Biodiversity monitoring method and system and storage medium

Abstract

The invention relates to the technical field of biodiversity monitoring, in particular to a biodiversity monitoring method and system and a storage medium. Biodiversity multi-source data of a target area are collected in real time through a distributed sensor network and a mobile terminal device, the multi-source data comprise species image data, environmental factor time sequence data, voiceprint feature data and geographic space trajectory data, and encryption transmission and format standardization preprocessing are carried out on the data. Through a dynamic weight distribution mechanism driven by an environmental factor influence degree, a self-adaptive mapping relation between the sensor data entropy and the species ecological niche width is constructed, the problem of feature space mismatching of a traditional method in a non-stationary environment is solved, and the ecological interpretation of a species-environment joint feature vector is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of biodiversity monitoring, and in particular to a biodiversity monitoring method, system and storage medium. Background Art

[0002] Current biodiversity monitoring technologies mainly rely on the combination of static sensor networks and manual surveys, and are insufficient in adapting to spatio-temporal heterogeneity in complex habitats. Due to the non-linear and multi-scale characteristics of the dynamic coupling relationship between environmental factors and species distribution, it is difficult to accurately characterize the real-time impact of environmental mutations (such as extreme climate events, human interference pulses) on species habitat selection when traditional methods use fixed weights to fuse multi-source data, resulting in a semantic gap in the data representation level of classification models. Especially in scenarios with drastic changes in light conditions or overlapping sound print characteristics, existing convolutional neural networks are vulnerable to interference from abnormal data, and the false detection rate increases significantly. In addition, existing monitoring systems lack spatio-temporal correlation modeling of species migration trajectories and ecological security thresholds, and cannot effectively predict the diffusion paths of invasive species and the trend of habitat fragmentation, resulting in protection decisions lagging behind the actual degradation rate of the ecosystem. Summary of the Invention

[0003] In order to overcome at least one defect existing in the prior art, the present disclosure proposes a biodiversity monitoring method, system and storage medium, aiming to construct a multi-source data fusion mechanism with environmental adaptability, and simultaneously improve the adaptability of species recognition and the timeliness of ecological threat assessment.

[0004] To achieve the above object, the technical solutions disclosed by the present invention are as follows: According to one aspect of the present disclosure, a biodiversity monitoring method is provided, and the steps of the monitoring method include: Real-time collect multi-source data of biodiversity in a target area through a distributed sensor network and a mobile terminal device, the multi-source data includes species image data, environmental factor time series data, sound print feature data and geospatial trajectory data, and perform encrypted transmission and format standardization preprocessing on the data; Based on a dynamic weight allocation algorithm, perform feature fusion on the multi-source data to generate a species-environment joint feature vector, and the dynamic weight allocation algorithm adjusts the weight ratio of sensor data in real time according to the influence degree of the environmental factor time series data on species distribution; Use a deep convolutional adversarial network to construct a classification model to perform species classification and recognition on the species-environment joint feature vector to obtain a classification result, and the classification model is pre-trained through transfer learning and embedded with an abnormal data filtering module to eliminate false detection data caused by abnormal light, motion blur or noise interference; Construct a spatio-temporal association map based on the classification results, count the distribution density of species with different protection levels, the population migration trajectories, and the hotspots of human interference, and dynamically update them to the biodiversity database; Trigger a multi-level early warning mechanism according to the preset ecological security threshold, and generate a visual monitoring report and a regulation strategy. The visual monitoring report includes the species endangerment index, the invasion species diffusion model, and the ecological restoration plan.

[0005] Furthermore, the steps of the dynamic weight allocation algorithm include: Define the environmental factor influence function: , where T is the temperature change rate, P is the precipitation intensity, S i is the sensitivity coefficient of the i-th type of species, n is the total number of species types, D i is the species distribution dispersion, and α, β, γ are trainable parameters; Calculate the local entropy value of each sensor data through a sliding time window, and dynamically allocate weights according to the entropy value ratio: , where λ is the environmental response coefficient, E k is the environmental factor influence degree of the k-th sensor; Use the gradient descent method to optimize the allocation result of the weight ratio until the within-class distance variance of the species-environment joint feature vector is less than the preset threshold.

[0006] Furthermore, the optimization steps of the deep convolutional adversarial network include: Construct a two-channel feature extractor to process the texture features of the image data and the frequency domain features of the voiceprint data respectively, and fuse the features through a cross-attention mechanism; Design an adversarial loss function: , where G is the generator, D is the discriminator, E is the expected value, x is the real data, and η is the reconstruction loss weight; Embed an interpretability module in the output layer of the classifier to generate a species recognition confidence heat map and an analysis report on the reasons for misdetection.

[0007] Furthermore, the steps of constructing a spatio-temporal association map based on the classification results include: Map the species distribution data to the map nodes, and the node attributes include the protection level, the population quantity, and the genetic diversity index; Calculate the association strength between nodes, and the calculation expression is: , where C ij is the species symbiosis coefficient, d ij is the geographical distance, Δt is the time interval, and σ, τ are attenuation factors; Reduce the high-dimensional node features to a low-dimensional space through graph embedding technology to generate a visualized ecological security situation map.

[0008] Furthermore, the steps of the pre-training of transfer learning include: Construct a hierarchical transfer strategy, freeze the parameters of the first three convolutional kernels of the pre-trained model, and only fine-tune the target area species features for the fully connected layer; Introduce an adversarial domain adaptation mechanism, insert a gradient reversal layer between the feature extractor and the classifier to force the network to learn the species essential features independent of geographical location; Design a diversity-enhanced sample library, use neural style transfer technology to generate virtual training samples with different lighting conditions and vegetation coverage, and expand the generalization ability of the pre-trained model in the rainy season and dry season; Implement a dynamic curriculum learning strategy, unlock the classifier neurons in stages according to the order of species appearance frequency from high to low; Deploy a model drift detection module. When the F1-score of the validation set drops by more than the preset value for continuously set days, trigger the incremental learning process and update the model parameters of the edge computing node.

