A data analysis-based forestry pest control risk assessment system

Abstract

The application provides a forestry pest control risk assessment system based on data analysis, and belongs to the technical field of intelligent forestry pest control, and comprises the following steps: constructing a whole-cycle management and control logic map and adopting three-layer progressive coupling processing, and fusing multi-modal data of site, insect situation and environment; generating high-quality samples through multi-factor environment compensation correction, feature completion and computing power self-adaptive parallel preprocessing; adopting multi-granularity loss constraint to train a spatio-temporal graph convolution network model, realizing risk quantification and grading; completing disaster tracing based on a global knowledge graph, and outputting a targeted prevention and control scheme through three-source decision fusion and multi-objective conflict resolution. The application improves the early warning window period to 15-30 days, greatly reduces data deviation and false positive rate, significantly improves processing efficiency and prevention and control effect, can realize whole-cycle, high-precision and self-adaptive active prevention and control of forestry pests, and is suitable for normal management and control of various stands and pests.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for forest pests, and in particular to a risk assessment system for forest pest control based on data analysis. Background Technology

[0002] Forest pests are one of the core threats to my country's forestry ecological security. While current risk assessment technologies for forest pest control have gradually incorporated intelligent methods such as machine learning and big data analysis, existing risk assessments often focus on emergency warnings after a disaster occurs, failing to cover the entire life cycle of pests. This results in insufficient foresight, an inability to identify pre-emptive causes and manage the entire process, poor training sample quality, and consequently, low assessment efficiency, failing to meet the needs of forest area assessments.

[0003] Therefore, this invention proposes a data analysis-based risk assessment system for the control of forest pests. Summary of the Invention

[0004] This invention provides a data analysis-based risk assessment system for the prevention and control of forest pests, in order to solve the aforementioned technical problems.

[0005] This invention provides a data analysis-based risk assessment system for forest pest control, comprising:

[0006] The map construction module is used to construct a full-cycle management logic map of forest pest control that matches the target forest area. Each management node in the full-cycle management logic map is processed in three layers. The first layer is used to lock the site conditions and basic attributes of the forest stand. The second layer is used to lock the population dynamics of pests and disaster occurrence monitoring data. The third layer is used to lock the meteorological, hydrological and environmental stress factor data of the forest area.

[0007] The sample acquisition module is used to acquire multimodal full-cycle historical datasets of sample forest areas with multiple different forest stand types, and combine them with the processed full-cycle management logic map to generate a training sample set for model training.

[0008] The model training module is used to perform supervised training on the neural network model using the training sample set to obtain a converged forestry pest control risk assessment model.

[0009] The model analysis module is used to collect real-time multimodal monitoring data of the target forest area, perform three-layer processing on the real-time multimodal monitoring data, input the processed real-time data into the converged forest pest control risk assessment model, and output the real-time control risk quantification value and standardized risk classification result of the target forest area.

[0010] The prevention and control decision module is used to match the real-time prevention and control risk quantification value with the standardized risk classification result based on the pre-constructed forest pest control knowledge graph, identify the core risk factors corresponding to the target forest area, and output targeted prevention and control decisions that match the core risk factors.

[0011] Preferably, the full-domain knowledge graph stores biological characteristic data of all types of forest pests, data on the causes of disasters throughout their entire lifecycle, data on standardized prevention and control technical procedures in the forestry industry, correlation mapping data between disaster events of different forest stand types and corresponding prevention and control schemes, and data on the correspondence between prevention and control operation parameters and control effects.

[0012] The full-cycle management logic diagram includes multiple interconnected full-cycle management nodes divided according to forestry standard prevention and control technical regulations.

[0013] Preferably, the sample acquisition module includes:

[0014] The bias correction submodule is used to construct a multi-factor environmental compensation correction model based on the altitude, aspect, canopy density, and extreme weather event data of the sample forest area. The multi-factor environmental compensation correction model is used to correct the bias of environmental stress factor data, pest population dynamics, and disaster occurrence monitoring data in the multimodal full-cycle historical dataset.

[0015] The completion submodule is used to complete control nodes with dimensions lower than the maximum effective feature dimension of the same data layer of all control nodes in the processed full-cycle control logic graph. For control nodes with dimensions lower than the benchmark, the module uses the global feature mean of the corresponding forest area sample in the corrected multimodal full-cycle historical dataset to complete the control nodes, so that all three layers of all control nodes have a unified feature dimension. Only when the missing effective feature data rate in the corresponding layer exceeds a preset threshold is the module used to complete the control nodes with zero values.

[0016] The initial set construction submodule is used to construct a spatiotemporal heterogeneous feature tensor based on the corrected multimodal full-cycle historical dataset and the completed full-cycle management logic map, integrating the spatial geographic dimension of forest area, the time dimension of disaster occurrence, and the three-layer feature dimension. At the same time, using the full-cycle record data of historical disaster events and the verification data of prevention and control effect corresponding to the sample forest area as label data, labeled training samples are generated to form the initial training sample set.

[0017] The data acquisition submodule is used to collect the hardware operating status data of the current computing device in real time and determine the computing load saturation. The computing load saturation is a weighted value obtained by combining CPU utilization, memory usage, and parallel computing unit load rate.

[0018] The first judgment submodule is used to directly perform single-process normalization and standardization preprocessing on the spatiotemporal heterogeneous feature tensor of the initial training sample set if the computing power load saturation is less than the preset load threshold, so as to obtain the final training sample set.

[0019] The second judgment submodule is used to distribute the data corresponding to the spatiotemporal heterogeneous feature tensor to each independent processing process to perform parallel normalization and standardization preprocessing if the computing power load saturation is greater than or equal to a preset load threshold, and merge the output results of all independent processing processes to obtain the final training sample set.

[0020] Preferably, the second judgment submodule includes:

[0021] The dependency unit is used to construct an initial hierarchical graph of depth 1, with each control node of the completed full-cycle control logic graph as the core node, and the direct dependency relationship corresponding to each core node as the directed edge, and to perform structural optimization on the initial hierarchical graph to obtain the dependency graph.

[0022] The communication pool construction unit is used to determine the optimal number of parallel subtasks based on the hardware operating status data, create independent processing processes equal to the optimal number of parallel subtasks, and establish a high-speed communication pool based on shared memory for each process.

[0023] The sharding and process correspondence unit is used to perform targeted sharding of spatiotemporal heterogeneous feature tensors according to the node business attributes based on the business logic and data volume of each control node in the dependency graph and the optimal number of parallel subtasks, so that each data shard corresponds one-to-one with an independent processing process, and the shard data volume matches the hardware load capacity of the corresponding independent processing process.

[0024] The retrieval unit is used to extract feature anchor points of each data shard in parallel, calculate the data balance of different data shards, optimize the shard allocation strategy with the goal of maximizing the data balance, and schedule mapping tasks and reduction tasks with the same task label to the same processing process.

[0025] The data sharding and distribution unit is used to distribute each data shard to the corresponding independent processing process based on the high-speed communication pool, perform parallel normalization preprocessing and risk feature extraction, and merge the output results of all independent processing processes to obtain the final training sample set.

[0026] Preferably, the dependent unit includes:

[0027] The distribution subunit is used to traverse the business description text of each core node in the initial hierarchical diagram, extract several logical objects, and find the first distribution that has a parent relationship with each logical object and the second distribution that has a child relationship from the business description text of each core node.

[0028] The refinement subunit is used to refine the first distribution and the second distribution of the same logical object respectively to obtain the corresponding third distribution and the fourth distribution, and to calculate the first expansion coefficient of the first distribution and the third distribution, and the second expansion coefficient of the second distribution and the fourth distribution.

[0029] The extended subunit is used to determine the number of extensions of the corresponding logical object based on the first extension coefficient and the second extension coefficient, which is regarded as the first number, and to extract the first number of dimensions with the highest confidence from the corresponding refinement processing results as the object extension dimension.

[0030] The description subunit is used to perform overlapping analysis on the extended dimensions of all objects under the same core node to obtain the semantic extended dimensions of the corresponding core node. Combined with the data scale characteristics, computing load characteristics and upstream and downstream dependency characteristics of the corresponding core node, the extended feature description set of the corresponding core node is obtained.

[0031] The mining subunit is used to mine implicit and explicit business relationships between nodes based on the extended feature description set of each core node, and to construct node optimization functions according to the implicit logical relationship of the implicit business relationship and the explicit logical relationship of the explicit business relationship.

[0032] The iterative calculation subunit is used to iteratively calculate the node optimization function according to the input logical query conditions and the output logical query conditions, determine the optimization parameter values, and obtain supplementary associated nodes by combining the parameter types of each optimization parameter value;

[0033] The position determination subunit is used to determine the initial optimization solution of the supplementary associated nodes and each core node, and combined with the main node logic of the corresponding core node, to obtain the position of the supplementary associated node of the corresponding supplementary associated node, and to supplement the corresponding supplementary associated node into the initial hierarchical graph according to the position of the supplementary node to obtain the dependency graph.

[0034] Preferably, the prevention and control decision module includes:

[0035] The boundary determination submodule is used to determine that there is no risk when the risk quantification value is lower than the minimum preset threshold, and output a normalized monitoring and early warning decision.

