Mechanism-constrained deep learning inversion method and system for regional ecological monitoring parameters

Abstract

This invention discloses a deep learning inversion method and system for regional ecological monitoring parameters constrained by mechanisms, belonging to the interdisciplinary field of ecological environment remote sensing monitoring and artificial intelligence. The method includes: acquiring basic data of the target area and constructing multi-dimensional input features; constructing ecological mechanism constraints and generating simulated samples based on these constraints; establishing a deep learning inversion model, pre-training it using simulated samples, and then adding an ecological mechanism constraint term to the loss function for training; inputting the multi-dimensional input features into the trained model and outputting regional ecological monitoring parameters. This invention embeds ecological process mechanisms into the main deep learning training process in the form of mathematical constraints. Through a phased strategy combining simulated sample pre-training and measured sample fine-tuning, it effectively alleviates the problem of sample scarcity, achieves integrated joint inversion of ecosystem structure, quality, and service function parameters, and significantly improves the ecological rationality, regional adaptability, and operational practicality of the inversion results.

Description

Technical Field

[0001] This invention relates to the field of ecological environment remote sensing monitoring and artificial intelligence, and in particular to a mechanism-constrained deep learning inversion method and system for regional ecological monitoring parameters. Background Technology

[0002] Regional ecological monitoring parameter inversion is a crucial technical foundation for ecological environment assessment, evaluation of ecological restoration effectiveness, and dynamic resource monitoring. In arid and semi-arid regions, ecosystem structures are fragile, hydrothermal conditions are significantly constrained, and ecological processes are influenced by multiple factors, including vegetation type, topography, spatial and temporal precipitation distribution, and anthropogenic disturbances. Consequently, regional ecological parameters typically exhibit strong spatial heterogeneity and temporal fluctuations. While traditional ground-based survey methods offer high accuracy, their limited scope, long update cycles, and high costs make them unsuitable for continuous regional-scale monitoring. Inversion methods based on single remote sensing indices or empirical statistical relationships are relatively simple to implement, but they fail to adequately express the coupling relationships between ecosystem structure, ecological quality, and ecosystem service functions. Ecological process models possess strong mechanistic foundations, but they often suffer from numerous parameters, complex regional calibration, and high operating costs, limiting their widespread and rapid application.

[0003] Existing research has enabled tasks such as vegetation type identification, forest structure parameter extraction, and change detection. While some technologies have proposed overall training approaches based on multi-source data, simplified physical constraints, and deep learning coupling, the following shortcomings remain: First, mechanistic information is mostly confined to result interpretation or posterior validation, and has not yet been fully integrated into the main model training process. Second, regarding the issue of insufficient samples, there is a lack of a complete implementation path for generating simulated samples using mechanistic relationships and co-training them with real samples. Third, existing methods do not sufficiently disclose the correspondence between input features, mechanistic constraints, training procedures, and output parameters, making it difficult to form a stable and reusable technical solution for inverting regional ecological monitoring parameters. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the current field of ecological environment remote sensing monitoring and artificial intelligence, this invention provides a mechanism-constrained deep learning inversion method and system for regional ecological monitoring parameters. It can deeply embed the ecological process mechanism into the deep learning training process in the form of mathematical constraints, and at the same time use simulated samples for pre-training and apply constraints to the ecological process mechanism through loss function, thereby maintaining the strong fitting ability of deep learning.

[0005] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0006] A mechanism-constrained deep learning inversion method for regional ecological monitoring parameters includes the following steps:

[0007] Acquire basic data for the target region and construct multidimensional input features;

[0008] Ecological mechanism constraints are constructed, and simulation samples are generated based on the ecological mechanism constraints; the ecological mechanism constraints express the value range relationship, conservation relationship or monotonicity relationship between ecological parameters in the form of mathematical inequalities or equations;

[0009] A deep learning inversion model is established, the model is pre-trained using the simulated samples, the ecological mechanism constraint term is added to the loss function of the model, and the pre-trained model parameters are used for training.