[0009] Furthermore, the steps of the monitoring method further include: Perform multi-dimensional cross-validation of the classification results with an authoritative species database, including morphological feature matching degree, habitat suitability index, and breeding cycle matching degree; Construct a credibility scoring model: , where Ψ represents the credibility score, ω k is the weight of the kth verification dimension, D k is the database feature, O k is the observed feature, P is the classification probability distribution, and geo match is the geographical distribution matching degree; When the credibility score Ψ is less than the preset score, initiate an expert review process, and send the multi-source data of the suspected species to the verification terminals of multiple domain experts; Dynamically adjust the classification model threshold parameters according to the expert feedback results, and update the species distribution baseline data of the biodiversity database.

[0010] Furthermore, the steps of setting the ecological security threshold include: Construct a threshold dynamic prediction framework based on the LSTM-GARCH hybrid model, and input historical species fluctuation data, climate change trend, and human activity intensity indicators; Define the threshold adaptive adjustment rule: , where θ t is the current threshold, θbase is the baseline threshold, N t is the current population size, ΔH t is the change rate of human disturbance intensity, and a and b are adjustment coefficients; Implement threshold sensitivity analysis, and calculate the balance points of false alarm rates and missed alarm rates for each warning level within the threshold preset error fluctuation range through Monte Carlo simulation; When the same geographical grid continuously triggers a set number of warnings within a preset time, start the threshold emergency re-evaluation program, and call high-resolution satellite remote sensing data for re-evaluation of habitat integrity.

[0011] Furthermore, the monitoring method further includes ecological modeling, and the steps of the ecological modeling include: Establish an invasive species diffusion model, and comprehensively predict the diffusion path based on the wind speed vector, soil moisture gradient, and the distribution of competing species; Construct a multi-objective optimization algorithm for vegetation restoration, balance ecological benefits, engineering costs, and species diversity gain, and output the Pareto optimal solution set; Design a three-dimensional planning tool for ecological corridors, and automatically generate the topological map of the corridor path according to the hydrogeological permeability coefficient and animal migration law.

[0012] According to another aspect of the present disclosure, there is provided a biodiversity monitoring system for implementing the biodiversity monitoring method as described above, and the monitoring system includes: A multi-source data acquisition module, including a distributed sensor array deployed in the monitoring area, a mobile terminal equipped with high-precision GPS, and a voiceprint recorder, for real-time acquisition of species images, environmental parameters, geographical trajectories, and bioacoustic data; A dynamic weight fusion module, provided with an FPGA accelerator and a DDR4 memory matrix, performs calculation of the influence degree of environmental factors and weight assignment of the entropy value of sensor data, and is used to output a species-environment joint feature vector; An identification engine, integrating a dual-channel deep convolutional adversarial network and an interpretability analysis unit, and equipped with a CUDA parallel computing card to achieve real-time species classification; A spatio-temporal map construction module, implemented based on the graph database Neo4j, includes an ant colony optimization processor and a three-dimensional visualization engine, and dynamically generates a species distribution heat map and an ecological association network; A multi-level warning decision-making module, having an edge computing gateway, to achieve dynamic optimization of warning thresholds, generation of emergency strategies, and distribution of instructions to multiple departments; The multi-source data acquisition module is connected to the dynamic weight fusion module through the LoRaWAN protocol, the output end of the identification engine is connected to the streaming computing interface of the spatio-temporal map construction module, and the multi-level warning decision-making module performs data interaction with an external law enforcement system through the REST API.

[0013] According to another aspect of the present disclosure, there is provided a computer-readable storage medium storing a computer program, which when executed by a processor implements the biodiversity monitoring method as described above.

[0014] The beneficial effects of the present invention are as follows: The present invention constructs an adaptive mapping relationship between the entropy of sensor data and the niche breadth of species through a dynamic weight allocation mechanism driven by the influence degree of environmental factors, overcomes the problem of feature space mismatch of traditional methods in non-stationary environments, and improves the ecological interpretability of the species-environment joint feature vector.

[0015] Furthermore, the deep convolutional adversarial network integrates the cross-attention mechanism and the interpretability module, while strengthening the discriminant boundary of multi-modal features, realizes the collaborative filtering of abnormal data in the joint frequency-spatial domain, and effectively suppresses semantic-level misdetection in complex habitats.

[0016] Furthermore, the spatio-temporal correlation map reveals the multi-scale coupling law of species migration paths and human disturbance hotspots through node embedding technology regulated by attenuation factors, provides high-fidelity spatio-temporal constraints for the path prediction of invasion diffusion models, and combines the threshold dynamic prediction framework and the credibility closed-loop verification mechanism to realize the paradigm shift of ecological security situation from passive monitoring to active regulation, and enhances the targeted intervention ability of ecological restoration plans for the degradation of habitat heterogeneity.

[0017] The above description is only an overview of the technical solution of the present invention. In order to understand the technical means of the present invention more clearly and implement it according to the content of the specification, the following describes the preferred embodiments of the present invention in detail with reference to the accompanying drawings. Description of the Drawings

[0018] Figure 1 is a flowchart of the biodiversity monitoring method in an embodiment of the present invention; Figure 2 is a schematic diagram of biodiversity fluctuation data in an embodiment of the present invention; Figure 3 is a schematic diagram of an invasive species diffusion model in an embodiment of the present invention; Figure 4 is a schematic diagram of constructing multi-objective optimization of vegetation restoration in an embodiment of the present invention; Figure 5 is a schematic topological diagram of an ecological corridor path in an embodiment of the present invention; Figure 6 is a schematic diagram of feature fusion of multi-source data in an embodiment of the present invention; Figure 7 is a schematic diagram of a spatio-temporal correlation network in an embodiment of the present invention; Figure 8This is the confidence heat map in an embodiment of the present invention. Detailed implementation manners

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] The term "including" and any of its variations in the description and claims of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices. In addition, the use of "and / or" in the description and claims means at least one of the connected objects. For example, A and / or B means including three cases: A alone, B alone, and both A and B existing.

