Landscaping pest and disease damage intelligent monitoring and early warning system based on AI visual identification

Abstract

The invention discloses an AI visual identification-based landscaping pest and disease damage intelligent monitoring and early warning system, and particularly relates to the technical field of AI visual identification pest and disease damage, and the system comprises the steps: obtaining time sequence visual image data, space-time labels and environment sensing data of a target garden area; inputting the time sequence visual image data into a pre-trained disease and insect pest AI identification model, and outputting an identification result; calculating initial occurrence density and local diffusion trend parameters of disease and pest species, and generating a spatial distribution thermodynamic diagram; constructing a multi-factor risk assessment model; dynamically generating a graded early warning signal; according to the invention, the landscaping pest and disease damage intelligent monitoring and early warning system is constructed through the multi-source heterogeneous data acquisition module, the feature extraction and identification module, the data aggregation and space association module, the multi-factor risk assessment and time sequence prediction module and the dynamic early warning and decision support module; the problems of single monitoring dimension, weak data fusion and space association capability, fragmentation of decision support information and the like are solved.

Description

Technical Field

[0001] This invention relates to the field of AI visual recognition technology for pests and diseases, and more specifically, to an intelligent monitoring and early warning system for pests and diseases in landscaping based on AI visual recognition. Background Technology

[0002] As a core component of the urban ecosystem, the health of landscaping directly impacts the quality of the urban ecological environment and residents' living experience. Pests and diseases, as major threats to the growth of landscaping vegetation, are characterized by rapid spread, wide impact, and strong concealment of outbreaks. If monitoring is not timely or control measures are inappropriate, large-scale withering and death of vegetation can easily occur, causing serious ecological and economic losses. Therefore, achieving accurate monitoring, timely early warning, and scientific control of landscaping pests and diseases has become a key technological requirement in the field of landscaping management. Existing technologies use high-definition cameras deployed at fixed locations to collect images of plant leaves and use convolutional neural network (CNN) models to analyze single static images to identify the presence of specific pests and diseases. However, in practical use, this approach still has technical shortcomings: First, the monitoring dimensions are limited and the spatiotemporal continuity is insufficient: existing vision-based solutions focus on static recognition of single-frame images and lack analysis of time-series image sequences of the same monitoring point. This makes it difficult to capture the dynamic development process of pests and diseases, such as the spread of lesions and the movement of insects, resulting in a weak ability to perceive the early stage and evolution trend of pests and diseases. Second, the data fusion and spatial correlation capabilities are weak: the existing solutions fail to effectively correlate and aggregate visual recognition results, environmental sensing data and high-precision geospatial information (GIS) for in-depth analysis; the recognition results are isolated and scattered, and cannot form a heat map that reflects the overall spatial distribution density and diffusion pattern of pests and diseases, which limits the ability to assess risks from a global perspective. Third, the risk assessment models are simple and lack predictive capabilities: existing early warning systems mostly rely on the current identification results or simple environmental thresholds. The models fail to effectively integrate historical time-series data, current spatial distribution characteristics, and multi-dimensional environmental factors to conduct comprehensive and quantitative risk assessments. In particular, they lack the ability to predict the future spread probability of pests and diseases based on time-series patterns, resulting in delayed early warnings and an inability to provide decision support for proactive prevention and control. Fourth, decision support information is fragmented: existing systems mostly output simple lists of identification results or early warning signals, failing to automatically generate structured decision support reports that integrate pest and disease types, spatial distribution visualization, risk levels, diffusion area identification, and targeted prevention and control suggestions. The level of intelligence and the effectiveness of decision support need to be improved.

[0003] Therefore, there is an urgent need for an intelligent monitoring and early warning system for garden and greening pests and diseases that can achieve temporal visual monitoring, deep fusion of multi-source data, spatial aggregation analysis, dynamic risk assessment and prediction, and provide integrated decision support. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent monitoring and early warning system for garden greening pests and diseases based on AI visual recognition, which solves the problems mentioned in the background art through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent monitoring and early warning system for garden pests and diseases based on AI visual recognition, comprising: Multi-source heterogeneous data acquisition module: acquires temporal visual image data of the target garden area, as well as spatiotemporal labels and environmental sensing data corresponding to the target garden area. The spatiotemporal labels include the acquisition timestamp and the coordinates of the acquisition point. Feature extraction and recognition module: Input the time-series visual image data into the pre-trained pest and disease AI recognition model, extract the time-series visual image features, and output the recognition results, which include pest and disease types, confidence levels, and location coordinates; Data aggregation and spatial association module: Aggregates and analyzes the identification results of the same collection point, calculates the initial occurrence density and local spread trend parameters of the pests and diseases, and associates the identification results of the collection point with its corresponding location coordinates to generate a spatial distribution heat map of pests and diseases on the GIS map layer; Multi-factor risk assessment and time-series prediction module: This module integrates the spatial distribution heatmap with the environmental sensor data to construct a multi-factor risk assessment model. Based on historical data and the currently integrated data, this model performs the following operations: D1: Assess the overall risk level of current pests and diseases; D2: Calculate the probability of the pest spreading to a nearby healthy area within a preset time period in the future; Dynamic Early Warning and Decision Support Module: Based on the overall risk level and the diffusion probability, dynamically generates tiered early warning signals, and generates and outputs a structured decision support report based on the early warning signals. The report includes: E1: Heat map of pest and disease types, spatial distribution, and visualization of risk levels; E2: Identifies areas where pests and diseases occur and areas with high risk of spread; E3: A list of recommended prevention and control measures based on the types of pests and diseases, initial occurrence density, and environmental sensor data.

