MoE-based ecological restoration evaluation index adaptive screening method
By constructing a hierarchical hybrid expert network and generative adversarial network based on MoE, the nonlinear and multi-scale problems of indicator selection in ecological restoration assessment are solved, and adaptive selection of ecological restoration assessment indicators is realized, thereby improving the scientificity and applicability of the assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-04-10
AI Technical Summary
Existing ecological restoration assessment methods rely on expert experience, lack a systematic screening process, struggle to handle the nonlinear and multi-scale coupling relationships of ecosystems, and have a weak ability to respond to complex ecological situations, resulting in highly subjective assessment results and limited applicability.
A hierarchical hybrid expert network based on MoE is constructed, combined with generative adversarial networks for data augmentation, and a multi-objective loss function is used for joint training to dynamically select ecological restoration assessment indicators. The adaptive selection of indicators is achieved through a dynamic gating network.
It has improved the intelligence level of ecological restoration assessment, reduced reliance on human experience, enhanced the stability and generalization ability of the model, adapted to the diverse needs of different ecological scenarios, and strengthened the scientificity and accuracy of the assessment system.
Smart Images
Figure CN121834283A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of ecological restoration assessment technology, and in particular to an adaptive screening method for ecological restoration assessment indicators based on MoE. Background Technology
[0002] Ecological restoration is a core means of improving the structure and function of damaged ecosystems, and its assessment process directly affects the adjustment of restoration strategies and the scientific nature of resource allocation. Traditional ecological restoration assessments often use methods such as expert scoring or the analytic hierarchy process (AHP) to establish evaluation indicator systems, covering multiple aspects such as biodiversity, soil quality, hydrological regulation capacity, and landscape pattern. However, these methods have significant limitations: first, they are highly dependent on indicators but lack a systematic screening process, easily introducing redundant or low-relevance indicators; second, they have weak responsiveness to complex ecological situations and struggle to reflect the differences between different restoration objects or regional characteristics; and third, the expert-driven approach inevitably introduces subjective bias, reducing the objectivity and transferability of the assessment system.
[0003] Some studies have attempted to optimize and screen ecological assessment indicators through statistical analysis or machine learning, such as using principal component analysis (PCA), correlation analysis, and cluster analysis to reduce the dimensionality or classify the indicator set. However, these methods are mostly based on linear assumptions or static data structures, making it difficult to handle the complex relationships of highly nonlinear and multi-scale coupling within ecosystems, and their generalization ability to new samples is insufficient. In addition, although some neural network models have nonlinear modeling capabilities, their training stability and interpretability are still insufficient in dealing with high-dimensional, low-sample scenarios.
[0004] In recent years, Mixture of Experts (MoE) models have demonstrated excellent performance in multi-task learning, feature selection, and subspace modeling due to their modular structure and dynamic routing capabilities. These models dynamically allocate subtasks to different expert modules through gating networks, effectively addressing the challenges of data heterogeneity and task diversity, and exhibiting good generalization performance and model compression capabilities. However, the application of MoE in ecological restoration assessment is still in its early stages, and a complete modeling paradigm for indicator selection scenarios has not yet been established. Furthermore, ecological data often exhibits spatiotemporal sparsity and imbalance, placing higher demands on the gating strategies and expert selection mechanisms of MoE models. Therefore, how to integrate the structural advantages of MoE and adapt it to the actual needs of ecological assessment indicator selection is a key challenge in improving the intelligent assessment level of ecological restoration effectiveness. Summary of the Invention
[0005] The purpose of this invention is to address the problems of existing ecological restoration assessment indicator systems being complex, relying on expert experience for indicator selection, and lacking intelligent dynamic screening mechanisms, which lead to highly subjective assessment results and limited applicability. This invention provides an adaptive screening method for ecological restoration assessment indicators based on MoE.
[0006] The above-mentioned objective of this application is achieved through the following technical solution: S1: Construct a dataset of original ecological restoration assessment indicators based on four categories of indicators: ecosystem structure, ecosystem quality, ecosystem services, and drivers of ecosystem change; S2: Use generative adversarial networks to augment the original ecological restoration assessment index dataset and construct a complete dataset; S3: Design a hierarchical hybrid expert network structure model; S4: Use a multi-objective loss function and a complete dataset to jointly train and optimize the hierarchical hybrid expert network; S5: Obtain new ecological restoration scenario features and input them into the trained hierarchical hybrid expert network structure model. Quickly output the importance ranking of each indicator in the scenario. Through normalization processing and threshold screening strategies, select the most representative and guiding key indicators and construct an indicator set to complete the adaptive screening of ecological restoration assessment indicators based on MoE.
