Multi-source data fusion early warning and precise pesticide application method and system for pine wood nematode disease

Abstract

This invention discloses a multi-source data fusion early warning and precision application method and system for pine wilt disease, belonging to the field of smart forestry technology. It includes a multi-source data acquisition module, a spatiotemporal data fusion module, a health index calculation module, an external feature extraction module, a dual-channel coupled diagnosis module, a graded early warning module, an application urgency assessment module, an application strategy planning module, and an adaptive optimization module. This invention collects and fuses multi-source data such as satellite remote sensing data, UAV inspection data, ground-based IoT data, and forest meteorological data. It utilizes a dual-channel coupled deep learning model to achieve probabilistic diagnosis of the multidimensional state of individual trees and to extrapolate the spread of the disease. Based on the diagnostic results, it generates graded differentiated application strategies and finally achieves adaptive closed-loop optimization of the calculation parameters and model through a control effect feedback mechanism.

Description

Technical Field

[0001] This invention relates to the field of smart forestry technology, specifically to a method and system for multi-source data fusion early warning and precise application of pesticides for pine wilt disease. Background Technology

[0002] Pine wilt disease, known as the cancer of pine trees, is characterized by its high pathogenicity, rapid spread, and high mortality rate, posing a significant threat to my country's forest ecological security.

[0003] In existing technologies, the detection of pine wilt disease mainly relies on the following two methods: one is manual ground patrol, in which forest rangers go deep into the forest area to identify diseased trees by visually observing whether the pine needles have changed color, and combine this with traps to count the number of longhorn beetles; the other is single remote sensing monitoring, which mainly uses the spectral characteristics of satellite remote sensing images to screen for dead pine trees on a large scale by interpreting changes in the color of the tree crown. In terms of prevention and control decisions, the diseased tree clearing area is usually delineated or large-scale uniform spraying is carried out based on the distribution range of the diseased trees found, lacking a refined treatment plan for individual trees.

[0004] However, the aforementioned existing technology has the following two key drawbacks:

[0005] First, the monitoring methods are outdated and cannot effectively identify diseased trees in the incubation period. Whether it is manual inspection or satellite remote sensing, the core criterion is the discoloration of the tree's appearance. However, after the pine wood nematode infects the tree, there is a long incubation period. During this time, the tree's appearance remains green, but its internal physiological indicators have become abnormal. The existing technology lacks in-depth fusion analysis of the multidimensional physiological spectral characteristics of individual trees and external environmental factors, resulting in a large number of latent sources of infection being missed, thus missing the golden window of opportunity to block the spread of the epidemic.

[0006] Second, the prevention and control strategies are extensive and lack closed-loop optimization, resulting in low resource utilization efficiency. Traditional prevention and control often adopts a one-size-fits-all approach, failing to quantify the rate of disease deterioration and spatial spread risk of individual trees. This leads to the failure to prioritize the treatment of high-risk core sources of infection, while excessive use of pesticides in low-risk areas, resulting in a waste of human and material resources.

[0007] Therefore, a method and system for early warning and precise application of pine wilt disease based on multi-source data fusion, which can accurately identify diseased trees in the latent stage and achieve graded and precise application of pesticides based on multi-dimensional coupled diagnosis, is needed to solve the above problems. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for multi-source data fusion early warning and precise application of pesticides for pine wilt disease.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: a method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease, comprising:

[0010] S1. Simultaneously collect multi-source data of the target forest area, including satellite remote sensing data, drone inspection data, ground IoT data, and forest area meteorological data;

[0011] S2. Extract the center coordinates of individual trees based on the UAV inspection data, map the multi-source data to the center coordinates of each individual tree through inverse distance weighted interpolation, perform time alignment, and output standardized individual tree data.

[0012] S3. Based on the standardized single tree data, extract the physiological indicators of the single tree and perform weighted calculations to output the single tree health index that represents the internal physiological state of each single tree.

[0013] S4. Based on the standardized single tree data, extract the feature vectors representing the external pathogenic pressure of each single tree in parallel, including the structural feature vector that quantifies the tree size, the longhorn beetle threat vector that quantifies the external risk, and the environmental induction vector that quantifies the induction conditions.

[0014] S5. The single tree health index is used as the core state vector, and features are concatenated with the structural feature vector, the longhorn beetle threat vector and the environmental inducement vector. The concatenation is then input into a pre-trained dual-channel coupled diagnostic model, which outputs the probability distribution of a single tree in a multi-dimensional state and a heat map of the probability of epidemic spread within a preset time period.

[0015] S6. Based on the probability distribution of the multidimensional state, set a dynamic threshold, perform treatment strategy judgment for each individual tree, mark the cleaning target, the pesticide application target, the monitoring target and the health target, perform graded early warning according to the target type, generate corresponding suggested measures and push them to the management terminal.

[0016] S7. For individual trees marked as targets for pesticide application, based on the probability distribution of the multidimensional state and the heat map of the epidemic spread probability, calculate the pesticide application urgency index of the individual trees marked as targets for pesticide application and generate a pesticide application priority sequence.

[0017] S8. Based on the application priority sequence, generate a graded application strategy and output an instruction file containing single plant coordinates, operation mode, operation path and operation parameters to the application equipment.

[0018] S9. Monitor the execution of the operation and compare the status before and after the application of the pesticide to calculate the control effect index. If the control effect index is lower than the preset threshold, then reverse the urgency calculation parameters and the weight coefficient and judgment threshold of the diagnostic model.

[0019] A multi-source data fusion early warning and precision application system for pine wilt disease includes:

[0020] Multi-source data acquisition module: used to simultaneously collect multi-source data of the target forest area, including satellite remote sensing data, UAV inspection data, ground IoT data and forest area meteorological data;

[0021] Spatiotemporal data fusion module: used to extract the center coordinates of individual trees based on the UAV inspection data, map the multi-source data to the center coordinates of each individual tree through inverse distance weighted interpolation, perform time alignment, and output standardized individual tree data;

[0022] Health Index Calculation Module: Based on the standardized individual tree data, extract the physiological indicators of individual trees and perform weighted calculations to output the individual tree health index that represents the internal physiological state of each individual tree.