[0036] When the risk quantification value is higher than the highest preset threshold, it is judged as a major emergency risk, and an emergency prevention and control decision is output.

[0037] When the risk quantification value is between the lowest preset threshold and the highest preset threshold, it is determined to be a normal preventable risk and enters the multi-source decision fusion optimization stage;

[0038] The decision generation submodule is used to output preliminary handling decisions and corresponding decision confidence levels for routine preventable risks through three analysis sources. Combined with the prevention and control effect mapping data in the pre-constructed global knowledge graph, the multi-source decision results are weighted and fused and conflict resolution is processed to generate global targeted prevention and control decisions.

[0039] Preferably, the first analysis source is a site suitability analysis unit obtained by matching forest stand basic attributes with risk factors, corresponding to the feature dimensions of the first site foundation layer; the second analysis source is a targeted prevention and control analysis unit obtained by matching pest population characteristics with disaster development trends, corresponding to the feature dimensions of the second disaster core layer; and the third analysis source is an environmental intervention analysis unit obtained by matching environmental stress factors with disaster diffusion conditions, corresponding to the feature dimensions of the third environmental stress layer.

[0040] The global targeted prevention and control decision includes the prevention and control operation plan, decision confidence level, risk severity level, unique geographic identifier of the corresponding forest area plot, and threshold range of operation parameters. The decision confidence level is calculated by weighting the proportion of effective data in the corresponding analysis source feature dimension and the matching degree with historical successful cases in the global knowledge graph. The weight of the weighted fusion is the normalized value of the corresponding decision confidence level.

[0041] The conflict resolution aims to achieve the highest prevention and control efficiency, the least environmental impact, and the lowest prevention and control cost, and prioritizes and adjusts the results of multi-source decisions.

[0042] Preferably, the model training module includes:

[0043] The mechanism construction submodule is used to construct a multi-granularity loss constraint mechanism by using the spatiotemporal heterogeneous feature tensor of the training sample set as the model input of the neural network model and the historical disaster level of forestry pests and the control failure label as the supervision signal.

[0044] The gradient backpropagation submodule is used to independently backpropagate the feature channels corresponding to each control node of the processed full-cycle control logic graph during the forward propagation phase, and synchronously execute distributed gradient aggregation according to the hardware parallel process allocation results.

[0045] The learning rate adjustment submodule is used to dynamically adjust the learning rate based on the dual indicators of risk prediction error and control node feature fitting degree of the neural network model on the validation set after each iteration. When the dual indicators are consistently lower than the preset convergence threshold and gradient oscillations no longer occur in multiple consecutive iterations, the model is determined to have reached the convergence state, and a forestry pest control risk assessment model adapted to the heterogeneous data characteristics of forest areas is obtained.

[0046] Compared with the prior art, the beneficial effects of this application are as follows:

[0047] It has achieved proactive full-cycle management of pests and significantly improved the risk warning window: It has constructed a full-cycle management logic map based on the dual-dimensional coupling of pest life history and forestry control process. Through three-layer progressive coupling processing, it has realized the business association of multi-source data, which solves the core defect of existing technology that only focuses on warnings after disasters occur and lacks foresight. It can advance the warning of pest risks by 15-30 days, reduce the false alarm rate, and realize the transformation from "passive emergency response" to "proactive prevention and control".

[0048] The data bias caused by the heterogeneity of forest areas has been resolved, and the generalization ability and prediction accuracy of the model have been greatly improved: a multi-factor environmental compensation correction model based on the biological characteristics of harmful organisms has been constructed, and a unified representation of multi-source heterogeneous data in forest areas has been achieved through spatiotemporal heterogeneous feature tensors, improving the quality of training samples by more than 80%; the constructed multi-granularity loss-constrained ST-GCN model has solved the problems of sample imbalance and insufficient feature fitting, improving the accuracy of risk prediction, but the accuracy decreases when applied across forest areas.

[0049] It achieves efficient data preprocessing with adaptive computing power and solves the dependency conflict problem of parallel processing: Based on the business dependency graph, it realizes targeted data sharding, divides nodes with strong dependencies into the same shard, completely avoids the dependency conflict of parallel processing, and adaptively selects the preprocessing mode based on hardware computing power load. The data processing efficiency is more than 12 times higher than that of single process and more than 5 times higher than that of random sharding parallel processing.

[0050] It achieves deep coupling of risk tracing and targeted decision-making, significantly improving prevention and control effectiveness and economy: Based on the full-domain knowledge graph, it realizes accurate tracing of risk causes. Through the integration of three-source decision-making with three-layer feature dimensions, it ensures accurate matching between decision-making and causes. Through conflict resolution of multi-objective optimization, it solves the conflict problem of multi-source decision-making. The output prevention and control plan improves the reduction rate of pest populations, while reducing prevention and control costs, taking into account prevention and control effectiveness, ecological security and economy.

[0051] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0052] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a structural diagram of a forestry pest control risk assessment system based on data analysis, as described in an embodiment of the present invention.

[0055] Figure 2 This is a structural diagram of the dependent unit in an embodiment of the present invention. Detailed Implementation

[0056] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0057] This invention provides a data analysis-based risk assessment system for forest pest control, such as... Figure 1 As shown, it includes:

[0058] The map construction module is used to construct a full-cycle management logic map of forest pest control that matches the target forest area. Each management node in the full-cycle management logic map is processed in three layers. The first layer is used to lock the site conditions and basic attributes of the forest stand. The second layer is used to lock the population dynamics of pests and disaster occurrence monitoring data. The third layer is used to lock the meteorological, hydrological and environmental stress factor data of the forest area.

[0059] The sample acquisition module is used to acquire multimodal full-cycle historical datasets of sample forest areas with multiple different forest stand types, and combine them with the processed full-cycle management logic map to generate a training sample set for model training.

[0060] The model training module is used to perform supervised training on the neural network model using the training sample set to obtain a converged forestry pest control risk assessment model.

[0061] The model analysis module is used to collect real-time multimodal monitoring data of the target forest area, perform three-layer processing on the real-time multimodal monitoring data, input the processed real-time data into the converged forest pest control risk assessment model, and output the real-time control risk quantification value and standardized risk classification result of the target forest area.

[0062] The prevention and control decision module is used to match the real-time prevention and control risk quantification value with the standardized risk classification result based on the pre-constructed forest pest control knowledge graph, identify the core risk factors corresponding to the target forest area, and output targeted prevention and control decisions that match the core risk factors.

[0063] Preferably, the full-domain knowledge graph stores biological characteristic data of all types of forest pests, data on the causes of disasters throughout their entire lifecycle, data on standardized prevention and control technical procedures in the forestry industry, correlation mapping data between disaster events of different forest stand types and corresponding prevention and control schemes, and data on the correspondence between prevention and control operation parameters and control effects.

[0064] The full-cycle management logic diagram includes multiple interconnected full-cycle management nodes divided according to forestry standard prevention and control technical regulations.

[0065] In this embodiment, the minimum sample size required for model training is only monitoring data from 3 forest areas of the same type and 2 complete life cycle periods, which is lower than the requirement of dozens of sample forest areas and more than 5 years of data required by existing technologies.

[0066] In this embodiment, the minimum sample size required for model training is only monitoring data from 3 forest areas of the same type and 2 complete life cycle periods, which is far less than the requirement of dozens of sample forest areas and more than 5 years of data required by existing technologies.

[0067] In this embodiment, the full life cycle refers to the complete life history cycle of forest pests from overwintering recovery → population expansion → disaster occurrence → control operations → effectiveness verification → overwintering dormancy, synchronously matching the entire annual prevention and control process of forestry departments. The multimodal full-cycle historical dataset contains four core modalities: numerical modalities (structured monitoring data such as site, pest, and meteorological data); spatial geographic modalities (forest area GIS vector data, plot grid data); image modalities (UAV remote sensing images, ground sample plot images); and textual modalities (forest survey records, control operation logs, disaster event reports).

[0068] The specific method for constructing the full-cycle control logic graph is as follows:

[0069] Initial nodes: overwintering generation monitoring node, population dynamics monitoring node, disaster early warning node, prevention and control operation management node, prevention and control effect verification node, and pre-overwintering monitoring node. For each management node, upstream and downstream directed dependencies are established, forming an initial full-cycle management logic graph with a directed acyclic graph (DAG) structure, as shown in Table 1.

[0070] Table 1 Initial Node Table

[0071]

[0072] The target forest area adaptation process involves adjusting the initial node database based on the forest stand type, dominant tree species, and dominant pest species of the target forest area. The general adjustment rules are as follows:

[0073] Rules for adding nodes: When there are specific monitoring needs for quarantine pests, vector insects, and damage to fruits in economic forests in the target forest area, corresponding specific monitoring nodes shall be added. The added nodes shall clearly define the upstream and downstream dependencies, triggering conditions and standardized attributes. For example, for pure pine forests in the pine wilt disease epidemic area, a new monitoring node for the vector insect pine sawyer beetle, N01-1, shall be added. The upstream node is the N01 overwintering generation monitoring node, and the downstream node is the N02 population dynamic monitoring node. For walnut economic forests, a new monitoring node for fruit damage, N03-1, shall be added. The upstream node is the N03 disaster occurrence early warning node, and the downstream node is the N04 prevention and control operation management node.