[0006] The technical effects and advantages of this invention are as follows: 1. This invention uses temporal visual image acquisition and combines a YOLOv8 temporal enhancement model that integrates CNN and LSTM for feature extraction, which can capture the dynamic changes of pests and diseases in the time dimension. Compared with the existing single-frame static recognition scheme, this system can detect the occurrence trend of pests and diseases earlier and enhance the perception of the evolution process of pests and diseases, thereby providing support for early intervention. 2. This invention integrates the identification results with GIS geographic information and environmental sensor data through a data aggregation and spatial association module, and generates a spatial distribution heat map using a kernel density estimation algorithm; thereby assessing the occurrence of pests and diseases from a global perspective and improving the ability to accurately locate risk areas. 3. This invention employs a hybrid model combining XGBoost and LSTM, integrating spatial distribution, environmental factors, and historical time-series data. This model can assess the overall risk level of current pests and diseases, predict their future spread to adjacent healthy areas, improve the timeliness and accuracy of early warning, and support proactive prevention and control decisions. 4. This invention automatically generates a structured report containing pest and disease types, spatial heat maps, risk levels, diffusion area identifiers, and targeted prevention and control suggestions through a dynamic early warning and decision support module, and supports GIS visualization interaction. Compared with the fragmented information output by existing systems, this system integrates multi-dimensional information, provides integrated and operable decision support, and significantly enhances the practicality of the system and the level of intelligence in garden management. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the overall system structure of the present invention.

[0008] Figure 2 This is a schematic diagram of the feature extraction and recognition module of the present invention.

[0009] Figure 3 This is a schematic diagram of the data aggregation and spatial association module of the present invention.

[0010] Figure 4 This is a schematic diagram of the multi-factor risk assessment and time series prediction module of the present invention.

[0011] Figure 5 This is a schematic diagram of the dynamic early warning and decision support module of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Please see Figures 1-5 As shown in the figure, this invention provides an intelligent monitoring and early warning system for garden pests and diseases based on AI visual recognition. The system includes: Multi-source heterogeneous data acquisition module: acquires temporal visual image data of the target garden area, as well as spatiotemporal labels and environmental sensing data corresponding to the target garden area. The spatiotemporal labels include the acquisition timestamp and the coordinates of the acquisition point. Feature extraction and recognition module: Input the time-series visual image data into the pre-trained pest and disease AI recognition model, extract the time-series visual image features, and output the recognition results, which include pest and disease types, confidence levels, and location coordinates; Data aggregation and spatial association module: Aggregates and analyzes the identification results of the same collection point, calculates the initial occurrence density and local spread trend parameters of the pests and diseases, and associates the identification results of the collection point with its corresponding location coordinates to generate a spatial distribution heat map of pests and diseases on the GIS map layer; Multi-factor risk assessment and time-series prediction module: This module integrates the spatial distribution heatmap with the environmental sensor data to construct a multi-factor risk assessment model. Based on historical data and the currently integrated data, this model performs the following operations: D1: Assess the overall risk level of current pests and diseases; D2: Calculate the probability of the pest spreading to a nearby healthy area within a preset time period in the future; Dynamic Early Warning and Decision Support Module: Based on the overall risk level and the diffusion probability, dynamically generates tiered early warning signals, and generates and outputs a structured decision support report based on the early warning signals. The report includes: E1: Heat map of pest and disease types, spatial distribution, and visualization of risk levels; E2: Identifies areas where pests and diseases occur and areas with high risk of spread; E3: A list of recommended prevention and control measures based on the types of pests and diseases, initial occurrence density, and environmental sensor data.

