Ground-air cooperative confidence grading early warning method and system for alien weed invasion

Abstract

The invention relates to the technical field of plant protection monitoring, in particular to an alien weed invasion ground-air cooperative confidence grading early warning method and system, and the method comprises the following steps: collecting plant multi-source sensing data of a target region through plant ground-air cooperative sensing equipment; performing preliminary identification on the plant multi-source sensing data to output screening confidence, and screening out a suspected foreign weed region; finely sensing the suspected alien weed area by using the plant ground-air cooperative sensing equipment to obtain multi-dimensional feature data, and determining alien weed plants through classification and recognition; constructing a weed invasion risk diffusion prediction model, and obtaining a normalized invasion risk index of the exotic weed plant; and dividing weed invasion grades according to the normalized invasion risk index, and obtaining early warning confidence of the weed invasion grades to generate foreign weed invasion early warning information. According to the invention, precise monitoring, dynamic early warning and reliable decision making of foreign weed invasion are realized through ground-air cooperation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of plant protection monitoring, and in particular to an air-ground collaborative confidence grading early warning method and system for alien weed invasion. BACKGROUND

[0002] In the prior art, the monitoring and early warning of alien weeds mainly rely on remote sensing technology of a single platform or ground artificial investigation. Common practices include using satellite remote sensing images for large-scale vegetation type identification, or using unmanned aerial vehicles to carry visible light or multispectral cameras for regional aerial photography, and then combining image processing algorithms or manual interpretation to identify suspicious vegetation. In terms of risk assessment, the existing methods mostly use statistical models or simple ecological models based on historical data for static assessment of invasion risk. The above technical path constitutes the mainstream scheme in the current plant protection monitoring field, which to some extent realizes the identification and rough risk judgment of alien weeds.

[0003] However, the existing technology has some limitations. First, the perception scheme based on a single platform has a defect in data dimension: high-altitude remote sensing cannot obtain plant detail features, which is prone to misidentification; ground investigation is inefficient and limited in scope, and the singleness of the data source directly leads to an inherent contradiction between identification accuracy and coverage range; second, the existing identification methods mostly rely on simple analysis of apparent features, lack a multi-level verification mechanism from macro to micro, and have insufficient ability to distinguish similar species, resulting in a high misjudgment rate of early warning results; at the same time, the traditional risk assessment model is mostly static analysis, which cannot simulate the dynamic diffusion process of weed invasion behavior, resulting in a large deviation between the early warning result and the actual situation; finally, the existing early warning method can only output the risk level, and cannot evaluate the reliability of the early warning conclusion itself, making it difficult for managers to judge the credibility of the early warning information and develop differentiated prevention and control strategies. The above defects result in the lack of accuracy, forward-looking and practicality of the existing alien weed invasion early warning method. SUMMARY

[0004] In view of the defects in the prior art, the present application provides an air-ground collaborative confidence grading early warning method and system for alien weed invasion.

[0005] In order to achieve the above object, in a first aspect, the application provides a method for ground-air cooperative confidence grading early warning of alien weed invasion, comprising the following steps: collecting plant multi-source perception data of a target area by a plant ground-air cooperative perception device; performing preliminary identification output screening confidence on the plant multi-source perception data, and screening out a suspected alien weed area based on the screening confidence; using the plant ground-air cooperative perception device to perform fine perception on the suspected alien weed area to obtain multi-dimensional feature data, and determining an alien weed plant through classification and identification; constructing a weed invasion risk diffusion prediction model, and obtaining a normalized invasion risk index of the alien weed plant according to the weed invasion risk diffusion prediction model; dividing a weed invasion level according to the normalized invasion risk index, and obtaining early warning confidence of the weed invasion level, to generate alien weed invasion early warning information. The application obtains multi-source data through ground-air cooperative perception, accurately locates a suspected area in combination with screening confidence, generates grading early warning through fine identification and risk prediction, improves the comprehensiveness and accuracy of weed invasion monitoring and early warning, and is helpful for efficient prevention and control.

[0006] Optionally, the collecting plant multi-source perception data of a target area by a plant ground-air cooperative perception device comprises that the plant ground-air cooperative perception device is composed of a high-altitude unmanned aerial vehicle, a forest unmanned aerial vehicle and a ground robot; the high-altitude unmanned aerial vehicle is used to obtain plant high-altitude image data of the target area; the forest unmanned aerial vehicle is used to collect plant low-altitude multi-angle image data of the target area; the ground robot is used to obtain plant ground high-definition image data of the target area; and the plant high-altitude image data, the plant low-altitude multi-angle image data and the plant ground high-definition image data are used as the plant multi-source perception data. The application uses high-altitude, forest unmanned aerial vehicles and ground robots to cooperatively collect multi-dimensional image data, realizes stereoscopic coverage of plant information of a target area, and improves data richness and integrity through multi-source data complementation, thereby providing a high-quality basis for subsequent identification and enhancing perception reliability.