[0007] Optionally, step S1 includes: S11: Collect on-site measurement data of the demonstration area and ecological restoration assessment index survey forms filled out by experts; the on-site measurement data and ecological restoration assessment index survey forms cover typical ecological restoration target area types, typical ecological problems, and restoration project types; the typical ecological restoration target area types include: grassland, wetland, and forest; the typical ecological problems include: soil erosion, vegetation degradation, and soil pollution; the restoration project types include: returning farmland to forest, wetland restoration, and habitat reconstruction; S12: Establish ecological restoration effectiveness indicators that match typical ecological restoration target area types, typical ecological problems, and restoration project types; S13: Based on ecosystem attributes, the indicators of ecological restoration effectiveness are divided into four categories, including: ecosystem structure, ecosystem quality, ecosystem services, and ecosystem change drivers. Each type of ecological restoration effectiveness indicator is further subdivided into primary and secondary indicators, and its applicable ecological scenarios and response characteristics are labeled to form a dataset of original ecological restoration assessment indicators with category labels. Optionally, step S2 includes: The discriminator is trained using training samples from the original ecological restoration assessment index dataset; the generator is then trained in reverse using the feedback gradient signal from the discriminator, resulting in a generator model capable of generating various ecological scenario index configurations. High-quality simulated samples are generated in batches using the trained generator and merged with the original training samples to build a complete dataset for subsequent training.
[0008] Optionally, step S3 includes: S31: Based on the classification characteristics of the four types of indicators, four sub-MoE networks are constructed respectively; each sub-MoE network adopts a hierarchical design, with the first layer of experts corresponding to the first-level indicators and the second layer of experts corresponding to the second-level indicators. S32: Input the ecological restoration scenario features into four sub-networks, and identify the key indicator categories and specific indicator items that should be focused on in the current scenario through the high-dimensional mapping relationship between the input information and the expert database; The characteristics of an ecological restoration scenario include: regional attributes, restoration measures, terrain type, and ecological parameters; specific indicators are either primary or secondary indicators. S33: Each sub-MoE network contains a gating network; each gating network dynamically selects a subset of expert networks based on the characteristics of the input ecological restoration scenario, and outputs the importance probability distribution of that type of indicator, thereby achieving adaptive selection of indicators.
[0009] Optionally, step S4 includes: S41: Train the four sub-MoE networks separately using the complete dataset. The goal is to minimize the difference between the importance of the prediction metrics and the manually labeled data, while optimizing the gating parameters and expert network weights. S42: Introduce cross-entropy loss, auxiliary load balancing loss, and multi-router loss into the loss function and perform joint optimization; S43: After training is complete, save the optimal model weights for the four types of hybrid expert networks.
[0010] Optionally, step S5 includes: During the inference phase, new ecological restoration scenario features are input into the trained model to calculate the importance weights of each indicator in the current scenario. By using normalization and threshold screening strategies, combined with importance weights, key indicators are selected from four categories of indicators and an indicator set is constructed.
[0011] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform an adaptive screening method for ecological restoration assessment indicators based on MoE.
[0012] A computer-readable storage medium storing instructions that, when executed, perform an adaptive screening method for MoE-based ecological restoration assessment indicators.
[0013] The beneficial effects of the technical solution provided in this application are: This invention introduces a generative adversarial network (GAN) data augmentation mechanism to effectively address the scarcity of ecological restoration assessment samples, significantly improving the stability and generalization ability of model training. It employs a hierarchical hybrid expert network structure combined with a dynamic gating network to achieve refined correlation modeling and dynamic weight allocation of multi-level, multi-channel indicators, enhancing the accuracy of indicator selection and adaptability to ecological scenarios. A MoE multi-objective loss joint optimization strategy is designed to avoid over-reliance on partial expert networks during training, enhancing the model's robustness and diverse feature representation. This invention achieves automatic recommendation and dynamic selection of ecological restoration assessment indicator systems, reducing reliance on human experience and improving the scientific rigor and intelligence of the ecological assessment system. Attached Figure Description
[0014] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a step diagram of an embodiment of this application; Figure 2 This is a schematic diagram illustrating the construction of the training dataset in an embodiment of this application; Figure 3 This is a network architecture diagram of the indicator selection model in the embodiments of this application; Figure 4 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation
[0015] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0016] The embodiments of this application provide an adaptive screening method for ecological restoration assessment indicators based on MoE.