[0023] External feature extraction module: used to extract feature vectors representing the external pathogenic pressure of each tree in parallel based on the standardized single tree data, including structural feature vectors that quantify tree size, longhorn beetle threat vectors that quantify external risks, and environmental induction vectors that quantify induction conditions.

[0024] Dual-channel coupled diagnostic module: used to take the single tree health index as the core state vector, and perform feature concatenation with the structural feature vector, longhorn beetle threat vector and environmental inducement vector, input to the pre-trained dual-channel coupled diagnostic model, and output the probability distribution of the single tree in a multi-dimensional state and the heat map of the epidemic spread probability within a preset time period;

[0025] The graded early warning module is used to set dynamic thresholds based on the probability distribution of the multidimensional state, perform treatment strategy judgment for each individual tree, mark the cleaning target, the pesticide application target, the monitoring target and the health target, perform graded early warning according to the target type, generate corresponding suggested measures and push them to the management terminal.

[0026] Application urgency assessment module: Based on the probability distribution of the multidimensional state and the heat map of the epidemic spread probability, it calculates the application urgency index of the individual trees marked as application targets and generates an application priority sequence.

[0027] Application strategy planning module: Used to generate a graded application strategy based on the application priority sequence, and output an instruction file containing the coordinates of a single plant, operation mode, operation path and operation parameters to the application equipment;

[0028] Adaptive optimization module: Used to monitor the execution of operations and compare the state before and after application to calculate the control effect index. If the control effect index is lower than the preset threshold, the urgency calculation parameters, as well as the weight coefficients and judgment thresholds of the diagnostic model, are corrected in reverse.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1. This invention constructs a multi-source data fusion network integrating sky, earth, and air. It captures microscopic spectral anomalies of individual trees using UAV hyperspectral imaging, combines three-dimensional structural features extracted by lidar with longhorn beetle infestation and meteorological data monitored by the Internet of Things on the ground, and deeply integrates internal physiological state and external pathogenic pressure through a dual-channel coupled deep learning model. This can effectively identify latent diseased trees that appear normal but have abnormal physiological indicators, reducing the false negative rate and shifting prevention and control work from post-event cleanup to pre-event intervention, fundamentally curbing the hidden spread of the epidemic.

[0031] 2. This invention innovatively proposes a pesticide application urgency index calculation model. This model comprehensively considers the probability of disease deterioration in a single tree, the risk of external spread, and spatial aggregation effects, prioritizing all target trees in the forest area. Based on this, the system generates differentiated, tiered pesticide application strategies. This not only ensures that high-risk sources of infection are treated first, blocking the transmission chain, but also significantly reduces the input of pesticides and manpower in inefficient areas, improving the efficiency of prevention and control funds and operational effectiveness. By comparing the changes in tree status before and after pesticide application, the system quantifies the prevention and control effect index. If the effect does not meet expectations, the system automatically corrects the weight coefficients, dynamic judgment thresholds, and urgency calculation parameters of the diagnostic model, enabling the system to self-iterate and upgrade, ensuring the accuracy of diagnosis and the effectiveness of strategies in long-term operation, and realizing sustainable management of smart forestry. Attached Figure Description

[0032] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0033] Figure 1 This is a schematic diagram of the modules of the present invention;

[0034] Figure 2 This is a flowchart of the method of the present invention;

[0035] Figure 3 This is the target determination diagram of the present invention. Detailed Implementation

[0036] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0037] like Figure 1As shown, the multi-source data fusion early warning and precision application system for pine wilt disease includes a multi-source data acquisition module, a spatiotemporal data fusion module, a health index calculation module, an external feature extraction module, a dual-channel coupled diagnosis module, a graded early warning module, an application urgency assessment module, an application strategy planning module, and an adaptive optimization module.

[0038] like Figure 2 As shown, the specific implementation steps of the present invention include the following steps:

[0039] S1. Simultaneously collect multi-source data from the target forest area, including satellite remote sensing data, drone inspection data, ground IoT data, and forest area meteorological data.

[0040] It should be specifically noted that the satellite remote sensing data specifically refers to: acquiring high-resolution satellite images of the target forest area, including multispectral images and high-resolution optical images, wherein the multispectral images are used for macro background analysis and the high-resolution optical images are used for auxiliary positioning; focusing on acquiring geographic coordinate information, red-edge band, near-infrared band and short-wave infrared band data.

[0041] The drone inspection data specifically refers to: using drones equipped with hyperspectral imagers and lidar scanners to fly during fixed time periods with good lighting conditions to acquire hyperspectral images of individual tree crowns, which are used to capture chlorophyll fluorescence and water absorption characteristics; and using lidar to scan and acquire three-dimensional point cloud data of the forest area, which are used to extract tree height, crown width and understory topography.

[0042] The specific ground-based IoT data refers to: deploying an intelligent insect monitoring network in the target forest area, with each network point containing a special trap for the pine sawyer beetle and a high-resolution automatic photography device. The trap has a built-in AI image recognition module that automatically identifies and counts the pine sawyer beetles, and collects the insect population density data of the intelligent pine sawyer beetle traps on the ground, including the daily number of traps and the GPS coordinates of the traps.

[0043] The specific meteorological data for the forest area includes: automatic weather stations are set up in the forest area to collect meteorological data in real time, including station coordinates, temperature, relative humidity, rainfall, wind direction and speed, and light intensity.

[0044] All data acquisition tasks are set with a unified time window, and the timestamps of all devices are unified to the millisecond level through the NTP protocol to ensure data spatiotemporal consistency.

[0045] S2. Based on the UAV inspection data, extract the center coordinates of individual trees, map the multi-source data to the center coordinates of each individual tree through inverse distance weighted interpolation, perform time alignment, and output standardized individual tree data.