[0021] In the embodiments of the present invention, words such as "exemplary" or "for example" are used to represent examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0022] The present invention provides the following preferred embodiments: Embodiment 1; To solve the problems of environmental factor dynamic response lag and multi-source data semantic gap in existing biodiversity monitoring, this embodiment proposes a biodiversity monitoring method, as Figure 1 shown, the method includes the following steps: S100. Real-time collect multi-source data of biodiversity in the target area through a distributed sensor network and a mobile terminal device. The multi-source data includes species image data, environmental factor time-series data, voiceprint feature data, and geospatial trajectory data, and perform encrypted transmission and format standardization preprocessing on the data.

[0023] S200. Perform feature fusion on the multi-source data based on the dynamic weight allocation algorithm to generate a species-environment joint feature vector. The dynamic weight allocation algorithm adjusts the weight ratio of the sensor data in real time according to the influence degree of the environmental factor time-series data on the species distribution.

[0024] S300. A classification model is constructed using a deep convolutional adversarial network to classify and identify species in the species-environment joint feature vector, and a classification result is obtained. The classification model is pre-trained through transfer learning and embedded with an abnormal data filtering module to eliminate misdetected data caused by abnormal lighting, motion blur, or noise interference.

[0025] S400. Based on the classification result, a spatio-temporal association map is constructed to statistically analyze the distribution density, population migration trajectory, and human interference hotspots of species with different protection levels, and dynamically update them to the biodiversity database.

[0026] S500. Trigger a multi-level early warning mechanism according to the preset ecological security threshold, generate a visual monitoring report and a regulation strategy. The visual monitoring report includes the species endangerment index, the invasive species spread model, and the ecological restoration plan.

[0027] Specifically, multi-source data on biodiversity in the target area are collected in real time through a distributed sensor network and mobile terminal devices, including species image data, environmental factor time-series data, voiceprint feature data, and geospatial trajectory data, and these data are encrypted and transmitted and preprocessed for format standardization.

[0028] Furthermore, a comprehensive coverage of the target area is achieved by deploying a high-precision distributed sensor array and mobile terminal devices equipped with GPS. The sensor array includes high-definition cameras, temperature and humidity sensors, voiceprint recorders, etc., which can capture different types of biodiversity data. The mobile terminal devices are used to assist in manual surveys to ensure the diversity and accuracy of the data. During the data collection process, the AES-256 encryption algorithm is used to encrypt and transmit the data to protect the security of the data. At the same time, the data is preprocessed for format standardization before transmission to ensure the consistency and efficiency of subsequent processing.

[0029] Furthermore, based on the dynamic weight allocation algorithm, feature fusion of multi-source data is performed to generate a species-environment joint feature vector, as Figure 6 shown. This algorithm adjusts the weight ratio of sensor data in real time according to the influence degree of environmental factor time-series data on species distribution. Specifically, by analyzing environmental factors such as temperature change rate, precipitation intensity, species sensitivity coefficient, and species distribution dispersion, the weights of various sensor data are dynamically adjusted. For example, when the temperature changes rapidly or the precipitation intensity increases significantly in a certain area, the data weights of the relevant sensors will be increased accordingly. This dynamic weight allocation mechanism can effectively cope with the impact of environmental mutations on species distribution and improve the accuracy of feature fusion.

[0030] Furthermore, a classification model is constructed using a deep convolutional adversarial network to classify and identify species from the species-environment joint feature vectors. The classification model is pre-trained through transfer learning and embedded with an abnormal data filtering module to eliminate misdetected data caused by abnormal lighting, motion blur, or noise interference. Specifically, the classification model uses a dual-channel feature extractor to process the texture features of image data and the frequency-domain features of voiceprint data, and fuses these features through a cross-attention mechanism. In addition, an interpretability module is embedded in the output layer of the classifier to generate a species recognition confidence heatmap and an analysis report on the reasons for misdetection, such as Figure 8 shown. The confidence heatmap shows the confidence level of the model for each classification result, while the analysis report on the reasons for misdetection details the factors that may cause misdetection, such as poor lighting conditions, excessive background noise, etc.

[0031] Furthermore, a spatio-temporal association map is constructed based on the classification results, as Figure 7 shown, to statistically analyze the distribution density, population migration trajectories, and hotspots of human interference of species with different protection levels, and dynamically update them to the biodiversity database. Specifically, species distribution data is mapped to map nodes, and the node attributes include protection level, population size, and genetic diversity index. By calculating the association strength between nodes, the symbiotic relationships between species and their relationships with geographical distance and time interval are revealed. This spatio-temporal association map can reveal the multi-scale coupling law of species migration paths and hotspots of human interference, providing a scientific basis for ecological protection decision-making.

[0032] Furthermore, a multi-level warning mechanism is triggered according to a preset ecological security threshold to generate a visual monitoring report and a regulation strategy. As Figure 2 shown, the ecological security threshold is dynamically adjusted based on historical species fluctuation data, climate change trends, and human activity intensity indicators. When the number of species in a certain area shows abnormal fluctuations or the intensity of human interference exceeds the preset threshold, the system will automatically trigger a multi-level warning mechanism to generate a visual monitoring report containing the species endangerment index, the invasion species diffusion model (as Figure 3 shown), and an ecological restoration plan. These reports not only provide detailed information on the current ecological situation but also propose specific regulation strategies to help relevant departments take timely measures to prevent further degradation of the ecosystem.

[0033] The benefits of this embodiment are as follows: Through the coupled design of the dynamic weight allocation algorithm and the environmental factor influence function, the ecologically driven adjustment of sensor data weights is achieved, overcoming the feature drift problem of traditional methods under extreme weather conditions such as heavy rain. The dual-channel architecture and frequency-domain to spatial-domain joint filtering mechanism of the deep convolutional adversarial network effectively suppress cross-modal misdetection caused by acoustic reverberation and motion blur. The attenuation factor regulation strategy of the spatio-temporal correlation map accurately depicts the spatio-temporal coupling law of human interference hotspots and species migration trajectories, providing a high-resolution data basis for the dynamic calibration of ecological safety thresholds. Through the synergistic effect of the above technical solutions, the adaptability and decision-making support capabilities of the monitoring system in complex habitats are improved.