[0074] Node deletion rules: When there is no history of corresponding pest occurrences or related business needs in the target forest area, non-core nodes can be deleted. After deletion, it is necessary to ensure that the business links of the remaining nodes are complete and there are no logical breaks. For example, for protective forests without fruit damage needs, fruit damage monitoring nodes can be deleted.

[0075] Node parameter adaptation rules: Adjust the business execution time limit and trigger condition threshold of each node according to the climate zone and life history of pests in the target forest area; for example, for the Northeast cold temperate forest area, the start time of the overwintering generation monitoring node is adjusted to late April, and for the South China tropical forest area, it is adjusted to early February; for quarantine pests, the trigger threshold is lowered by 20% to improve the sensitivity of early warning.

[0076] Map Verification and Consolidation: The adapted map is verified for completeness. The verification rules are as follows: there are no breaks in the business chain, and all nodes have clear upstream and downstream dependencies; there are no logical conflicts in the triggering conditions and no circular dependencies; the attributes of all nodes are complete and conform to forestry industry standards; after the verification is passed, it is consolidated into a full-cycle management logic map that matches the target forest area.

[0077] In this embodiment, a unified three-layer progressive coupling process is performed on each control node in the full-cycle control logic graph. The specific execution steps, coupling logic, and output standards are as follows:

[0078] Each layer of processing performs four standardized steps: data acquisition, validity verification, feature selection, and dimension mapping. Locking the data refers to outputting the standardized valid feature dataset for that layer after completing these four steps, as detailed below:

[0079] The first layer of processing (site foundation layer) locks in the site conditions and basic attributes of the forest stand. The core feature dimensions are fixed at 8 items, namely: altitude, aspect, slope, canopy closure, dominant tree species, age structure, stand density, soil type, and soil fertility (the baseline dimension is 8, of which the dominant tree species is a categorical feature using one-hot encoding, and the rest are continuous features); the data sources are forest resource survey data, UAV aerial survey data, and ground fixed sample plot survey data; invalid feature values ​​are replaced with the feature mean of adjacent plots in the same forest area.

[0080] The second layer of processing (core disaster layer) focuses on pest population dynamics and disaster occurrence monitoring data. The core feature dimensions are fixed at six items: pest population density, infested plant rate, disease-susceptible plant rate, pest developmental progress, disaster-affected area, and disaster severity. Data sources include ground-based fixed plot monitoring data, pheromone trap monitoring data, UAV hyperspectral remote sensing data, and forest survey data. Outliers are handled using... Criteria for elimination are followed by completion using the mean of adjacent monitoring points in the same time series.

[0081] The third layer of processing (environmental stress layer) focuses on meteorological and hydrological data and environmental stress factors in the forest area. The core feature dimensions are fixed at 7 items: ten-day average temperature, ten-day cumulative precipitation, relative humidity, intensity of extreme weather events, wind speed, forest light intensity, and groundwater level. The data sources are automatic weather station data in the forest area, shared data from the National Meteorological Science Data Center, and hydrological monitoring station data. The intensity of extreme weather events is quantified by: the number of days of the event × the standard deviation of the meteorological element from the historical average for the same period, with a value range of 0 to 5.

[0082] In this embodiment, the three layers of data in the three-layer progressive coupling logic do not exist independently, but rather undergo progressive coupling processing to achieve deep association of business logic. The specific coupling rules are as follows:

[0083] The coupling between the first and second layers: Based on the data of the first layer (foundation layer), the sampling threshold of the second layer (hazard core layer) is determined. The general calculation formula is as follows: In the formula: This is the sampling interval threshold for this node (unit: m). C represents the basic sampling interval specified by forestry industry standards, and C represents the actual canopy closure of the target forest stand. H represents the optimal canopy closure for this harmful organism, where H is the actual elevation (in meters) of the target forest area. This is the optimal altitude for this harmful organism. The unit is m; this formula enables adaptive adjustment of sampling strategies under different site conditions, ensuring the representativeness of the second-layer monitoring data.

[0084] The coupling between the third and second layers: Using the data from the third layer of environmental stress as a correction factor, the population dynamics data from the second layer of disaster core layer are trend-corrected to restore the true development trend of the invasive species population. The correction formula is as follows: In the formula: This is the corrected population dynamics trend value. These are the original monitoring values, where T is the ten-day average temperature during the monitoring period. The temperature is the optimal temperature for the development of this harmful organism, and P is the cumulative precipitation during the monitoring period. This is the optimal precipitation level for the development of this harmful organism; this formula eliminates the interference of short-term fluctuations in environmental factors on the judgment of population trends.

[0085] The output standard for the three-layer processing is as follows: After each control node completes the three-layer processing, it outputs a feature dataset in a unified format. The feature dimension is fixed at 8+6+7=21 dimensions, which is completely consistent with the dimensions of subsequent sample generation and model input. The output data must pass the dimension consistency check to ensure that the three-layer feature dimensions of all control nodes are completely consistent.

[0086] In this embodiment, the global knowledge graph serves as the knowledge foundation of the entire system, forming a dual-graph mapping relationship with the full-cycle management logic graph. The complete construction method is as follows:

[0087] Ontology layer construction: Defines standardized core entity types, entity attributes, and semantic relationships between entities, as follows:

[0088] Core entity types: harmful organism entities, host tree species entities, environmental factor entities, disaster event entities, control technology entities, control agents entities, operational equipment entities, and forest site entities;

[0089] Entity attributes: Define standardized attributes for each entity. For example, the attributes of pest entities include scientific name, taxonomic status, host range, life history, biological characteristics, suitable environmental threshold, and control indicators; the attributes of control agents entities include active ingredient, registered target pests, dosage, application method, safety interval, environmental toxicity, and applicable insect stage.

[0090] Semantic relationships between entities: Define standardized business relationships, including host relationships (pests → host tree species), suitability relationships (pests → environmental factors), triggering relationships (environmental factors → disaster events), control relationships (control technologies → pests), adaptation relationships (control agents → operating equipment), applicability relationships (control technologies → forest site), and effect relationships (control plans → control effects).

[0091] Data layer construction: Named entity recognition is completed using the BERT-BiLSTM-CRF model, and relation extraction is completed using the PCNN model. After knowledge fusion, entity disambiguation, and rule reasoning, a full-domain knowledge graph is constructed. The graph contains more than 120,000 entities and more than 380,000 relations, covering all mainstream species of forest pests and standardized prevention and control schemes.

[0092] Dual-graph mapping association rules: Each control node in the full-cycle control logic graph establishes a one-to-one mapping relationship with an entity in the full-domain knowledge graph. For example, the overwintering generation monitoring node is mapped to the overwintering stage and suitable environment attributes of the harmful organism entity; the prevention and control operation control node is mapped to the entities of prevention and control technology, prevention and control agents, and operation equipment. Through the mapping relationship, the three-layer processing data of each node in the full-cycle control logic graph can be matched, retrieved, and reasoned in the full-domain knowledge graph, providing knowledge support for subsequent sample correction, risk tracing, and decision output.

[0093] In this embodiment, the specific implementation process and effect verification of the risk assessment for the control of pine caterpillars are as follows:

[0094] Overview of the target forest area: Area of ​​the target forest area The forest stand type is a pure Pinus tabuliformis forest, with Pinus tabuliformis as the dominant tree species. The average tree age is 18 years and the average canopy closure is 0.7. The main pest is Pinus tabuliformis caterpillar, which is a key area for the prevention and control of forest pests.

[0095] Construction of a full-cycle management logic map: Based on the life history of the Pinus tabuliformis caterpillar (overwintering generation recovers in April, pupates in May-June, adults lay eggs in July-August, and larvae overwinter in September), 6 core management nodes were constructed, and a special monitoring node for the population density of the overwintering generation of Pinus tabuliformis caterpillar was added to complete the construction of a full-cycle management logic map adapted to the target forest area.

[0096] Dataset and Model Training: A multimodal, full-cycle historical dataset from 2020 to 2023 was obtained from 20 pure Pinus tabuliformis forest samples. Data correction, feature completion, and tensor construction were performed to generate a training sample set. The ST-GCN model was trained using this training sample set. The model reached convergence after 42 iterations, and the validation set MAPE was 3.2%. It is 0.972;

[0097] Real-time risk assessment and verification: Real-time monitoring data of the target forest area was collected in April 2024. After inputting the data into the model, the output risk quantification value was 62, which was determined to be a conventional and preventable risk. The core risk factors were identified as: high overwintering insect population density, rapid rise in spring temperature, and high forest canopy closure.

[0098] Targeted Decision Output and Effect Verification: A multi-source decision fusion approach was used to output the following control measures: Site-appropriate measures: Sanitary felling, clearing weak trees, and reducing stand canopy closure to 0.6; Targeted control measures: Releasing 300,000 Trichogramma wasps / For young larvae, spray with 1.2% nicotine·matrine emulsifiable concentrate diluted 1500 times; environmental intervention measures: increase the monitoring frequency to once every 5 days, and set up sex traps to monitor adult dynamics;

[0099] Results: After the control operation was implemented, the percentage of trees infested with pine caterpillars decreased from 18.2% to 2.1%, and the pest population reduction rate reached 92.3%. No large-scale disasters occurred, and the control effect was significantly better than the local conventional control plan (the conventional plan has a pest population reduction rate of about 70%).

