Natural reserve ecological damage identification and disposal decision system based on big data

Abstract

The application discloses a natural reserve ecological damage identification and disposal decision system based on big data and belongs to the technical field of ecological intelligent supervision. The system comprises the following modules: a model construction training module, which is used for building a multi-modal ecological damage sample data set and training a multi-modal identification quantification model; a collection and scheduling module, which is used for acquiring multi-source original monitoring data through artificial configuration regular global inspection and network big data searching public opinion clue directional collection; a data processing module, which is used for generating standardized model input feature data after processing the multi-source original monitoring data; an intelligent identification and evaluation module, which is used for determining an ecological damage type and generating complete ecological damage event information by matching a responsible subject based on the multi-modal identification quantification model; a disposal decision generation module, which is used for building a disposal decision model, automatically generating a disposal scheme and pushing the disposal scheme to the responsible subject; and a closed-loop tracking and control module, which is used for continuously tracking the rectification progress and completing rectification verification. The application solves the pain points of traditional ecological supervision lag and non-standard disposal.

Description

Technical Field

[0001] This invention relates to the field of ecological intelligent supervision technology, specifically to a decision-making system for identifying and addressing ecological damage in nature reserves based on big data. Background Technology

[0002] Nature reserves are vast and have complex topography, making them prone to ecological damage such as illegal logging and unauthorized construction. Traditional monitoring relies mainly on manual on-site patrols, which have limited manpower coverage and leave long-term blind spots in remote areas. This delay in detecting destructive activities can easily exacerbate ecological losses. Existing remote sensing monitoring relies solely on identifying single image patches, failing to distinguish between natural surface changes and deliberate human damage, nor can it quantify the level of damage. Ground-based IoT monitoring points are isolated and have not been integrated with remote sensing data, resulting in a limited monitoring dimension. Meanwhile, online public opinion reports rely on manual screening of each report, lacking automatic capture, semantic recognition, and spatial positioning capabilities, leading to low efficiency in verifying reports.

[0003] Furthermore, the existing regulatory system lacks integrated intelligent support for handling incidents. After a breach is discovered, the handling plan, legal basis, and division of responsibilities all rely on manual review and compilation, resulting in inconsistent standards and low efficiency in resource allocation. Moreover, there is no routine automatic tracking and monitoring mechanism after the incident is issued; the effectiveness of rectification is judged solely by on-site follow-ups, making it impossible to quantify and assess the progress of rectification. This makes it difficult to build a closed-loop regulatory process from clue discovery, intelligent identification, tiered handling, to rectification review and closure, and the overall level of digitalization and intelligence in ecological supervision is insufficient. Summary of the Invention

[0004] This invention provides a big data-based decision-making system for identifying and addressing ecological damage in nature reserves, which solves the problems of lagging ecological supervision and identification, single monitoring dimensions, inefficient clue verification, lack of standardized intelligent support for disposal, and lack of closed-loop management throughout the entire process in traditional nature reserves.

[0005] To address the aforementioned issues, this invention discloses a big data-based system for identifying and responding to ecological damage in nature reserves, comprising a model building and training module, a data acquisition and scheduling module, a data processing module, an intelligent identification and evaluation module, a response decision generation module, and a response decision generation module. The model building and training module is used to build an ecological damage sample dataset and train an ecological damage multimodal identification and quantification model based on the sample dataset. The data acquisition and scheduling module is used to perform regular full-area inspections to collect multi-source raw monitoring data according to manual configuration, and to obtain ecological clues based on network big data retrieval and collect multi-source raw monitoring data of the target areas corresponding to the ecological clues in a targeted manner. The data processing module is used to preprocess multi-source raw monitoring data to generate standardized model input feature data. The intelligent identification and assessment module is used to input standardized model input feature data into the ecological damage multimodal identification and quantification model to determine the type of ecological damage; to conduct ecological damage quantification assessment of ecological damage targets; and to match the responsible jurisdiction subject with spatial geographic ownership data to generate complete ecological damage event information. The decision generation module is used to build a decision generation model, automatically generate a response plan based on complete information about the ecological damage incident, and push it to local management departments and front-line law enforcement units. The closed-loop tracking and control module is used to continuously collect multi-source raw monitoring data of ecological damage targets and dynamically monitor the rectification process of ecological damage events corresponding to ecological damage targets in order to complete the verification of rectification effectiveness.

[0006] Preferably, the multimodal identification and quantification model for ecological damage includes a visual feature extraction branch, a temporal feature extraction branch, a cross-modal attention fusion module, and a multi-task joint output head; The visual feature extraction branch adopts the Swin-Transformer v2 combined with the Feature Pyramid Network (FPN) backbone structure to extract spatial features from the standardized model input feature data and output a visual feature map. The spatial features are composed of spectral index features and texture spatial features spliced ​​and fused together. The temporal feature extraction branch uses a bidirectional long short-term memory network Bi-LSTM combined with a multi-head attention mechanism to extract temporal statistical features and frequency domain features from the standardized model input feature data and output a temporal feature vector. The cross-modal attention fusion module is used to map visual feature maps and temporal feature vectors to the same feature space, and realize bidirectional information interaction and feature fusion through a cross-attention mechanism, and output a fused feature tensor. The multi-task joint output head includes a classification output subheader and an instance segmentation output subheader, which are used to output the probability of ecological damage type and the pixel-level mask of the damage area.

