Insect situation forecasting system based on image recognition and deep learning

Abstract

The invention particularly relates to an insect condition forecasting system based on image recognition and deep learning, and relates to the technical field of agricultural insect disease monitoring. A disaster damage intelligent identification and quantification module; a yield loss and economic loss evaluation module; and an automatic claim decision and execution module. According to the invention, a three-dimensional data acquisition network breaks through the limitation of a traditional single data source, unmanned aerial vehicle mesoscopic aerial photography is combined with ground microscopic acquisition, details of pest feeding traces, pest body characteristics and the like are accurately captured, satellite macroscopic data locks regional pest situation boundaries, and point-to-surface full-coverage monitoring is realized; task scheduling adopts a double-trigger mechanism, response time is shortened under extreme insect conditions, real-time data support is provided for accurate prevention and control of diseases and insect pests, and crop yield loss is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of agricultural pest and disease monitoring technology, and in particular to a pest monitoring and forecasting system based on image recognition and deep learning. Background Technology

[0002] Insect pest monitoring is a core aspect of agricultural production management; its accuracy and timeliness directly determine the effectiveness of pest and disease control and crop yield security. Traditional insect pest monitoring relies on manual inspections and counting with traps, which has significant drawbacks: Manual patrols are inefficient, and full-coverage monitoring of large areas of farmland takes several days to several weeks, which cannot meet the real-time needs during the outbreak of pests; the accuracy of judgment depends on the experience of personnel, and the estimation of pest species, age and density has large subjective bias, which is prone to misjudgment and missed judgment. The data collection is limited to a single dimension, only able to obtain local point data, making it difficult to form a macro-level understanding of the distribution and spread trends of pests in a region; moreover, the data lacks an effective evidence preservation mechanism, and the correlation between the monitoring and reporting results and subsequent agricultural insurance loss assessment and control effect evaluation is weak, which can easily lead to disputes.

[0003] While existing insect pest monitoring technologies based on image recognition have made some progress, they still have many shortcomings: most systems rely on a single data source (such as images of fixed traps) and cannot take into account both macroscopic areas and microscopic insect pest details; the recognition models mostly use general visual algorithms and do not integrate agricultural expertise, resulting in insufficient accuracy in identifying similar pest species, young larvae, and insect pests in complex field backgrounds; and the monitoring results are difficult to directly serve subsequent production decisions and insurance claims.

[0004] Therefore, an insect pest monitoring and forecasting system based on image recognition and deep learning is proposed to address the aforementioned problems. Summary of the Invention

[0005] The purpose of this invention is to propose an insect pest monitoring and forecasting system based on image recognition and deep learning to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Insect pest monitoring and forecasting system based on image recognition and deep learning includes: The multi-source data acquisition and collaborative sensing module is configured to schedule tasks based on dual triggers of farmer reports and system early warnings, and to collect data through a three-dimensional network to construct a three-dimensional evidence chain; The intelligent disaster damage identification and quantification module is configured with a customized model library based on agronomic knowledge to identify disasters, accurately quantify damage indicators, and solidify evidence through blockchain to achieve damage assessment. The yield loss and economic loss assessment module is configured to combine an agronomic knowledge base with a fusion model to establish a correlation between damage data and yield loss, and then scientifically calculate the compensation amount based on insurance terms and market data. The automated claims decision-making and execution module is configured to generate standardized reports, achieve intelligent approval through a rules engine, and link up to complete automatic payment and dispute handling.

[0007] Preferably, the multi-source data acquisition and collaborative sensing module specifically includes: When farmers report a case, the system will initiate a step-by-step guided form filling process: Automatically obtains mobile phone GPS location and displays satellite image of the land plot, allowing farmers to manually select the damaged boundary; A pop-up disaster type tree menu, along with a diagram illustrating disaster characteristics, helps farmers make accurate matches; The system connects to data sources including: regional insect monitoring stations, real-time APIs from meteorological departments, and satellite remote sensing vegetation index monitoring; Drone aerial photography unit: The system automatically generates a grid flight path based on the plot boundaries to ensure no blind spots in shooting; the flight altitude is intelligently adjusted according to crop height and resolution requirements. After aerial photography is completed, orthophoto maps and multispectral composite maps are automatically generated by stitching together, with a unified data format. The ground-based mobile unit focuses on capturing images of areas blind spots for drones, including diseases, pests, and physical damage. Each photo is automatically associated with the plot ID, sample point coordinates, shooting time, and surveyor ID, and the annotation content is stored in a bound manner with the image data; Satellite remote sensing unit, disaster boundary confirmation: By comparing satellite images from 30 days before the disaster with those from 1-3 days after the disaster, the boundaries of the affected area were determined using NDVI difference maps, and cross-validated with plot-level data from drone aerial photography.

[0008] Preferably, the intelligent disaster damage identification and quantification module specifically includes: Construct an image database of agricultural disasters, with each sample containing: basic information and annotation content; Optimize the U-Net semantic segmentation model: Introduce an attention mechanism; Add residual connections; The model output is a classification map, where each pixel is labeled as healthy crop, disease damage, insect damage, physical damage, or non-crop area. Model branch functionality refinement: Pest damage recognition branch: trained based on the feeding characteristics of different pests; Disease identification branch: Focuses on the color, shape, and texture of lesions; Physical damage identification branch: based on morphological features and spatial relationship identification; Healthy tissue identification branch: Simultaneously segment healthy crop areas as a benchmark for quantitative comparison, excluding interference from non-crop areas.