[0100] The beneficial effects of the above technical solution are: it constructs a logical map of the whole life cycle management of forest pests and a core architecture of three-layer progressive coupling processing, opens up a closed loop of the whole business from data processing, model training, risk assessment to targeted decision-making, significantly shortens the pest risk warning window from 3-7 days to 15-30 days, solves the core pain points of existing technology such as lagging early warning and passive prevention and control, and realizes a fundamental transformation from passive emergency response to proactive prevention and control.

[0101] This invention provides a data analysis-based risk assessment system for forest pest control, wherein the sample acquisition module includes:

[0102] The bias correction submodule is used to construct a multi-factor environmental compensation correction model based on the altitude, aspect, canopy density, and extreme weather event data of the sample forest area. The multi-factor environmental compensation correction model is used to correct the bias of environmental stress factor data, pest population dynamics, and disaster occurrence monitoring data in the multimodal full-cycle historical dataset.

[0103] The completion submodule is used to complete control nodes with dimensions lower than the maximum effective feature dimension of the same data layer of all control nodes in the processed full-cycle control logic graph. For control nodes with dimensions lower than the benchmark, the module uses the global feature mean of the corresponding forest area sample in the corrected multimodal full-cycle historical dataset to complete the control nodes, so that all three layers of all control nodes have a unified feature dimension. Only when the missing effective feature data rate in the corresponding layer exceeds a preset threshold is the module used to complete the control nodes with zero values.

[0104] The initial set construction submodule is used to construct a spatiotemporal heterogeneous feature tensor based on the corrected multimodal full-cycle historical dataset and the completed full-cycle management logic map, integrating the spatial geographic dimension of forest area, the time dimension of disaster occurrence, and the three-layer feature dimension. At the same time, using the full-cycle record data of historical disaster events and the verification data of prevention and control effect corresponding to the sample forest area as label data, labeled training samples are generated to form the initial training sample set.

[0105] The data acquisition submodule is used to collect the hardware operating status data of the current computing device in real time and determine the computing load saturation. The computing load saturation is a weighted value obtained by combining CPU utilization, memory usage, and parallel computing unit load rate.

[0106] The first judgment submodule is used to directly perform single-process normalization and standardization preprocessing on the spatiotemporal heterogeneous feature tensor of the initial training sample set if the computing power load saturation is less than the preset load threshold, so as to obtain the final training sample set.

[0107] The second judgment submodule is used to distribute the data corresponding to the spatiotemporal heterogeneous feature tensor to each independent processing process to perform parallel normalization and standardization preprocessing if the computing power load saturation is greater than or equal to a preset load threshold, and merge the output results of all independent processing processes to obtain the final training sample set.

[0108] In this embodiment, a multi-factor environmental compensation correction model is used to correct the bias in the historical dataset, eliminating systematic data bias caused by site heterogeneity in the forest area. The specific implementation method is as follows:

[0109] Model input and output definitions: Input factors: elevation H (unit: m), aspect A, canopy closure C, and extreme weather event intensity E of the sample forest area; aspect A is quantified using the aspect cosine value, with 1 for sunny slopes (south-facing) and -1 for shady slopes (north-facing), and semi-sunny / semi-shady slopes are linearly converted by angle; extreme weather event intensity E is quantified using "event duration days × standard deviation of meteorological elements from the historical average for the same period", with a value range of 0~5; Output parameters: environmental compensation correction coefficient K, and corrected data on the dynamics of harmful organism populations and disaster occurrence monitoring, and environmental stress factor data.

[0110] Correction factor calculation formula: In the formula:

[0111] The altitude correction factor is calculated using the following formula: , The optimal average altitude for corresponding harmful organisms is derived from the global knowledge graph.

[0112] The slope aspect correction factor is calculated using the following formula: The setting is based on the biological characteristics of sunny slopes, where temperatures are higher and harmful organisms develop faster.

[0113] The canopy closure correction factor is calculated using the following formula: , The optimal canopy closure for corresponding pests is derived from the whole-domain knowledge graph;

[0114] The correction factor for extreme weather events is calculated using the following formula: The setting is based on the suppressive effect of extreme weather events on harmful organism populations;

[0115] =0.25、 =0.15、 =0.3、 =0.3 represents the weight of each correction coefficient, which was determined by combining the scores from 15 senior experts in the field of forestry pest control using the analytic hierarchy process (AHP). The sum of the weights is 1. In this embodiment, the expert scores were given using the 1-9 scale and passed the consistency test (CR=0.07<0.1) to ensure the scientific validity and repeatability of the weights.

[0116] In this embodiment, data deviation correction is performed as follows:

[0117] Environmental stress factor data correction formula: ,in, These are the original monitoring values. This is the average value for the same period in the area where the forest is located;

[0118] Formula for correcting population dynamics data of harmful organisms: This will eliminate biases in monitoring data caused by environmental factors and restore the true development trend of harmful organism populations.

[0119] In this embodiment, the baseline dimension is determined as follows: the maximum effective feature dimension of the same data layer of all control nodes in the processed full-cycle control logic graph is used as the baseline, specifically: the first layer of the foundation layer is the baseline dimension 8, the second layer of the disaster core layer is the baseline dimension 6, and the third layer of the environmental stress layer is the baseline dimension 7. All three data layers of all control nodes need to be unified to the corresponding baseline dimension.

[0120] Missing value completion rules: First, calculate the missing rate of valid feature data within the corresponding layer: Missing rate = Number of missing features / Dimension of the baseline feature of this layer; the preset missing rate threshold is 30%, and the following completion rules are executed:

[0121] When the missing rate is ≤30%, the mean global feature of the corresponding whole forest area sample in the corrected multimodal full-cycle historical dataset is used for imputation. The formula for calculating the mean global feature is as follows: ,in, is the effective value of the feature corresponding to the i-th sample forest area, and N is the number of effective samples;

[0122] When the missing rate is greater than 30%, it is determined that there is insufficient effective data in this layer, and zero values ​​are used for imputation. At the same time, in the subsequent tensor construction, a weight decay coefficient of 0.5 is set for the feature channels of this node. This coefficient is applied to the feature channel weight update during the model training phase to reduce the interference of invalid data on model training. The weight decay is implemented by multiplying the weight update gradient of this channel by 0.5 during the backpropagation phase to slow down the weight update magnitude of invalid features.

[0123] Dimension verification rules: After completion, verify the feature dimensions of the three data layers of all control nodes to ensure that they are completely consistent with the baseline dimensions. After the verification is passed, the feature dimension unification is completed.

[0124] In this embodiment, a spatiotemporal heterogeneous feature tensor is constructed to generate an initial training sample set. The specific implementation method is as follows:

[0125] The spatiotemporal heterogeneous feature tensor is defined as follows: A 5-dimensional spatiotemporal heterogeneous feature tensor is constructed, with the dimension format as follows: ,in, The sample size dimension corresponds to the number of sample forest areas, and its value is the total number of training samples. Using time as the time dimension, with ten days as the time step, it covers three complete life cycles of harmful organisms, with a total of 36 time steps, which are aligned with the business time nodes of the control nodes. For the spatial dimension, the corresponding forest area plot grid code is used. Each grid corresponds to a unique geographic identifier (GIS plot number). The grid resolution is 30m×30m, which is consistent with the resolution of the Class II forest resource survey data. To control the node hierarchy, the number of core control nodes corresponding to the full-cycle control logic graph is 6-dimensional. The number of new nodes corresponds to the expansion dimensions. The feature dimension is a unified feature dimension corresponding to the three-layer processing of each control node, which is fixed at 21 dimensions (8+6+7).

[0126] In this embodiment, the spatiotemporal data alignment method uses the land parcel grid code as the spatial anchor point and the ten-day period as the time anchor point to align the feature data of the corrected multimodal full-cycle historical dataset and the completed full-cycle control logic map to the corresponding spatiotemporal anchor points. The alignment rule is that feature data of the same spatial anchor point and the same time anchor point are mapped to a unique control node and feature dimension to ensure that the feature data of each spatiotemporal location corresponds one-to-one with the control node, without misalignment or missing data.

[0127] Tensor fusion and initial sample set generation directly splice and fuse the aligned spatial geographical dimension, disaster occurrence time dimension, three-layer feature dimension, and control node level dimension to construct a spatio-temporal heterogeneous feature tensor; at the same time, using the full-cycle record data of historical disaster events corresponding to the sample forest area and the prevention and control effect verification data as label data, the label format is: [risk quantification value, disaster level, prevention and control efficiency, prevention and control failure label], generate labeled training samples, and all samples form an initial training sample set.