[0007] Preferably, the method for obtaining the ecological clues is as follows: The distributed crawler cluster is used to crawl publicly available text data from public public opinion data sources across the entire Internet in real time, and after deduplication and cleaning, a candidate text set is generated. A pre-trained BERT text classification model is used to perform binary classification of ecological damage relevance on candidate texts in the candidate text set. Location entities, damage type entities, and time entities are extracted from the candidate texts through named entity recognition. The extracted location entities are geocoded and converted to generate spatial coordinates. These coordinates are then spatially overlaid and filtered with the boundaries of nature reserves to retain spatial matching clues within the nature reserves and form candidate clues. A clue credibility assessment model is constructed, which calculates the clue credibility score of candidate clues by comprehensively considering the authority of the publishing entity, the completeness of information, the accuracy of historical reports, and the cross-validation of multiple sources. Candidate clues with a credibility score higher than the preset credibility threshold are identified as ecological clues, thereby triggering targeted collection tasks in the corresponding target areas.

[0008] Preferably, the multi-source raw monitoring data includes at least remote sensing images; when IoT monitoring sensors are deployed in the target area or inspection area, the multi-source raw monitoring data also includes IoT time-series data, which is at least one of water quality monitoring data, atmospheric environment monitoring data, and soil environment monitoring data.

[0009] Preferably, the data preprocessing is as follows: The remote sensing images are sequentially subjected to radiometric calibration, atmospheric correction, geometric fine correction, cloud shadow removal and mosaicking and cropping to obtain the spatial data of the remote sensing images. At the same time, the IoT time series data are sequentially subjected to outlier removal, missing value completion and uniform granularity resampling to obtain standardized time series monitoring data. The spatial data of remote sensing images are unified to a preset standard geographic coordinate system, and standardized time-series monitoring data for the corresponding time period are extracted based on the remote sensing image capture time to complete the spatiotemporal alignment and matching of the data. Spatial features are extracted from the spatial data of the spatiotemporally aligned remote sensing images. At the same time, temporal statistical features and frequency domain features are extracted from the standardized time-series monitoring data after spatiotemporal alignment. Based on these features, they are integrated to form standardized model input feature data. If there is no IoT time-series data, then zeros are uniformly padded to fill the time-domain statistical features and frequency-domain features.

[0010] Preferably, the method for conducting a quantitative assessment of ecological damage is as follows: Based on pixel-level masks of the damaged areas, the actual area, center point coordinates, and boundary vectors of the damaged patches are calculated in combination with the resolution of remote sensing images, so as to complete the quantitative measurement of the spatial range of ecological damage targets. A differentiated quantitative indicator system is constructed, and the comprehensive index method is used to calculate the comprehensive ecological damage score of ecological damage targets. The comprehensive ecological damage score is mapped and divided into mild, moderate and severe ecological damage levels to complete the quantitative assessment of the damage intensity level of ecological damage targets. Simultaneously, for ecological damage targets classified as illegal hunting, a population impact assessment model of hunting activities is constructed by combining species distribution data and habitat connectivity data of the areas where the ecological damage targets are located. This model is used to extrapolate the potential impact range of hunting activities on species populations and the degree of habitat disturbance, so as to complete a quantitative assessment of the overall impact of illegal hunting on ecological damage targets.

[0011] Preferably, the differentiated quantitative indicator system is as follows: If the ecological damage target corresponds to the type of ecological damage, such as vegetation felling, then calculate the vegetation coverage decline rate, NDVI attenuation difference, and biomass loss index. If the ecological damage target corresponds to the type of ecological damage, such as illegal construction in industrial and mining development, then calculate the indicators of surface hardening ratio, surface disturbance intensity, and estimated earthwork excavation volume. If the ecological damage target corresponds to the type of ecological damage, namely illegal hunting, then the frequency of hunting activities is calculated. Number and poaching disturbance density index of trail development.

[0012] Preferably, the process of constructing the disposal decision model and generating the disposal plan is as follows: An ecological big data knowledge base is pre-built, including a legal and regulatory database, a disposal process database, a historical case database, and a law enforcement resource database, and an ecological disposal knowledge graph is constructed based on entity relationships; Based on complete information on ecological damage events, risk level assessments of these events are conducted from the dimensions of damage intensity, ecological sensitivity, and social impact. The analytic hierarchy process (AHP) combined with the entropy weight method is used to calculate the comprehensive risk value of the ecological damage events and classify them into risk levels. The rule engine matches the illegal clauses and legal handling requirements corresponding to ecological damage events, the knowledge graph retrieves similar historical ecological damage event handling cases to optimize handling details, and the spatial analysis engine schedules local law enforcement resources and law enforcement action paths. The system automatically generates standardized handling plans that include the facts of the violation, legal basis, handling measures, rectification requirements, completion deadlines, and responsible units, and pushes them out in a tiered manner according to risk level, simultaneously distributing them to the corresponding local management departments and front-line law enforcement units.

[0013] Preferably, the method for verifying the effectiveness of the rectification is as follows: Differentiated monitoring frequencies are configured according to the risk level of ecological damage events, and monitoring equipment is dispatched to continuously collect multi-source raw monitoring data of ecological damage targets, and the multi-source raw monitoring data is automatically collected into the corresponding ecological damage event rectification file; Based on the rectification archives of ecological damage incidents, quantitative indicators for the rectification of ecological damage incidents are calculated. Combined with pre-set early warning nodes for rectification progress, reminders are automatically sent before the deadline for rectification. For ecological damage incidents that are not completed by the deadline, the supervision level is automatically upgraded. After the responsible unit submits an application for completion of rectification, the effectiveness of the rectification will be verified by combining the current multi-source original monitoring data with the on-site verification results. Once the verification is passed, the incident will be closed out.