[0009] Preferably, the method further includes quantifying the degree of damage: Crop region extraction: First, extract all crop pixels from the image using a crop segmentation model; Damaged pixel statistics: The number of damaged pixels is counted according to the categories of disease, pests, and physical damage; Error correction: For areas with blurred edges (such as the transition zone between lesions and healthy tissue), fuzzy mathematics algorithms are used to calculate the attribution probability, reducing human judgment errors, and the final result is accurate to two decimal places; After obtaining the total number of damaged pixels of each type, divide it by the total number of crop pixels to obtain the percentage of damaged area. The damaged areas are presented in the form of heat maps and color blocks. The damaged area is divided into k damage levels, and each damage level is assigned a corresponding weight factor. After determining all damage levels of the crop, the product of each damaged area and its corresponding weight factor is calculated, and the sum is divided by the total damaged area to obtain the multi-level damage index. By combining multispectral data and meteorological forecasts, we can determine whether the disaster will spread. It adopts a consortium blockchain architecture, with nodes including insurance companies, agricultural and rural affairs bureaus, third-party notary institutions, and technology providers.

[0010] Preferably, the method further includes: Select core data from the entire loss assessment process and put it on the blockchain: Original evidence: Key images acquired by drones / ground / satellite; Key results: image recognition and annotation map, percentage of damaged area, and multi-level damage index; Process data: model calculation parameters, operation logs; Data encryption: The data is encrypted using an encryption algorithm to generate a unique hash value; On-chain storage: Hash values ​​and data metadata are uploaded to blockchain nodes, and each node stores them synchronously, so that a single node cannot tamper with the data; Evidence storage certificate: Generates a blockchain evidence storage number and a public query link; Insurance company / farmer verification: Enter the certificate number to view the data hash value, generation time, and node storage record. Compare the hash value with the local data to confirm whether the data has been tampered with. Judicial verification: The evidence data is connected to the Internet court's evidence storage platform.

[0011] Preferably, the production loss and economic loss assessment module specifically includes: A structured knowledge base is constructed based on crop type, growth period, and disaster type, storing a preset number of agronomic parameter combinations; A fusion model combining gradient boosting trees and neural networks is adopted: Input features include: Key characteristics: percentage of damaged area, multi-level damage index, crop growth period, disaster type, disaster duration, and crop variety resistance level; Auxiliary features: average yield over the past 3 years, soil fertility level, irrigation conditions, fertilization records, weather forecasts, and planting density; Model training: The model is trained using a preset amount of historical loss assessment-actual production data combination, and the parameters are optimized through 5-fold cross-validation to control prediction error; Prediction output: Generates the yield loss rate and confidence interval.

[0012] Preferably, the method further includes economic value calculation: Input data and calculations: Key input data include: yield loss rate, expected normal yield, and crop market price; Daily synchronization with three major data sources: government agricultural and rural affairs bureau guidance price, local agricultural product wholesale market average price of the previous week, and futures market average price of corresponding commodities of the previous week. The average of the three prices is used as the benchmark price. If the price fluctuates drastically, the average price of the 15 days prior to the loss determination is used to ensure that the price is fair. Insurance terms and conditions parameters: automatically retrieved from the insurance company's core business system, including deductible, reimbursement ratio, and sum insured per unit area.

[0013] Preferably, the automated claims decision-making and execution module specifically includes: It automatically generates two types of reports: a professional version and a farmer version, which respectively meet the needs of insurance companies for internal approval, regulatory verification, and farmers' right to know, avoiding communication costs caused by vague report content and missing key data; Using a visual rule configuration platform, insurance companies can flexibly set approval rules based on insurance type, regional policies, and claim amount. The rule engine supports multiple condition combinations, priority sorting, and dynamic updates. Rule matching: The system automatically matches the loss assessment results, farmer information, and policy information with the rule base and completes the judgment within a preset time. The system automatically generates a claim approval form, updates the policy status to "claim pending payment" synchronously, pushes a payment instruction to the financial system, and sends a claim approval notification to the farmer. The system will push the report to the workflow interface of the corresponding reviewer.

[0014] Preferably, the method further includes: The system automatically links to the agricultural subsidy account reserved by the farmer when applying for insurance, and verifies that the account holder's name and ID number match the insured's information.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention breaks through the limitations of traditional single data sources by using a three-dimensional data acquisition network. It combines drone aerial photography with ground micro-collection to accurately capture details such as pest feeding traces and insect characteristics, while satellite macro data locks the boundaries of regional pest infestations, achieving full coverage monitoring from point to area. The task scheduling adopts a dual-trigger mechanism, which shortens the response time under extreme pest conditions, provides real-time data support for precise pest control, and effectively reduces crop yield losses.

[0016] 2. This invention uses blockchain technology to store core data such as original images and recognition results on the blockchain, generating a unique hash value to ensure immutability and solving the problem of insufficient credibility of traditional monitoring and reporting data. The quantitative output indicators such as insect density and the proportion of damaged area can be directly connected to the agricultural insurance loss assessment system, providing an objective basis for yield loss assessment and reducing claims disputes. At the same time, by combining meteorological data with the spread trend prediction of crop growth period, it provides farmers with targeted prevention and control suggestions and insurance companies with risk warnings, realizing the extension of monitoring and reporting data to multiple scenarios such as production decision-making and risk protection, and promoting the transformation of agricultural insect management from passive response to proactive prediction. Attached Figure Description

[0017] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0018] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0019] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0020] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0021] Appendix Figure 1 The diagram below illustrates the structure of an insect pest monitoring system based on image recognition and deep learning, as provided in this embodiment of the invention. It shows the connection between the multi-source data acquisition and collaborative sensing module and the automated claims decision-making and execution module, and marks the main functional interaction flow of each module.