[0046] It should be specifically explained that the radar point cloud data is filtered, denoised, and normalized to obtain standardized point cloud data. The watershed algorithm based on the canopy height model is used in combination with the standardized point cloud data to achieve single tree segmentation. For each segmented tree, the convex hull algorithm is used to construct a 3D model of the canopy for extraction. The 3D canopy outline is extracted, the tree height is extracted by the difference between the highest point of the canopy and the ground elevation, and the center coordinates are extracted by the geographic coordinates of the center point of the trunk base.

[0047] Using the center coordinates as the feature reference, the multi-source data is mapped to each individual tree using the inverse distance weighted interpolation method. Specifically, the satellite image mapping is as follows: for the center coordinates P(x, y) of each tree, a circular region with P as the center and a radius of r is extracted from the satellite image. The weighted average of the spectral values ​​of all pixels in the region is calculated. The weight is inversely proportional to the distance, and the value of r is 3-5 meters, which is dynamically adjusted according to the size of the tree crown.

[0048] The hyperspectral image mapping specifically involves: registering UAV hyperspectral images with point cloud data, extracting the hyperspectral reflectance curve of a single tree based on the center coordinates of the trunk, and taking the average spectral value of all pixels within the canopy range as the spectral feature of the single tree for the hyperspectral image.

[0049] The insect population density data mapping is specifically as follows: taking the location of each trap as a known point, the insect population density raster surface of the entire forest area is generated by using the inverse distance weighted interpolation method, and then the insect population density value corresponding to the center coordinates of each individual tree is extracted.

[0050] Specifically, the meteorological data mapping involves using the same inverse distance weighted interpolation method to interpolate the temperature, relative humidity, rainfall, wind direction and speed, and light intensity data of each meteorological station onto the coordinates of each individual tree.

[0051] Due to differences in the collection times of different data sources, they need to be unified to the same reference time slice: the collection time of the drone flight is used as the reference time. For insect population density data, the average insect population density of the 3 days before and after the reference time is taken as the insect population density value of the reference time. For meteorological data, the moving average of the 24 hours before and after the reference time is taken as the meteorological data of the reference time. For satellite imagery, if the collection time differs from the reference time by more than 7 days, it will not be included in this analysis and will only be used as a historical trend reference.

[0052] S3. Based on the standardized single tree data, extract the physiological indicators of the single tree and perform weighted calculations to output the single tree health index that represents the internal physiological state of each single tree.

[0053] It should be specifically noted that satellite imagery was used for initial screening, and the normalized vegetation index and normalized water index were calculated to mark discolored and dead areas as the background for high-risk areas.

[0054] Detailed diagnosis is performed based on UAV hyperspectral data, extracting the average reflectance curve of a single tree crown and calculating key physiological indicators, including:

[0055] Red edge position: that is, the wavelength corresponding to the maximum value of the first derivative spectrum, which represents the chlorophyll content. In early-stage infected trees, the red edge will undergo a blue shift, that is, a shift towards shorter wavelengths.

[0056] Moisture index: Indicates the water content of leaves and reflects the degree of obstruction of water transport in the trunk. It is more than 15% lower than the average value of healthy trees.

[0057] Pigment index: indicates anthocyanin content;

[0058] Photochemical reflectance index: Indicates the efficiency of light energy utilization and reflects early stress on photosynthetic efficiency.

[0059] The single-tree health index is specifically a weighted sum of physiological indicators, including red-edge position, moisture index, pigment index and photochemical reflectance index, after normalization. The physiological weight coefficients are obtained through training with historical samples. The single-tree health index ranges from 0 to 1, with higher values ​​indicating healthier trees.

[0060] S4. Based on the standardized single tree data, extract the feature vectors representing the external pathogenic pressure of each single tree in parallel, including the structural feature vector that quantifies the tree size, the longhorn beetle threat vector that quantifies the external risk, and the environmental induction vector that quantifies the inducing conditions.

[0061] It should be specifically noted that the pine sawyer beetle prefers to lay its eggs on trees that are weak, have a large diameter at breast height (DBH), and are isolated at the forest edge. The structural feature vector quantifies the probability that a tree will be selected as a host by the beetle, and specifically includes:

[0062] Tree age / diameter at breast height (DBH) grade: DBH is estimated based on tree height, crown width, and lidar echo characteristics, and then tree age is estimated. Middle-aged and older pine trees with DBH exceeding a certain value have fully developed resin ducts, large breeding space for nematodes, and many bark cracks, making them more susceptible to longhorn beetle infestation.

[0063] Canopy compactness: Based on lidar point cloud computing, the ratio of canopy projection area to canopy outline area. High compactness leads to poor ventilation and light penetration within the forest, which is conducive to the survival and reproduction of nematodes and longhorn beetles; low compactness leads to more active flight activities of longhorn beetles.

[0064] Stand density: The number of trees within a 10-meter radius of the target tree is counted and normalized. High density leads to competition for light and roots, resulting in weak tree vigor; low density results in high wind speed and strong transpiration, both of which are high-risk areas.

[0065] Trunk straightness: The curvature or tilt angle of the central axis of the trunk generated by fitting radar point cloud. Tilted or damaged trunks are often accompanied by a decrease in resin secretion capacity and are prone to bark cracks, which are the preferred locations for longhorn beetle egg-laying and also the channels for nematode invasion.

[0066] The flight dispersal of the pine sawyer beetle exhibits distance attenuation and wind-directed characteristics. The beetle threat vector quantifies the probability of the beetle carrying nematodes reaching the tree per unit time, specifically including:

[0067] Insect population density: Obtain the insect population density of the three smart traps closest to the target tree, and sum them using inverse distance weighting. Insect population density directly reflects the size of the longhorn beetle population in the area. The higher the density, the greater the probability that the tree will be bitten and inoculated with nematodes.

[0068] Diffusion resistance coefficient: The Euclidean distance from the target tree to the nearest known diseased tree is calculated and combined with the density of understory shrubs extracted by radar as the resistance coefficient. The closer to the diseased tree and the fewer the understory obstacles, the higher the probability of being bitten by longhorn beetles carrying nematodes.

[0069] Diffusion Risk Index: Combining real-time weather wind direction and the azimuth of the target tree relative to the nearest insect source / disease tree, the pine sawyer beetle has a downwind flight habit, and trees located upwind are more likely to be infested by beetles carrying nematodes.