[0128] In this embodiment, the calculation power load saturation calculation method collects the hardware operation status data of the current computing device in real time, including the CPU utilization rate Ucpu, the memory occupancy rate Umem, and the parallel computing unit (GPU / NPU) load rate Ugpu. The calculation power load saturation S0 is the weighted value of the three, and the calculation formula is: S0 = ω01×Ucpu + ω02×Umem + ω03×Ugpu, where: the weight coefficients ω01 = 0.3, ω02 = 0.3, ω03 = 0.4, determined based on the calculation power consumption characteristics of data preprocessing. The parallel computing unit has the greatest impact on the preprocessing efficiency, so the weight is the highest; the value range of S0 is 0~1, and the preset load threshold Sth = 0.7; if there is no parallel computing unit such as GPU / NPU, then Ugpu = 0, and the weights are reallocated as ω01 = 0.5, ω02 = 0.5 to ensure generality in different hardware environments.

[0129] Adaptive preprocessing branch logic

[0130] The first judgment sub-module: when S0 < Sth, it is determined that the calculation power is sufficient, and directly perform single-process normalization and standardization preprocessing on the spatio-temporal heterogeneous feature tensor of the initial training sample set; the normalization uses Min-Max normalization to map the feature values to the [0,1] interval, formula: Xnorm = (X - Xmin) / (Xmax - Xmin); the standardization uses Z-Score standardization, formula: Xstd = (X - μ) / σ, where μ is the feature mean and σ is the feature standard deviation; after the preprocessing is completed, the final training sample set is obtained.

[0131] The second judgment sub-module: when S0 ≥ Sth, it is determined that the calculation power load is high, start the parallel preprocessing process, construct a business dependency graph through the dependency unit, complete data directional sharding and parallel processing, and obtain the final training sample set after merging the results.

[0132] In this embodiment, starting 8-process parallel processing on a common desktop computer, the actual measured time consumption is only 22 seconds, and the efficiency is increased by 12.5 times compared with the 276 seconds of single-process processing, instead significantly reducing the preprocessing time consumption and improving the operation efficiency.

[0133] The beneficial effects of the above technical solution are as follows: through a complete sample processing flow of multi-factor environmental compensation correction, standardized feature completion, and adaptive preprocessing, the systematic bias of monitoring data caused by site heterogeneity in forest areas is eliminated from the root. At the same time, the efficient use of hardware computing power is realized, the quality of training samples is improved, and a core data foundation is laid for the high-precision prediction and strong generalization ability of the risk assessment model.

[0134] This invention provides a data analysis-based risk assessment system for forest pest control, wherein the second judgment submodule includes:

[0135] The dependency unit is used to construct an initial hierarchical graph of depth 1, with each control node of the completed full-cycle control logic graph as the core node, and the direct dependency relationship corresponding to each core node as the directed edge, and to perform structural optimization on the initial hierarchical graph to obtain the dependency graph.

[0136] The communication pool construction unit is used to determine the optimal number of parallel subtasks based on the hardware operating status data, create independent processing processes equal to the optimal number of parallel subtasks, and establish a high-speed communication pool based on shared memory for each process.

[0137] The sharding and process correspondence unit is used to perform targeted sharding of spatiotemporal heterogeneous feature tensors according to the node business attributes based on the business logic and data volume of each control node in the dependency graph and the optimal number of parallel subtasks, so that each data shard corresponds one-to-one with an independent processing process, and the shard data volume matches the hardware load capacity of the corresponding independent processing process.

[0138] The retrieval unit is used to extract feature anchor points of each data shard in parallel, calculate the data balance of different data shards, optimize the shard allocation strategy with the goal of maximizing the data balance, and schedule mapping tasks and reduction tasks with the same task label to the same processing process.

[0139] The data sharding and distribution unit is used to distribute each data shard to the corresponding independent processing process based on the high-speed communication pool, perform parallel normalization preprocessing and risk feature extraction, and merge the output results of all independent processing processes to obtain the final training sample set.

[0140] In this embodiment, the optimal number of parallel subtasks is determined based on hardware operating status data. The calculation formula is: , Value range: 2~32; Create and Equal independent processing processes are established for each process, and a high-speed communication pool based on shared memory is established for each process. The communication pool adopts a memory mapping method to avoid the performance loss of data copying between processes. The read and write bandwidth of the communication pool is not less than 10GB / s.

[0141] In this embodiment, the sharding rules are as follows: nodes with strong dependencies are assigned to the same shard; the amount of data in each shard is matched with the hardware load capacity of the corresponding process, and the process with higher hardware load capacity is allocated a larger amount of data; the number of shards is equal to the number of parallel subtasks, ensuring that each data shard corresponds one-to-one with an independent processing process.

[0142] In this embodiment, the feature anchor points are: the core control node ID, time range, and spatial range of each segment;

[0143] In this embodiment, the formula for calculating the balance is: ,in, The standard deviation of the data volume of each partition. The average data volume of each partition is used; with the goal of maximizing data balance, a genetic algorithm is used to optimize the partition allocation strategy, with 50 iterations; at the same time, mapping tasks and reduction tasks with the same task label are scheduled to the same processing process to reduce inter-process communication overhead.

[0144] The beneficial effects of the above technical solution are: the parallel preprocessing scheme based on business dependency graph-oriented sharding, high-speed communication pool construction, and load balancing optimization divides the control nodes with strong business dependencies into the same data shard, avoids cross-process dependency conflicts in parallel processing, improves data processing efficiency compared to single process, and solves the problems of processing errors and low efficiency caused by random sharding in existing technologies.

[0145] This invention provides a data analysis-based risk assessment system for forest pest control, wherein the dependent unit, such as... Figure 2 As shown, it includes:

[0146] The distribution subunit is used to traverse the business description text of each core node in the initial hierarchical diagram, extract several logical objects, and find the first distribution that has a parent relationship with each logical object and the second distribution that has a child relationship from the business description text of each core node.

[0147] The refinement subunit is used to refine the first distribution and the second distribution of the same logical object respectively to obtain the corresponding third distribution and the fourth distribution, and to calculate the first expansion coefficient of the first distribution and the third distribution, and the second expansion coefficient of the second distribution and the fourth distribution.

[0148] The extended subunit is used to determine the number of extensions of the corresponding logical object based on the first extension coefficient and the second extension coefficient, which is regarded as the first number, and to extract the first number of dimensions with the highest confidence from the corresponding refinement processing results as the object extension dimension.

[0149] The description subunit is used to perform overlapping analysis on the extended dimensions of all objects under the same core node to obtain the semantic extended dimensions of the corresponding core node. Combined with the data scale characteristics, computing load characteristics and upstream and downstream dependency characteristics of the corresponding core node, the extended feature description set of the corresponding core node is obtained.

[0150] The mining subunit is used to mine implicit and explicit business relationships between nodes based on the extended feature description set of each core node, and to construct node optimization functions according to the implicit logical relationship of the implicit business relationship and the explicit logical relationship of the explicit business relationship.

[0151] The iterative calculation subunit is used to iteratively calculate the node optimization function according to the input logical query conditions and the output logical query conditions, determine the optimization parameter values, and obtain supplementary associated nodes by combining the parameter types of each optimization parameter value;

[0152] The position determination subunit is used to determine the initial optimization solution of the supplementary associated nodes and each core node, and combined with the main node logic of the corresponding core node, to obtain the position of the supplementary associated node of the corresponding supplementary associated node, and to supplement the corresponding supplementary associated node into the initial hierarchical graph according to the position of the supplementary node to obtain the dependency graph.

[0153] In this embodiment, the business description text of each core node in the initial hierarchical graph is traversed, and several logical objects are extracted using Jieba word segmentation and part-of-speech tagging. The logical objects are defined as core entities in the node's business process (such as monitoring data, prevention and control operations, harmful organisms, and triggering conditions). From the business description text of each core node, the first distribution with a parent relationship and the second distribution with a child relationship with each logical object are searched. Here, the parent relationship refers to the upstream input source of the logical object, and the child relationship refers to the downstream output destination of the logical object. The first distribution and the second distribution are both probability distributions of the frequency, position, and association weight of the logical object in the text.

[0154] Calculate the first expansion coefficient of the first distribution and the third distribution. The second expansion coefficient of the second and fourth distributions The formula for calculating the expansion factor is: In the formula: KL divergence is used to measure the difference between two distributions. For the original distribution, This is the refined distribution. The preset maximum KL divergence threshold is set to 2; the expansion coefficient... The value range is 0~1. The larger the value, the more scalable dimensions the logical object has.

[0155] In this embodiment, the first quantity , This is a rounding function. The value ranges from 1 to 5; from the corresponding refined processing results, extract the top results with the highest confidence level. Each dimension serves as an extended dimension for the object, and the confidence level is determined by refining the peak probability of the distribution.

[0156] In this embodiment, an overlap analysis is performed on the extended dimensions of all objects under the same core node, and the cosine similarity between dimensions is calculated. Dimensions with similarity ≥ 0.7 are merged to obtain the semantic extended dimensions of the corresponding core node. Combining the data scale characteristics (data volume), computing power load characteristics (computing power required for processing), and upstream and downstream dependency characteristics (input and output nodes) of the corresponding core node, an extended feature description set of the corresponding core node is generated. The description set includes four dimensions: semantic features, data features, computing power features, and dependency features of the node.