[0010] The multidimensional input features are input into the trained deep learning inversion model, which outputs regional ecological monitoring parameters.

[0011] According to one aspect of the present invention, the ecological mechanism constraints include one or more of carbon sink constraints, water balance constraints, and windbreak and sand fixation constraints.

[0012] According to one aspect of the invention, the carbon sink constraint is that net primary productivity (NPP) is not less than a lower limit determined based on vegetation type, temperature, and precipitation, and not greater than an upper limit determined based on a light energy utilization model.

[0013] According to one aspect of the present invention, the water balance constraint is that the water conservation capacity is not less than zero and not greater than the difference between precipitation and evapotranspiration, and not greater than the maximum water holding capacity of the soil.

[0014] According to one aspect of the present invention, the windbreak and sand-fixing constraint is: the amount of windbreak and sand-fixing remains monotonically constant with the increase of vegetation coverage and with the increase of wind speed, and is not greater than the maximum amount of sand-fixing determined by wind speed and soil properties.

[0015] According to one aspect of the present invention, the generation of simulation samples based on the ecological mechanism constraints specifically includes:

[0016] Within the target area, vegetation type combinations are generated by random sampling based on land use maps;

[0017] Within a preset reasonable range, structural parameters and environmental factors are randomly generated. The structural parameters include one or more of tree height, canopy closure, and leaf area index, and the environmental factors include precipitation, temperature, and wind speed.

[0018] The corresponding ecological parameters are calculated using a lightweight ecological process model, and the input features are combined with the corresponding ecological parameters to form a simulated sample dataset.

[0019] According to one aspect of the present invention, the lightweight ecological process model is selected from one or more of the carbon module, water balance module, or windbreak and sand fixation module of the InVEST model.

[0020] According to one aspect of the present invention, the pre-training of the model using simulated samples specifically includes: in the pre-training stage, only the simulated samples are used, and the weight of the ecological mechanism constraint term in the loss function is zero or approaches zero; the step of adding the ecological mechanism constraint term to the loss function of the model for training is a fine-tuning stage, and the fine-tuning stage specifically includes: using simulated samples and real samples in combination, and the weight of the ecological mechanism constraint term is greater than zero.

[0021] According to one aspect of the present invention, the ratio of the number of simulated samples to real samples used in the fine-tuning phase is from 1:1 to 5:1, the weight of the ecological mechanism constraint term ranges from 0.01 to 0.5, and the weight of the ecological mechanism constraint term is dynamically adjusted with the training rounds.

[0022] A mechanism-constrained deep learning inversion system for regional ecological monitoring parameters, comprising:

[0023] The acquisition module obtains basic data of the target region and constructs multidimensional input features;

[0024] The constraint module constructs ecological mechanism constraints and generates simulation samples based on the ecological mechanism constraints; the ecological mechanism constraints express the value range relationship, conservation relationship or monotonicity relationship between ecological parameters in the form of mathematical inequalities or equations;

[0025] The inversion module establishes a deep learning inversion model, pre-trains the model using the simulated samples, adds the ecological mechanism constraint term to the loss function of the model, and trains the model using the pre-trained model parameters.

[0026] The output module takes the multi-dimensional input features and inputs them into the trained deep learning inversion model to output regional ecological monitoring parameters.

[0027] The advantages of this invention are as follows: It deeply embeds ecological process mechanisms into the main deep learning training process in the form of mathematical constraints, fundamentally solving the problem in existing methods where mechanism information is only used for posterior validation and cannot guide model optimization. Simultaneously, it utilizes mechanism constraints to generate simulated samples for pre-training and then mixes them with measured samples for fine-tuning, effectively alleviating the model instability problem caused by the scarcity of samples in arid and semi-arid regions. Furthermore, the joint inversion framework for multiple ecological monitoring parameters avoids the repetitive work of independently modeling single indicators, ensuring the synergistic consistency among parameters of ecosystem structure, quality, and service functions. Ultimately, this invention significantly improves the ecological rationality, regional adaptability, and operational scalability of the inversion results while maintaining the strong fitting ability of deep learning. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating the mechanism-constrained deep learning inversion method for regional ecological monitoring parameters as described in this invention.