[0070] Reproduction suitability: The current date is calculated based on meteorological data to determine its proximity to the peak emergence period of the pine sawyer beetle. During the peak emergence period, the beetle is most active, with the most active feeding and egg-laying behaviors, and the trees are at the highest risk of infection at this time.

[0071] High temperature and drought are key triggering factors for pine wilt disease outbreaks, accelerating nematode reproduction within the tree and weakening tree resistance. The environmental triggering vectors focus on environmental factors affecting nematode reproduction, longhorn beetle activity, and tree disease resistance, specifically including:

[0072] Water stress index: It is a standardized precipitation evapotranspiration index calculated by combining precipitation data from meteorological stations with soil moisture retrieved from satellites over the past 60 days. The larger the negative value, the more severe the drought, which leads to reduced resin secretion in trees, a sharp decline in disease resistance, and makes them more susceptible to being eaten by longhorn beetles and invaded by nematodes.

[0073] Extreme temperature risk index: Statistics on the number of days with extreme low temperatures below -10℃ and extreme high temperatures above 35℃ in the past 30 days, as well as the cumulative duration. Extreme low temperatures can kill some overwintering nematodes and longhorn beetle larvae, while extreme high temperatures affect the activity of longhorn beetles.

[0074] Temperature and humidity stress index: Based on the average daily temperature and relative humidity over the past 30 days, a suitable index for nematode reproduction was constructed. Pine wood nematodes reproduce fastest at around 25℃, and relative humidity >70% is conducive to nematode activity and longhorn beetle emergence.

[0075] All of the above features are Z-Score standardized to eliminate the influence of dimensions.

[0076] S5. The single tree health index is used as the core state vector, and features are concatenated with the structural feature vector, the longhorn beetle threat vector and the environmental inducement vector. The concatenation is then input into a pre-trained dual-channel coupled diagnostic model, which outputs the probability distribution of a single tree in a multi-dimensional state and a heat map of the probability of epidemic spread within a preset time period.

[0077] It should be specifically noted that the dual-channel coupled diagnostic model adopts a hybrid neural network architecture, including a temporal evolution channel and a spatial propagation channel, and fuses and classifies the output through a spatiotemporal attention mechanism, specifically as follows:

[0078] Constructing the input layer: For each tree in the forest area, the single tree health index is concatenated with the feature vector of the external pathogenic pressure to form an initial feature vector;

[0079] Temporal evolution channel: The time series of a single tree is processed using an LSTM long short-term memory network to learn the rate of change of tree health indicators and output the time-state hidden vector of the single tree representing the internal disease evolution trend.

[0080] Spatial propagation channel: A graph neural network (GNN) is used to pass messages on the constructed forest topology map and aggregate the states of neighboring trees. If multiple trees around a certain tree have been identified as being in the disease stage, even if the tree currently has a high health index, its probability of being infected will increase due to the spatial clustering effect. The output is a single-tree spatial state hidden vector representing the degree of external threat it faces.

[0081] Spatiotemporal attention fusion and classification output: A spatiotemporal attention mechanism is introduced to dynamically allocate and fuse the weights of the temporal and spatial hidden vectors. The result is input into a fully connected classification layer and outputs the probability distribution of whether the plant belongs to the healthy stage, latent stage, disease stage, or withering stage.

[0082] The generation of the epidemic spread probability heatmap specifically involves: dividing the forest area into M×N regular grids, with each grid serving as a cell; mapping the diagnosis result of a single tree to the corresponding grid; if there is a diseased tree within a grid, the initial state of that cell is the source of infection; the infection intensity is proportional to the number of diseased trees and the probability of disease onset; and the state transition rules are learned from historical data through a TCN temporal convolutional network. ,in: For the cell state at the next time step, This represents the current cell state. For neighboring cell states, As environmental factors, The density of media insects is used; the cellular automaton is iteratively run for k steps, representing the next k days, to record the cumulative probability of each grid being infected during the simulation period. A threshold is set and different colored areas are used to represent different risk levels to generate a heat map of the epidemic spread probability, including high-risk, medium-risk, and low-risk areas.

[0083] S6. Based on the probability distribution of the multidimensional state, set a dynamic threshold, perform treatment strategy judgment for each individual tree, mark the cleaning target, the pesticide application target, the monitoring target and the health target, perform graded early warning according to the target type, generate corresponding suggested measures and push them to the management terminal.

[0084] It should be specifically noted that the incidence of pine wilt disease is seasonal, and fixed thresholds may lead to underreporting or winter outbreaks. Therefore, dynamic thresholds are constructed, and baseline judgment thresholds are set for each state. The phenological stage is obtained based on the current date. If the phenological stage is the peak emergence period of longhorn beetles, the incubation period threshold is lowered to expand the scope of preventive application. If the phenological stage is the high-temperature outbreak period, the disease threshold is raised to reduce false positives caused by drought. If the phenological stage is the winter cleanup period, the dead period threshold is lowered to ensure that all suspected dead trees are included in the cleanup plan. The location of a single tree in the diffusion heat map is read. If it is in the highest risk area, the thresholds for all disease states are automatically lowered, and upgraded control measures are implemented.

[0085] like Figure 3 As shown, based on dynamic thresholds, hierarchical discrimination logic is executed. Logical judgments are performed according to the following priority order; once a condition is met, subsequent judgments cease, and the tree is directly categorized, marking each tree as one of the following: a cleanup target, a pesticide application target, a monitoring target, or a health target. Specifically:

[0086] If the probability of a tree being in the dead period is greater than or equal to the dead period threshold, it is marked as a target for cleanup, indicating that the tree is dead or in the late stage of the disease and has lost its value for treatment, and is the core source of infection.

[0087] If the probability of a tree being in the dead period is less than the dead period threshold, and the probability of it being in the disease period is greater than or equal to the disease period threshold or the probability of it being in the latent period is greater than or equal to the latent period threshold, then it is marked as a target for drug application, indicating that the tree has become diseased or infected but still has a chance of survival.