[0017] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of an adaptive screening method for ecological restoration assessment indicators based on MoE, as described in an embodiment of this application, including: S1: Construct a dataset of original ecological restoration assessment indicators based on four categories of indicators: ecosystem structure, ecosystem quality, ecosystem services, and drivers of ecosystem change; Step S1 includes: S11: Collect on-site measurement data of the demonstration area and ecological restoration assessment index survey forms filled out by experts; the on-site measurement data and ecological restoration assessment index survey forms cover typical ecological restoration target area types, typical ecological problems, and restoration project types; the typical ecological restoration target area types include: grassland, wetland, and forest; the typical ecological problems include: soil erosion, vegetation degradation, and soil pollution; the restoration project types include: returning farmland to forest, wetland restoration, and habitat reconstruction; S12: Establish ecological restoration effectiveness indicators that match typical ecological restoration target area types, typical ecological problems, and restoration project types; S13: Based on ecosystem attributes, the indicators of ecological restoration effectiveness are divided into four categories, including: ecosystem structure, ecosystem quality, ecosystem services, and ecosystem change drivers. Each type of ecological restoration effectiveness indicator is further subdivided into primary and secondary indicators, and its applicable ecological scenarios and response characteristics are labeled to form a dataset of original ecological restoration assessment indicators with category labels. As one example, based on ecosystem attributes, the indicators are divided into four categories: ecosystem structure, ecosystem quality, ecosystem services, and drivers of ecosystem change. These are further subdivided into primary and secondary indicators, clarifying their applicable scope and response characteristics, and forming labeled training samples, such as... Figure 2 As shown.
[0018] S2: Use generative adversarial networks to augment the original ecological restoration assessment index dataset and construct a complete dataset; Step S2 includes: The discriminator is trained using training samples from the original ecological restoration assessment index dataset; the generator is then trained in reverse using the feedback gradient signal from the discriminator, resulting in a generator model capable of generating various ecological scenario index configurations. High-quality simulated samples are generated in batches using the trained generator and merged with the original training samples to build a complete dataset for subsequent training.
[0019] As one embodiment, to alleviate the problem of limited ecological restoration sample quantity, this invention employs Generative Adversarial Networks (GANs) for data augmentation. During training, the discriminator passes some text features to the generator in a hidden manner, maintaining structural consistency and semantic continuity during generation, thereby improving the generation quality of long texts. Through alternating training of the generator and discriminator, the generative model achieves stable convergence on the ecological assessment text generation task, ultimately obtaining a generator model capable of generating various ecological scenario indicator configurations. The trained generator is used to generate high-quality simulated samples in batches, which are then merged with the original demonstration area samples to construct a complete dataset for subsequent training.
[0020] S3: Design a hierarchical hybrid expert network structure model; Step S3 includes: S31: Based on the classification characteristics of the four categories of indicators, four sub-MoE networks are constructed respectively; each sub-MoE network consists of two layers: the first layer of experts corresponds to the first-level indicators and is also a MoE network; the second layer of experts corresponds to the second-level indicators and is also a MoE network; the hierarchical hybrid expert network structure model is represented as follows:
[0021]
[0022]
[0023] Where Y represents a hierarchical hybrid expert network, This represents the MoE network, a primary indicator. Represents the MoE network as a secondary indicator; 、 and Indicates the other three and Sub-MoE networks with identical structures; and These represent the number of experts in the first-level MoE and the second-level MoE, respectively. Indicates the summation index; This represents the exponent of the output score of the gating network of the i-th expert in the first layer of MoE; This represents the exponentiation of the output score of the t-th expert in the first-layer MoE gating network. This represents the exponent of the gating network output score of the j-th expert in the second-layer MoE; This represents the exponent of the gating network output score of the t-th expert in the second-layer MoE; This represents the j-th expert model in the second-layer MoE; S32: Input the ecological restoration scenario features into four sub-networks, and identify the key indicator categories and specific indicator items that should be focused on in the current scenario through the high-dimensional mapping relationship between the input information and the expert database; The characteristics of an ecological restoration scenario include: regional attributes, restoration measures, terrain type, and ecological parameters; specific indicators are either primary or secondary indicators. S33: Each sub-MoE network contains a gating network; each gating network dynamically selects a subset of the expert network based on the characteristics of the input ecological restoration scenario, and outputs the importance probability distribution of that type of indicator, realizing adaptive selection of indicators. The gating network is represented as follows:
[0024]
[0025]
[0026] in, This indicates that the hierarchical hybrid expert network is divided into four sub-MoE networks, each corresponding to a different gated network. and This represents the gated network corresponding to the MoE network, a primary index. , , and This represents the gated network corresponding to the second-level index MoE network; This represents the gated network for the i-th index in the first-level index MoE network; This represents the i-th expert model in the MoE network, which is a first-level index. Indicates the number of primary indicators; This represents the gated network corresponding to the MoE network, a primary index. This represents the weight matrix of the gated network; Feature embeddings representing the characteristics of ecological restoration scenarios; This represents the gated network corresponding to the second-level index MoE network.