[0030] Figure 2 This is a schematic diagram of the structure of a mechanism-constrained deep learning inversion system for regional ecological monitoring parameters according to the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1

[0033] To address the problems in existing regional ecological monitoring parameter inversion methods, such as the failure of mechanistic information to effectively participate in model training, insufficient samples leading to poor regional adaptability of the model, and lack of synergistic consistency in multi-parameter inversion, this embodiment provides a mechanism-constrained deep learning inversion method for regional ecological monitoring parameters. By constructing multi-dimensional input features and ecological mechanism constraints, simulated samples are generated based on the mechanism constraints for pre-training, and a mechanism constraint term is introduced into the loss function for fine-tuning training, thereby achieving integrated joint inversion of ecological parameters.

[0034] like Figure 1 As shown, a mechanism-constrained deep learning inversion method for regional ecological monitoring parameters includes the following steps:

[0035] Step S1: Obtain basic data of the target region and construct multidimensional input features; Step S1 in this embodiment includes the following steps:

[0036] Step S11: Obtain basic data for the target area.

[0037] The basic data includes remote sensing observation data, ground observation data, and environmental auxiliary data. Among them, the remote sensing observation data includes at least multi-temporal optical remote sensing imagery, and radar, thermal infrared, or high-resolution imagery data may be further integrated if necessary; the ground observation data includes one or more of the following: plot survey data, flux observation data, and verification sample data; the environmental auxiliary data includes one or more of the following: precipitation, temperature, radiation, evapotranspiration, wind speed, topography, soil, or land use data.

[0038] Step S12: Preprocess the basic data.

[0039] The preprocessing includes coordinate unification, outlier removal, spatiotemporal registration, resolution coordination, standardization, and missing value handling, forming the basic dataset required for model input.

[0040] Step S13: Construct multidimensional input features based on the preprocessed basic data.

[0041] The multidimensional input features include at least the following four categories: spectral and index features, used to characterize land cover status and vegetation growth differences; structural features, used to characterize stand height, canopy closure, biomass level or other structural attributes; variation features, used to characterize information on regional ecological state recovery, degradation or fluctuation; and environmental driving features, used to characterize external constraints such as temperature, moisture, wind field, topography and soil.

[0042] Preferably, the above features are organized hierarchically, including bottom-level pixel features, mid-level structural features, and high-level semantic features, and input into the model in a unified data cube format. Furthermore, independent encoding paths can be set for different feature groups to allow for differentiated processing of features from different sources in subsequent fusion stages.

[0043] Step S2: Construct ecological mechanism constraints and generate simulation samples based on the ecological mechanism constraints;

[0044] Step S2 in this embodiment includes the following steps:

[0045] Step S21: Construct ecological mechanism constraints.

[0046] The ecological mechanism constraints are expressed in the form of mathematical inequalities or equations, representing the relationships between the ranges of ecological parameters, conservation relationships, or monotonic relationships. Specifically, they include one or more of the following constraints:

[0047] (1) Carbon sink constraint For a given vegetation type v and leaf area index LAI, net primary productivity (NPP) should satisfy:

[0048] NPP min (v,T,P)≤NPP≤NPP max (v,T,P);

[0049] Where v represents vegetation type, LAI represents leaf area index, T represents average temperature during the statistical period, and P represents cumulative precipitation during the statistical period.

[0050] The lower and upper limits of NPP are calculated as follows: NPP min (v,T,P)=ε min (v)×f T (T)×f P (P)×fPAR×PAR, NPP max (v,T,P)=ε max (v)×f T (T)×f P (P)×fPAR×PAR.

[0051] Where, ε min (v) and ε max (v) represents the minimum and maximum light energy utilization rates corresponding to vegetation type v, which can be determined based on the 5th and 95th percentiles of the actual light energy utilization rates of historical samples of the same type of vegetation; f T (T) is the temperature stress coefficient, f P (P) is the water stress coefficient, and both values ​​range from [0,1].