[0088] If the probability of a tree being in the dead period, disease period, or incubation period is less than its respective threshold and the probability of being in the healthy period is greater than or equal to the healthy period threshold, then it is marked as a healthy target, indicating that the tree's various indicators are normal and it is in a healthy state.

[0089] If a tree does not meet all of the above conditions, it is marked as a monitoring target, indicating that the tree's condition is suspected to be abnormal but the evidence is insufficient, meaning that the probability of any of the four conditions has not reached the threshold.

[0090] The tiered early warning system specifically refers to:

[0091] For the target to be cleared, a red alert is issued, the coordinates are marked and sent to the logging team, requiring them to complete the felling within a specified time, and to carry out on-site crushing, burning or fumigation treatment to cut off the source of infection;

[0092] For the target area, an orange alert is issued, and immediate drug intervention is recommended to save the trees or stop the internal spread.

[0093] For the monitored target, a yellow alert is issued, and it is recommended not to apply pesticides immediately, but to list it as a key focus, increase the frequency of inspections of the tree, and conduct targeted manual re-inspections;

[0094] No warnings will be issued for the health targets mentioned; it is recommended to maintain routine inspections.

[0095] S7. For individual trees marked as targets for pesticide application, based on the probability distribution of the multidimensional state and the heat map of the epidemic spread probability, calculate the pesticide application urgency index of the individual trees marked as targets for pesticide application and generate a pesticide application priority sequence.

[0096] It should be specifically noted that the drug administration urgency index is a weighted sum of the internal disease deterioration rate, the external spread risk probability, and the aggregation coefficient.

[0097] The internal disease deterioration rate is based on a probability distribution. It reads the probability of a single tree marked as the target for treatment being in the disease outbreak or incubation period. If the probability of the disease outbreak is higher than that of the incubation period, it indicates that the disease has broken out and is rapidly progressing towards death, indicating high urgency. If the probability of the incubation period is higher than that of the disease outbreak, it indicates that the disease is in its early stages, representing a golden window to stop the spread of the epidemic, also indicating high urgency, requiring control before symptoms appear. If both probabilities are close to a threshold, it indicates that the disease is relatively stable, with moderate urgency. This internal disease deterioration rate ensures that trees nearing death and newly infected trees receive high priority, avoiding the situation of only treating severely diseased trees while missing early sources of infection. Specifically:

[0098] ;

[0099] Where H represents the rate of internal disease deterioration. This represents the probability of onset. The threshold for the onset of illness, The probability of incubation period. This is the incubation period threshold. If the probability is close to the threshold, this term is close to 0; if the probability is close to 1, this term is close to 1. and To preset the internal disease weighting coefficient, This is a crisis amplification factor, ensuring that as long as the probability of either the onset or the latent state is high, the total score will increase.

[0100] The probability of external spread risk is based on a heatmap of spread probability. The value of a single tree marked as a target for pesticide application is read from the heatmap. If the tree is located in a high-risk area, it means that a large area of ​​surrounding forest will be affected, indicating a high degree of urgency; if the tree is located in a low-risk area, it means that the risk of outward spread is low, indicating a low degree of urgency. This probability of external spread risk reflects a global perspective, prioritizing the protection of key areas to prevent widespread outbreaks. Specifically:

[0101] ;

[0102] Where W represents the probability of external spread risk, and P represents the original probability value on the epidemic spread probability heatmap. This is the maximum value of the current forest area heat map, used for dynamic normalization.

[0103] The clustering coefficient, based on spatial distribution, analyzes whether there are other targets for pesticide application or cleanup within a certain range around the tree. If multiple diseased trees are densely distributed around the tree, it indicates that this is the epicenter of the outbreak and must be prioritized for pesticide application, indicating high urgency. If the tree is a single, isolated diseased tree surrounded by healthy trees, the urgency is low. The clustering coefficient is used to optimize operational efficiency, prioritizing the treatment of contiguous areas and reducing the cost of round-trip travel for pesticide application equipment. Specifically:

[0104] ;

[0105] Where J is the aggregation coefficient, which quantifies the density of surrounding diseased trees; N is the total number of trees marked as targets for pesticide application or clearing within a radius R centered on the tree; and K is a saturation constant, which sets the aggregation effect to its maximum when the number of surrounding diseased trees reaches K.

[0106] The three normalization factors mentioned above are weighted and summed, and mapped to a 0-100 score system. The allocation of the urgency weight coefficient follows the principle of dynamic adaptation, with multiple built-in weight configuration templates corresponding to three typical scenarios: early spring blocking, midsummer outbreak control, and autumn / winter clearing. The initial allocation defaults to the expert experience value as the initial startup parameter. The system automatically switches to the corresponding weight template based on the current season and the global mean of the heat map. Combined with the feedback mechanism of step S9, if a certain weight configuration causes the prevention and control effect index to be lower than expected for two consecutive cycles, the system will trigger a weight fine-tuning algorithm to perturb the weight within a range of ±0.1 until the optimal solution for the current forest area environment is found.

[0107] The specific steps for generating the application priority sequence are as follows: Based on the calculated application urgency index score, all application targets are sorted from high to low and divided into three priority levels to generate the final task sequence, specifically:

[0108] Level 1: Emergency Interception Sequence, targeting the top-ranked sequences based on urgency. The pesticide application equipment flies to these locations regardless of distance to carry out operations on the trees.

[0109] Level 2: Priority Treatment Sequence, urgency index ranking When planning routes for trees, these points are connected to ensure coverage, thereby improving operational efficiency, and are scheduled to be executed immediately after the completion of the primary task.

[0110] Level 3: Prevention and Consolidation Sequence, Urgency Index Ranking Trees, as targets for preventative pesticide application, will only be treated for the remaining time after the completion of Level 1 and Level 2 tasks.

[0111] S8. Based on the application priority sequence, generate a graded application strategy and output an instruction file containing the coordinates of a single plant, the operation mode, the operation path, and the operation parameters to the application equipment.