[0022] In this embodiment, it includes: The multi-source data acquisition and collaborative sensing module is configured to schedule tasks based on dual triggers of farmer reports and system early warnings. It collects data through an air-ground-space three-dimensional network and stores it in encryption, thus constructing a three-dimensional evidence chain to solidify the foundation for loss assessment. Specifically, it includes: When farmers report a case through the dedicated app, the system will initiate a step-by-step guided form filling process: Automatically acquires mobile phone GPS location (accuracy ±1 meter) and displays satellite image of the plot simultaneously, allowing farmers to manually select the damaged boundary (supports zooming in to a 1:100 scale, error ≤5 square meters). A pop-up tree menu of disaster types (first-level classification: diseases / insects / physical disasters; second-level classification: such as diseases, which are divided into 20 categories including rice blast and powdery mildew; third-level classification: such as rice blast, which are divided into leaf blast and neck blast) is provided, along with a diagram of disaster characteristics (such as leaf blast, which is a spindle-shaped lesion) to help farmers match accurately. The app prompts users to upload initial photos taken on-site (at least 3: a panoramic view of the plot, a close-up of the damage, and a crop variety label). The app will automatically mark the shooting time and location to prevent later tampering. After submission, the system immediately generates a report receipt, which includes the task number and the estimated inspection time (e.g., if an inspection is conducted by drone, the inspection will be completed within 2 hours), alleviating farmers' anxiety.

[0023] The system connects to three major data sources to enable early risk identification, including: Regional pest monitoring stations: These stations connect to the pest monitoring and forecasting data from the Agriculture and Rural Affairs Bureau. When the density of a target pest (such as locusts) in a certain insured area is greater than or equal to the extreme threshold (500 aphids per 100 plants for wheat, 1000 planthoppers per 100 clumps for rice) for three consecutive days, the system automatically marks high-risk plots, calculates the estimated affected area, and sends pre-inspection suggestions to the insurance company. Meteorological Department Real-time API: Synchronize minute-level meteorological data from the Central Meteorological Observatory and local meteorological bureaus. When extreme weather such as hail (diameter ≥5mm), rainstorm (daily precipitation ≥50mm), or frost (temperature ≤0℃) is detected in the insured area and lasts for ≥30 minutes, the system immediately matches the insured plots in the area and generates an emergency survey task with the highest priority. Satellite remote sensing vegetation index monitoring: Daily data from Sentinel-2 satellite is used to calculate the NDVI (Normalized Difference Vegetation Index, reflecting crop growth) of insured plots. When the NDVI value of a plot decreases by ≥30% compared to the average of the previous 30 days, and normal fluctuation factors such as irrigation and fertilization are excluded (based on farmers' planting records), the system marks the plot with abnormal growth and recommends initiating ground verification and drone aerial photography.

[0024] The system has a built-in task priority algorithm: emergency tasks (within 24 hours after extreme weather) > early warning tasks (insect infestation / abnormal growth) > ordinary reporting tasks. At the same time, it combines the location of the surveyor and the drone's battery life to intelligently dispatch tasks. For example, if there are 3 survey tasks in a certain area, the system will prioritize assigning them to the surveyor who is closest to the area and whose drone has ≥80% battery power, thus reducing the time spent idling.

[0025] A three-tiered data collection network—macro, meso, and micro—is constructed, with each level having clear technical indicators and collection standards to ensure that the data can be cross-verified.