[0052] In one implementation, f T (T)=max{0,1-|TT opt (v)| / ΔT v}, f P (P)=min{1,max[0,P / P req (v)]}, where T opt (v) represents the optimum temperature for vegetation type v, ΔT v To allow for temperature deviation, P req (v) Water requirements to meet normal growth; fPAR is the proportion of photosynthetically active radiation absorbed by vegetation, which can be obtained from remote sensing products or calculated as fPAR=1-exp[-k v Calculate k using [×LAI] v PAR is the extinction coefficient of vegetation type v; PAR is the incident photosynthetically active radiation during the statistical period.

[0053] (2) The water balance constraint water conservation capacity WY should meet the following requirements:

[0054] 0≤WY≤P-ET;

[0055] Where P is the cumulative precipitation during the statistical period, and ET is the evapotranspiration during the same period; further, WY ≤ SW can be constrained. max , of which SWmax To determine the maximum effective water holding capacity within the soil layer, when using volumetric water content, SW max =(θ FC -θ WP )×D×1000, where θ FC and θ WP These represent field water holding capacity and volumetric water content at the wilting point, respectively. D represents the statistical soil layer thickness, and the unit of the calculated result is mm.

[0056] (3) Windbreak and sand fixation constraints. The amount of windbreak and sand fixation SF is related to the vegetation coverage f. c The relationship between wind speed u and the monotonicity constraint is satisfied:

[0057]

[0058] There is also an upper limit constraint: SF≤SF max (u,soil). Where SF represents the amount of windbreak and sand fixation per unit area within a unit statistical period, and f... c SF represents vegetation cover, u represents the representative wind speed exceeding the sand-lifting threshold during the statistical period, and soil represents a set of soil properties, including soil texture, moisture content, and surface roughness. max (u,soil) is determined according to the potential wind erosion under no-vegetation conditions, and can be expressed as: SF max (u,soil)=K soil (soil)×max[0,u²-u t [(soil)²]×Δt. Where K soil (soil) is the wind erosion coefficient determined by soil properties, u t (soil) represents the threshold wind speed for sand initiation, and Δt represents the length of the statistical period; K soil and u t It can be determined based on wind tunnel tests, field observations, or preset soil parameter tables.

[0059] The aforementioned ecological mechanism constraints are transformed into differentiable constraint violation penalty functions during the training phase. For value ranges or upper and lower bound relationships, only the portion exceeding the allowable range is penalized; for equations or approximate conservation relationships, the deviation between the predicted value and the right-hand side of the constraint relationship is penalized; for monotonic relationships, the direction of the partial derivative or the degree of violation of the sample pair ranking relationship is penalized.

[0060] Step S22: Generate simulated samples based on the ecological mechanism constraints.

[0061] This step S22 specifically includes the following steps:

[0062] Step S221: Within the target area, generate vegetation type combinations by random sampling based on land use or cover maps;

[0063] Step S222: Randomly generate structural parameters and environmental factors within a reasonable range. The structural parameters include tree height, canopy closure, and leaf area index. The environmental factors include precipitation, temperature, and wind speed.

[0064] Step S223: Calculate the corresponding ecological parameters using a lightweight ecological process model. The ecological parameters include net primary productivity, water conservation capacity, and windbreak and sand fixation capacity. The lightweight ecological process model is selected from one or more of the carbon module, water balance module, or windbreak and sand fixation module of the InVEST model. Other lightweight modules or regularized calculation processes with mechanism expression capabilities can also be used.

[0065] Step S224: Combine the input features with the corresponding ecological parameters to form a simulated sample dataset.

[0066] Step S3: Establish a deep learning inversion model, pre-train the model using the simulated samples, add the ecological mechanism constraint term to the loss function of the model, and train the model using the pre-trained model parameters;

[0067] Step S3 in this embodiment includes the following steps:

[0068] Step S31: Establish a deep learning inversion model.