[0112] It should be specifically explained that, based on the application level sequence and the performance parameters of the application equipment, a graded application strategy library is constructed. The optimal operation mode and parameters are matched for each tree in the sequence, the globally optimal operation path is planned, and finally, a standardized executable instruction file is generated and sent to the application equipment.

[0113] The tiered application strategy library automatically matches differentiated operation plans based on the priority level of the trees, specifically:

[0114] For the Level 1 emergency blocking sequence, a dual-mode operation of treatment and blockade is adopted. Individual trees are treated and sprayed within a set distance around the tree canopy to form an isolation zone to block the migration of longhorn beetles. The concentration and dosage of the pesticide solution are x times and y times the preset standard concentration and dosage to ensure rapid killing of nematodes inside the tree. The pesticide solution is applied close to the tree canopy and the downward pressure wind field is used to force the pesticide into the canopy. The values ​​of x and y are between 1.1 and 1.3.

[0115] For the secondary key treatment sequence, a precision treatment mode is adopted, which involves targeted removal of individual trees, taking into account both effectiveness and efficiency. Pre-set standard concentrations and dosages of pesticides are used, and a grid-shaped full-coverage path is adopted to ensure that the entire tree canopy is treated.

[0116] For the three-level preventive consolidation sequence, a preventive coverage mode is adopted to spray the designated area to form a protective barrier. The concentration of the pesticide solution is z times the preset standard concentration and the preset basic dosage is used to preferentially connect the contiguous areas, where the value of z is between 0.7 and 0.9.

[0117] After determining the operation parameters for each tree, the system performs path planning based on a priority sequence, specifically as follows:

[0118] Using primary trees as core anchor points, regardless of their geographical location, priority is given to planning paths to primary points, and the shortest flight path is planned while meeting the priority. Secondary and tertiary trees are clustered according to the principle of geographical proximity to form several operation clusters. When generating the main path, the application equipment is required to visit all primary points first, and then visit secondary and tertiary points in sequence. If the primary task is not completed but the power / pesticide supply is insufficient, the instruction file includes a forced return point and marks the remaining primary points as the highest priority for the next equipment, ensuring that emergency tasks are not interrupted or missed. Combined with high-precision maps, obstacles such as high-voltage lines and tall buildings are automatically avoided, and independent obstacle avoidance profiles are generated for different flight altitudes.

[0119] The generation of standardized executable instruction files encapsulates the aforementioned strategies and paths into instruction files that can be directly parsed by the application equipment, specifically including:

[0120] Task header information includes task ID, generation time, estimated total time, and total amount of potions required;

[0121] Single-tree operation instruction list: includes single-tree coordinates, operation mode code, action sequence and operation parameter set, where the operation parameter set includes flow rate, flight speed, relative canopy height and recommended pesticide ratio;

[0122] Emergency strategy package: includes low battery return coordinates, hovering or continuation logic after communication loss, and emergency landing point in case of sudden wind and rain.

[0123] For complex terrains inaccessible to spraying equipment, simplified navigation instructions are generated to guide manual personnel carrying spray guns to the corresponding locations. The recommended number of spray holes and amount of pesticide for the tree are displayed directly on the screen.

[0124] S9. Monitor the execution of the operation and compare the status before and after the application of the pesticide to calculate the control effect index. If the control effect index is lower than the preset threshold, then reverse the urgency calculation parameters and the weight coefficient and judgment threshold of the diagnostic model.

[0125] It should be noted that the intelligent spraying equipment reads the individual tree operation parameters in the instruction file to perform differentiated operations. The equipment records and transmits the actual spraying trajectory, actual flight speed, instantaneous flow rate, cumulative amount of pesticide used, and operation timestamp in real time. It compares the planned parameters with the actual execution data and generates an operation compliance report. If the actual amount of pesticide used on a tree does not reach the standard threshold, it is automatically marked as needing to be re-sprayed and inserted into the next round of task queue.

[0126] In the next inspection cycle after the pesticide application is completed, the new multidimensional state probability distribution generated by S5 is called and compared with the state before pesticide application on a plant-by-plant basis. The control effect index is calculated. By calculating the change in the multidimensional state probability and the direction of the dominant state transition, the control effect is comprehensively judged. The dominant state is the maximum value of the multidimensional state probability, specifically:

[0127] If the dominant state of a tree reverses from the disease stage or latent stage to the healthy stage, or if the probability growth rate of the dead stage is lower than a preset threshold, it is considered valid.

[0128] If the dominant state of a tree changes from the latent period to the disease-prone period, or from the disease-prone period to the dormant period, it is considered invalid.

[0129] If the dominant state of a tree does not meet any of the above valid or invalid conditions, it is recorded as equal;

[0130] The specific prevention and control efficacy index is as follows:

[0131] In this formula, E represents the control efficacy index, Y represents the number of effective plants, C represents the number of plants with the same efficacy, W represents the number of ineffective plants, and Z represents the total number of plants treated. Control efficacy index values ​​are calculated for different levels to evaluate the effectiveness of different strategies.

[0132] If the prevention and control effectiveness index of a certain type of target falls below a preset threshold, the current strategy is deemed ineffective, triggering a reverse correction process, which specifically includes:

[0133] If the primary tree strategy fails, it means that the dead-end threshold is set too high or the risk of spatial spread is underestimated. The dead-end probability threshold should be lowered to include more trees in the late stage of the disease but not yet completely dead in the cleanup scope to prevent them from becoming sources of infection. The coefficient of the weight of adjacent diseased trees in the spatial transmission channel should be increased to enhance the sensitivity to clustered outbreaks. If there are dead trees around, the dead / high-risk probability of the central tree should be increased. The weight of the spatial clustering coefficient in the urgency calculation should be increased to give priority to trees located in the center of the outbreak to ensure that the core source of infection is dealt with first.