[0157] In this embodiment, explicit business relationships are the existing upstream and downstream directed edges in the initial hierarchical graph, while implicit business relationships are those between nodes that are not directly connected but share data and collaborate (such as the relationship between the annual insect population base of overwintering monitoring nodes and pre-overwintering monitoring nodes). Based on the explicit and implicit logical relationships, a node optimization function is constructed: In the formula: For the number of core nodes, Let J be the association weight between node j3 and node j4. To determine the business similarity between nodes, The distance between nodes representing business dependencies. is the regularization coefficient, with a value of 0.1; the optimization objective is to maximize... This achieves optimal matching of relationships between nodes.

[0158] In this embodiment, the node optimization function is iteratively calculated using the gradient ascent algorithm based on the input logic query conditions (upstream input requirements of the node) and the output logic query conditions (downstream output requirements of the node). The iteration termination condition is: after 10 consecutive iterations... The change is less than 1e-6; after the iteration is completed, the optimal association weight parameter value is determined, and the association type corresponding to the parameter value is combined to obtain the supplementary association node. The supplementary association node is the intermediate carrier node of the implicit association between nodes, which is used to eliminate cross-node dependency conflicts during parallel processing.

[0159] In this embodiment, the least squares method is used to determine the initial optimization solution of the supplementary associated nodes and each core node. Combined with the main node logic of the corresponding core node, the position of the supplementary associated node is determined. The supplementary position must meet the following requirements: it does not disrupt the upstream and downstream order of the original business link, and it can convert long-distance dependencies across nodes into short-distance dependencies between adjacent nodes. The supplementary associated nodes are added to the initial hierarchical graph to finally obtain the dependency graph.

[0160] The dependency graph clarifies the business dependencies between various control nodes. When performing parallel data sharding, nodes with strong dependencies must be assigned to the same data shard to avoid cross-process dependency conflicts and ensure the accuracy of parallel preprocessing results.

[0161] The beneficial effects of the above technical solution are as follows: Through the dependency graph automatic construction method of semantic analysis, implicit association mining and node iterative optimization, it can automatically identify implicit business dependencies between management nodes that have not been manually defined, eliminate long-distance dependency conflicts by supplementing related nodes, and build a complete dependency graph without the need for manual review of business rules. This greatly improves the completeness of the graph and the accuracy of parallel processing, and adapts to the personalized management needs of different forest areas.

[0162] This invention provides a data analysis-based risk assessment system for forest pest control, wherein the control decision-making module includes:

[0163] The boundary determination submodule is used to determine that there is no risk when the risk quantification value is lower than the minimum preset threshold, and output a normalized monitoring and early warning decision.

[0164] When the risk quantification value is higher than the highest preset threshold, it is judged as a major emergency risk, and an emergency prevention and control decision is output.

[0165] When the risk quantification value is between the lowest preset threshold and the highest preset threshold, it is determined to be a normal preventable risk and enters the multi-source decision fusion optimization stage;

[0166] The decision generation submodule is used to output preliminary handling decisions and corresponding decision confidence levels for routine preventable risks through three analysis sources. Combined with the prevention and control effect mapping data in the pre-constructed global knowledge graph, the multi-source decision results are weighted and fused and conflict resolution is processed to generate global targeted prevention and control decisions.

[0167] Preferably, the first analysis source is a site suitability analysis unit obtained by matching forest stand basic attributes with risk factors, corresponding to the feature dimensions of the first site foundation layer; the second analysis source is a targeted prevention and control analysis unit obtained by matching pest population characteristics with disaster development trends, corresponding to the feature dimensions of the second disaster core layer; and the third analysis source is an environmental intervention analysis unit obtained by matching environmental stress factors with disaster diffusion conditions, corresponding to the feature dimensions of the third environmental stress layer.

[0168] The global targeted prevention and control decision includes the prevention and control operation plan, decision confidence level, risk severity level, unique geographic identifier of the corresponding forest area plot, and threshold range of operation parameters. The decision confidence level is calculated by weighting the proportion of effective data in the corresponding analysis source feature dimension and the matching degree with historical successful cases in the global knowledge graph. The weight of the weighted fusion is the normalized value of the corresponding decision confidence level.

[0169] The conflict resolution aims to achieve the highest prevention and control efficiency, the least environmental impact, and the lowest prevention and control cost, and prioritizes and adjusts the results of multi-source decisions.

[0170] In this embodiment, according to the three-layer processing rules of step S1, real-time multimodal monitoring data of the target forest area is collected and processed to ensure that the feature dimensions of the real-time data are completely consistent with the feature dimensions of the model training samples.

[0171] The processed real-time data is input into the converged risk assessment model, and the model outputs the real-time prevention and control risk quantification value of the target forest area, with a value range of [0,100].

[0172] Standardized risk grading: Based on the pre-set risk grading thresholds of the full-domain knowledge graph, the quantified risk value is divided into 3 levels:

[0173] Low risk (no risk): Risk quantification value < 30 (minimum preset threshold);

[0174] Medium risk (routinely preventable risk): 30 ≤ risk quantification value ≤ 70 (between the minimum and maximum thresholds);

[0175] High risk (major emergency risk): Risk quantification value > 70 (maximum preset threshold).

[0176] In this embodiment, the specific implementation method for disaster source tracing and targeted decision output is as follows:

[0177] Based on the full-domain knowledge graph, disaster source matching is performed on the real-time prevention and control risk quantification value and the classification result: the real-time data of the target forest area after three layers of processing is matched with the full-cycle disaster occurrence and development cause data in the full-domain knowledge graph using cosine similarity, and the top 3 causes with the highest similarity are identified as core risk causes.

[0178] Based on the risk classification results, implement differentiated decision outputs:

[0179] When the risk quantification value is <30, it is determined to be no risk, and the following routine monitoring and early warning decision is output: maintain the existing fixed sample plot monitoring frequency, collect forest monitoring data once every ten days, continuously track the dynamics of harmful organism populations and changes in environmental factors, and output a risk assessment report once a month.

[0180] When the risk quantification value is greater than 70, it is judged as a major emergency risk, and an emergency prevention and control decision is output: based on the emergency prevention and control technical procedures for the corresponding pests in the full-domain knowledge graph, an emergency response plan including the types of emergency agents, application dosage, operation methods, operation time limits, and blockade and isolation measures is output and simultaneously pushed to the local forestry pest control authorities.

[0181] When 30 ≤ risk quantification value ≤ 70, it is determined to be a normal preventable risk and enters the multi-source decision fusion optimization stage.

[0182] Definition and output of three analysis sources: The three analysis sources correspond to the feature dimensions of the three processing layers, achieving a one-to-one correspondence between business dimensions.

[0183] First source of analysis: Site suitability analysis unit, corresponding to the first site foundation layer. Based on the matching degree between the basic attributes of the forest stand and the core risk factors, it outputs the preliminary treatment decisions for site suitability (such as forest stand tending, sanitary logging, and tree species structure adjustment) and the decision confidence level C1.

[0184] Second source of analysis: Targeted prevention and control analysis unit, corresponding to the second disaster core layer. Based on the characteristics of pest populations, the development progress of insect stages, and the development trend of disasters, it outputs the initial targeted treatment decisions and decision confidence C2 for biological control, chemical control, and physical control.

[0185] The third source of analysis: the environmental intervention analysis unit, corresponding to the third layer of environmental stress, outputs preliminary disposal decisions and decision confidence level C3 based on environmental stress factors and disaster diffusion conditions, including environmental regulation, monitoring frequency adjustment, and diffusion blocking.

[0186] Decision confidence calculation method: The decision confidence of each analysis source is calculated by weighting the effective data ratio of the corresponding feature dimension of the analysis source and the matching degree with historical successful cases in the whole domain knowledge graph. The formula is: C0=0.6×PM+0.4×PD, where PM is the matching degree with historical successful cases, PD is the effective data ratio of the corresponding feature dimension, and the confidence value range is [0,1].

[0187] Weighted fusion and conflict resolution:

[0188] Weighted fusion: The preliminary decision on the three sources of analysis is weighted and fused according to the confidence level. The fusion weight is the normalized value of the confidence level of the corresponding decision. The formula is: Ji2=Ci2 / (C1+C2+C3), i2=1,2,3;

[0189] Conflict Resolution: When there is a conflict between decisions from two or more analytical sources (such as a temporal conflict between chemical and biological control, or a spatial conflict between sustenance measures and spread blocking), the following resolution process is executed: Step 1: Based on the control effect mapping data in the global knowledge graph, extract the historical control effectiveness rate corresponding to the conflicting decisions; Step 2: With "highest control effectiveness rate, least environmental impact, and lowest control cost" as the optimization objective, construct a multi-objective optimization function to prioritize the conflicting decisions; Step 3: Retain the highest priority decision and adjust the remaining conflicting decisions temporally or spatially to completely eliminate the conflict.

[0190] Global targeted prevention and control decision generation: After integration, a global targeted prevention and control decision is generated, which includes: prevention and control operation plan, decision confidence level, risk severity level, unique geographic identifier of corresponding forest area plot, and threshold range of operation parameters.