[0014] The multi-source heterogeneous data acquisition module acquires temporal visual image data of the target garden area, as well as spatiotemporal labels and environmental sensing data corresponding to the target garden area. The spatiotemporal labels include the acquisition timestamp and the coordinates of the acquisition point. A101: Data Acquisition Preparation Phase Based on a GIS geographic information system, the boundary coordinates of the target garden area are input, which are the latitude and longitude range in the WGS84 coordinate system, to generate a digital map of the garden area; the area is divided into vegetation type sub-regions and key geographic features are marked to provide a basis for the planning of data collection points; the vegetation type sub-regions include tree areas, shrub areas, lawn areas, and flower areas; the key geographic features include roads, water systems, buildings, and steep slope areas with a slope greater than 15 degrees; The collection points are deployed according to the principle of even distribution with increased density in key areas. The even distribution is adjusted according to the area of ​​the garden. Take 10 meters, area The data collection point is divided into grids on the digital map at intervals of 10 to 50 meters, with the center point of each grid serving as the basic collection point. For the densification deployment, the spacing between the densification collection points is reduced to 5 to 10 meters in areas with high vegetation density, areas with a high incidence of historical pests and diseases, areas around water systems, and low-lying areas with poor ventilation. Invalid areas are excluded, and grids occupied by roads and buildings are removed to avoid collecting invalid data. The visual image acquisition equipment deploys a fixed high-definition camera at each acquisition point, with a resolution of 1920×1080 pixels, a frame rate of 30fps, and support for autofocus; the environmental sensing equipment deploys multi-parameter sensors synchronously at each acquisition point, integrating functions for acquiring temperature (°C), humidity (RH), soil moisture content (vol), light intensity (lux), and precipitation (H); for positioning and timing, all acquisition devices have built-in GPS positioning and NTP network time synchronization to ensure the accuracy of spatiotemporal tags.

[0015] A102: Data Acquisition and Execution Phase For time-series visual image data acquisition, a fixed high-definition camera is installed 2 to 3 meters above the ground, with the lens tilted vertically downwards at 30 to 45 degrees, focusing on the vegetation leaves and branches within a 2-meter radius of the acquisition point. Each acquisition continuously captures 3 to 5 frames of images to avoid recognition errors caused by blurry single-frame images. The acquisition frequency is set: basic acquisition points are acquired once every 2 hours, and encrypted acquisition points are acquired once every 30 minutes. Exposure parameters are automatically adjusted during acquisition to adapt to different lighting conditions, and the image format is uniformly PNG. Environmental sensing data acquisition: Sensors monitor environmental parameters in real time, trigger data sampling at the acquisition point, automatically remove outliers that exceed the measurement range, and temporarily store the valid data in the device's local cache after binding it with spatiotemporal tags; The acquisition frequency is synchronized with visual image acquisition, with basic acquisition points once every 2 hours and encrypted acquisition points once every 30 minutes, and the data output format is JSON format; Spatiotemporal tag generation: After each image acquisition or environmental data acquisition is completed, the location and timing synchronously output the acquisition timestamp and the latitude and longitude coordinates of the acquisition point to form a spatiotemporal tag; Spatiotemporal tag synchronization configuration: Each time an image or a set of environmental data is acquired, the location is automatically triggered to record the latitude and longitude coordinates of the current acquisition point and synchronize the NTP timestamp. For data transmission, the acquisition device transmits visual image data, spatiotemporal tags, and environmental sensor data to the system server via a 5G network. The transmission process uses AES-256 encryption to prevent data leakage.

[0016] A103: Data Cleaning and Format Standardization Image data cleaning: The server uses image processing algorithms to remove blurry, overexposed, and underexposed invalid images, retaining clear and valid images; all valid images are uniformly scaled to 1920×1080 pixels and converted to PNG format. Environmental data cleaning, using Outliers in environmental sensor data are removed in principle. For example, if the humidity data of a certain collection point exceeds the average humidity of the area for the same period by ±3 times the standard deviation, it is considered an outlier and removed. Missing data is filled in by linear interpolation. Spatiotemporal labels are standardized by converting the latitude and longitude coordinates of all collection points to the WGS84 coordinate system and the timestamps to the UTC time format.

[0017] The feature extraction and recognition module inputs the temporal visual image data into a pre-trained pest and disease AI recognition model, extracts temporal visual image features, and outputs recognition results, which include pest and disease types, confidence levels, and location coordinates. B101: Input Data Reception and Integrity Verification It receives data transmitted from multi-source heterogeneous data acquisition modules and parses it into: time-series visual image data, 3 to 5 standardized images continuously acquired from the same acquisition point, with a resolution of 1920×1080, PNG format, and WGS84 coordinate system spatiotemporal label binding; it verifies data integrity and removes invalid data with blurred images, missing spatiotemporal labels, or abnormal device parameters.