[0069] The model includes at least an input layer, a spatial feature extraction module, a temporal feature extraction module, a feature fusion module, and a multi-task output module. The spatial feature extraction module extracts the spatial distribution and structural patterns of ground features at different scales, preferably using convolutional neural networks, encoder-decoder networks, or residual network structures. The temporal feature extraction module extracts dynamic change information from multi-temporal observation data, preferably using recurrent neural networks, temporal convolutional networks, or Transformer structures. The feature fusion module jointly models features from different sources and at different levels, preferably using splicing fusion, weighted fusion, or attention-guided fusion methods. The multi-task output module outputs one or more of the following: ecosystem structure parameters, ecosystem quality parameters, and ecosystem service function parameters. Preferably, the model uses a shared backbone and task-specific header structures to support the collaborative inversion of different parameters.

[0070] Step S32: Pre-train the model using the simulated samples.

[0071] During the pre-training phase, only simulated sample datasets are used, and the loss function adopts mean squared error or mean absolute error. At this time, the loss function does not contain mechanistic constraint terms, that is, the weight of the mechanistic constraint terms is zero or tends to zero, so that the model learns the basic mapping relationship between input features and ecological parameters.

[0072] Step S33: Add the ecological mechanism constraint term to the loss function of the model.

[0073] The total loss function is designed as follows: L total =L data +λ×L phy Among them, L total L is the total loss function; data λ is the data fitting term, using mean squared error or mean absolute error to calculate the deviation between the model's predicted values ​​and the measured samples; λ is the weighting coefficient of the mechanistic constraint term, ranging from 0.01 to 0.5, and can be dynamically adjusted with training epochs; L phy This is a mechanism constraint term used to calculate the degree to which the prediction results violate the ecological mechanism constraints constructed in step S21.

[0074] Specifically, let the predicted values ​​of net primary productivity, water conservation, and windbreak and sand fixation for the i-th sample be NPP, respectively. pred,i WY pred,i and SF pred,i The corresponding vegetation type, temperature, precipitation, evapotranspiration, vegetation cover, wind speed, and soil properties in the input features are used to calculate NPP. min NPP max SW max and SF max Equal constraint boundaries. The mechanistic constraint term can be expressed as:

[0075] L phy =β1×L NPP +β2×L WY +β3×L SF,bound +β4×L SF,mono ;

[0076] Among them, β1, β2, β3 and β4 are the weight coefficients of various mechanism constraint sub-items. Each weight can be set according to the importance of regional ecological monitoring parameters, sample reliability or training rounds.

[0077] Carbon sink constraint loss L NPP The penalty for situations where the predicted net primary productivity is below the ecological lower limit or above the ecological upper limit can be expressed as: L NPP =(1 / N)Σ i {[max(0,NPP min (v i ,T i ,P i )-NPP pred,i )]²+[max(0,NPP pred,i -NPP max (v i,T i ,P i ))]²};

[0078] Where N is the number of samples in the current training batch, and i is the sample index; NPP pred,i v is the predicted net primary productivity value for the i-th sample; i T i and P i These represent the vegetation type, average temperature during the statistical period, and cumulative precipitation for the i-th sample, respectively; NPP min (v i ,T i ,P i ) and NPP max (v i ,T i ,P i ) Calculate according to the method described in step S21, that is, use ε respectively min (v i ) and ε max (v i ) and f T (T i ), f P (P i fPAR i and PAR i The product is determined; fPAR i and PAR i These represent the absorbed photosynthetically active radiation ratio and the incident photosynthetically active radiation amount for the i-th sample, respectively.

[0079] Water balance constraint loss L WY The penalty for situations where the predicted water conservation capacity is less than zero, exceeds the difference between precipitation and evapotranspiration, or exceeds the maximum soil water holding capacity can be expressed as: L WY =(1 / N)Σ i {[max(0,-WY pred,i )]²+[max(0,WY pred,i -(P i -ET i ))]²+[max(0,WY pred,i -SW max,i )]²};wherein, WY pred,i Let P be the predicted water conservation capacity for the i-th sample. i and ET i These represent the cumulative precipitation and evapotranspiration for the same period, respectively. max,i The maximum effective water holding capacity is calculated based on the soil properties and statistical soil layer thickness of the i-th sample.