[0027] As one implementation, four sub-MoE networks are constructed for four types of indicators. Each sub-network contains two expert layers: the first layer corresponds to the primary indicator, and the second layer corresponds to the secondary indicator. Each layer contains multiple expert sub-networks to enhance expressive power. By inputting features related to the ecological restoration scenario (such as regional attributes, restoration measures, ecological parameters, etc.), the model uses a gating network to dynamically activate relevant experts and outputs the probability distribution of the importance of the corresponding indicator in the current scenario, achieving adaptive selection of indicators. Figure 3 As shown.
[0028] S4: Use a multi-objective loss function and a complete dataset to jointly train and optimize the hierarchical hybrid expert network; Step S4 includes: S41: Train the four sub-MoE networks separately using the complete dataset. The goal is to minimize the difference between the importance of the prediction metrics and the manually labeled data, while optimizing the gating parameters and expert network weights. S42: Introduce cross-entropy loss, auxiliary load balancing loss, and multi-router loss into the loss function and perform joint optimization to achieve the total loss. It is represented as follows;
[0029]
[0030]
[0031]
[0032] in Represents cross-entropy loss, Indicates the true label, Predict the probability of the label for the model; This indicates the loss from auxiliary load balancing. This represents the weight assigned to the e-th expert by the gating network; This represents the routing loss of multiple parallel gated networks. This represents the original weights assigned to the e-th expert by the gating network; express?; express?; express?; express?.
[0033] As one example, to avoid the model's over-reliance on a certain part of the expert sub-network during training, which would cause other experts to remain in a "dormant" state for a long time and be unable to learn effective features, auxiliary load balancing loss and multi-router loss are introduced into the loss function to improve the activation balance of each expert sub-network.
[0034] S43: After training is complete, save the optimal model weights for the four types of hybrid expert networks.
[0035] As one example, after training, the optimal model weights are saved for dynamic indicator selection and intelligent recommendation in actual ecological restoration assessment.
[0036] S5: Obtain new ecological restoration scenario features and input them into the trained hierarchical hybrid expert network structure model. Quickly output the importance ranking of each indicator in the scenario. Through normalization processing and threshold screening strategies, select the most representative and guiding key indicators and construct an indicator set to complete the adaptive screening of ecological restoration assessment indicators based on MoE.
[0037] Step S5 includes: During the inference phase, new ecological restoration scenario features are input into the trained model to calculate the importance weights of each indicator in the current scenario. By employing normalization and threshold screening strategies, combined with importance weights, key indicators are selected from four categories of indicators to construct an indicator set. Unlike indicator screening methods that rely on expert knowledge and statistics, this application proposes an adaptive screening method for ecological restoration assessment indicators based on MoE (Modal of the Ecosystem) to address the problem of scarce ecological restoration samples by enhancing data through generative adversarial networks. Employing a hierarchical MoE structure and dynamic gating mechanism, it can model and screen the importance of multi-level indicators under different ecological scenarios. Simultaneously, by combining multi-objective loss functions (cross-entropy, load balancing, and routing constraints), it ensures the stability of model training and the balanced activation of expert sub-networks, improving robustness and interpretability. This technology not only reduces reliance on expert experience and manual statistical analysis but is also more suitable for achieving dynamic adaptation and accurate screening of indicators in complex and diverse ecological restoration scenarios.
[0038] This application also discloses an electronic device. (See reference...) Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.
[0039] The communication bus 502 is used to enable communication between these components.
[0040] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.
[0041] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).
[0042] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the aforementioned adaptive screening method for MoE-based ecological restoration assessment indicators.
[0043] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.
[0044] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.