[0026] The drone aerial photography unit (mid-level data acquisition) determined the overall landscape of the site: Equipment configuration: All devices use DJI M300RTK, Zenmuse P1 camera, and multispectral sensor. The RTK positioning accuracy is ±1cm. The RGB camera has a resolution of 20 megapixels. The multispectral sensor covers four core bands: blue (450nm), green (530nm), red (670nm), and near-infrared (820nm) (corresponding to different characteristics of crop chlorophyll, water, and pests and diseases). Flight planning: The system automatically generates grid flight paths based on plot boundaries, with an overlap of ≥70% between adjacent flight paths, and each photo covers an area of ​​10m×10m to ensure no blind spots; the flight altitude is intelligently adjusted according to crop height and resolution requirements. The flight altitude is 50 meters for short-stalked crops such as wheat and rice, and 80 meters for tall-stalked crops such as corn and fruit trees, ensuring a pixel resolution of 5cm / pixel (which can clearly identify lesions of 1 square centimeter). Data processing: After aerial photography is completed, the system automatically stitches together orthophotos (correcting distortions caused by terrain undulations) and multispectral composite images (overlaying data from four bands to highlight damaged areas; for example, areas affected by pests and diseases show a significant decrease in reflectivity in the near-infrared band, resulting in a darker tone). The data format is uniformly TIFF with a resolution of ≥300dpi. Ground-based mobile unit (microscopic data acquisition) focuses on details that drones cannot recognize: Equipment and Operation: Surveyors carry intelligent data acquisition terminals (with built-in macro lenses, supplementary lighting, and GPS modules), and select five 1m x 1m sample areas (four corners plus the center) on the plot using the diagonal five-point sampling method. They focus on capturing images in areas blind spots for the drone, including: Diseases: lesions on the underside of leaves (such as spore masses of powdery mildew), vascular bundles in stems (such as brown mycelium of rice sheath blight), and cross-section of fruits (such as brown rot of tomato late blight). Pests: Live pests (photographs of insect characteristics for species identification), feeding traces (such as the notched edges of armyworms, honeydew of aphids); Physical damage: cross-section of the stem break (distinguishing between natural disasters and man-made breakage), hail damage on the fruit surface (record the number and size); Each photo is automatically associated with the plot ID, sample point coordinates, shooting time, and surveyor ID. The terminal supports voice annotation (e.g., sample point 3, corn stalks are completely broken, with no insect damage). The annotation content is stored in conjunction with the image data. Satellite remote sensing units (macro-level data acquisition) for boundary confirmation of large-scale disasters: Data sources: Priority is given to Gaofen-6 (a domestic satellite, with a resolution of 2 meters and a revisit period of 2 days) and Sentinel-2 (a European satellite, with a resolution of 10 meters and a revisit period of 5 days). MODIS data (with a resolution of 250 meters and strong real-time performance) is overlaid when the disaster area is greater than 1,000 mu. Core function: By comparing satellite images from 30 days before the disaster with those from 1 to 3 days after the disaster, the boundaries of the disaster area are identified through NDVI difference maps (areas with a difference of <0 are damaged areas). This data is then cross-validated with plot-level data from drone aerial photography to prevent drones from missing edge plots or farmers from falsely reporting the extent of the disaster. For example, satellite data shows that 5,000 mu of land in a certain township has been affected by the disaster, and drones can be used to survey an insured plot of land located in the core disaster area, further quantifying the proportion of damage.

[0027] All collected data is uploaded to a distributed database (supporting petabyte-level storage), categorized and indexed by task number, plot ID, and data type, and is subject to triple encryption: transmission encryption (HTTPS), storage encryption (AES-256), and access encryption (hierarchical access control: surveyors can only view their own task data, while insurance company administrators can view all data).

[0028] The disaster damage intelligent identification and quantification module is configured with a customized AI model library based on agronomic knowledge to achieve disaster identification, accurately quantify damage indicators, and solidify evidence through blockchain to achieve AI-driven accurate damage assessment. Specifically, it includes: An agricultural disaster image database was constructed, covering 15 mainstream crops (rice, wheat, corn, soybeans, fruits and vegetables, etc.) and 20 common disasters (10 diseases, 6 pests, and 4 physical disasters). Each sample includes: Basic information: crop type, variety, growth period, shooting location, shooting time; Labeling content: disaster type, damage level, and boundary of damaged area (pixel level); Data augmentation: Expand the sample by means of rotation, scaling, brightness adjustment, adding noise, etc. (e.g., generate 10 samples from different angles from 1 rice blast photo) to avoid model overfitting.

[0029] Optimize the U-Net semantic segmentation model: Introducing a focus attention mechanism (CBAM): This allows the model to automatically focus on disaster-featured areas and reduce background interference such as soil and weeds; for example, when identifying wheat aphids, the model prioritizes focusing on small black dots (aphids) on the leaves rather than shadows on the ground. Adding residual connections (ResNet): solves the gradient vanishing problem in deep networks and improves the ability to identify minute damage; for example, it can identify early lesions as small as 0.5 square centimeters on leaves (which are easily missed by ordinary models).

[0030] The model output is a classification map, where each pixel is labeled as healthy crop, disease damage, insect damage, physical damage, or non-crop area. Model branch functionality refinement: Pest damage recognition branch: trained based on the feeding characteristics of different pests; Chewing pests (locusts, armyworms): Identify leaf holes and notches, count the number of holes (only holes ≥1mm are counted) and the length of notches (the cumulative length of notches as a percentage of the leaf circumference). Piercing-sucking pests (aphids, rice planthoppers): Identify chlorotic spots and honeydew marks on leaves, and combine multispectral data (piercing-sucking pests cause a reduction in chlorophyll on leaves and an increase in red band reflectance) to improve identification accuracy; Disease identification branch: Focuses on the color, shape, and texture of lesions; Rice blast: Identify spindle-shaped lesions with brown edges and grayish-white centers to differentiate between leaf blast (seedling stage) and neck blast (grain-filling stage). Powdery mildew: Identify the white powdery covering by distinguishing between the initial stage (sparse powder) and the peak stage (dense powder) by texture features. Physical damage identification branch: based on morphological features and spatial relationship identification; Lodging: Calculate the angle between the crop stem and the ground (<30° is considered lodging), and distinguish between root lodging (the whole plant is tilted) and stem lodging (the stem is bent). Stem breakage: Identify areas of sudden grayscale changes in stem pixels (the broken area is darker), and combine this with cross-sectional photos collected from the ground to confirm whether it was caused by a disaster; Healthy tissue identification branch: Simultaneously segment healthy crop areas as a benchmark for quantitative comparison, excluding interference from non-crop areas; The system has a built-in model iteration engine: after processing a preset number of damage assessment cases, it automatically filters out samples with an identification error >10% (such as the model misjudging mechanical damage as hail damage), pushes them to the manual review platform, and adds them to the training set after being re-labeled by agronomic experts. The model parameters are updated through fine-tuning to ensure the adaptability of the model in different regions (rice in the south vs. wheat in the north) and different climates (drought or rainy).