[0018] B102: Temporal visual image data adaptation input Single-frame temporal visual image processing, the image processing includes: color gamut conversion, converting the RGB image to the HSV color gamut, and enhancing the contrast between the diseased and pest areas and the background, the diseased and pest areas including lesions and insect bodies; Normalization scales pixel values ​​from [0, 255] to [0, 1]. The normalization formula is as follows: ,in Represents the average pixel value of the training set, The standard deviation of the training set pixels is represented; noise reduction is achieved by using Gaussian filtering to remove image noise and avoid interference with feature extraction. Temporal visual image sequence construction, which involves collecting images from the same acquisition point. Frame images are processed and then stitched together to form a temporal tensor. Dimensions are unified as ,in Indicates the length of the image time sequence frame, This indicates that the horizontal image resolution is 1920 pixels. The vertical image resolution is 1080 pixels, and 3 indicates the number of channels.

[0019] B103: Loading and Parameter Configuration of Pre-trained AI Model for Pest and Disease Identification The model architecture adopts a YOLOv8 temporal augmentation model with a CNN+LSTM hybrid architecture, adapted to the temporal image features of garden pests and diseases. The pre-trained weights are adjusted based on the labeled dataset of garden pests and diseases. The garden pests include: powdery mildew / black spot / leaf blight; the garden pests include: aphids / red spider mites / longhorn beetles, etc. Inference parameters are configured such that the initial confidence threshold is 0.5, and recognition results below this value are temporarily filtered out; the non-maximum suppression (NMS) threshold is 0.45, which is used to eliminate duplicate detection boxes; the batch processing volume is grouped by acquisition point, and each batch processes a temporal visual image sequence of 1 acquisition point.

[0020] B104: Temporal Visual Image Feature Extraction Spatial feature extraction is performed based on a CNN backbone network for each frame of the temporal tensor. Spatial features are extracted using YOLOv8's C2f module, and the core convolutional layers are computed: , in This indicates that the convolutional kernel weights are obtained from pre-training. Indicates the bias term, Represents convolution operation, Indicates activation function, in Indicates the number of channels; Temporal visual image feature extraction based on LSTM, Frame Image Spatial Feature Set Input a bidirectional LSTM layer to extract dynamic features of pests and diseases over time. These dynamic features include insect movement and the spread of lesions. The LSTM calculation formula is as follows: , in These are the input gate, forget gate, candidate gate, and output gate of LSTM. Represented as the temporal characteristic state at time t, Indicates the state of vegetation cells, Indicates the weights of LSTM, This represents the bias term of the LSTM; Spatiotemporal feature fusion, splicing the spatial features of the last frame With final time series characteristics , obtain fusion features ,in The function is used to concatenate strings.

[0021] B105: Pest and Disease Identification and Initial Results Output The classification of pests and diseases will integrate characteristics. Input the classification header, and calculate the probability of each type of pest or disease through a fully connected layer: ,in Represents classification weights, Indicates classification bias, n is the number of pest and disease categories. For the first The probability of a class; Confidence score is calculated as the product of the class probability and the probability of the target's existence. The specific formula is as follows: , in Indicates the existence weight of the target. Indicates target bias, It is the softmax function; Initial position coordinate regression, by detecting the head, regresses the bounding box coordinates of pest targets in the image. Position coordinates are determined, and the bounding box coordinates include... The calculation formula is: , in Indicates the bias of the model prediction. Indicates the center pixel coordinates of the anchor frame. Indicates the width and height of the anchor frame; B106: Processing identification results through confidence screening and coordinate calibration. Confidence level screening, elimination The recognition results are used to retain high-confidence results; NMS deduplication is performed on duplicate detection boxes at the same acquisition point using the NMS formula: ,in This indicates two duplicate detection boxes. When necessary, retain the boxes with higher confidence levels; Pixel coordinate to geographic coordinate calibration converts image pixel coordinates into actual geographic latitude and longitude using the following formula: , ,in The reference latitude and longitude of the data collection point This indicates that the horizontal image resolution is 1920 pixels. This indicates that the vertical image resolution is 1080 pixels. This indicates the latitude and longitude span of the area covered by the image.

[0022] The data aggregation and spatial association module aggregates and analyzes the identification results of the same collection point, calculates the initial occurrence density and local spread trend parameters of the pests and diseases, and associates the identification results of the collection point with its corresponding location coordinates to generate a spatial distribution heat map of pests and diseases on the GIS map layer. This module receives the structured recognition results output by the feature extraction and recognition module, including the collection point, pest / disease type, confidence level, and calibrated geographic coordinates. Collect timestamps, aggregate and analyze density and diffusion trends at the same collection point, identify the correlation between the results and spatial coordinates, generate GIS heat maps, and finally transfer the aggregated data and heat map data to the multi-factor risk assessment and time series prediction module. C101: Input Data Reception and Preprocessing The system receives the recognition results transmitted from the feature extraction and recognition module, performs deduplication based on collection point, pest / disease type, and collection timestamp; and filters the data, retaining the confidence level. The effective identification results are further eliminated to remove low-confidence results and reduce aggregation error; Data grouping: The first group is grouped by collection point and pest type, summarizing the valid identification results of the same type of pest at the same collection point, which is used to calculate the initial occurrence density. ; The second group is grouped by collection point, type of pest or disease, and time dimension. The time dimension is divided into time windows of 24 hours to match the collection frequency. The time-series identification data of the same collection point and the same type of pest or disease are summarized to calculate the local spread trend parameters. The third group is grouped according to the target garden area and the type of pests and diseases, and the effective identification results of the associated collection points are linked with the corresponding geographical coordinates for spatial association and heat map generation.