[0134] If the secondary tree strategy fails, it means that the timing of pesticide application is delayed, or the impact of external characteristics on the disease is underestimated. The probability threshold boundary W between the disease onset period and the incubation period should be dynamically adjusted. If W mainly comes from trees originally judged to be in the incubation period, the incubation period threshold should be lowered and the pesticide application strategy upgraded. If W comes from trees originally judged to be in the disease onset period, the disease onset period threshold should be raised, and these trees should be directly converted into cleanup targets instead of being treated. The weight coefficients of longhorn beetle infestation data and water stress data in the external feature extraction module should be increased, indicating that the simple spectral health index is insufficient to reflect the true risk, and more reliance should be placed on pest density and environmental pressure to predict the deterioration trend. The weight of the internal disease deterioration rate should be increased. For trees with declining physiological indicators, a higher degree of urgency should be given, and early intervention should be carried out.

[0135] If the three-level tree strategy fails, lower the probability threshold for the healthy period to make it easier for monitored trees to be included in the target range for pesticide application, reducing the ambiguity of the monitored targets. Raise the entry threshold for pesticide application targets to avoid ineffective pesticide application to low-risk trees. Increase the weight of the historical rate of change in the time evolution channel. Even if the current absolute value is normal, if the historical trend shows a decline, an early warning should be issued. Introduce the dynamic weight of meteorological inducing factors. In seasons prone to outbreaks such as high temperature and high humidity, automatically lower the intervention threshold for three-level trees and temporarily upgrade some monitored targets to pesticide application targets.

[0136] After each work cycle, the evaluation program runs automatically. If the control effect index is greater than or equal to the threshold, the current parameters are maintained and the data is archived as a historical baseline. If the control effect index is less than the threshold, a parameter correction package is generated and the configuration database of S5 / S6 / S7 is updated. The corrected parameters take effect immediately in the next inspection and application cycle. All automatically corrected parameter changes generate logs and are pushed to the expert terminal. Experts can review and confirm or manually roll back the changes to ensure the system is safe and controllable.

[0137] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A multi-source data fusion early warning and precision application method for pine wilt disease, characterized in that, Specifically: S1. Simultaneously collect multi-source data of the target forest area, including satellite remote sensing data, drone inspection data, ground IoT data, and forest area meteorological data; S2. Extract the center coordinates of individual trees based on the UAV inspection data, map the multi-source data to the center coordinates of each individual tree through inverse distance weighted interpolation, perform time alignment, and output standardized individual tree data. S3. Based on the standardized single tree data, extract the physiological indicators of the single tree and perform weighted calculations to output the single tree health index that represents the internal physiological state of each tree. S4. Based on the standardized single tree data, extract the feature vectors representing the external pathogenic pressure of each tree in parallel, including the structural feature vector that quantifies the tree size, the longhorn beetle threat vector that quantifies the external risk, and the environmental induction vector that quantifies the induction conditions. S5. The single tree health index is used as the core state vector, and features are concatenated with the structural feature vector, the longhorn beetle threat vector and the environmental inducement vector. The concatenation is then input into a pre-trained dual-channel coupled diagnostic model, which outputs the probability distribution of a single tree in a multi-dimensional state and a heat map of the probability of epidemic spread within a preset time period. S6. Based on the probability distribution of the multidimensional state, set a dynamic threshold, perform treatment strategy judgment for each tree, mark the cleaning target, the pesticide application target, the monitoring target and the health target, perform graded early warning according to the target type, generate corresponding suggested measures and push them to the management terminal; S7. For individual trees marked as targets for pesticide application, based on the probability distribution of the multidimensional state and the heat map of the epidemic spread probability, calculate the pesticide application urgency index of the individual trees marked as targets for pesticide application and generate a pesticide application priority sequence. S8. Based on the application priority sequence, generate a graded application strategy and output an instruction file containing single plant coordinates, operation mode, operation path and operation parameters to the application equipment. S9. Monitor the execution of the operation and compare the status before and after the application of the pesticide to calculate the control effect index. If the control effect index is lower than the preset threshold, then reverse the urgency calculation parameters and the weight coefficient and judgment threshold of the diagnostic model.

2. The method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease according to claim 1, characterized in that: The single-tree health index is specifically a weighted sum of physiological indicators, including red-edge position, moisture index, pigment index and photochemical reflectance index, after normalization. The physiological weight coefficients are obtained through training with historical samples. The single-tree health index ranges from 0 to 1, with higher values ​​indicating healthier trees.

3. The method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease according to claim 1, characterized in that: The structural feature vectors include: tree age / diameter at breast height (DBH) grade, canopy compactness, stand density, and trunk straightness; the longhorn beetle threat vectors include: beetle population density, diffusion resistance coefficient, diffusion risk index, and reproductive suitability; the environmental induced vectors include: water stress index, extreme temperature risk index, and combined temperature and humidity stress index.

4. The method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease according to claim 1, characterized in that: The dual-channel coupled diagnostic model adopts a hybrid neural network architecture, which includes a temporal evolution channel and a spatial propagation channel, and fuses and classifies the output through a spatiotemporal attention mechanism. The temporal evolution channel uses an LSTM long short-term memory network to process the time series of a single tree, learns the rate of change of tree health indicators, and outputs a single tree temporal hidden vector representing the internal disease evolution trend. The spatial propagation channel uses a graph neural network (GNN) to pass messages on the constructed forest topology map, aggregating the states of neighboring trees and outputting a single tree spatial state hidden vector representing the degree of external threat. The spatiotemporal attention fusion and classification output introduces a spatiotemporal attention mechanism, dynamically allocates the weights of the temporal hidden vector and the spatial hidden vector, fuses them, inputs them to the fully connected classification layer, and outputs the probability distribution of whether the single tree belongs to the healthy stage, the incubation stage, the disease stage, or the death stage.