[0034] Embodiment 2; To solve the problem of insufficient adaptability to spatio-temporal heterogeneity of existing biodiversity monitoring technologies in complex habitats, this embodiment further optimizes the dynamic weight allocation algorithm. Specifically, by defining the environmental factor influence function and dynamically adjusting the weight ratio of sensor data, the accuracy of feature fusion is improved.

[0035] Furthermore, define the environmental factor influence function: , where T is the temperature change rate, representing the temperature change within a certain time period; P is the precipitation intensity, representing the precipitation within a certain time period; S i is the sensitivity coefficient of the i-th species, representing the sensitivity of this species to environmental changes; n is the total number of species, D i is the species distribution dispersion, representing the spatial distribution range of this species; α, β, γ are trainable parameters, which are optimized through the training process. It should be understood that these parameters together reflect the influence degree of environmental factors on species distribution, thus providing a basis for subsequent weight allocation.

[0036] Furthermore, calculate the local entropy value of each sensor data through a sliding time window, and dynamically allocate weights according to the entropy value ratio. Specifically, use a sliding time window, such as 30 minutes or 1 hour, to calculate the local entropy value of each sensor data, reflecting the uncertainty of the data. Then, dynamically allocate weights according to the entropy value ratio: , where λ is the environmental response coefficient, representing the influence degree of the environmental factor influence on weight allocation; E k is the environmental factor influence degree of the k-th sensor; E j represents the environmental factor influence degree of the j-th sensor, where j traverses all sensors from 1 to m, so m represents the total number of sensors. This dynamic weight allocation mechanism can effectively cope with the impact of environmental mutations on species distribution and improve the accuracy of feature fusion.

[0037] Through the method of this embodiment, the weight allocation formula can effectively reflect the importance of each sensor under the current environmental conditions and ensure that the calculation of weights is based on the overall situation of all sensors, thereby improving the accuracy of feature fusion.

[0038] Furthermore, the gradient descent method is used to optimize the allocation result of the weight ratio until the within-class distance variance of the species-environment joint feature vector is less than a preset threshold. Specifically, the weight ratio is continuously adjusted by the gradient descent method, so that the generated species-environment joint feature vector has a smaller variance within the class, that is, the difference between samples of the same class is smaller. This helps to improve the accuracy of the classification model. It can be understood that this optimization method ensures the effectiveness and consistency of the feature vector and improves the performance of the overall system.

[0039] The benefit of this embodiment is that through the dynamic weight allocation algorithm, the weight ratio of sensor data can be adjusted in real time, effectively coping with the impact of environmental mutations on species distribution, improving the accuracy of feature fusion, and providing a high-quality data basis for subsequent classification and recognition.

[0040] Embodiment 3; To solve the problem of insufficient adaptability to spatio-temporal heterogeneity of existing biodiversity monitoring technologies in complex habitats, this embodiment further optimizes the construction and optimization steps of the deep convolutional adversarial network. Specifically, through a dual-channel feature extractor, a cross-attention mechanism, and an interpretability module, the discriminative ability of multi-modal features and the interpretability of the model are improved.

[0041] In this embodiment, a dual-channel feature extractor is first constructed to process the texture features of image data and the frequency domain features of voiceprint data respectively. Specifically, the image data extracts texture features through a convolutional neural network (CNN), and the voiceprint data extracts frequency domain features through the short-time Fourier transform (STFT). These two types of features respectively reflect the visual and auditory information of species. The two types of features are fused through the cross-attention mechanism to enhance the discriminative ability of the model for multi-modal data. The cross-attention mechanism dynamically adjusts the importance of features by calculating the correlation between different features, so as to better capture the comprehensive features of species.

[0042] Furthermore, an adversarial loss function is designed: , where G is the generator responsible for generating realistic data samples; D is the discriminator responsible for distinguishing real data from generated data; η is the reconstruction loss weight to balance the losses of the generator and the discriminator; E is the expected value. In this embodiment, E[logD(G(x))] represents the expected value of the logarithm of the output of the generated data G(x) through the discriminator D, while E[log(1−D(x))] represents the expected value of the logarithm of the output of the real data x through the discriminator D. Through adversarial training, the generator and the discriminator play against each other, gradually improving the authenticity of the generated data and the discrimination ability of the discriminator. It can be understood that this adversarial training mechanism enhances the model's ability to distinguish abnormal data and improves the accuracy of classification.

[0043] Furthermore, an interpretability module is embedded in the output layer of the classifier to generate a confidence heatmap for species recognition and an analysis report on the reasons for misdetection. The confidence heatmap shows the confidence level of the model for each classification result, helping users understand the decision-making process of the model. The analysis report on the reasons for misdetection details the factors that may lead to misdetection, such as poor lighting conditions, excessive background noise, etc. This interpretability module not only enhances the transparency of the model but also improves the user's trust. It can be understood that through this module, users can better understand the decision-making logic of the model, timely discover and correct potential problems.

[0044] The benefits of this embodiment are that through the dual-channel feature extractor, cross-attention mechanism, and interpretability module, the discrimination ability of multi-modal features and the interpretability of the model are improved, providing more reliable technical support for biodiversity monitoring.

[0045] Embodiment Four; to address the problem of insufficient adaptability to spatio-temporal heterogeneity of existing biodiversity monitoring technologies in complex habitats, this embodiment further optimizes the steps of constructing a spatio-temporal association map based on classification results. Specifically, by mapping species distribution data into map nodes and calculating the association strength between nodes, a visualized ecological security situation map is generated.

[0046] Furthermore, the species distribution data is mapped into map nodes, and the node attributes include protection level, population quantity, and genetic diversity index. Specifically, each species distribution point is mapped into a map node, and the attributes of the node include the protection level of the species, such as endangered, vulnerable, least concern, population quantity, and genetic diversity index. These attributes reflect the ecological importance and survival status of the species. In this way, the distribution characteristics of the species can be comprehensively described, providing basic data for subsequent analysis.