[0031] It also includes the quantification of the degree of damage: Crop region extraction: First, extract all crop pixels from the image using a crop segmentation model (excluding soil, weeds, and field ridges) to ensure accurate calculation basis; Damaged pixel statistics: The number of damaged pixels is counted according to the categories of disease, pests, and physical damage. It also supports viewing the percentage of a certain type of disaster. Error correction: For areas with blurred edges (such as the transition zone between lesions and healthy tissue), fuzzy mathematics algorithms are used to calculate the attribution probability, reducing human judgment errors, and the final result is accurate to two decimal places; After obtaining the total number of damaged pixels of each type, divide it by the total number of crop pixels to obtain the percentage of damaged area. The damaged areas are presented using heat maps and color-coded blocks: Heat map: The degree of damage is represented by the color depth (dark red = severe, bright red = moderate, light red = mild), which visually shows the concentrated areas of the disaster within the plot (e.g., the northwest corner of the plot is the most severely damaged). Color-coded labels: Different colors represent disaster types (blue = physical damage, green = insect pests, yellow = diseases), which are overlaid on the orthophoto map. Clicking on any area will show the percentage of damage and the level of damage in that area.

[0032] The damaged area is divided into k (k takes 3-5) damage levels, and each damage level is assigned a corresponding weight factor. After determining all damage levels of the crop, the product of each damaged area and its corresponding weight factor is calculated, and the sum is divided by the total damaged area to obtain the multi-level damage index. By combining multispectral data and meteorological forecasts, we can determine whether the disaster will spread. For example, by using multispectral data to discover that the near-infrared reflectance of the lesion area continues to decline and that there will be rainfall in the next 3 days, the system predicts that the disease will spread by 15%-20% within 72 hours, providing insurance companies with a reference for whether to prepay part of the compensation. The system supports manual verification: surveyors can manually select suspected misjudged areas on the orthophoto map, and the system will recalculate the damage percentage of the area and compare it with the AI ​​result. If the error is greater than 5%, the model will issue an early warning and the sample will be included in the incremental learning set.

[0033] The blockchain adopts a consortium blockchain architecture, with nodes including insurance companies, agricultural and rural affairs bureaus, third-party notary institutions (such as notary offices), and technology providers, to ensure data credibility and avoid problems such as damage assessment data being tampered with and claims disputes having no verifiable evidence.

[0034] Also includes: The evidence preservation process and content are as follows: Data screening for evidence preservation: Select core data from the entire loss assessment process and upload it to the blockchain to avoid redundancy: Original evidence: Key images collected by drones / ground / satellite (compressed and encrypted, size ≤100MB / image); Key results: image recognition and annotation map, percentage of damaged area, and multi-level damage index; Process data: model calculation parameters (such as pixel statistical formulas), operation logs (who modified the data and when).

[0035] Hash Encryption and On-Chain: Data encryption: The above data is encrypted using the SHA-256 algorithm to generate a unique hash value (such as "a8f5e2d3..."). Even if the data content is modified by just one pixel, the hash value will be completely different. On-chain storage: The hash value and data metadata (generation time, operation subject ID) are uploaded to the blockchain node, and each node stores them synchronously. A single node cannot tamper with the data (more than 51% of the nodes must agree to modify it, which is almost impossible in a consortium blockchain). Evidence storage certificate: Generates a blockchain evidence storage number and a public query link, which can be accessed through a browser without the need to install special software.

[0036] Data tracing and verification: Insurance company / farmer verification: Enter the certificate number to view the data hash value, generation time, and node storage record. Compare the hash value with the local data to confirm whether the data has been tampered with. Judicial verification: The evidence data is connected to the Internet court's evidence storage platform and can be used directly as judicial evidence without the need for additional notarization, thus reducing the cost of dispute resolution.

[0037] The yield loss and economic loss assessment module is configured to combine an agronomic knowledge base with a fusion model to establish a correlation between damage data and yield loss, and then scientifically calculate the compensation amount based on insurance terms and market data. Specifically, it includes: In traditional damage assessment, the simple conversion that 30% of the damaged area equals 30% of the yield loss completely violates agronomic principles. For the same crop, the losses vary greatly depending on the growth stage and the type of disaster (e.g., if 30% of the leaves are damaged during the rice seedling stage, the yield loss is only 2%; if 30% of the leaves are damaged during the grain-filling stage, the yield loss reaches 28%).

[0038] A structured knowledge base is built according to crop type, growth period, and disaster type, storing a preset number of agronomic parameter combinations; it is connected to the annual agronomic research report of the Agriculture and Rural Affairs Bureau and is supplemented with new crop varieties and new disaster type parameters (such as new diseases and pests) every year. A hybrid model combining gradient boosting trees (XGBoost) and neural networks (MLP) is employed. Input features include: Core features (6 dimensions): percentage of damaged area, multi-level damage index, crop growth period (coded as 0-5, corresponding to sowing to maturity), disaster type (coded as 0-19), duration of disaster (days), and crop variety resistance level (levels 1-5, with level 1 being the most resistant). Auxiliary features (6 dimensions): average yield over the past 3 years, soil fertility level (1-3), irrigation conditions (present / absent), fertilization records (whether fertilizer was applied in the last 30 days), weather forecast (rainfall / sunshine in the next 10 days), and planting density. Model training: The model is trained using a preset amount of historical loss assessment and actual yield data (e.g., in 2023, 40% of wheat in a certain plot was damaged by stripe rust during the jointing stage, and the actual yield loss was 38%). The parameters are optimized through 5-fold cross-validation to control the prediction error. Forecast output: Generates the yield loss rate and confidence interval. The confidence interval reflects the reliability of the forecast; the more sufficient the data, the narrower the confidence interval. The model prediction results were revised based on the actual conditions of the land parcel: If farmers can provide yield records and fertilization and irrigation certificates for the past three years, the system will recalculate these data as additional features. If the plot is used for specialty crops (such as organic vegetables or Chinese medicinal herbs), the system will call the specialty crop agronomic database to supplement specific parameters (such as the loss pattern of ginseng root rot) to avoid bias in the general model.