[0080] Upper limit constraint loss L for windbreak and sand fixationSF,bound The penalty for cases where the predicted value of windbreak and sand fixation is less than zero or exceeds the maximum sand fixation amount determined by wind speed and soil properties can be expressed as: L SF,bound =(1 / N)Σ i {[max(0,-SF pred,i )]²+[max(0,SF pred,i -SF max (u i soil i ))]²};wherein, SF pred,i Let u be the predicted value of windbreak and sand fixation for the i-th sample. i As a representative wind speed, soil i For soil properties, SF max (u i soil i Calculate the potential wind erosion amount according to the formula described in step S21.

[0081] windbreak and sand fixation monotonic constraint loss L SF,mono This is used to penalize situations where the amount of windbreak and sand-fixing capacity decreases with increasing vegetation cover or wind speed. When automatic differentiation is used, it can be expressed as: Among them, f c,i Let be the vegetation cover of the i-th sample. and These represent the partial derivatives of the predicted windbreak and sand-fixing volume with respect to vegetation coverage and wind speed, respectively.

[0082] When partial derivatives are not directly calculated, sample pairs can be constructed under the same or similar soil conditions. If the predicted windbreak and sand-fixing amount corresponding to the sample with higher vegetation cover or wind speed is smaller, then the penalty is calculated based on the prediction difference. In the above terms, max(0,·) indicates that only the part that violates the constraint is included in the loss; when the prediction result meets the corresponding ecological mechanism constraint in step S21, the penalty value of this constraint sub-item is zero or close to zero. The greater the degree of violation, the greater the penalty for updating the model parameters. Through the above method, L phy The carbon sink constraint, water balance constraint, and windbreak and sand fixation constraint in step S21 are transformed into loss terms that can participate in backpropagation training, so that the model can maintain ecological rationality while fitting real samples.

[0083] Step S34: Train the model using the pre-trained model parameters.

[0084] In this step, the training using the pre-trained model parameters is divided into a fine-tuning stage: a mixture of simulated and real samples is used, with the ratio of simulated to real samples ranging from 1:1 to 5:1. At this point, λ is greater than zero, so that the model can fit real observations while maintaining ecological rationality.

[0085] Step S4: Input the multidimensional input features into the trained deep learning inversion model and output regional ecological monitoring parameters;

[0086] Using the model trained in step S3, the multidimensional input features of the target area are processed to output one or more of the following: ecosystem structure parameters, ecological quality parameters, and ecosystem service function parameters. The output results include, but are not limited to: vegetation structure-related parameters, ecological quality-related parameters, carbon sink-related parameters, water conservation-related parameters, windbreak and sand fixation-related parameters, and their spatial distribution results. These output results are used for regional ecological monitoring, ecological engineering effectiveness evaluation, ecological resource surveys, and ecological governance decision-making.

[0087] Example 2

[0088] This embodiment uses a typical region in Ningxia as an example to illustrate the specific application of the method of the present invention.

[0089] B1: Obtain basic data of the target region and construct multi-dimensional input features;

[0090] Acquire remote sensing data, ground observation data, and environmental auxiliary data for the target area. After preprocessing such as coordinate unification, anomaly removal, spatiotemporal registration, and standardization, construct spectral and exponential features, structural features, variation features, and environmental driving features to form a multidimensional input feature set.

[0091] B2: Construct ecological mechanism constraints and generate simulation samples based on the ecological mechanism constraints;

[0092] Based on the characteristics of the regional ecosystem, one or more of the following constraints are constructed: carbon sink constraints, water balance constraints, and windbreak and sand fixation constraints. Using these constraints, vegetation types, structural parameters, and environmental factors are randomly generated within a reasonable range. A lightweight ecological process model is then used to calculate the corresponding ecological parameters, forming a simulated sample dataset.