[0007] Optionally, the preliminary identification and screening confidence score of the multi-source sensing data of the plant, and the screening of suspected invasive weed areas based on the screening confidence score, includes: performing large-scale semantic segmentation on the aerial image data of the plant to obtain preliminary abnormal plant areas; performing multi-view feature fusion enhancement on the preliminary abnormal plant areas based on the low-altitude multi-angle image data of the plant and the high-definition ground image data of the plant to extract plant morphological texture fusion features; dividing the preliminary abnormal plant areas into abnormal plant sub-regions according to the plant morphological texture fusion features; calculating the screening confidence score of the abnormal plant sub-regions, and comparing the screening confidence score with a preset threshold to determine the suspected invasive weed areas. This invention first determines the preliminary abnormal areas through semantic segmentation, then extracts key features through multi-view feature fusion, and combines the screening confidence score to screen suspected areas, quickly narrowing the monitoring range, reducing redundant data processing, and improving the efficiency and accuracy of preliminary identification.

[0008] Optionally, calculating the screening confidence of the abnormal sub-region of the plant includes: obtaining an initial confidence of the abnormal sub-region of the plant based on the plant morphology and texture fusion features using a lightweight classifier; acquiring a local plant prior knowledge base for the target region, the local plant prior knowledge base including plant phenological stages and plant growth characteristics; matching and comparing the plant morphology and texture fusion features with the prior knowledge base to obtain a prior conformity score; and combining the initial confidence and the prior conformity score to obtain the screening confidence. This invention combines the initial confidence obtained from the lightweight classifier with the prior conformity score obtained from the local plant prior knowledge base to calculate the screening confidence, fusing data features and regional prior knowledge, reducing the false positive rate, and making the identification of suspected invasive weed areas more reliable.

[0009] Optionally, the step of using the plant-ground-air collaborative sensing device to perform fine sensing of the suspected invasive weed area to obtain multi-dimensional feature data, and identifying the invasive weed plants through classification, includes: scheduling the forest drone to collect low-altitude multi-angle image data of the plants in the suspected invasive weed area; generating a three-dimensional point cloud model based on the low-altitude multi-angle image data of the plants to obtain three-dimensional structural features, and marking key plants to be identified; under the guidance of the forest drone, planning the autonomous path of the ground robot to the optimal observation point, and collecting stem detail features of the key plants to be identified; using the three-dimensional structural features and the stem detail features as the multi-dimensional feature data; based on the multi-dimensional feature data, performing discrimination through a cascaded classification model, including a first-level classifier obtaining a coarse screening identification result based on the three-dimensional structural features, and a second-level classifier combining the coarse screening identification result and the stem detail features to obtain a fine identification result; and jointly analyzing the coarse screening identification result and the fine identification result to determine the invasive weed plants. This invention uses fine perception to acquire three-dimensional structural features and stem detail features of suspected areas. Through joint analysis using a cascaded classification model, it achieves recognition from coarse to fine, significantly improving the accuracy of identifying invasive weeds, reducing missed and false judgments, and ensuring the reliability of the recognition results.

[0010] Optionally, the construction of the weed invasion risk diffusion prediction model includes: obtaining a set of invasion risk assessment factors for the invasive weed plants, including invasion source intensity factors, environmental driving force factors, and land type resistance factors; and coupling cellular automata and multilayer perceptrons based on the set of invasion risk assessment factors to construct the weed invasion risk diffusion prediction model. This invention integrates invasion risk assessment factors and couples cellular automata and multilayer perceptrons to construct a prediction model, taking into account spatiotemporal dynamics and the influence of complex environments, making the weed invasion risk diffusion prediction more consistent with reality and improving the scientific rigor of the model.

[0011] Optionally, obtaining the normalized invasion risk index of the invasive weeds based on the weed invasion risk diffusion prediction model includes: using the weed invasion risk diffusion prediction model to iteratively simulate and obtain a spatiotemporal invasion risk probability map of the invasive weeds within a preset future time period; performing multi-dimensional statistical analysis on the spatiotemporal invasion risk probability map to obtain core weed invasion indicators, including the potential maximum impact area, average spatial diffusion speed, and overlap of ecologically sensitive areas; obtaining the dynamic weight coefficients of the core weed invasion indicators; and performing linear weighted fusion of the core weed invasion indicators based on the dynamic weight coefficients to obtain the normalized invasion risk index. This invention obtains a spatiotemporal invasion risk probability map through simulation, extracts core indicators, and combines them with dynamic weights to obtain a normalized risk index, quantifying invasion risk from multiple dimensions. The dynamic weights adapt to different scenarios, making the index more objectively reflect the true risk level.

