether the physical constraint satisfaction rate of the current round has reached the pre-configured satisfaction rate threshold. If it has not reached the satisfaction rate threshold, the penalty weight parameter corresponding to the physical constraint penalty term is increased and the model is retrained until the satisfaction rate threshold limit is finally met.

6. The method according to claim 1, characterized in that, The hydrological model is a distributed hydrological model; The transformation and derivation of time-varying parameters and the simulation process of control variables are all run one by one on each independent computational grid unit divided by the spatial grid time series dataset, so as to reflect the spatial heterogeneity of vegetation and hydrological element evolution within the watershed.

7. The method according to claim 1, characterized in that, After quantifying and outputting the quantitative contribution rates of different vegetation restoration models to hydrological effects, the following is also included: Based on the quantitative contribution rate output by each computational grid unit, a spatial distribution map of the watershed hydrological effect attribution is generated and rendered to visualize the spatial differences in the impact of climate change and human intervention on regional water resources.

8. The method according to claim 1, characterized in that, The study basin includes areas where ecological restoration projects such as converting farmland back to forest and grassland have been implemented; We extract the independent driving forces of different restoration mechanisms that cause vegetation changes, and use them to assess the water resource and vegetation carrying capacity under the dual driving forces of natural succession and human restoration.

Images