[0023] C102: Aggregation analysis of the same collection point, calculation of initial occurrence density. Target quantity statistics: For the same collection point ID, the same pest / disease class, and the same time window, count the number of duplicate detection boxes after NMS deduplication. ; Based on the latitude and longitude span of the image coverage of the acquisition points Longitude range The actual land surface area is calculated using the Harvey sine formula within the latitude range. The surface area ,in The radius representing the area of ​​the image captured at the point of acquisition; Initial occurrence density The density is calculated as the number of pests and diseases per unit area. ; C103: Aggregation analysis of the same data collection point to calculate local diffusion trend parameters. Grouped by collection point, pest / disease type, and time dimension, with the time dimension divided into 24-hour time series windows to match the collection frequency, the time series identification data of the same collection point and the same type of pest / disease are summarized, and three core diffusion trend parameters are calculated: diffusion rate ,in This indicates the maximum diameter of the pest distribution in the second sampling. This indicates the maximum diameter of the pest and disease distribution in the first collection. This indicates the time interval between two samplings, used to reflect the speed of the spread of pests and diseases within the sampling area; diffusion range change rate ,in This indicates the area of ​​pest and disease coverage collected in the second sampling. This indicates the area of ​​pest and disease coverage collected in the first survey, used to reflect the increase / decrease in the coverage area of ​​pests and diseases. Diffusion direction , Indicates the difference in longitude between the target centers of the two data collections. This indicates the difference in latitude between the centers of the two sampling points, used to reflect the main direction of the spread of pests and diseases within the sampling point; C104: Correlation of recognition results with spatial coordinates The spatial association core aggregates the results of steps C102 to C103. Geographic coordinates of the collection point Bind the GIS map data of the target garden area, including vegetation type, topography, and road and water system layers; Generate a GIS spatial distribution heat map, load a GIS map, load an electronic map of the target garden area into the system's GIS engine, and overlay auxiliary layers such as vegetation type, roads, and water system; With initial occurrence density As a core weight in heatmaps, higher density corresponds to greater thermal intensity, supplemented by diffusion rate. Make corrections when At that time, the weighting coefficient increased by 1.2 times, highlighting the rapidly spreading area; Kernel density estimation (KDE) calculation is performed on the associated data of all collection points. The kernel density estimation algorithm is used to generate a smooth heat map layer. The core parameter settings are: the search radius R is 1.5 times the average spacing of the collection points, the kernel function is a Gaussian kernel function, and the output grid size is 1 meter × 1 meter. Further explanation is needed regarding the calculation of the kernel density estimation KDE, specifically the following formula: , in This represents the kernel density value of a grid point (x,y), i.e., thermal intensity. Indicates the number of data collection points involved in the calculation. Indicates the radius of the sampling point, Represents the Gaussian kernel function; Heatmap rendering in stages, based on density values The heatmap is divided into four levels, corresponding to different color gradients, to enhance the visualization effect: Low risk, when At that time, the heat map is light blue, indicating sporadic occurrence of pests and diseases; Medium risk, when At that time, the heat map is yellow, indicating that the disease or pest has occurred locally; High risk, when At that time, the heat map is orange, indicating a concentrated occurrence of pests and diseases; Extremely high risk, when At that time, the heat map is dark red, indicating a large-scale outbreak of pests and diseases; The calculated data from C101 to C104 are synchronously transmitted to the multi-factor risk assessment and time series prediction module for subsequent risk level assessment and diffusion probability calculation.