[0014] Compared with the prior art, the present invention has the following advantages: 1. This invention constructs a multimodal recognition and quantification model that integrates remote sensing imagery and IoT time-series data. It relies on a visual and temporal dual-branch cross-attention fusion mechanism to achieve pixel-level segmentation, automatic classification, and damage grading of ecological damage. This overcomes the shortcomings of traditional single remote sensing monitoring, which is unable to distinguish between natural fluctuations and human subjective damage and has limited monitoring dimensions. It also solves the pain points of lagging ecological damage recognition and lack of quantitative assessment.

[0015] 2. This invention integrates distributed crawling, BERT semantic recognition, geocoding spatial filtering, and clue credibility assessment models to automatically crawl and filter Internet ecological public opinion clues, and to trigger monitoring and collection in key areas. It replaces manual screening of public opinion one by one, fills the regulatory blind spots of manual patrols in remote protected areas, and greatly improves the efficiency of uncovering hidden ecological damage clues.

[0016] 3. This invention constructs an ecological disposal knowledge graph containing regulations, cases, and law enforcement resources, and is equipped with a full-process closed-loop tracking and control mechanism. It can automatically generate standardized hierarchical disposal plans, dynamically monitor rectification progress, issue early warnings and supervision, and verify and close cases, forming an integrated intelligent supervision link. This solves the problems of non-standardized traditional supervision and disposal processes, lack of automatic tracking and closed loop, and low efficiency of manual disposal. Attached Figure Description

[0017] Figure 1 This is a system architecture diagram of the big data-based ecological damage identification and treatment decision-making system for nature reserves, based on the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Examples, such as Figure 1 As shown in this embodiment, the big data-based ecological damage identification and response decision-making system for nature reserves is deployed on a cloud platform and adopts a modular, layered architecture. From top to bottom, it includes a model building and training module, a data acquisition and scheduling module, a data processing module, an intelligent identification and assessment module, a response decision generation module, and a closed-loop tracking and control module. Each module achieves data interaction and business linkage through standardized data interfaces, jointly forming a full-chain ecological supervision system.

[0020] The model building and training module is used to construct an ecological damage sample dataset and train an ecological damage multimodal identification and quantification model based on the sample dataset. The specific implementation method is as follows: First, an ecological damage sample dataset is constructed. In this embodiment, the ecological damage sample dataset covers scenarios such as no ecological damage, vegetation clearing, illegal mining, illegal construction, and illegal hunting. The ecological damage samples in the ecological damage sample dataset come from publicly available remote sensing image datasets, archived historical ecological damage event data, historical IoT monitoring data, and manually annotated supplementary data. Each ecological damage sample carries a classification label corresponding to the ecological damage type and a pixel-level mask annotation of the damaged area. The classification label is used for model classification branch training, and the pixel-level mask annotation is used for instance segmentation branch training.

[0021] Then, a multimodal recognition and quantification model for ecological damage is constructed, which includes a visual feature extraction branch, a temporal feature extraction branch, a cross-modal attention fusion module, and a multi-task joint output head.

[0022] The visual feature extraction branch employs a Swin-Transformer v2 combined with a Feature Pyramid Network (FPN) backbone structure to extract spatial features from the standardized model input feature data and output visual feature maps. In this embodiment, the Swin-Transformer v2 uses a basic configuration, containing four stages of feature extraction layers and outputting four different scale feature maps with strides of 4, 8, 16, and 32. The Feature Pyramid Network (FPN) performs top-down feature fusion and lateral connections on the four different scale feature maps, outputting a unified-dimensional multi-scale visual feature map. The spatial features of the input visual branch are the stitched spectral index features and texture spatial features. Among them, the spectral index features include the Normalized Difference Vegetation Index (NDVI) reflecting vegetation growth and coverage, the Normalized Difference Water Index (NDWI) reflecting the distribution range of water bodies, the Normalized Difference Building Index (NDBI) reflecting the proportion of artificially hardened surfaces, the Modified Normalized Difference Water Index (MNDWI) reflecting the elimination of vegetation interference in water body identification, and the Soil Adjusted Vegetation Index (SAVI) reflecting soil background correction for vegetation monitoring. The texture spatial features are extracted based on the Gray-Level Co-occurrence Matrix (GLCM) of a 5×5 pixel window, including contrast, correlation, energy, entropy, mean, variance, homogeneity, and variability. In this embodiment, the feature calculation window is set to 5×5 pixels.

[0023] The temporal feature extraction branch employs a bidirectional long short-term memory (Bi-LSTM) network paired with a multi-head attention mechanism to extract temporal statistical features and frequency domain features from the standardized model input feature data, outputting a temporal feature vector. In this embodiment, the Bi-LSTM contains two layers of bidirectional recurrent networks with a hidden layer dimension of 256, used to capture the forward and backward dependencies of the input temporal data; the multi-head attention mechanism is set to 8 heads to strengthen the feature weights of key time nodes. The features of the input temporal branch are a concatenated vector of temporal statistical features and frequency domain features. The temporal statistical features include mean, maximum, minimum, standard deviation, coefficient of variation, kurtosis, skewness, and range; the frequency domain features are obtained by performing frequency domain decomposition on the spatiotemporally aligned standardized temporal monitoring data using Fast Fourier Transform (FFT), extracting 16-dimensional frequency domain features corresponding to the amplitude and phase of the first 8 dominant frequencies. The hidden state output by the Bi-LSTM is weighted by the multi-head attention mechanism to output the temporal feature vector.