Claims
1. A MoE-based ecological restoration evaluation index adaptive screening method, characterized in that, The method comprises the following steps: S1: Constructing an original ecological restoration evaluation index dataset based on four types of indexes, including ecosystem structure, ecosystem quality, ecosystem service and ecosystem change driving force; S2: Using a generative adversarial network to perform data enhancement on the original ecological restoration evaluation index dataset to construct a complete dataset; S3: Designing a hierarchical mixed expert network structure model; S4: Using a multi-objective loss function and the complete dataset to jointly train and optimize the hierarchical mixed expert network; S5: Obtaining new ecological restoration scene features and inputting them into the trained hierarchical mixed expert network structure model to quickly output the importance ranking of each index in the scene, and through normalization processing and threshold screening strategy, screening out the most representative key indicators and constructing an index set to complete the adaptive screening of ecological restoration evaluation indexes.
2. The MoE-based ecological restoration evaluation index adaptive screening method according to claim 1, characterized in that, Step S1 comprises: S11: Collecting field measurement data and expert-filled ecological restoration evaluation index survey tables in the demonstration area; the field measurement data and the ecological restoration evaluation index survey tables cover typical ecological restoration target area types, typical ecological problems and restoration engineering types; the typical ecological restoration target area types include grassland, wetland and forest; the typical ecological problems include water and soil loss, vegetation degradation and soil pollution; the restoration engineering types include returning farmland to forest, wetland restoration and habitat reconstruction; S12: Establishing ecological restoration effectiveness indexes matched with the typical ecological restoration target area types, typical ecological problems and restoration engineering types; S13: According to the properties of the ecosystem, the ecological restoration effectiveness indexes are divided into four categories, including ecosystem structure, ecosystem quality, ecosystem service and ecosystem change driving force; Each type of ecological restoration effectiveness index is subdivided into primary and secondary indexes, and is labeled with its applicable ecological scene and response characteristics to form an original ecological restoration evaluation index dataset with category labels.
3. The MoE-based ecological restoration evaluation index adaptive screening method according to claim 1, characterized in that, Step S2 comprises: Training a discriminator using training samples of the original ecological restoration evaluation index dataset; and using the feedback gradient signal of the discriminator to reversely train a generator to obtain a generator model capable of generating various ecological scene index configurations; Using the trained generator to batch generate high-quality simulation samples, and merging them with the original training samples to construct a complete dataset for subsequent training.
4. The MoE-based ecological restoration evaluation index adaptive screening method according to claim 2, characterized in that, Step S3 comprises: S31: According to the classification characteristics of the four types of indexes, four sub-MoE networks are constructed; each sub-MoE network adopts a hierarchical design, with the first layer of experts corresponding to primary indexes and the second layer of experts corresponding to secondary indexes; S32: Inputting ecological restoration scene features into the four sub-networks to identify the index categories and specific index items that should be focused on in the current scene through the high-dimensional mapping relationship between the input information and the expert library; The ecological restoration scene features include regional attributes, restoration measures, terrain types and ecological parameters; the specific index items are primary indexes or secondary indexes; S33: Each sub-MoE network includes a gating network; each gating network dynamically selects a subset of expert networks according to the input ecological restoration scene features, outputs the importance probability distribution of this type of index, and realizes adaptive screening of the indexes.
5. The MoE-based ecological restoration evaluation index adaptive screening method according to claim 4, characterized in that, Step S4 comprises: S41: Train the four sub-MoE networks respectively using the complete data set, the goal is to minimize the difference between the predicted indicator importance and the human-labeled label, while optimizing the gating parameters and expert network weights; S42: Introduce cross-entropy loss, auxiliary load balancing loss and multi-router loss into the loss function and jointly optimize them; S43: After training, save the optimal model weights of the four types of hybrid expert networks respectively.
6. The MoE-based ecological restoration evaluation index adaptive screening method according to claim 1, characterized in that, Step S5 includes: In the inference stage, input the new ecological restoration scene features into the trained model to calculate the importance weight of each indicator in the current scene; Through normalization processing and threshold screening strategy, combined with the importance weight, select the key indicators from the four types of indicators and construct the indicator set.
7. An electronic device, comprising: The electronic device includes a processor, a memory, a user interface, and a network interface. The memory is configured to store instructions. The user interface and the network interface are configured to communicate with other devices. The processor is configured to execute the instructions stored in the memory to cause the electronic device to perform the method of any one of claims 1-6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions, when the instructions are executed by a computer, the method of any one of claims 1-6 is executed.