[0047] Furthermore, calculate the association strength between nodes, and the calculation expression is: , where C ijis the species symbiosis coefficient, representing the symbiotic relationship between species i and species j; d ij is the geographical distance, representing the spatial distance between species i and species j; Δt is the time interval, representing the observation time difference between species i and species j; σ and τ are attenuation factors, used to control the attenuation degree of the association strength between regulatory nodes with time and distance. This calculation method of association strength can reveal the spatio-temporal coupling relationship between species and provide a scientific basis for ecological protection decision-making.

[0048] Furthermore, the high-dimensional node features are reduced to a low-dimensional space through graph embedding technology to generate a visualized ecological security situation map. Specifically, graph embedding technology (such as Node2Vec or GraphSAGE) is used to reduce the high-dimensional node features to a low-dimensional space for easy visualization. The generated ecological security situation map shows the hotspots of species distribution, migration paths, and human interference hotspots, providing intuitive ecological security situation information for relevant departments. It can be understood that this visualization method helps to quickly identify ecological threats and formulate effective protection measures.

[0049] Example Five: To solve the problem of insufficient spatio-temporal heterogeneity adaptation ability of existing biodiversity monitoring technologies in complex habitats, this example further optimizes the steps of transfer learning pre-training. Specifically, through a hierarchical transfer strategy, an adversarial domain adaptation mechanism, a diversity-enhanced sample library, and a dynamic curriculum learning strategy, the generalization ability of the model in the target area is improved.

[0050] Furthermore, a hierarchical transfer strategy is constructed to freeze the parameters of the first three convolutional kernels of the pre-trained model and only fine-tune the species features in the target area for the fully connected layer. Specifically, the pre-trained model is trained on a large-scale dataset to obtain good feature extraction ability. During the fine-tuning process in the target area, the parameters of the first three convolutional kernels are frozen, and only the fully connected layer is fine-tuned to adapt to the species features in the target area. This hierarchical transfer strategy can improve the recognition accuracy of the model in a specific area while maintaining its generalization ability.

[0051] Furthermore, an adversarial domain adaptation mechanism is introduced, and a gradient reversal layer is inserted between the feature extractor and the classifier to force the network to learn the essential species features independent of geographical location. Specifically, the gradient reversal layer makes the features learned by the feature extractor consistent among different regions through backpropagation gradients. This mechanism can reduce the feature drift caused by geographical location differences. It can be understood that the adversarial domain adaptation mechanism helps the model maintain stable performance in different environments.

[0052] Furthermore, a diversity-enhanced sample library is designed. Using neural style transfer technology, virtual training samples with different lighting conditions and vegetation coverage are generated to expand the generalization ability of the pre-trained model in the rainy and dry seasons. Specifically, through neural style transfer technology, virtual samples with various lighting conditions and vegetation coverage are generated to enrich the training dataset. This diversity-enhanced sample library can simulate species characteristics under different environmental conditions and improve the generalization ability of the model in different seasons and environments. It can be understood that diverse training samples help the model better adapt to complex environments in practical applications.

[0053] Furthermore, a dynamic curriculum learning strategy is implemented. The classifier neurons are unlocked in stages according to the frequency of species appearance from high to low. Specifically, in the initial stage, only the species with higher appearance frequencies are trained, and the neurons of other species are gradually unlocked. This strategy can gradually increase the complexity of the model, improve the training efficiency and stability. It can be understood that the dynamic curriculum learning strategy helps the model gradually learn and adapt to the characteristics of different species.

[0054] Furthermore, a model drift detection module is deployed. When the F1-score of the validation set drops by more than a preset value for a continuously set number of days, an incremental learning process is triggered and the model parameters of the edge computing node are updated. Specifically, by regularly evaluating the F1-score of the validation set, the performance change of the model is monitored. When the performance drops by more than the preset threshold, the incremental learning process is started to update the model parameters to adapt to environmental changes. It can be understood that the model drift detection module can timely detect and correct the problem of model performance decline and maintain the long-term effectiveness of the model.

[0055] Example Six: To solve the credibility problem of classification results in biodiversity monitoring, this example further optimizes the verification and adjustment mechanism of classification results. Specifically, the classification results are cross-verified with an authoritative species database in multiple dimensions, including morphological feature matching degree, habitat suitability index, and reproductive cycle matching degree.

[0056] Furthermore, a credibility scoring model is constructed: , where Ψ represents the credibility score; ω k is the weight of the k-th verification dimension, used to balance the importance of different verification dimensions; D k is the feature value in the database, representing the species characteristics recorded in the authoritative database; O k is the observed feature value, representing the species characteristics in the actual monitoring data; entropy(P) is the entropy value of the classification probability distribution, reflecting the uncertainty of the classification result; geo match is the geographical distribution matching degree, representing the consistency of the observed species and the geographical distribution recorded in the database.

[0057] It should be understood that sin(D k ,O k ) is used to calculate the similarity between database features and observed features, and its value range is in [−1,1]. When D k and O k are exactly the same, sin(D k ,O k ) = 1; when they are completely inconsistent, sin(D k ,O k ) = −1. Through this similarity calculation method, the feature matching degree of different dimensions can be quantified.

[0058] Furthermore, when the credibility score Ψ is less than the preset score, an expert review process is initiated, and the multi-source data of the suspected species is sent to the verification terminals of multiple domain experts. The experts can review these data through a dedicated verification platform and provide feedback. According to the expert feedback results, the threshold parameters of the classification model are dynamically adjusted to improve the classification accuracy. At the same time, the species distribution baseline data in the biodiversity database is updated to ensure the timeliness and accuracy of the database.

[0059] Furthermore, in order to improve the verification efficiency and accuracy, machine learning algorithms can also be introduced to automatically identify and mark high-risk classification results and give priority to sending them to experts for review. In addition, by regularly updating the expert database and the database, the reliability of the verification process is ensured.

[0060] The benefit of this embodiment is that through multi-dimensional cross-verification and the credibility scoring model, the accuracy and credibility of the classification results can be effectively improved. At the same time, combined with the expert review mechanism, it ensures that the classification results are professionally verified, further improving the usability of the monitoring system.

[0061] Embodiment Seven; to solve the problem of setting the ecological security threshold, this embodiment further refines the dynamic prediction and adaptive adjustment mechanism of the ecological security threshold. Specifically, a threshold dynamic prediction framework based on the LSTM-GARCH hybrid model is constructed, and historical species fluctuation data, climate change trends, and human activity intensity indicators are input.