[0024] The cross-modal attention fusion module maps visual feature maps and temporal feature vectors to the same feature space, and achieves bidirectional information interaction and feature fusion through a cross-attention mechanism, outputting a fused feature tensor. In this embodiment, firstly, a convolutional layer maps the number of channels of the visual feature map to a predetermined dimension, and a fully connected layer maps the temporal feature vector to a predetermined dimension, achieving feature dimension unification; then, a bidirectional cross-attention mechanism is used to calculate the attention weights of the visual feature map to the temporal feature vector and the attention weights of the temporal feature vector to the visual feature map: using the visual feature map as the query and the temporal feature vector as the key, visually guided temporally weighted features are calculated; using the temporal feature vector as the query and the visual feature map as the key, temporally guided visually weighted features are calculated; finally, the two weighted features are concatenated and fused through a convolutional layer to output a fused feature tensor.

[0025] The multi-task joint output head includes a classification output subheader and an instance segmentation output subheader, used to output the probability of ecological damage types and pixel-level masks of damaged areas. In this embodiment, the classification output subheader uses a global average pooling layer followed by two fully connected layers, and finally outputs the probability values ​​of various ecological damage types through a Softmax activation function. Ecological damage types include no ecological damage, vegetation clearing, illegal mining, illegal construction, and illegal hunting. The instance segmentation output subheader uses a fully convolutional network structure, which upsamples the fused feature tensor to restore it to the original resolution of the input visual feature map, and outputs the probability value of each pixel belonging to the ecological damage area through a Sigmoid activation function, generating a pixel-level mask of the damaged area.

[0026] Finally, the ecological damage multimodal identification and quantification model is trained based on the constructed ecological damage sample dataset. A multi-task loss function that combines classification loss and segmentation loss is adopted, and the AdamW optimizer is combined with a cosine annealing learning rate decay strategy. Training is stopped after the validation set loss has not decreased for 10 consecutive rounds. The optimal model weights are saved, and the trained ecological damage multimodal identification and quantification model is finally obtained.

[0027] The data acquisition and scheduling module is used to perform regular full-area inspections and collect multi-source raw monitoring data according to manual configuration, and to obtain ecological clues based on network big data retrieval and collect multi-source raw monitoring data of target areas corresponding to the ecological clues. The multi-source raw monitoring data includes at least remote sensing images. When IoT monitoring sensors are deployed in the target area or inspection area, the multi-source raw monitoring data also includes IoT time-series data. The IoT time-series data is at least one of water quality monitoring data, atmospheric environment monitoring data, and soil environment monitoring data. The specific implementation method is as follows: In this embodiment, the manually configured periodic full-area inspection strategy is as follows: an inspection area is delineated for the nature reserves within the jurisdiction, and a monthly remote sensing inspection is performed to obtain remote sensing images. If there are IoT monitoring sensors deployed in the inspection area, the IoT time series data of the inspection area is obtained.

[0028] Simultaneously, a distributed crawler cluster is used to crawl publicly available text data from public public opinion data sources across the entire internet in real time, and after deduplication and cleaning, a candidate text set is generated. In this embodiment, the distributed crawler cluster is built on the Scrapy-Redis distributed framework and deployed on several cloud server nodes to achieve parallel crawling across multiple cloud server nodes; the public public opinion data sources include mainstream social media platforms, short video platforms, local government message boards, official reporting platforms of ecological and environmental departments, local news websites, and other publicly accessible channels. The crawling strategy adopts real-time incremental crawling. The crawled public text data is first deduplicated by using the SimHash algorithm to calculate text fingerprints and remove duplicate content; then, irrelevant characters are cleaned using regular expressions, and the text encoding format is unified to finally generate a candidate text set.

[0029] Subsequently, a pre-trained BERT text classification model is used to perform binary classification of the candidate texts in the candidate text set to determine their relevance to ecological damage. Named entity recognition is then used to extract location entities, damage type entities, and time entities from the candidate texts. In this embodiment, the text classification model is trained based on a pre-trained Chinese BERT-base model, with the output layer set to binary classification, labeled as "relevant to ecological damage" and "unrelated to ecological damage." The training data consists of historical ecological complaint texts and ecological news texts, achieving a classification accuracy of no less than 92%. Named entity recognition is implemented using a BERT-BiLSTM-CRF model, with entity types including location entities (such as village, mountain, river, protected area name), damage type entities (such as tree felling, illegal construction, etc.), and time entities (such as yesterday, last week, corresponding year and month); the entity recognition F1 score is no less than 88%. After classification and filtering, only candidate texts judged as "relevant to ecological damage" are retained, and the extracted entity information is output.