[0039] It also includes economic value calculation: Input data and calculations: Core input data (4 types), including: Production loss rate; Expected normal production: Priority will be given to the target yield declared by farmers when they apply for insurance (proof of yield for the previous 3 years, such as grain sales receipts, is required); If the declared value deviates by more than 30% from the average yield of the same crop over three years published by the local agricultural department, the average yield shall be used (to prevent farmers from falsely reporting yields to fraudulently claim insurance). Crop market prices: Daily synchronization with three major data sources: government agricultural and rural affairs bureau guidance price, local agricultural product wholesale market average price of the previous week, and futures market average price of corresponding commodities of the previous week. The average of the three prices is used as the benchmark price. If the price fluctuates drastically (e.g., a single-day increase of more than 10%), the average price of the 15 days prior to the loss determination is used to ensure fair pricing. Insurance terms and conditions parameters: automatically retrieved from the insurance company's core business system, including deductible (e.g., a deductible of 500 yuan for a single accident), reimbursement ratio (e.g., 80% of the loss exceeding 5% will be reimbursed), and insured amount per unit area (e.g., 1200 yuan per mu for rice).

[0040] The calculation example is as follows: A rice paddy field with an insured area of ​​5 mu (approximately 0.33 hectares), an expected normal yield of 1000 jin (approximately 500 catties) per mu (approximately 600 kg / hectare), a yield loss rate of 32.5%, a crop market price of 1.5 yuan per jin (approximately 0.067 yuan / catties), a deductible of 500 yuan, and a compensation Calculate the expected production loss: Estimated lost output = Expected normal output × Output loss rate × Insured area; That is: 1000 jin / mu × 32.5% × 5 mu = 1625 jin; Calculate the theoretical compensation amount (take the lower of the two methods, following the principle of loss compensation): Method 1 (based on yield loss): 1625 jin × 1.5 yuan / jin = 2437.5 yuan; Method 2 (based on insured amount): Insured amount per unit area × insured area × yield loss rate = 1200 yuan / mu × 5 mu × 32.5% = 1950 yuan; Theoretical compensation amount = min(2437.5, 1950) = 1950 yuan; Calculate the final claim amount: Final claim amount = max(0, (theoretical claim amount - deductible) × claim ratio); That is: (1950-500)×80%=1450×80%=1160 yuan; If the theoretical payout is less than the deductible (e.g., the theoretical payout is 400 yuan), then the final payout will be 0.

[0041] The calculation results are compiled into a loss accounting table, which details the calculation formula, input parameters, and source of values ​​for each step. For example, the crop market price of 1.5 yuan / jin comes from the average price of agricultural products wholesale market in XX City from XX month XX day to XX day in 2024, ensuring that farmers and insurance companies can clearly trace the basis of the calculation.

[0042] The automated claims decision-making and execution module is configured to generate standardized reports, achieve intelligent approval through a rules engine, and link up to complete automatic payment and objection handling. Specifically, it includes: It automatically generates two types of reports: a professional version and a farmer version, which respectively meet the needs of insurance companies for internal approval, regulatory verification, and farmers' right to know, avoiding communication costs caused by vague report content and missing key data; The professional version report includes the following: Task Overview: Task number, reporting time, investigation time, coordinates of the affected area, crop information (type, variety, growth period); Data collection instructions: UAV flight path map, ground sample point distribution table, satellite image source and comparison map; Image recognition results: Comparison of original image and identified annotation map (with boundary markings of damaged areas), accuracy rate of various disaster identifications; Damage quantification indicators: calculation process of damaged area percentage (including raw pixel statistics data), damage heat map, and multi-level damage index; Yield loss assessment: Model input parameter table, yield loss rate prediction process, and confidence interval explanation; Economic loss accounting: loss accounting sheet, price basis, insurance terms and conditions parameters; Blockchain-based evidence storage information includes: evidence storage number, query link, and hash value list. Risk warning: Disaster spread trend prediction and subsequent disaster prevention recommendations (such as recommending the application of XX pesticide to prevent the spread of diseases); The farmer version of the report includes the following: Key results: Photos of damaged plots (damaged areas marked), percentage of damaged area, estimated yield loss, and final compensation amount; Basis for claims: Reference to the corresponding clause in the insurance policy (e.g., according to Article X of your policy, the reimbursement ratio is 80%), and price basis (e.g., crop price is 1.5 yuan / jin, referencing the average local market price). User Guide: How to check the status of compensation payments, channels for handling objections (telephone, app access), and methods for checking blockchain evidence; Disaster prevention recommendations: Follow-up response measures for this disaster (such as timely support of lodged rice seedlings and application of foliar fertilizer).