[0024] The multi-factor risk assessment and time-series prediction module integrates the spatial distribution heat map and the environmental sensor data to construct a multi-factor risk assessment model. This model, based on historical data and current integrated data, performs the following operations: By integrating the spatial distribution heatmap with the environmental sensor data, a multi-factor risk assessment model is constructed. The model adopts a hybrid model of static multi-factor fusion XGBoost and time-series prediction LSTM, taking into account both the static correlation of spatial and environmental factors and the time-series patterns of historical data. Spatial factors: Kernel density estimation (KDE), initial occurrence density Density level, local diffusion trend parameters Spatial attribute tags, vegetation type, terrain slope; Environmental factors: temperature (°C), humidity (RH), soil moisture content (vol), light intensity (lux), precipitation (H); Historical factors: Access the system's historical database to extract historical risk levels, historical spread probabilities, and historical environmental data for the same type of pests and diseases from the same collection point within the past three months; Sub-model 1, XGBoost is responsible for assessing the overall risk level of the current pests and diseases in D1. It takes the spatial distribution heatmap kernel density estimation (KDE) and the environmental sensing data as inputs and outputs a risk score. Sub-model 2, LSTM is responsible for calculating the future diffusion probability of D2, inputting the time-series key factor sequence, and outputting the diffusion probability of the future preset time period; D1: Assess the overall risk level of current pests and diseases; The weights of each key factor are output through XGBoost sub-model 1, and the overall risk score is calculated by weighting these factors. The formula is: ,in Indicates the first The weights of the key factors, with a sum of weights of 1. Indicates the first Historical values ​​of key factors; Based on risk score The value range is [0,1], used to classify the overall risk level. At that time, it corresponds to low risk in the graded rendering of heatmaps; when At that time, it corresponds to medium risk in the heatmap hierarchical rendering; when At that time, it corresponds to high risk in the graded rendering of heatmaps; when At that time, it corresponds to an extremely high risk in the graded rendering of heatmaps.

[0025] D2: Calculate the probability of the pest spreading to a nearby healthy area within a preset time period in the future; Preset time period configuration, supporting user-defined or system default, splitting by time series window, the LSTM sub-model outputs the diffusion probability of each future time series window, such as within 24 hours. The adjacent healthy area is divided into sub-regions according to the vegetation type and the distance from the concentrated pest and disease occurrence area. The sub-regions include: Sub-region 1: Tree area, distance ≤ 50m; Sub-region 2: Shrub area, distance 50m < ≤ 100m; Sub-region 3: Lawn area, distance > 100m. Based on environmental and spatial factors of each sub-region, the diffusion probability of each sub-region is calculated and corrected. ,in Represents the diffusion probability within the time window t. , This represents a correction factor based on environmental factors of a sub-region, such as a sub-region with high humidity (RH). It should be larger than the sub-region with low RH. ; The average diffusion probability of each sub-region Mapping to diffusion levels facilitates subsequent early warning: when At this time, the spread level is low probability, and pests and diseases basically do not spread; when At that time, the spread level was low to medium probability, and the pests and diseases spread slightly; when At that time, the spread level was medium to high probability, and the pests and diseases spread moderately; when At that time, the spread level is high probability, indicating a severe spread of pests and diseases.

[0026] The dynamic early warning and decision support module dynamically generates tiered early warning signals based on the overall risk level and the diffusion probability, and generates and outputs a structured decision support report based on the early warning signals. The report includes: E1: Heat map of pest and disease types, spatial distribution, and visualization of risk levels; Basic information is solidified, clearly indicating the types of pests and diseases, warning levels, and risk scores. Data collection time range; heat map visualization rendering, calling previous heat map data to generate interactive GIS visualization components, overlaying risk level indicators, such as adding a red flashing border to extremely high-risk areas; risk level visualization, with progress bars corresponding to risk scores. The level icons, from 1 to 4, correspond to different colored icons for a visual display, and indicate the influencing factors, such as the high risk caused by humidity among environmental factors.

[0027] E2: Identifies areas where pests and diseases occur and areas with high risk of spread; Based on initial occurrence density Mark the core occurrence area on the GIS layer. Secondary occurrence area Use solid lines of different colors to define the boundaries and mark the area and average density of the collection points; Screening diffusion probability The sub-regions are defined on the GIS layer with dashed boxes, and the sub-region type is marked, such as the tree area within 50m, the diffusion probability, and the expected diffusion time. The regional correlation display overlays the boundaries of the occurrence area and high-risk diffusion area on the heat map, and uses arrows to indicate the direction of diffusion.

[0028] E3: A list of recommended prevention and control measures based on the types of pests and diseases, initial occurrence density, and environmental sensor data; Based on the types of pests and diseases, the system's control measures library is accessed to select suitable basic measures, such as physical control of aphids: sticky traps; biological control: ladybugs; chemical control: imidacloprid. Based on initial occurrence density Adjust the intensity of measures when When, prioritize physical or biological control; when At this time, physical and biological control methods are the main approach, with chemical control used in some areas; when At the same time, a combination of biological and chemical control methods is used to achieve full coverage. Following the logical order of E1 visualization information, E2 area identification, and E3 prevention and control measures, the report is encapsulated into a standardized JSON data format; the output is adapted to generate HTML heatmaps and GIS area identification components that can be embedded in the front-end page; the list of measures is presented in tabular form; the structured report is synchronized to the system front-end interface and the administrator's mobile APP, triggering early warning signal pop-up reminders, and supporting one-click export or printing.