[0062] Furthermore, define the threshold adaptive adjustment rule: , where θ t is the current threshold; θ base is the baseline threshold, representing the initially set security threshold; N t is the current population quantity; ΔH tis the change rate of human interference intensity; a and b are adjustment coefficients used to control the sensitivity of threshold adjustment. In this way, the ecological security threshold can be dynamically adjusted according to the changes in the current population size and the impact of human activities.

[0063] Furthermore, perform threshold sensitivity analysis, and calculate the balance points of false alarm rates and missed alarm rates for each warning level within the error fluctuation range of threshold preset through Monte Carlo simulation. Monte Carlo simulation evaluates the warning effects under different threshold settings through multiple random samplings, so as to find the best balance points of false alarm rates and missed alarm rates.

[0064] Furthermore, when the same geographical grid continuously triggers a set number of warnings within a preset time, start the threshold emergency re-evaluation program, and call high-resolution satellite remote sensing data for re-evaluation of habitat integrity. High-resolution satellite remote sensing data can provide detailed habitat information, helping to more accurately evaluate the health status of the ecosystem, so as to adjust the threshold in a timely manner to prevent false alarms or missed alarms.

[0065] It should be understood that the LSTM-GARCH hybrid model combines the time series prediction ability of the long short-term memory network (LSTM) and the volatility modeling ability of the generalized autoregressive conditional heteroskedasticity model (GARCH), and can more accurately capture the trends of species fluctuations and environmental changes. It can be understood that in this embodiment, by introducing the human activity intensity index, the impact of human activities on the ecosystem can be better reflected, thereby improving the scientificity and rationality of threshold setting.

[0066] Furthermore, in order to improve the real-time performance and accuracy of threshold adjustment, edge computing technology can be introduced to deploy some computing tasks on edge devices close to the data source, reduce data transmission delay, and improve the response speed. At the same time, centrally manage data and models through the cloud platform to achieve global optimization and collaborative scheduling.

[0067] Through this embodiment, the ecological security threshold can be dynamically adjusted, the accuracy and reliability of the warning system can be improved, and the impacts of environmental changes and human activities on the ecosystem can be effectively addressed.

[0068] Embodiment Eight; To solve the problems of invasive species spread and vegetation restoration, this embodiment further optimizes the ecological modeling method. Specifically, establish an invasive species spread model to predict the spread path by comprehensively considering the wind speed vector, soil moisture gradient, and the distribution of competing species.

[0069] Furthermore, as Figure 4As shown in the figure, a multi-objective optimization algorithm for vegetation restoration is constructed to balance ecological benefits, engineering costs, and species diversity gain, and a Pareto optimal solution set is output. This algorithm uses methods such as genetic algorithms or particle swarm optimization to find solutions that achieve the best balance among multiple objectives. For example, by maximizing ecological benefits such as increasing vegetation cover and biodiversity, and minimizing engineering costs such as planting and maintenance costs, a series of feasible vegetation restoration plans are generated.

[0070] Furthermore, a three-dimensional planning tool for ecological corridors is designed to automatically generate a topological map of the corridor path according to the hydrological permeability coefficient and the animal migration pattern, as Figure 5 shown in the figure. This tool uses GIS technology and three-dimensional visualization technology, combines terrain, hydrological, and animal behavior data, and generates the optimal ecological corridor path. In this way, the effectiveness and connectivity of the corridor can be ensured, promoting species migration and ecosystem restoration.

[0071] It should be understood that the invasive species spread model can more accurately predict the spread path of invasive species by comprehensively considering various environmental factors, providing a scientific basis for prevention and control measures. The multi-objective optimization algorithm for vegetation restoration can generate vegetation restoration plans that meet multiple needs by balancing multiple objectives, improving the effect of ecological restoration. The three-dimensional planning tool for ecological corridors intuitively displays the corridor path through three-dimensional visualization technology, facilitating decision-makers to understand and implement.

[0072] Furthermore, to improve the accuracy and practicality of the model, machine learning algorithms can be introduced to continuously optimize the model parameters. For example, through deep learning methods, learn the patterns of species spread and vegetation restoration from a large amount of historical data to improve the prediction accuracy of the model. At the same time, through a user-friendly interface design, non-professionals can also easily use these tools, improving the popularity of the application.

[0073] Through this embodiment, the spread path of invasive species can be effectively predicted, the vegetation restoration plan can be optimized, and an efficient ecological corridor can be designed, thereby improving the restoration effect and stability of the ecosystem.

[0074] Embodiment Nine: To solve the problems of multi-source data collection, fusion, and processing in the biodiversity monitoring system, this embodiment provides a biodiversity monitoring system. Specifically, the monitoring system includes a multi-source data collection module, a dynamic weight fusion module, an identification engine, a spatio-temporal atlas construction module, and a multi-level early warning decision-making module.

[0075] Furthermore, the multi-source data acquisition module includes a distributed sensor array deployed in the monitoring area, a mobile terminal equipped with high-precision GPS, and a voiceprint recorder, which are used to obtain species images, environmental parameters, geographical trajectories, and bioacoustic data in real time. The distributed sensor array can cover a wide monitoring area and collect various environmental parameters and species data in real time. The high-precision GPS mobile terminal and the voiceprint recorder provide location information and acoustic data, enriching the data sources.

[0076] Furthermore, the dynamic weight fusion module is equipped with an FPGA accelerator and a DDR5 memory matrix, which perform the calculation of the influence degree of environmental factors and the weight distribution of the entropy value of sensor data, and are used to output the species-environment joint feature vector. The FPGA accelerator can significantly improve the calculation speed, and the DDR5 memory matrix provides high-speed data storage and access capabilities. By calculating the influence degree of environmental factors and the entropy value of sensor data, the weights of each sensor data are dynamically adjusted to ensure the accuracy of the feature vector.