[0043] Within 10 minutes of the report being generated, it will be pushed to the insurance company’s core business system (for approval), the surveyor’s App (for on-site confirmation), and the farmer’s App (for viewing). Using a visual rule configuration platform (no coding required), insurance companies can flexibly set approval rules based on insurance type (such as wheat insurance / rice insurance), regional policies (such as poverty alleviation policies for impoverished counties), and claim amount. The rule engine supports multiple condition combinations, priority sorting, and dynamic updates. Rule matching: The system automatically matches the loss assessment results (loss rate, claim amount), farmer information (historical records, region), and policy information with the rule base and completes the judgment within a preset time. The system automatically generates a claim approval form, updates the policy status to "claim pending payment" synchronously, pushes a payment instruction to the financial system, and sends a claim approval notification to the farmer. The system will push the report to the workflow interface of the corresponding reviewer, highlighting key verification items (such as "If the claim amount exceeds 100,000, production certificate needs to be verified"). The reviewer will then provide online comments (approve / reject / require supplementary materials). Rejection requires an explanation of the reasons. If a rule conflict occurs (e.g., a case satisfies two rules simultaneously), the system will automatically trigger the highest priority rule and send a rule conflict alert to the administrator, suggesting that the rule be optimized.

[0044] The system automatically links to the agricultural subsidy account reserved by farmers when they apply for insurance (data is connected with the government's agricultural department), and verifies that the account holder's name and ID number match the insured's information to avoid incorrect payments; If an account status is abnormal (such as being frozen), the system will immediately send an account abnormality alert to guide farmers to update their account information; For scenarios where multiple farmers in the same area are assessed for losses at the same time (such as after a hail disaster), the system supports batch generation of payment instructions. After the finance department confirms with one click, the batch transfer is completed through the bank's API interface, reducing repetitive operations. Payment tracking: After the transfer, the system receives bank payment receipts in real time (including transaction number and arrival time). If the payment fails (e.g., the account does not exist), it will automatically retry and send a payment failure notification. Status Update: After successful payment, the system immediately updates the policy status to "claim completed" and synchronizes the transaction serial number to the farmer's App and the insurance company's system.

[0045] Within one minute of receiving the compensation payment, a notification will be sent to the farmer via SMS and App push notification. The notification will include: Your agricultural insurance compensation of XX yuan has been received, transaction number XXX, arrival time XXX, and a link to view the loss assessment report.

[0046] If farmers disagree with the loss assessment results, they can submit an application through the objection feedback portal on the App, uploading supplementary photos, production certificates (such as grain sales receipts), opinions from agronomic experts, and other materials. The system will then automatically generate an objection work order.

[0047] The system promises a response within 24 hours and a review result within 48 hours. Objection work orders are given priority to the review team composed of agronomic experts and damage assessment specialists. The review team can retrieve the original data again, conduct supplementary investigations, and provide opinions on maintaining the original result / correcting the result / supplementing materials.

[0048] The review results are notified to farmers via the app and SMS. If the results are corrected, the system will automatically recalculate the compensation amount and initiate the payment process.

[0049] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0050] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0051] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "includes a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0052] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0053] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0054] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0055] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0056] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0058] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An insect pest monitoring and forecasting system based on image recognition and deep learning, characterized in that, include: The multi-source data acquisition and collaborative sensing module is configured to schedule tasks based on dual triggers of farmer reports and system early warnings, and to collect data through a three-dimensional network to construct a three-dimensional evidence chain; The intelligent disaster damage identification and quantification module is configured with a customized model library based on agronomic knowledge to identify disasters, accurately quantify damage indicators, and solidify evidence through blockchain to achieve damage assessment. The yield loss and economic loss assessment module is configured to combine an agronomic knowledge base with a fusion model to establish a correlation between damage data and yield loss, and then scientifically calculate the compensation amount based on insurance terms and market data. The automated claims decision-making and execution module is configured to generate standardized reports, achieve intelligent approval through a rules engine, and link up to complete automatic payment and dispute handling.

2. The insect pest monitoring system based on image recognition and deep learning according to claim 1, characterized in that, The multi-source data acquisition and collaborative sensing module specifically includes: When farmers report a case, the system will initiate a step-by-step guided form filling process: Automatically obtains mobile phone GPS location and displays satellite image of the land plot, allowing farmers to manually select the damaged boundary; A pop-up disaster type tree menu, along with a diagram illustrating disaster characteristics, helps farmers make accurate matches; The system connects to data sources including: regional insect monitoring stations, real-time APIs from meteorological departments, and satellite remote sensing vegetation index monitoring; Drone aerial photography unit: The system automatically generates a grid flight path based on the plot boundaries to ensure no blind spots in shooting; the flight altitude is intelligently adjusted according to crop height and resolution requirements. After aerial photography is completed, orthophoto maps and multispectral composite maps are automatically generated by stitching together, with a unified data format. The ground-based mobile unit focuses on capturing images of areas blind spots for drones, including diseases, pests, and physical damage. Each photo is automatically associated with the plot ID, sample point coordinates, shooting time, and surveyor ID, and the annotation content is stored in a bound manner with the image data; Satellite remote sensing unit, disaster boundary confirmation: By comparing satellite images from 30 days before the disaster with those from 1-3 days after the disaster, the boundaries of the affected area were determined using NDVI difference maps, and cross-validated with plot-level data from drone aerial photography.