[0093] B3: Establish a deep learning inversion model, pre-train the model using the simulated samples, then add the ecological mechanism constraint term to the loss function of the model, and train the model using the pre-trained parameters;

[0094] A deep learning inversion model is constructed, comprising an input layer, a spatial feature extraction module, a temporal feature extraction module, a feature fusion module, and a multi-task output module. First, pre-training is performed using only simulated samples, with the loss function excluding mechanistic constraints. Then, simulated samples are mixed proportionally with real samples, and mechanistic constraints are added to the total loss function. The pre-trained model is then fine-tuned.

[0095] Among them, the mechanism constraint term in the total loss function described in B3 is constructed in the manner described in step S33 of Example 1, that is, the carbon sink constraint, water balance constraint and windbreak and sand fixation constraint are respectively converted into L NPP L WY L SF,bound and L SF,mono And sum them according to their weights to form L phy When the prediction result meets the corresponding ecological constraint, the constraint term does not generate or generates almost no penalty; when the prediction result violates the constraint, a penalty is generated according to the extent of the violation.

[0096] B4: Input the multidimensional input features into the trained deep learning inversion model and output regional ecological monitoring parameters;

[0097] The multidimensional input features of the target region are fed into the trained model, which outputs the spatial distribution results of ecosystem structure parameters, ecosystem quality parameters, and ecosystem service function parameters. The output results can be used in operational scenarios such as regional ecological monitoring and ecological engineering effectiveness evaluation.

[0098] The advantages of this invention are: it fully utilizes multi-temporal remote sensing imagery, meteorological data, and ground-based sample plot observation data of the region to construct multi-dimensional input characteristics and constraints related to carbon sink, water balance, and windbreak and sand fixation mechanisms, adaptable to the ecological environment of arid and semi-arid regions. The results of the implementation examples show that the ecological parameters such as net primary productivity, water conservation capacity, and windbreak and sand fixation capacity generated by this method are highly consistent with known ecological functional zoning in terms of spatial distribution, and effectively avoid prediction results that violate ecological laws, such as water conservation capacity exceeding precipitation and sand fixation capacity decreasing with increasing vegetation cover. This implementation example verifies the practicality and superiority of this invention in ecological monitoring and ecological restoration effectiveness assessment in arid and semi-arid regions.

[0099] Example 3

[0100] like Figure 2 As shown, a mechanism-constrained deep learning inversion system for regional ecological monitoring parameters includes:

[0101] The acquisition module M1 is used to acquire basic data of the target area and construct multidimensional input features, output multidimensional input features to the inversion module M3, and provide environmental auxiliary data to the constraint module M2.

[0102] The constraint module M2 receives environmental auxiliary data provided by the acquisition module M1 to construct ecological mechanism constraints, and generates simulation samples based on the ecological mechanism constraints and the environmental auxiliary data, and outputs the generated simulation samples to the inversion module M3; the ecological mechanism constraints express the value range relationship, conservation relationship or monotonicity relationship between ecological parameters in the form of mathematical inequalities or equations.

[0103] The inversion module M3 receives the multi-dimensional input features output by the acquisition module M1 and the simulated samples output by the constraint module M2, respectively, establishes a deep learning inversion model, pre-trains the model using the simulated samples, adds the ecological mechanism constraint term to the loss function of the model, trains the model using the pre-trained model parameters, and outputs the trained model to the output module M4.

[0104] The output module M4 receives the trained model output by the inversion module M3, inputs the multi-dimensional input features output by the acquisition module M1 into the model, and outputs regional ecological monitoring parameters.

[0105] The specific working process is as follows: Module M1 preprocesses the basic data of the target area, constructs multi-dimensional input features, and separately transmits environmental auxiliary data to the constraint module M2; after receiving the environmental auxiliary data, the constraint module M2 constructs ecological mechanism constraints, generates vegetation type combinations, structural parameters, and environmental factors within a reasonable range, calls a lightweight ecological process model to calculate the corresponding ecological parameters, forms a simulated sample dataset, and outputs it to the inversion module M3; after receiving the multi-dimensional input features and the simulated sample dataset, the inversion module M3 first uses the simulated samples to pre-train the model, then adds a mechanism constraint term to the loss function and mixes real samples for fine-tuning training, and transmits the trained model parameters to the output module M4; the output module M4 uses the trained model to infer the multi-dimensional input features to be predicted, and outputs the spatial distribution results of the regional ecological monitoring parameters.