[0029] This module receives structured data from the multi-factor risk assessment and time-series prediction module, including pest and disease types, current overall risk level and score, adjacent healthy sub-regions, and the spread probability and level of each sub-region. It also calls upon spatial distribution heatmap data and initial occurrence density from the preceding module. Using environmental sensor data and GIS spatial attribute data, we can generate graded early warning signals, construct the core content of structured reports, and package and output the reports. Receive core input data, risk and prediction data, pest and disease types, overall risk level 1 to 4, and risk score. Average diffusion probability of sub-region and rating, model confidence level; Related data retrieval includes complete spatial distribution heatmap data and initial occurrence density at each collection point. Environmental sensor data and GIS sub-area coordinate range; Data is categorized and integrated into several types: visualization data, including heatmap rendering parameters, risk level indicators, and sub-region boundary coordinates; core information data, including pest and disease types, risk levels and scores, spread probability and levels, and occurrence density; and decision support data, including environmental sensor data, sub-region spatial attributes such as vegetation type, and distance from the pest and disease occurrence area. Based on the dual dimensions of overall risk level and spread probability, a warning level mapping rule is established. The area with a high probability of spread is compared with the total area of ​​the healthy zone to generate a corresponding warning signal. When the area accounts for less than 10%, the warning signal is of general concern; when the area accounts for 10% to 30%, the warning signal is of vigilance; when the area accounts for 30% to 50%, the warning signal is of emergency prevention and control; and when the area accounts for more than 50%, the warning signal is of full-area control.

[0030] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent monitoring and early warning system for garden pests and diseases based on AI visual recognition, characterized in that, include: Multi-source heterogeneous data acquisition module: acquires temporal visual image data of the target garden area, as well as spatiotemporal labels and environmental sensing data corresponding to the target garden area. The spatiotemporal labels include the acquisition timestamp and the coordinates of the acquisition point. Feature extraction and recognition module: Input the time-series visual image data into the pre-trained pest and disease AI recognition model, extract the time-series visual image features, and output the recognition results, which include pest and disease types, confidence levels, and location coordinates; Data aggregation and spatial association module: Aggregates and analyzes the identification results of the same collection point, calculates the initial occurrence density and local spread trend parameters of the pests and diseases, and associates the identification results of the collection point with its corresponding location coordinates to generate a spatial distribution heat map of pests and diseases on the GIS map layer; Multi-factor risk assessment and time-series prediction module: This module integrates the spatial distribution heatmap with the environmental sensor data to construct a multi-factor risk assessment model. Based on historical data and the currently integrated data, this model performs the following operations: D1: Assess the overall risk level of current pests and diseases; D2: Calculate the probability of the pest spreading to a nearby healthy area within a preset time period in the future; Dynamic Early Warning and Decision Support Module: Based on the overall risk level and the diffusion probability, dynamically generates tiered early warning signals, and generates and outputs a structured decision support report based on the early warning signals. The report includes: E1: Heat map of pest and disease types, spatial distribution, and visualization of risk levels; E2: Identifies areas where pests and diseases occur and areas with high risk of spread; E3: A list of recommended prevention and control measures based on the types of pests and diseases, initial occurrence density, and environmental sensor data.

2. The intelligent monitoring and early warning system for garden greening pests and diseases based on AI visual recognition as described in claim 1, characterized in that, The multi-source heterogeneous data acquisition module includes: A101: In the data collection and preparation stage, based on the WGS84 coordinate system latitude and longitude range of the target garden area input into the GIS geographic information system, a digital map of the garden area is generated, vegetation type sub-regions are divided and collection points are planned; the collection points are laid out according to the principle of uniform distribution and densification in key areas. A102: During the data acquisition execution phase, the high-definition camera is 2 to 3 meters above the ground with the lens tilted vertically downwards at 30 to 45 degrees, focusing on a 2-meter radius around the acquisition point. Each time, 3 to 5 frames of PNG format images are captured continuously. Basic acquisition points are acquired once every 2 hours, and encrypted acquisition points are acquired once every 30 minutes. The environmental sensor data acquisition frequency is synchronized with the visual image acquisition. After each acquisition is completed, a spatiotemporal tag containing latitude and longitude coordinates and an NTP timestamp is generated synchronously. A103: In the data cleaning and format standardization stage, invalid images are removed on the server side and uniformly scaled to 1920×1080 pixels in PNG format; outliers in environmental data are removed using the 3-times-standard-deviation principle, and missing data is filled in using linear interpolation; latitude and longitude coordinates are uniformly converted to the WGS84 coordinate system, and timestamps are uniformly converted to UTC time format.

3. The intelligent monitoring and early warning system for garden pests and diseases based on AI visual recognition as described in claim 1, characterized in that, The feature extraction and recognition module, and the pre-trained AI model for pest and disease recognition, include: The training dataset for the pre-trained AI model for identifying diseases and pests includes: garden diseases (powdery mildew, black spot, leaf blight); garden pests (aphids, spider mites, longhorn beetles); vegetation type sub-regions (tree area, shrub area, lawn area, flower area); and geographical features (roads, water systems, buildings, and steep slopes with a gradient greater than 15 degrees).