3. The insect pest monitoring system based on image recognition and deep learning according to claim 1, characterized in that, The intelligent disaster damage identification and quantification module specifically includes: Construct an image database of agricultural disasters, with each sample containing: basic information and annotation content; Optimize the U-Net semantic segmentation model: Introduce an attention mechanism; Add residual connections; The model output is a classification map, where each pixel is labeled as healthy crop, disease damage, insect damage, physical damage, or non-crop area. Model branch functionality refinement: Pest damage recognition branch: trained based on the feeding characteristics of different pests; Disease identification branch: Focuses on the color, shape, and texture of lesions; Physical damage identification branch: based on morphological features and spatial relationship identification; Healthy tissue identification branch: Simultaneously segment healthy crop areas as a benchmark for quantitative comparison, excluding interference from non-crop areas.

4. The insect pest monitoring system based on image recognition and deep learning according to claim 3, characterized in that, It also includes the quantification of the degree of damage: Crop region extraction: First, extract all crop pixels from the image using a crop segmentation model; Damaged pixel statistics: The number of damaged pixels is counted according to the categories of disease, pests, and physical damage; Error correction: For areas with blurred edges (such as the transition zone between lesions and healthy tissue), fuzzy mathematics algorithms are used to calculate the attribution probability, reducing human judgment errors, and the final result is accurate to two decimal places; After obtaining the total number of damaged pixels of each type, divide it by the total number of crop pixels to obtain the percentage of damaged area. The damaged areas are presented in the form of heat maps and color blocks. The damaged area is divided into k damage levels, and each damage level is assigned a corresponding weight factor. After determining all damage levels of the crop, the product of each damaged area and its corresponding weight factor is calculated, and the sum is divided by the total damaged area to obtain the multi-level damage index. By combining multispectral data and meteorological forecasts, we can determine whether the disaster will spread. It adopts a consortium blockchain architecture, with nodes including insurance companies, agricultural and rural affairs bureaus, third-party notary institutions, and technology providers.

5. The insect pest monitoring system based on image recognition and deep learning according to claim 4, characterized in that, Also includes: Select core data from the entire loss assessment process and put it on the blockchain: Original evidence: Key images acquired by drones / ground / satellite; Key results: image recognition and annotation map, percentage of damaged area, and multi-level damage index; Process data: model calculation parameters, operation logs; Data encryption: The data is encrypted using an encryption algorithm to generate a unique hash value; On-chain storage: Hash values ​​and data metadata are uploaded to blockchain nodes, and each node stores them synchronously, so that a single node cannot tamper with the data; Evidence storage certificate: Generates a blockchain evidence storage number and a public query link; Insurance company / farmer verification: Enter the certificate number to view the data hash value, generation time, and node storage record. Compare the hash value with the local data to confirm whether the data has been tampered with. Judicial verification: The evidence data is connected to the Internet court's evidence storage platform.

6. The insect pest monitoring system based on image recognition and deep learning according to claim 1, characterized in that, The module for assessing production losses and economic losses specifically includes: A structured knowledge base is constructed based on crop type, growth period, and disaster type, storing a preset number of agronomic parameter combinations; A fusion model combining gradient boosting trees and neural networks is adopted: Input features include: Key characteristics: percentage of damaged area, multi-level damage index, crop growth period, disaster type, disaster duration, and crop variety resistance level; Auxiliary features: average yield over the past 3 years, soil fertility level, irrigation conditions, fertilization records, weather forecasts, and planting density; Model training: The model is trained using a preset amount of historical loss assessment-actual production data combination, and the parameters are optimized through 5-fold cross-validation to control prediction error; Prediction output: Generates the yield loss rate and confidence interval.

7. The insect pest monitoring system based on image recognition and deep learning according to claim 6, characterized in that, It also includes economic value calculation: Input data and calculations: Key input data include: yield loss rate, expected normal yield, and crop market price; Daily synchronization with three major data sources: government agricultural and rural affairs bureau guidance price, local agricultural product wholesale market average price of the previous week, and futures market average price of corresponding commodities of the previous week. The average of the three prices is used as the benchmark price. If the price fluctuates drastically, the average price of the 15 days prior to the loss determination is used to ensure that the price is fair. Insurance terms and conditions parameters: automatically retrieved from the insurance company's core business system, including deductible, reimbursement ratio, and sum insured per unit area.

8. The insect pest monitoring system based on image recognition and deep learning according to claim 1, characterized in that, The automated claims decision-making and execution module specifically includes: It automatically generates two types of reports: a professional version and a farmer version, which respectively meet the needs of insurance companies for internal approval, regulatory verification, and farmers' right to know, avoiding communication costs caused by vague report content and missing key data; Using a visual rule configuration platform, insurance companies can flexibly set approval rules based on insurance type, regional policies, and claim amount. The rule engine supports multiple condition combinations, priority sorting, and dynamic updates. Rule matching: The system automatically matches the loss assessment results, farmer information, and policy information with the rule base and completes the judgment within a preset time. The system automatically generates a claim approval form, updates the policy status to "claim pending payment" synchronously, pushes a payment instruction to the financial system, and sends a claim approval notification to the farmer. The system will push the report to the workflow interface of the corresponding reviewer.

9. The insect pest monitoring system based on image recognition and deep learning according to claim 8, characterized in that, Also includes: The system automatically links to the agricultural subsidy account reserved by the farmer when applying for insurance, and verifies that the account holder's name and ID number match the insured's information.

Images