[0191] In this embodiment, based on the full-domain knowledge graph, disaster source tracing matching is performed on the real-time prevention and control risk quantification value and the classification result. The specific method is as follows: the real-time data of the target forest area after three-layer processing is matched with the disaster occurrence and development full-cycle inducing factor data in the full-domain knowledge graph using cosine similarity. The similarity calculation formula is: In the formula: This represents the real-time feature vector of the target forest area. The historical trigger feature vectors in the knowledge graph are used to identify the top 3 triggers with the highest similarity as core risk triggers, thereby achieving accurate risk tracing.

[0192] With the optimization objectives of maximizing prevention and control effectiveness, minimizing environmental impact, and minimizing prevention and control costs, a multi-objective optimization function is constructed: In the formula: To ensure effective prevention and control, To normalize the environmental impact level, To normalize prevention and control costs; =0.5、 =0.3、 =0.2 is the weighting coefficient, which is determined by combining the AHP method with expert scoring;

[0193] Based on a multi-objective optimization function, conflict decisions are prioritized, the highest priority decision is retained, and the remaining conflict decisions are adjusted in time or space (e.g., the interval between biological control and chemical control is adjusted to more than 15 days) to completely eliminate the conflict.

[0194] The beneficial effects of the above technical solution are as follows: through the three-source analysis and decision-making system that corresponds one-to-one with the three-layer feature dimensions, the business dimensions of risk causes and prevention and control decisions are accurately matched. At the same time, based on the dynamic weighted fusion of decision confidence and the resolution of conflicts in multi-objective optimization, the conflict problem of multi-source decision-making is completely solved. The output prevention and control solution can simultaneously improve the prevention and control effect, reduce the environmental impact and save prevention and control costs, taking into account both ecological security and economic benefits.

[0195] This invention provides a data analysis-based risk assessment system for forest pest control, wherein the model training module includes:

[0196] The mechanism construction submodule is used to construct a multi-granularity loss constraint mechanism by using the spatiotemporal heterogeneous feature tensor of the training sample set as the model input of the neural network model and the historical disaster level of forestry pests and the control failure label as the supervision signal.

[0197] The gradient backpropagation submodule is used to independently backpropagate the feature channels corresponding to each control node of the processed full-cycle control logic graph during the forward propagation phase, and synchronously execute distributed gradient aggregation according to the hardware parallel process allocation results.

[0198] The learning rate adjustment submodule is used to dynamically adjust the learning rate based on the dual indicators of risk prediction error and control node feature fitting degree of the neural network model on the validation set after each iteration. When the dual indicators are consistently lower than the preset convergence threshold and gradient oscillations no longer occur in multiple consecutive iterations, the model is determined to have reached the convergence state, and a forestry pest control risk assessment model adapted to the heterogeneous data characteristics of forest areas is obtained.

[0199] In this embodiment, a spatiotemporal graph convolutional neural network (ST-GCN) is used as the basic model of the neural network model. A multi-granularity loss constraint mechanism is constructed in combination with the characteristics of forestry operations to complete the supervised training of the model. The specific method is as follows:

[0200] Spatiotemporal Graph Convolutional Neural Network (ST-GCN) is used as the basic model. This model can simultaneously extract spatial correlation features and time series features of forest area data, and is adapted to the input format of spatiotemporal heterogeneous feature tensors. The complete architecture parameters are as follows:

[0201] Input layer: The input is a 5-dimensional spatiotemporal heterogeneous feature tensor, which is transformed into an ST-GCN-adapted format through dimensionality transformation: [batch size, time step, number of nodes, feature dimension], where the number of nodes is the number of control nodes, and the feature dimension is 21-dimensional;

[0202] Spatiotemporal graph convolutional blocks: There are 3 blocks in total. Each convolutional block contains 1 temporal convolutional layer, 1 graph convolutional layer, 1 batch normalization layer, 1 ReLU activation function layer, and 1 Dropout layer. The temporal convolutional layer has a 3×1 kernel size, a stride of 1, and padding of 1. The adjacency matrix of the graph convolutional layer is constructed based on the dependency graph, and the adjacency matrix elements... =1 indicates that node i2 and node i3 have a dependency relationship, otherwise it is 0; the dropout rate of the Dropout layer is set to 0.3;

[0203] Fully connected layers: There are 2 in total. The first fully connected layer has 128 neurons and the second fully connected layer has 64 neurons. The activation function is ReLU and the dropout rate is 0.3.

[0204] Output layer: There are 2 neurons, which output the risk quantification value of forestry pest control (0~100) and the risk classification probability (3 levels). The risk quantification value adopts the linear activation function and the risk classification adopts the Softmax activation function.

[0205] In this embodiment, the multi-granularity loss constraint mechanism is implemented as follows:

[0206] General formula for multi-granularity loss function: In the formula: =0.3、 =0.3、 =0.4, determined through 100 sets of comparative experiments based on the degree of influence of each loss term on the model performance, with a weight sum of 1.

[0207] Control node granularity loss The cross-entropy loss function is used to calculate the loss based on the insect development progress and risk trend prediction results of the feature channels of each control node, thus achieving independent loss constraints for each node. The formula is as follows: In the formula: To control the number of nodes, The number of node risk level classifications, The actual label for node i5. This represents the risk level probability of node i5 predicted by the model.

[0208] Data layer granularity loss The mean squared error (MSE) loss function is used to calculate the loss based on the risk contribution prediction results of the feature dimensions in the three-layer processing, achieving independent loss constraints for each layer of data. The formula is as follows: In the formula: It consists of three data layers. The contribution of this layer's features to the true risk level. This represents the risk contribution of the features predicted by the model for this layer.

[0209] Global granularity loss The FocalLoss function is used to calculate the loss based on the final risk level prediction result, thus addressing the problem of imbalanced disaster samples. The formula is as follows: In the formula: For the sample size, To balance the weights across categories, a weight of 0.75 is set for disaster samples and 0.25 for non-disaster samples. Set the focus parameter to 2; This represents the probability of the corresponding category predicted by the model.

[0210] In this embodiment, during the backpropagation phase, the feature channels corresponding to each control node in the processed full-cycle control logic graph are independently backpropagated to ensure that the features of each control node can be fully fitted. At the same time, based on the hardware parallel process allocation results, a distributed data parallelism (DDP) strategy is adopted to synchronously execute distributed gradient aggregation. Each process is responsible for the gradient calculation of the corresponding data slice. After the gradients of all processes are aggregated through the AllReduce operation, the global model parameters are updated to avoid the parameter inconsistency problem caused by multi-process training.

[0211] In this embodiment, the learning rate is dynamically adjusted to determine the model's convergence status. The complete implementation method is as follows:

[0212] Initial settings for training hyperparameters: The initial learning rate is set to 1e-4, the optimizer is AdamW, the weight decay coefficient is set to 1e-5, and the batch size is set to 32. The training sample set is divided into a training set and a validation set in an 8:2 ratio. The training set is used for updating model parameters, and the validation set is used for model performance evaluation and learning rate adjustment.

[0213] Dual core evaluation metrics: After each iteration, the computational model performs on the validation set using two core metrics:

[0214] Risk prediction error: Mean absolute percentage error MAPE = (1 / n0) × Σ∣(ytrue−ypred) / ytrue∣ × 100%;

[0215] Control node feature fit: coefficient of determination / This is used to measure the model's fit to the features of each control node.

[0216] Dynamic learning rate adjustment rule: When the MAPE decreases by less than 1% in two consecutive iterations, When the increase is less than 0.005, the learning rate is multiplied by 0.8 to decay, and the minimum learning rate is set to 1e-6 to avoid convergence stagnation caused by an excessively low learning rate.

[0217] Convergence criterion: When the MAPE of the validation set is consistently below 5% for 10 consecutive iterations, When the gradient norm is consistently above 0.95 and no further gradient oscillations occur (the change in gradient norm is less than 1e-6), the model is considered to have reached convergence, training is stopped, and the converged forestry pest control risk assessment model is obtained.

[0218] The beneficial effects of the above technical solution are as follows: through the model training scheme of multi-granularity loss constraint mechanism, independent gradient backpropagation of channels, and dynamic learning rate adjustment of dual business indicators, the accurate fitting of business characteristics of forestry full-cycle management nodes is achieved. At the same time, the problem of poor model generalization ability caused by imbalance of forestry disaster samples is solved, and the prediction accuracy, convergence stability and cross-forest area adaptability of risk assessment model are greatly improved.

[0219] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A data analysis-based risk assessment system for forest pest control, characterized in that, include: The map construction module is used to construct a full-cycle management logic map of forest pest control that matches the target forest area. Each management node in the full-cycle management logic map is processed in three layers. The first layer is used to lock the site conditions and basic attributes of the forest stand. The second layer is used to lock the population dynamics of pests and disaster occurrence monitoring data. The third layer is used to lock the meteorological, hydrological and environmental stress factor data of the forest area. The sample acquisition module is used to acquire multimodal full-cycle historical datasets of sample forest areas with multiple different forest stand types, and combine them with the processed full-cycle management logic map to generate a training sample set for model training. The model training module is used to perform supervised training on the neural network model using the training sample set to obtain a converged forestry pest control risk assessment model. The model analysis module is used to collect real-time multimodal monitoring data of the target forest area, perform three-layer processing on the real-time multimodal monitoring data, input the processed real-time data into the converged forest pest control risk assessment model, and output the real-time control risk quantification value and standardized risk classification result of the target forest area. The prevention and control decision module is used to match the real-time prevention and control risk quantification value with the standardized risk classification result based on the pre-constructed forest pest control knowledge graph, identify the core risk factors corresponding to the target forest area, and output targeted prevention and control decisions that match the core risk factors.