[0030] The extracted location entities are geocoded using a public geocoding service interface to convert the location names in the candidate text into latitude and longitude coordinates in the WGS84 coordinate system, i.e., spatial coordinates. Then, a GIS spatial intersection analysis tool is used to spatially overlay the spatial coordinates with the boundary of the nature reserve. If the spatial coordinates are within the boundary of the nature reserve, the candidate text is determined to be a clue within the nature reserve and is included in the candidate clues; clues that are completely outside the boundary of the nature reserve are directly filtered out.

[0031] A credibility assessment model for clues is constructed. In this embodiment, the credibility assessment model uses a weighted summation method to calculate the score. The weights of each indicator are as follows: the authority of the publishing entity is weighted at 0.3, with official accounts and government platforms receiving 1 point, verified personal accounts receiving 0.6 points, and anonymous accounts receiving 0.2 points; the information completeness weight is 0.25, with the inclusion of three types of entities—location entity, damage type entity, and time entity—receiving 1 point, including two types receiving 0.6 points, and including only one type receiving 0.2 points; the historical reporting accuracy weight is 0.25, the percentage of historical reports by the publishing entity that have been verified as true, with a 100% accuracy rate receiving 1 point, 50% receiving 0.5 points, and 0% receiving 0 points; the multi-source cross-verification weight is 0.2, with the same event being published by two or more independent sources receiving 1 point, and only one source receiving 0 points. Based on the clue credibility assessment model, the credibility score of candidate clues is calculated by comprehensively considering the authority of the issuing entity, the completeness of information, the accuracy rate of historical reports, and multi-source cross-verification. Candidate clues with credibility scores higher than a preset credibility threshold are identified as ecological clues, triggering targeted data collection tasks in the corresponding target areas. Satellites and surrounding IoT sensors are deployed to collect multi-source raw monitoring data in the target areas, and the collected data is then pushed to the data processing module.

[0032] The data processing module is used to preprocess multi-source raw monitoring data to generate standardized model input feature data. The specific implementation method is as follows: The remote sensing images are sequentially subjected to radiometric calibration, atmospheric correction, geometric fine correction, cloud shadow removal, and mosaicking and cropping to obtain spatial data of the remote sensing images. Among them, the cloud shadow removal is combined with the QA quality band of the remote sensing images, and the threshold method is used to identify cloud and shadow areas. Thin cloud areas are repaired using cloud removal algorithms, and thick cloud-covered areas are marked as invalid values. At the same time, the IoT time series data are sequentially subjected to outlier removal, missing value completion, and uniform granularity resampling to obtain standardized time series monitoring data.

[0033] Using Gauss-Kruger 3-dimensional projection, the spatial data of remote sensing images are unified to the preset standard geographic coordinate system CGCS2000 National Geodetic Coordinate System through coordinate transformation tools. Standardized time-series monitoring data corresponding to the time periods of 12 hours before and after the remote sensing image capture time are extracted, totaling 24 hours, to achieve temporal alignment with the remote sensing images. For remote sensing image areas without corresponding IoT time-series data, they are marked as having no time-series data.

[0034] Spatial features are extracted from the spatiotemporally aligned remote sensing image spatial data. Simultaneously, temporal statistical features and frequency domain features are extracted from the standardized time-series monitoring data after spatiotemporal alignment. These are then integrated to form the input feature data for a standardized model. If no IoT time-series data is available, zero-padding is applied to both the temporal statistical features and frequency domain features. In this embodiment, spatial feature extraction is performed pixel-by-pixel, window-by-window, generating a multi-channel feature map, including spectral index features and texture spatial features, i.e., spatial features. Temporal statistical features and frequency domain features are calculated for each IoT monitoring station. For areas without IoT time-series data, all temporal statistical features and frequency domain features are assigned a value of 0.

[0035] The intelligent identification and assessment module is used to input standardized model input feature data into the ecological damage multimodal identification and quantification model, determine the type of ecological damage, conduct ecological damage quantification assessment of ecological damage targets, and combine spatial geographic ownership data to match the responsible jurisdiction entity, generating complete ecological damage event information. The specific implementation method is as follows: The standardized model input feature data is input into the trained multimodal ecological damage identification and quantization model. After forward inference, the model outputs the probability value of the ecological damage type of each patch region in the remote sensing image through the classification output subheader. The category with the highest probability is selected as the determination type of the patch region. If the category of the patch region is determined to be non-ecological damage, then the patch region is determined to be an ecological damage target. At the same time, the instance segmentation output subheader outputs the pixel-level mask of the damage area corresponding to the ecological damage target for subsequent quantization calculation.

[0036] Then, a quantitative assessment of ecological damage is carried out on the ecological damage targets. Based on the pixel-level mask of the corresponding damage area, the actual area, center point coordinates and boundary vector of the damage patch are calculated in combination with the resolution of remote sensing image to complete the quantitative measurement of the spatial range of the ecological damage target. The actual area of ​​the damage patch is the product of the number of effective pixels in the mask and the ground area corresponding to a single pixel. The center point coordinates are the latitude and longitude coordinates of the geometric centroid of the mask area. The boundary vector is extracted by a raster vectorization tool.