4. The intelligent monitoring and early warning system for garden greening pests and diseases based on AI visual recognition as described in claim 3, characterized in that, The feature extraction and recognition module includes: B101: Receives data transmitted from the multi-source heterogeneous data acquisition module, parses the time-series visual image data and the bound spatiotemporal tags, and verifies and removes invalid data; B102: Performs color gamut conversion, normalization, and Gaussian filtering noise reduction on single-frame images. After processing 3 to 5 frames from the same acquisition point, the images are stitched together into a temporal tensor. ; B103: Load the YOLOv8 temporal augmentation model with a hybrid architecture of CNN and LSTM, adjust the pre-trained weights based on the garden pest and disease annotation dataset, configure the initial confidence threshold to be 0.5 and the NMS threshold to be 0.45, and process the temporal visual image sequences in batches according to the collection points. B104: Spatial features are extracted using the C2f module of YOLOv8. The spatial feature set is then input into a bidirectional LSTM layer to extract temporal dynamic features. The spatial features of the last frame are then concatenated with the final temporal features to obtain fused features. B105: Input the fused features into the classification head, calculate the probability and confidence of various pests and diseases, and regress the bounding box coordinates of pests and diseases through the detection head; B106: Remove recognition results with a confidence level below 0.5, perform NMS deduplication, and convert image pixel coordinates into actual geographic latitude and longitude coordinates.

5. The intelligent monitoring and early warning system for garden greening pests and diseases based on AI visual recognition as described in claim 1, characterized in that, The data aggregation and spatial association module includes: C101: Receive the identification results, deduplicate them by collection point, pest / disease type, and collection timestamp, retain valid results with confidence ≥ 0.5, and group the data; C102: Count the number of detection boxes at the same collection point, for the same type of pest or disease, and within the same time window. Calculate the surface area covered by the image at the collection point based on the Harvey sine formula, and calculate the number of pest or disease targets per unit area as the initial occurrence density. ; C103: Divide the time window into 24-hour intervals to calculate the diffusion rate. , diffusion range change rate , diffusion direction Three core diffusion trend parameters; C104: Binds the aggregation results to geographic coordinates, GIS map data, and auxiliary layers, based on the initial occurrence density. As the core weight, diffusion speed When the value exceeds the preset threshold, the weighting coefficient is increased by 1.2 times. A heat map with a grid size of 1 meter × 1 meter is generated using a kernel density estimation algorithm. The heat map is divided into four risk levels: low, medium, high, and extremely high, and rendered with different color gradients according to the density value.

6. The intelligent monitoring and early warning system for garden pests and diseases based on AI visual recognition as described in claim 1, characterized in that, The multi-factor risk assessment and time-series prediction module includes: The multi-factor risk assessment model adopts a hybrid model of XGBoost and LSTM; the XGBoost sub-model is used to assess the current overall risk level; and the LSTM sub-model is used to predict the probability of pests spreading to healthy sub-regions in the future.

7. The intelligent monitoring and early warning system for garden greening pests and diseases based on AI visual recognition as described in claim 5, characterized in that, The multi-factor risk assessment model includes: The input factors include spatial factors, environmental factors, and historical factors; D1: The weights of each key factor are output through the XGBoost sub-model, and the overall risk score is calculated by weighting. The overall risk level is divided into the corresponding heat map level according to the score range. D2: The preset time period is divided into time windows. The LSTM sub-model outputs the diffusion probability of each time window. The adjacent healthy area is divided into sub-regions according to the vegetation type and the distance from the pest and disease occurrence area. The diffusion probability of each sub-region is calculated by combining the environmental factors and spatial factors of the sub-region and mapped to four diffusion levels.

8. The intelligent monitoring and early warning system for garden greening pests and diseases based on AI visual recognition as described in claim 1, characterized in that, The dynamic early warning and decision support module includes: E1: Label pest and disease types, warning levels, and risk scores to generate an interactive GIS visualization heatmap; E2: On the GIS layer, use solid lines of different colors to delineate the core occurrence area and secondary occurrence area and label relevant information. Use dashed lines to delineate the diffusion probability, expected diffusion time, and diffusion direction arrows. E3: Select appropriate measures based on the types of pests and diseases, and adjust the measures according to the initial occurrence density.

9. The intelligent monitoring and early warning system for garden greening pests and diseases based on AI visual recognition according to claim 7, characterized in that, The dynamic early warning and decision support module includes: A warning level mapping rule is established based on two dimensions: risk score and spread probability. The mapping rule includes: when the high-risk spread area accounts for less than 10% of the total healthy area, the warning signal is general concern; when the proportion is 10% to 30%, the warning signal is precaution; when the proportion is 30% to 50%, the warning signal is emergency prevention and control; and when the proportion is greater than 50%, the warning signal is full-area control.

Images