[0106] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A mechanism-constrained deep learning inversion method for regional ecological monitoring parameters, characterized in that, Includes the following steps: Acquire basic data for the target region and construct multidimensional input features; Ecological mechanism constraints are constructed, and simulation samples are generated based on the ecological mechanism constraints; the ecological mechanism constraints express the value range relationship, conservation relationship or monotonicity relationship between ecological parameters in the form of mathematical inequalities or equations; A deep learning inversion model is established, the model is pre-trained using the simulated samples, the ecological mechanism constraint term is added to the loss function of the model, and the pre-trained model parameters are used for training. The multidimensional input features are input into the trained deep learning inversion model, which outputs regional ecological monitoring parameters.

2. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 1, characterized in that, The ecological mechanism constraints include one or more of the following: carbon sink constraints, water balance constraints, and windbreak and sand fixation constraints.

3. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 2, characterized in that, The carbon sink constraint is that net primary productivity (NPP) is not less than the lower limit determined based on vegetation type, temperature, and precipitation, and not greater than the upper limit determined based on the light energy utilization model.

4. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 2, characterized in that, The water balance constraints are: the water conservation capacity is not less than zero, and not greater than the difference between precipitation and evapotranspiration, and not greater than the maximum water holding capacity of the soil.

5. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 2, characterized in that, The constraints for windbreak and sand fixation are as follows: the amount of windbreak and sand fixation remains monotonically constant with the increase of vegetation coverage and with the increase of wind speed, and does not exceed the maximum amount of sand fixation determined by wind speed and soil properties.

6. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 1, characterized in that, The generation of simulation samples based on the aforementioned ecological mechanism constraints specifically includes: Within the target area, vegetation type combinations are generated by random sampling based on land use maps; Within a preset reasonable range, structural parameters and environmental factors are randomly generated. The structural parameters include one or more of tree height, canopy closure, and leaf area index, and the environmental factors include precipitation, temperature, and wind speed. The corresponding ecological parameters are calculated using a lightweight ecological process model, and the input features are combined with the corresponding ecological parameters to form a simulated sample dataset.

7. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 6, characterized in that, The lightweight ecological process model is selected from one or more of the carbon module, water balance module, or windbreak and sand fixation module of the InVEST model.

8. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 1, characterized in that, The pre-training of the model using simulated samples specifically includes: in the pre-training stage, only the simulated samples are used, and the weight of the ecological mechanism constraint term in the loss function is zero or approaches zero; the step of adding the ecological mechanism constraint term to the loss function of the model for training is the fine-tuning stage, which specifically includes: using simulated samples and real samples in combination, and the weight of the ecological mechanism constraint term is greater than zero.

9. The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters according to claim 8, characterized in that, The ratio of simulated samples to real samples used in the fine-tuning phase ranges from 1:1 to 5:1, the weight of the ecological mechanism constraint term ranges from 0.01 to 0.5, and the weight of the ecological mechanism constraint term is dynamically adjusted with each training round.

10. A mechanism-constrained deep learning inversion system for regional ecological monitoring parameters, characterized in that, The mechanism-constrained deep learning inversion method for regional ecological monitoring parameters based on any one of claims 1 to 9 includes: The acquisition module obtains basic data of the target region and constructs multidimensional input features; The constraint module constructs ecological mechanism constraints and generates simulation samples based on the ecological mechanism constraints; the ecological mechanism constraints express the value range relationship, conservation relationship or monotonicity relationship between ecological parameters in the form of mathematical inequalities or equations; The inversion module establishes a deep learning inversion model, pre-trains the model using the simulated samples, adds the ecological mechanism constraint term to the loss function of the model, and trains the model using the pre-trained model parameters. The output module takes the multi-dimensional input features and inputs them into the trained deep learning inversion model to output regional ecological monitoring parameters.

Images