[0077] Furthermore, the recognition engine integrates a dual-channel deep convolutional adversarial network and an interpretability analysis unit, and is equipped with a CUDA parallel computing card to achieve real-time species classification. The dual-channel deep convolutional adversarial network improves the classification accuracy and generalization ability by combining the generative adversarial network (GAN) and the convolutional neural network (CNN). The interpretability analysis unit provides an explanation of the classification results, enhancing the transparency and credibility of the system. The CUDA parallel computing card significantly improves the computing performance and realizes real-time classification.

[0078] Furthermore, the spatio-temporal atlas construction module is implemented based on the graph database Neo4j, and includes an ant colony optimization processor and a three-dimensional visualization engine, which dynamically generate a species distribution heat map and an ecological association network. The graph database Neo4j can efficiently store and query complex relational data. The ant colony optimization processor is used to optimize the atlas structure, and the three-dimensional visualization engine provides an intuitive visualization effect. Through these tools, a species distribution heat map and an ecological association network can be dynamically generated to help researchers and decision-makers better understand the state of the ecosystem.

[0079] Furthermore, the multi-level early warning decision-making module has an edge computing gateway to achieve dynamic optimization of early warning thresholds, generation of emergency strategies, and distribution of instructions to multiple departments. For example, the triggering conditions of the multi-level early warning mechanism can be set as follows: when the endangered species index exceeds the first threshold, initiate a first-level early warning and push it to the protected area management terminal; when the spread speed of invasive species exceeds the second threshold, initiate a second-level early warning and link it to the drone inspection system; when the frequency of human interference reaches the third threshold, initiate a third-level early warning and generate a joint instruction for the law enforcement department. The edge computing gateway can process local data in real time, reduce data transmission latency, and improve the response speed. By dynamically optimizing early warning thresholds and generating emergency strategies, sudden situations can be promptly addressed to ensure the safety of the ecosystem. The distribution of instructions to multiple departments enables cross-departmental collaborative work and improves the efficiency of emergency response.

[0080] It should be understood that the multi-source data acquisition module is connected to the dynamic weight fusion module through the LoRaWAN protocol, ensuring the stability and low power consumption of data transmission. The output end of the recognition engine is connected to the streaming computing interface of the spatio-temporal graph construction module, realizing real-time data processing and visualization. The multi-level early warning decision-making module conducts data interaction with the external law enforcement system through the REST API, ensuring the openness and scalability of the system.

[0081] Embodiment Ten; To solve the implementation problem of the biodiversity monitoring method, this embodiment provides a computer-readable storage medium. Specifically, the storage medium stores a computer program, and when the program is executed by a processor, it implements the biodiversity monitoring method.

[0082] Furthermore, the computer program includes multiple modules such as data acquisition, data fusion, species classification, ecological modeling, and early warning decision-making. The data acquisition module is responsible for obtaining real-time data from multi-source sensors and mobile terminals, including species images, environmental parameters, geographical trajectories, and bioacoustic data. The data fusion module dynamically adjusts the weights of each sensor data by calculating the influence degree of environmental factors and the entropy value of sensor data, and outputs a species-environment joint feature vector. The species classification module uses a dual-channel deep convolutional adversarial network and an interpretability analysis unit to achieve high-precision species classification. The ecological modeling module realizes the prediction and optimization of the ecosystem through an invasive species spread model, a multi-objective optimization algorithm for vegetation restoration, and a three-dimensional ecological corridor planning tool. The early warning decision-making module realizes efficient early warning and response by dynamically optimizing early warning thresholds and generating emergency strategies.

[0083] Furthermore, the computer program also includes a data management and a user interface module. The data management module is responsible for data storage, backup, and recovery, ensuring the security and integrity of the data. The user interface module provides a friendly operation interface, enabling users to conveniently view and manage monitoring data, and generate reports and charts.

[0084] It should be understood that the computer program adopts a modular design, and the various modules communicate through standard interfaces, improving the flexibility and scalability of the system. The program code has been strictly tested and optimized to ensure stable operation in various environments. In addition, by introducing containerization technologies such as Docker, rapid deployment and migration of the program can be achieved, simplifying the operation and maintenance work.

[0085] Through this embodiment, an efficient and reliable biodiversity monitoring method can be realized, providing comprehensive data collection, processing, and analysis functions, and supporting the development of scientific research and ecological protection work.

[0086] Although the present invention has been specifically described above with reference to the preferred embodiments of the present invention, it should be understood that the present invention is not limited to the embodiments described above. Rather, without departing from the essence of the present invention, those skilled in the art can make various modifications and changes, and these modifications and changes should fall within the scope defined by the appended claims and their equivalents.

Claims

1. A biodiversity monitoring method, characterized in that, The steps of the monitoring method include: Real-time collection of multi-source data on biodiversity in the target area through a distributed sensor network and mobile terminal devices. The multi-source data includes species image data, environmental factor time-series data, voiceprint feature data, and geospatial trajectory data, and the data is subjected to encrypted transmission and format standardization preprocessing; Based on the dynamic weight allocation algorithm, feature fusion is performed on the multi-source data to generate a species-environment joint feature vector. The dynamic weight allocation algorithm adjusts the weight ratio of sensor data in real time according to the influence degree of the environmental factor time-series data on species distribution; A classification model is constructed using a deep convolutional adversarial network to perform species classification and recognition on the species-environment joint feature vector to obtain a classification result. The classification model is pre-trained through transfer learning and an abnormal data filtering module is embedded to eliminate misdetected data caused by abnormal lighting, motion blur, or noise interference; Based on the classification result, a spatio-temporal association map is constructed, the distribution density of species with different protection levels, the population migration trajectory, and the hot spots of human interference are statistically analyzed, and the biodiversity database is dynamically updated; According to the preset ecological security threshold, a multi-level early warning mechanism is triggered to generate a visual monitoring report and a regulation strategy. The visual monitoring report includes the species endangerment index, the invasion species diffusion model, and the ecological restoration plan.