[0037] The system constructs a differentiated quantitative indicator system and uses the comprehensive index method to calculate the comprehensive ecological damage score of ecological damage targets. The objective weight of each sub-indicator is determined by the entropy weight method. After normalizing each indicator from 0 to 100, the weighted sum is calculated, and the score range is 0-100. A score range of 0-30 is judged as mild ecological damage, 30-70 as moderate ecological damage, and 70-100 as severe ecological damage. Based on this, the intensity level of ecological damage targets is quantitatively assessed. In this embodiment, the differentiated quantitative indicator system is set according to the type of damage: if the ecological damage target corresponds to the vegetation clearing type, then the vegetation cover decline rate, the normalized vegetation index (NVI) decay difference, and the biomass loss index are calculated. The vegetation cover decline rate is the ratio of the difference between the average vegetation cover before and after the damage to the average vegetation cover before the damage. The NVI decay difference is the difference between the average NVI of the same month in the twelve months before the damage and the NVI of the month of the damage. The biomass loss index... The estimated biomass per unit area is multiplied by the total area of ​​the damaged patch. If the ecological damage type corresponding to the ecological damage target is illegal construction due to industrial and mining development, then the surface hardening ratio index, surface disturbance intensity index, and earthwork excavation estimate index are calculated. The surface hardening ratio index is the ratio of the number of pixels in the hardened area to the total number of pixels in the patch. The surface disturbance intensity index is the result of the weighted average value of the normalized building index and the terrain slope coefficient. The earthwork excavation estimate index is the difference in elevation of the digital elevation model before and after the damage multiplied by the area of ​​the damaged area. If the ecological damage target corresponds to the type of ecological damage, namely illegal hunting, then the hunting activity frequency index and poaching disturbance path development density are calculated. The hunting activity frequency index is the cumulative number of illegal hunting incidents within a preset monitoring period, and the poaching disturbance path development density is the length of forest paths trampled or simply dug by humans and extracted based on the vectorization of remote sensing images, which is used to characterize the accessibility of poachers to nature reserves.

[0038] Simultaneously, for ecological damage targets classified as illegal hunting, a population impact assessment model for hunting activities is constructed by combining species distribution data and habitat connectivity data of the target areas. This model extrapolates the potential impact range and habitat disturbance degree of hunting activities on species populations, thereby completing a quantitative assessment of the overall impact of illegal hunting on ecological damage targets. In this embodiment, the habitat range of protected species is first delineated based on the regional species distribution layer of nature reserves. A human disturbance buffer zone is then delineated on the overlaid poaching disturbance patches to specifically determine the potential impact range of hunting activities on species populations. Simultaneously, a minimum cost distance model is used to analyze the degree of habitat connectivity disturbance, i.e., the degree of habitat disturbance, to assess the blocking effect of hunting activities on species migration corridors and gene flow. Finally, by integrating the results of the quantitative assessment of ecological damage and combining them with the spatial coordinates of the vector boundary corresponding to the ecological damage target, the local management department and front-line law enforcement unit are determined, and complete ecological damage event information for the ecological damage target is generated, with one ecological damage target corresponding to one ecological damage event.

[0039] The decision-making generation module is used to construct a decision-making model for handling ecological damage events. It automatically generates a response plan based on complete information about the ecological damage incident and pushes it to local management departments and frontline law enforcement units. The specific implementation method is as follows: First, a decision-making model for handling the situation is constructed, with a three-layer architecture consisting of a storage layer, an inference layer, and an output layer. The storage layer is built with an ecological big data knowledge base that includes a legal and regulatory database, a handling process database, a historical case database, and an enforcement resource database. An ecological handling knowledge graph is constructed based on the Neo4j graph database. The inference layer integrates AHP with an entropy weight risk assessment algorithm, a rule engine, and a spatial analysis engine. The output layer is responsible for generating standardized handling plan documents and connecting to hierarchical push interfaces.

[0040] Subsequently, based on complete information on ecological damage events, risk level assessments were conducted on ecological damage events from the dimensions of damage intensity, ecological sensitivity, and social impact. The analytic hierarchy process (AHP) combined with the entropy weight method was used to calculate the comprehensive risk value of ecological damage events and classify risk levels.

[0041] The rules engine matches the illegal clauses and legal requirements for handling ecological damage incidents, the knowledge graph retrieves similar historical ecological damage incident handling cases to optimize handling details, and the spatial analysis engine schedules local law enforcement resources and law enforcement action paths.

[0042] The system automatically generates standardized handling plans that include the facts of the violation, legal basis, handling measures, rectification requirements, completion deadlines, and responsible units, and pushes them out in a tiered manner according to risk level, simultaneously distributing them to the corresponding local management departments and front-line law enforcement units.

[0043] The closed-loop tracking and control module is used to continuously collect multi-source raw monitoring data of ecological damage targets, dynamically monitor the rectification progress of ecological damage events corresponding to ecological damage targets, and complete the verification of rectification effectiveness. The specific implementation method is as follows: The system allocates differentiated monitoring frequencies based on the risk level of ecological damage incidents, schedules monitoring equipment to continuously collect multi-source raw monitoring data of ecological damage targets, and automatically aggregates the multi-source raw monitoring data into the corresponding ecological damage incident rectification file. The system establishes an independent electronic rectification file for each ecological damage incident, automatically collecting all process information such as the disposal plan, monitoring data from previous sessions, feedback materials from responsible units, and on-site verification records, achieving full lifecycle traceability of the incident.