5. The method and system for multi-source data fusion early warning and precision application of pesticides for pine wilt disease according to claim 1, characterized in that: The specific method for determining the treatment strategy for each individual tree is as follows: Based on a dynamic threshold, a hierarchical discrimination logic is executed, that is, logical judgments are performed according to the following priority order. Once a certain condition is met, subsequent judgments are stopped, and the tree is directly classified, marking each tree as one of the following: a cleanup target, a pesticide application target, a monitoring target, or a health target. Specifically: If the probability of a tree being in the dead phase is greater than or equal to the dead phase threshold, it is marked as a cleanup target; if the probability of a tree being in the dead phase is less than the dead phase threshold, and the probability of it being in the disease phase is greater than or equal to the disease phase threshold or the probability of it being in the incubation phase is greater than or equal to the incubation phase threshold, it is marked as a pesticide application target; if the probability of a tree being in the dead phase, disease phase, and incubation phase is less than its respective threshold and the probability of it being in the healthy phase is greater than or equal to the healthy phase threshold, it is marked as a healthy target; if a tree does not meet all of the above conditions, that is, the probability of none of the four states reaches the threshold, it is marked as a monitoring target.

6. The method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease according to claim 1, characterized in that: The drug administration urgency index is specifically a weighted sum of the internal disease deterioration rate, the probability of external spread risk, and the clustering coefficient; The rate of internal disease deterioration is based on a probability distribution, specifically: ;in For the rate of internal disease deterioration, This represents the probability of onset. The threshold for the onset of illness, The probability of incubation period. This is the incubation period threshold; and This is the internal disease severity weighting coefficient. This is the crisis amplification factor. The probability of external diffusion risk is based on a diffusion probability heatmap, specifically: Where W is the probability of external spread risk, P is the original probability value on the epidemic spread probability heatmap, and Pmax is the maximum value of the current forest area heatmap. The clustering coefficient is based on spatial distribution, specifically: ; Where J is the aggregation coefficient, N is the total number of trees marked as the target for pesticide application or cleaning within a radius R centered on the tree, and K is the saturation constant, which is set to maximize the aggregation effect when the number of diseased trees in the surrounding area reaches K. The allocation of urgency weight coefficients follows the principle of dynamic adaptation and has multiple built-in weight configuration templates. For different scenarios, the initial allocation defaults to the expert experience value, and the system automatically switches to the corresponding weight template based on the current season and the global average of the heat map.

7. The method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease according to claim 6, characterized in that: The process of generating the application priority sequence is as follows: Based on the calculated application urgency index score, all application targets are sorted from high to low and divided into three priority levels, generating the final task sequence, specifically: Level 1: Emergency Interception Sequence, targeting the top-ranked targets by urgency index. Prioritize working on trees that are not in good condition; Level 2: Priority Treatment Sequence, urgency index ranking The trees are scheduled to be planted immediately after the first-level task is completed; Level 3: Prevention and Consolidation Sequence, Urgency Index Ranking Trees, as targets for preventative pesticide application, will only be treated for the remaining time after the completion of Level 1 and Level 2 tasks.

8. The method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease according to claim 1, characterized in that: The tiered application strategy library is as follows: For the first-level emergency blocking sequence, a dual-mode operation of treatment and blockade is adopted to treat individual trees and spray around the canopy to form an isolation zone; for the second-level key treatment sequence, a precision treatment mode is adopted to remove individual trees at specific locations; for the third-level preventive consolidation sequence, a preventive coverage mode is adopted to spray preventive measures and form a protective barrier.

9. The method for multi-source data fusion early warning and precise application of pesticides for pine wilt disease according to claim 1, characterized in that: The specific prevention and control efficacy index is as follows: Where E is the control effect index, Y is the number of effective trees, C is the number of trees with stable growth, W is the number of ineffective trees, and Z is the total number of trees treated. The control effect is comprehensively determined by calculating the change in the probability of the multidimensional state and the direction of the shift of the dominant state. The dominant state is the maximum value of the multidimensional state probability. Specifically: if the dominant state of the tree reverses from the disease stage or the latent stage to the healthy stage, or the growth rate of the probability of the dead stage is lower than a preset threshold, it is recorded as effective; if the dominant state of the tree changes from the latent stage to the disease stage, or from the disease stage to the dead stage, it is recorded as ineffective; if the dominant state of the tree does not meet any of the above conditions for effectiveness or ineffectiveness, it is recorded as stable.

10. A multi-source data fusion early warning and precision application system for pine wilt disease, characterized in that, Specifically, it includes: Multi-source data acquisition module: used to simultaneously acquire multi-source data of the target forest area, including satellite remote sensing data, UAV inspection data, ground IoT data and forest area meteorological data; Spatiotemporal data fusion module: used to extract the center coordinates of individual trees based on the UAV inspection data, map the multi-source data to the center coordinates of each individual tree through inverse distance weighted interpolation, perform time alignment, and output standardized individual tree data; Health Index Calculation Module: Based on the standardized single tree data, extract the physiological indicators of the single tree and perform weighted calculations to output the single tree health index that represents the internal physiological state of each tree. External feature extraction module: used to extract feature vectors representing the external pathogenic pressure of each tree in parallel based on the standardized single tree data, including structural feature vectors that quantify tree size, longhorn beetle threat vectors that quantify external risks, and environmental induction vectors that quantify induction conditions. Dual-channel coupled diagnostic module: used to take the single tree health index as the core state vector, and perform feature concatenation with the structural feature vector, longhorn beetle threat vector and environmental inducement vector, input to the pre-trained dual-channel coupled diagnostic model, and output the probability distribution of the single tree in a multi-dimensional state and the heat map of the epidemic spread probability within a preset time period; The graded early warning module is used to set dynamic thresholds based on the probability distribution of the multidimensional state, perform treatment strategy judgment for each tree, mark the cleaning target, the pesticide application target, the monitoring target and the health target, perform graded early warning according to the target type, and generate corresponding suggested measures and push them to the management terminal. Application urgency assessment module: Based on the probability distribution of the multidimensional state and the heat map of the epidemic spread probability, it calculates the application urgency index of the individual trees marked as application targets and generates an application priority sequence. Application strategy planning module: Used to generate a graded application strategy based on the application priority sequence, and output an instruction file containing the coordinates of a single plant, operation mode, operation path and operation parameters to the application equipment; Adaptive optimization module: Used to monitor the execution of operations and compare the state before and after application to calculate the control effect index. If the control effect index is lower than the preset threshold, the urgency calculation parameters, as well as the weight coefficients and judgment thresholds of the diagnostic model, are corrected in reverse.

Images