2. The biodiversity monitoring method according to claim 1, wherein The steps of the dynamic weight allocation algorithm include: Define the environmental factor influence degree function: , where T is the temperature change rate, P is the precipitation intensity, and S i is the sensitivity coefficient of the i-th type of species, n is the total number of species types, and D i is the species distribution dispersion, and α, β, and γ are trainable parameters; Calculate the local entropy value of each sensor data through a sliding time window, and dynamically allocate weights according to the entropy value ratio: , where λ is the environmental response coefficient, and E k is the environmental factor influence degree of the k-th sensor; Use the gradient descent method to optimize the allocation result of the weight ratio until the within-class distance variance of the species-environment joint feature vector is less than the preset threshold.

3. The biodiversity monitoring method according to claim 2, wherein The optimization steps of the deep convolutional adversarial network include: Construct a dual-channel feature extractor to process the texture features of image data and the frequency domain features of voiceprint data respectively, and fuse the features through a cross-attention mechanism; Design an adversarial loss function: , where G is the generator, D is the discriminator, E is the expected value, x is the real data, and η is the reconstruction loss weight; Embed an interpretability module in the output layer of the classifier to generate a heat map of species recognition confidence and an analysis report on the reasons for misdetection.

4. The biodiversity monitoring method according to claim 1, wherein The steps of constructing a spatio-temporal association map based on the classification result include: Map the species distribution data to map nodes, and the node attributes include protection level, population quantity, and genetic diversity index; Calculate the association strength between nodes, and the calculation expression is: , where C ij is the species symbiosis coefficient, d ij is the geographical distance, Δt is the time interval, and σ, τ are the attenuation factors; Reduce the high-dimensional node features to a low-dimensional space through map embedding technology to generate a visual ecological security situation map.

5. The biodiversity monitoring method according to claim 1, characterized in that, The steps of the transfer learning pre-training include: Construct a hierarchical transfer strategy, freeze the parameters of the first three convolutional kernels of the pre-trained model, and only perform fine-tuning training on the fully connected layer for the species features in the target area; Introduce an adversarial domain adaptation mechanism, and insert a gradient reversal layer between the feature extractor and the classifier to force the network to learn the essential species features independent of geographical location; Design a diversity-enhanced sample library, and use neural style transfer technology to generate virtual training samples with different lighting conditions and vegetation coverage to expand the generalization ability of the pre-trained model in the rainy season and dry season; Implement a dynamic curriculum learning strategy, and unlock the classifier neurons in stages according to the frequency of species appearance from high to low; Deploy a model drift detection module. When the F1-score of the validation set drops by more than a preset value for a continuously set number of days, trigger the incremental learning process and update the model parameters of the edge computing node.

6. The biodiversity monitoring method according to claim 1, wherein The steps of the monitoring method further include: Perform multi-dimensional cross-validation of the classification results with an authoritative species database, including morphological feature matching degree, habitat suitability index, and breeding cycle matching degree; Construct a credibility scoring model: , where Ψ represents the confidence score, ω k is the weight of the k-th verification dimension, D k is the database feature, O k is the observed feature, P is the classification probability distribution, geo match is the geographical distribution matching degree; When the credibility score Ψ is less than the preset score, initiate an expert review process and send multi-source data of suspected species to the verification terminals of multiple domain experts; Dynamically adjust the classification model threshold parameters according to the expert feedback results, and update the species distribution baseline data in the biodiversity database.

7. The biodiversity monitoring method according to claim 1, characterized in that, The steps for setting the ecological security threshold include: Construct a threshold dynamic prediction framework based on the LSTM-GARCH hybrid model, and input historical species fluctuation data, climate change trends, and human activity intensity indicators; Define the threshold adaptive adjustment rule: , where θ t is the current threshold, and θ base is the baseline threshold, N t is the current population size, and ΔH t is the change rate of human disturbance intensity, and a and b are adjustment coefficients; Implement threshold sensitivity analysis, and calculate the false alarm rate and missed alarm rate balance point of each warning level in the threshold preset error fluctuation range through Monte Carlo simulation; When the same geographical grid continuously triggers a set number of warnings within a preset time, start the threshold emergency re-evaluation program, and call high-resolution satellite remote sensing data for re-evaluation of habitat integrity.

8. The biodiversity monitoring method according to claim 1, characterized in that, The monitoring method further includes ecological modeling, and the steps of the ecological modeling include: Establish an invasive species diffusion model, and comprehensively predict the diffusion path based on wind speed vector, soil moisture gradient, and competing species distribution; Construct a multi-objective optimization algorithm for vegetation restoration, balance ecological benefits, project costs, and species diversity gain, and output the Pareto optimal solution set; Design a three-dimensional planning tool for ecological corridors, and automatically generate the topological map of the corridor path according to the hydrological permeability coefficient and animal migration rules.

9. A biodiversity monitoring system for implementing the biodiversity monitoring method according to any one of claims 1-8, characterized in that, The monitoring system includes: A multi-source data acquisition module, including a distributed sensor array deployed in the monitoring area, a mobile terminal equipped with high-precision GPS, and a voiceprint recorder, which is used to obtain species images, environmental parameters, geographical trajectories, and bioacoustic data in real time; A dynamic weight fusion module, equipped with an FPGA accelerator and a DDR5 memory matrix, which performs environmental factor influence degree calculation and sensor data entropy value weight allocation, and is used to output a species-environment joint feature vector; An identification engine, integrated with a dual-channel deep convolutional adversarial network and an interpretability analysis unit, and equipped with a CUDA parallel computing card to achieve real-time species classification; A spatio-temporal map construction module, implemented based on the graph database Neo4j, including an ant colony optimization processor and a three-dimensional visualization engine, which dynamically generates a species distribution heat map and an ecological association network; A multi-level early warning decision-making module, with an edge computing gateway, to achieve dynamic optimization of early warning thresholds, generation of emergency strategies, and distribution of instructions to multiple departments; The multi-source data acquisition module is connected to the dynamic weight fusion module through the LoRaWAN protocol. The output end of the identification engine is connected to the streaming computing interface of the spatio-temporal map construction module. The multi-level early warning decision-making module performs data interaction with an external law enforcement system through the RESTAPI.

10. A computer-readable storage medium storing a computer program, characterized in that, When the described program is executed by a processor, it implements the biodiversity monitoring method described in any one of claims 1-8.

Images