[0044] Based on the ecological damage incident rectification archives, quantitative indicators for the rectification of ecological damage incidents are calculated. Combined with pre-set rectification progress early warning nodes, reminders are automatically sent before the rectification deadline, and the supervision level is automatically escalated for ecological damage incidents that are not completed by the deadline. In this embodiment, the quantitative indicators for rectification are set according to the type of damage: for vegetation clearing, monitoring vegetation coverage and NDVI restoration rate; for illegal construction due to industrial and mining development, monitoring the surface hardening demolition rate and terrain restoration rate; for illegal hunting, monitoring the reduction rate of hunting activity frequency and the vegetation restoration coverage rate of forest poaching trails. The system automatically compares each monitoring data with the rectification target value and calculates the percentage of rectification progress.

[0045] The warning nodes are set at 50%, 80%, and 100% of the rectification deadline: a progress reminder is sent when the rectification progress reaches 50%; a sprint reminder is sent when it reaches 80%; and a follow-up reminder is sent 7 days before the deadline. If the responsible unit fails to submit a rectification completion application by the deadline, it is judged as overdue and the system automatically upgrades the supervision level from the original competent department to the superior competent department, and generates an overdue supervision order and pushes it to the corresponding department.

[0046] After the responsible unit submits its rectification completion application, the rectification effectiveness is verified by combining the current multi-source original monitoring data with the on-site verification results. Once the verification is passed, the incident is closed and delisted. In this embodiment, the rectification effectiveness verification is divided into automatic verification and on-site verification. Automatic verification involves retrieving the latest multi-source original monitoring data, calculating the quantitative indicators after rectification, and comparing them with the rectification target values. If all indicators meet the rectification requirements, such as vegetation coverage recovering to more than 90% of the pre-damage level, hunting activity frequency decreasing to the baseline level, and all hardened surfaces being removed, then the automatic verification is passed. On-site verification involves dispatching local law enforcement personnel to the site after the automatic verification is passed to conduct on-site verification, take on-site photos, fill in verification records, and upload them to the system. Once both automatic and on-site verifications are passed, the system generates a rectification verification approval letter, completing the event closure and archiving the event file. If the verification fails, the system automatically generates a secondary rectification notice, resets the rectification deadline, and initiates the next round of rectification tracking.

[0047] The above description is merely a specific embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A big data-based decision-making system for identifying and addressing ecological damage in nature reserves, characterized in that: include: The model building and training module is used to build an ecological damage sample dataset and train an ecological damage multimodal identification and quantification model based on the sample dataset. The data collection and scheduling module is used to perform regular full-area inspections to collect multi-source raw monitoring data according to manual configuration, as well as to obtain ecological clues based on network big data retrieval and collect multi-source raw monitoring data of the target areas corresponding to the ecological clues in a targeted manner. The data processing module is used to preprocess multi-source raw monitoring data and generate standardized model input feature data; The intelligent identification and assessment module is used to input standardized model input feature data into the ecological damage multimodal identification and quantification model to determine the type of ecological damage; Conduct quantitative assessments of ecological damage targets, and combine spatial geographic ownership data to match responsible jurisdiction entities, generating complete information on ecological damage events; The decision generation module is used to build a decision generation model for handling incidents. It automatically generates a handling plan based on complete information about the ecological damage incident and pushes it to local management departments and front-line law enforcement units. The closed-loop tracking and control module is used to continuously collect multi-source raw monitoring data of ecological damage targets, dynamically monitor the rectification progress of ecological damage events corresponding to ecological damage targets, and complete the verification of rectification effectiveness.

2. The big data-based decision-making system for identifying and addressing ecological damage in nature reserves according to claim 1, characterized in that, The ecological damage multimodal identification and quantification model includes a visual feature extraction branch, a temporal feature extraction branch, a cross-modal attention fusion module, and a multi-task joint output head; The visual feature extraction branch adopts the Swin-Transformer v2 combined with the Feature Pyramid Network (FPN) backbone structure to extract spatial features from the standardized model input feature data and output a visual feature map. The spatial features are composed of spectral index features and texture spatial features spliced ​​and fused together. The temporal feature extraction branch uses a bidirectional long short-term memory network (Bi-LSTM) combined with a multi-head attention mechanism to extract temporal statistical features and frequency domain features from the standardized model input feature data and output a temporal feature vector. The cross-modal attention fusion module is used to map visual feature maps and temporal feature vectors to the same feature space, and realize bidirectional information interaction and feature fusion through a cross-attention mechanism, and output a fused feature tensor. The multi-task joint output head includes a classification output subheader and an instance segmentation output subheader, which are used to output the probability of ecological damage type and the pixel-level mask of the damage area.

3. The big data-based decision-making system for identifying and addressing ecological damage in nature reserves according to claim 1, characterized in that, The method for obtaining the ecological clues is as follows: The distributed crawler cluster is used to crawl publicly available text data from public public opinion data sources across the entire Internet in real time, and after deduplication and cleaning, a candidate text set is generated. A pre-trained BERT text classification model is used to perform binary classification of ecological damage relevance on candidate texts in the candidate text set. Location entities, damage type entities, and time entities are extracted from the candidate texts through named entity recognition. The extracted location entities are geocoded and converted to generate spatial coordinates. These coordinates are then spatially overlaid and filtered with the boundaries of nature reserves to retain spatial matching clues within the nature reserves and form candidate clues. A clue credibility assessment model is constructed, which calculates the clue credibility score of candidate clues by comprehensively considering the authority of the publishing entity, the completeness of information, the accuracy of historical reports, and the cross-validation of multiple sources. Candidate clues with a credibility score higher than the preset credibility threshold are identified as ecological clues, thereby triggering targeted collection tasks in the corresponding target areas.