2. The forestry pest control risk assessment system based on data analysis according to claim 1, characterized in that, The full-domain knowledge graph stores biological characteristic data of all types of forest pests, data on the causes of disasters throughout their entire lifecycle, data on standardized prevention and control technical procedures in the forestry industry, correlation mapping data between disaster events and corresponding prevention and control schemes for different forest stand types, and data on the correspondence between prevention and control operation parameters and control effects. The full-cycle management logic diagram includes multiple interconnected full-cycle management nodes divided according to forestry standard prevention and control technical regulations.

3. The forestry pest control risk assessment system based on data analysis according to claim 1, characterized in that, The sample acquisition module includes: The bias correction submodule is used to construct a multi-factor environmental compensation correction model based on the altitude, aspect, canopy density, and extreme weather event data of the sample forest area. The multi-factor environmental compensation correction model is used to correct the bias of environmental stress factor data, pest population dynamics, and disaster occurrence monitoring data in the multimodal full-cycle historical dataset. The completion submodule is used to complete control nodes with dimensions lower than the maximum effective feature dimension of the same data layer of all control nodes in the processed full-cycle control logic graph. For control nodes with dimensions lower than the benchmark, the module uses the global feature mean of the corresponding forest area sample in the corrected multimodal full-cycle historical dataset to complete the control nodes, so that all three layers of all control nodes have a unified feature dimension. Only when the missing effective feature data rate in the corresponding layer exceeds a preset threshold is the module used to complete the control nodes with zero values. The initial set construction submodule is used to construct a spatiotemporal heterogeneous feature tensor based on the corrected multimodal full-cycle historical dataset and the completed full-cycle management logic map, integrating the spatial geographic dimension of forest area, the time dimension of disaster occurrence, and the three-layer feature dimension. At the same time, using the full-cycle record data of historical disaster events and the verification data of prevention and control effect corresponding to the sample forest area as label data, labeled training samples are generated to form the initial training sample set. The data acquisition submodule is used to collect the hardware operating status data of the current computing device in real time and determine the computing load saturation. The computing load saturation is a weighted value obtained by combining CPU utilization, memory usage, and parallel computing unit load rate. The first judgment submodule is used to directly perform single-process normalization and standardization preprocessing on the spatiotemporal heterogeneous feature tensor of the initial training sample set if the computing power load saturation is less than the preset load threshold, so as to obtain the final training sample set. The second judgment submodule is used to distribute the data corresponding to the spatiotemporal heterogeneous feature tensor to each independent processing process to perform parallel normalization and standardization preprocessing if the computing power load saturation is greater than or equal to a preset load threshold, and merge the output results of all independent processing processes to obtain the final training sample set.

4. The forestry pest control risk assessment system based on data analysis according to claim 3, characterized in that, The second judgment submodule includes: The dependency unit is used to construct an initial hierarchical graph of depth 1, with each control node of the completed full-cycle control logic graph as the core node, and the direct dependency relationship corresponding to each core node as the directed edge, and to perform structural optimization on the initial hierarchical graph to obtain the dependency graph. The communication pool construction unit is used to determine the optimal number of parallel subtasks based on the hardware operating status data, create independent processing processes equal to the optimal number of parallel subtasks, and establish a high-speed communication pool based on shared memory for each process. The sharding and process correspondence unit is used to perform targeted sharding of spatiotemporal heterogeneous feature tensors according to the node business attributes based on the business logic and data volume of each control node in the dependency graph and the optimal number of parallel subtasks, so that each data shard corresponds one-to-one with an independent processing process, and the shard data volume matches the hardware load capacity of the corresponding independent processing process. The retrieval unit is used to extract feature anchor points of each data shard in parallel, calculate the data balance of different data shards, optimize the shard allocation strategy with the goal of maximizing the data balance, and schedule mapping tasks and reduction tasks with the same task label to the same processing process. The data sharding and distribution unit is used to distribute each data shard to the corresponding independent processing process based on the high-speed communication pool, perform parallel normalization preprocessing and risk feature extraction, and merge the output results of all independent processing processes to obtain the final training sample set.

5. The forestry pest control risk assessment system based on data analysis according to claim 4, characterized in that, The dependent unit includes: The distribution subunit is used to traverse the business description text of each core node in the initial hierarchical diagram, extract several logical objects, and find the first distribution that has a parent relationship with each logical object and the second distribution that has a child relationship from the business description text of each core node. The refinement subunit is used to refine the first distribution and the second distribution of the same logical object respectively to obtain the corresponding third distribution and the fourth distribution, and to calculate the first expansion coefficient of the first distribution and the third distribution, and the second expansion coefficient of the second distribution and the fourth distribution. The extended subunit is used to determine the number of extensions of the corresponding logical object based on the first extension coefficient and the second extension coefficient, which is regarded as the first number, and to extract the first number of dimensions with the highest confidence from the corresponding refinement processing results as the object extension dimension. The description subunit is used to perform overlapping analysis on the extended dimensions of all objects under the same core node to obtain the semantic extended dimensions of the corresponding core node. Combined with the data scale characteristics, computing load characteristics and upstream and downstream dependency characteristics of the corresponding core node, the extended feature description set of the corresponding core node is obtained. The mining subunit is used to mine implicit and explicit business relationships between nodes based on the extended feature description set of each core node, and to construct node optimization functions according to the implicit logical relationship of the implicit business relationship and the explicit logical relationship of the explicit business relationship. The iterative calculation subunit is used to iteratively calculate the node optimization function according to the input logical query conditions and the output logical query conditions, determine the optimization parameter values, and obtain supplementary associated nodes by combining the parameter types of each optimization parameter value; The position determination subunit is used to determine the initial optimization solution of the supplementary associated nodes and each core node, and combined with the main node logic of the corresponding core node, to obtain the position of the supplementary associated node of the corresponding supplementary associated node, and to supplement the corresponding supplementary associated node into the initial hierarchical graph according to the position of the supplementary node to obtain the dependency graph.

6. The forestry pest control risk assessment system based on data analysis according to claim 1, characterized in that, The prevention and control decision-making module includes: The boundary determination submodule is used to determine that there is no risk when the risk quantification value is lower than the minimum preset threshold, and output a normalized monitoring and early warning decision. When the risk quantification value is higher than the highest preset threshold, it is judged as a major emergency risk, and an emergency prevention and control decision is output. When the risk quantification value is between the lowest preset threshold and the highest preset threshold, it is determined to be a normal preventable risk and enters the multi-source decision fusion optimization stage; The decision generation submodule is used to output preliminary handling decisions and corresponding decision confidence levels for routine preventable risks through three analysis sources. Combined with the prevention and control effect mapping data in the pre-constructed global knowledge graph, the multi-source decision results are weighted and fused and conflict resolution is processed to generate global targeted prevention and control decisions.

7. The forestry pest control risk assessment system based on data analysis according to claim 6, characterized in that, The first source of analysis is the site suitability analysis unit obtained by matching forest stand basic attributes with risk factors, corresponding to the feature dimensions of the first site foundation layer; the second source of analysis is the targeted prevention and control analysis unit obtained by matching pest population characteristics with disaster development trends, corresponding to the feature dimensions of the second disaster core layer. The third source of analysis is the environmental intervention analysis unit obtained based on environmental stress factors and disaster diffusion conditions, corresponding to the characteristic dimensions of the third layer of environmental stress. The global targeted prevention and control decision includes the prevention and control operation plan, decision confidence level, risk severity level, unique geographic identifier of the corresponding forest area plot, and threshold range of operation parameters. The decision confidence level is calculated by weighting the proportion of effective data in the corresponding analysis source feature dimension and the matching degree with historical successful cases in the global knowledge graph. The weight of the weighted fusion is the normalized value of the corresponding decision confidence level. The conflict resolution aims to achieve the highest prevention and control efficiency, the least environmental impact, and the lowest prevention and control cost, and prioritizes and adjusts the results of multi-source decisions.

8. The forestry pest control risk assessment system based on data analysis according to claim 1, characterized in that, The model training module includes: The mechanism construction submodule is used to construct a multi-granularity loss constraint mechanism by using the spatiotemporal heterogeneous feature tensor of the training sample set as the model input of the neural network model and the historical disaster level of forestry pests and the control failure label as the supervision signal. The gradient backpropagation submodule is used to independently backpropagate the feature channels corresponding to each control node of the processed full-cycle control logic graph during the forward propagation phase, and synchronously execute distributed gradient aggregation according to the hardware parallel process allocation results. The learning rate adjustment submodule is used to dynamically adjust the learning rate based on the dual indicators of risk prediction error and control node feature fitting degree of the neural network model on the validation set after each iteration. When the dual indicators are consistently lower than the preset convergence threshold and gradient oscillations no longer occur in multiple consecutive iterations, the model is determined to have reached the convergence state, and a forestry pest control risk assessment model adapted to the heterogeneous data characteristics of forest areas is obtained.

Images