4. The big data-based ecological damage identification and treatment decision-making system for nature reserves according to claim 1, characterized in that, The multi-source raw monitoring data includes at least remote sensing images; when IoT monitoring sensors are deployed in the target area or inspection area, the multi-source raw monitoring data also includes IoT time-series data, which is at least one of water quality monitoring data, atmospheric environment monitoring data, and soil environment monitoring data.

5. The big data-based decision-making system for identifying and addressing ecological damage in nature reserves according to claim 4, characterized in that, The data preprocessing is as follows: The remote sensing images are sequentially subjected to radiometric calibration, atmospheric correction, geometric fine correction, cloud shadow removal and mosaicking and cropping to obtain the spatial data of the remote sensing images. At the same time, the IoT time series data are sequentially subjected to outlier removal, missing value completion and uniform granularity resampling to obtain standardized time series monitoring data. The spatial data of remote sensing images are unified to a preset standard geographic coordinate system, and standardized time-series monitoring data for the corresponding time period are extracted based on the remote sensing image capture time to complete the spatiotemporal alignment and matching of the data. Spatial features are extracted from the spatial data of the spatiotemporally aligned remote sensing images. At the same time, temporal statistical features and frequency domain features are extracted from the standardized time-series monitoring data after spatiotemporal alignment. Based on these features, they are integrated to form standardized model input feature data. If there is no IoT time-series data, then zeros are uniformly padded to fill the time-domain statistical features and frequency-domain features.

6. The big data-based decision-making system for identifying and addressing ecological damage in nature reserves according to claim 1, characterized in that, The method for conducting quantitative assessment of ecological damage is as follows: Based on pixel-level masks of the damaged areas, the actual area, center point coordinates, and boundary vectors of the damaged patches are calculated in combination with the resolution of remote sensing images, so as to complete the quantitative measurement of the spatial range of ecological damage targets. A differentiated quantitative indicator system is constructed, and the comprehensive index method is used to calculate the comprehensive ecological damage score of ecological damage targets. The comprehensive ecological damage score is mapped and divided into mild, moderate and severe ecological damage levels to complete the quantitative assessment of the damage intensity level of ecological damage targets. Simultaneously, for ecological damage targets classified as illegal hunting, a population impact assessment model of hunting activities is constructed by combining species distribution data and habitat connectivity data of the areas where the ecological damage targets are located. This model is used to extrapolate the potential impact range of hunting activities on species populations and the degree of habitat disturbance, so as to complete a quantitative assessment of the overall impact of illegal hunting on ecological damage targets.

7. The big data-based decision-making system for identifying and addressing ecological damage in nature reserves according to claim 6, characterized in that, The differentiated quantitative indicator system is as follows: If the ecological damage target corresponds to the type of ecological damage, such as vegetation felling, then calculate the vegetation coverage decline rate, NDVI attenuation difference, and biomass loss index. If the ecological damage target corresponds to the type of ecological damage, such as illegal construction in industrial and mining development, then calculate the indicators of surface hardening ratio, surface disturbance intensity, and estimated earthwork excavation volume. If the ecological damage target corresponds to the type of ecological damage, namely illegal hunting, then the frequency of hunting activities is calculated. Number and poaching disturbance density index of trail development.

8. The big data-based decision-making system for identifying and addressing ecological damage in nature reserves according to claim 1, characterized in that, The process of constructing the disposal decision model and generating disposal solutions is as follows: An ecological big data knowledge base is pre-built, including a legal and regulatory database, a disposal process database, a historical case database, and a law enforcement resource database, and an ecological disposal knowledge graph is constructed based on entity relationships; Based on complete information on ecological damage events, risk level assessments of these events are conducted from the dimensions of damage intensity, ecological sensitivity, and social impact. The analytic hierarchy process (AHP) combined with the entropy weight method is used to calculate the comprehensive risk value of the ecological damage events and classify them into risk levels. The rule engine matches the illegal clauses and legal handling requirements corresponding to ecological damage events, the knowledge graph retrieves similar historical ecological damage event handling cases to optimize handling details, and the spatial analysis engine schedules local law enforcement resources and law enforcement action paths. The system automatically generates standardized handling plans that include the facts of the violation, legal basis, handling measures, rectification requirements, completion deadlines, and responsible units, and pushes them out in a tiered manner according to risk level, simultaneously distributing them to the corresponding local management departments and front-line law enforcement units.

9. The big data-based decision-making system for identifying and addressing ecological damage in nature reserves according to claim 1, characterized in that, The method for verifying the effectiveness of the rectification is as follows: Differentiated monitoring frequencies are configured according to the risk level of ecological damage events, and monitoring equipment is dispatched to continuously collect multi-source raw monitoring data of ecological damage targets, and the multi-source raw monitoring data is automatically collected into the corresponding ecological damage event rectification file; Based on the rectification archives of ecological damage incidents, quantitative indicators for the rectification of ecological damage incidents are calculated. Combined with pre-set early warning nodes for rectification progress, reminders are automatically sent before the deadline for rectification. For ecological damage incidents that are not completed by the deadline, the supervision level is automatically upgraded. After the responsible unit submits an application for completion of rectification, the effectiveness of the rectification will be verified by combining the current multi-source original monitoring data with the on-site verification results. Once the verification is passed, the incident will be closed out.

Images