[0012] Optionally, the step of classifying weed invasion levels based on the normalized invasion risk index and obtaining the warning confidence level of the weed invasion level to generate invasive weed early warning information includes: using a sliding window mechanism to dynamically set a warning level threshold by analyzing the changing trend of the normalized invasion risk index; mapping the normalized invasion risk index to obtain weed invasion levels, including attention level, alert level, action level, and emergency level, based on the warning level threshold; obtaining the warning confidence level based on the uncertainty component; constructing a warning decision matrix by combining the weed invasion level and the warning confidence level; and obtaining a differentiated warning strategy based on the warning decision matrix to generate the invasive weed early warning information. This invention dynamically sets warning thresholds to classify invasion levels, constructs a decision matrix based on the warning confidence level to generate differentiated strategies, adapts the dynamic thresholds to risk changes, and enhances the credibility of the warning by increasing the confidence level, making the warning information more accurate and the control strategies more operational.

[0013] Optionally, obtaining the early warning confidence level based on the uncertainty components includes: acquiring the uncertainty components, including data quality uncertainty components and model uncertainty components; and fusing the uncertainty components using Dempster-Shafer evidence theory to obtain the early warning confidence level. This invention fuses data and model uncertainty components based on Dempster-Shafer theory, quantifies uncertainty to obtain early warning confidence levels, scientifically assesses early warning reliability, reduces decision-making bias, and enhances the reference value of early warning information.

[0014] Secondly, this invention provides a ground-air collaborative confidence-based graded early warning system for invasive alien weeds. The system executes the ground-air collaborative confidence-based graded early warning method for invasive alien weeds provided by this invention. The system includes input devices, output devices, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions, and the processor is configured to invoke the program instructions. This invention, through high-performance hardware collaboration, efficiently executes the ground-air collaborative early warning method, achieving integrated data acquisition, identification, prediction, and early warning. It provides a systematic tool for the prevention and control of invasive alien weeds, improving monitoring and early warning efficiency and decision support capabilities. Attached Figure Description

[0015] Figure 1 This is a flowchart of a ground-air collaborative confidence-based hierarchical early warning method for invasive alien weeds according to an embodiment of the present invention; Figure 2 This is a framework diagram of a ground-air collaborative confidence-based hierarchical early warning system for invasive alien weeds, according to an embodiment of the present invention. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 One embodiment of the present invention provides a ground-air collaborative confidence-based graded early warning method for invasive alien weeds, the method comprising the following steps: S1. Collect multi-source sensing data of plants in the target area through plant-ground-air collaborative sensing equipment.

[0019] In this embodiment, a plant ground-air collaborative sensing device is established based on the collaboration of high-altitude drones, forest drones, and ground robots.

[0020] High-altitude drones can be selected from DJI Matrice 350 real-time dynamic positioning drones or equivalent industrial drones as flight platforms, equipped with full-frame aerial survey cameras and multispectral imaging systems. First, the survey area flight path is planned via a ground control station, with a flight altitude set at 120m-150m, a forward overlap of no less than 80%, a lateral overlap of no less than 70%, and a ground sampling distance of 3cm. During flight, centimeter-level accuracy positioning data is acquired through a real-time dynamic positioning module, while flight attitude parameters are recorded. The multispectral imaging system simultaneously acquires spectral data in five bands: blue, green, red, red-edge, and near-infrared. After each aerial survey acquisition, data quality is checked to ensure the images are free from cloud cover and significant distortion. All acquired high-altitude plant image data is transmitted in real-time to a ground data processing center via 4G / 5G networks. After radiometric calibration and atmospheric correction, orthophoto maps and digital surface models are generated to obtain preliminary plant inspection areas, serving as baseline data for subsequent analysis.

[0021] The drone used in the forest can be a DJI Mavic 3E Enterprise Edition small drone, which has precise hovering and obstacle avoidance capabilities. After receiving the initial coordinates of the plant inspection area from the high-altitude drone, the forest drone automatically flies to an altitude of 20m-30m above the target area. Using oblique photography technology, a five-lens camera array simultaneously collects low-altitude multi-angle image data of the plants from five angles: front, back, left, right, and vertical. The acquisition time for each point is controlled within 30 seconds to ensure consistent lighting conditions.

[0022] In an optional embodiment, data is acquired during the early morning or late afternoon to obtain optimal lighting and shadow effects for 3D modeling. The acquired image data is processed using a structure-of-motion (SOMO) algorithm to generate high-density 3D point cloud data. An iterative nearest-point algorithm is then used to precisely register the point cloud data with a digital surface model acquired by a high-altitude UAV, achieving geographic coordinate binding to establish the 3D point cloud model and subsequently identify key observation points.

[0023] The ground robot utilizes a tracked mobile platform equipped with a binocular stereo vision system and LiDAR. Upon receiving key observation points from a forest drone, the robot achieves autonomous navigation using Simultaneous Localization and Mapping (SLAM) technology. The robot first establishes an observation base station 2-3 meters away from the target plant. Using its robotic arm, it adjusts the observation angle and employs its onboard macro camera to capture microscopic features such as the front and back of leaves and vein textures at a spatial resolution of 0.1 millimeters. Simultaneously, a hyperspectral imager is used to collect spectral features of the plant within the 400nm-1000nm spectral range at a spectral resolution of 3nm. For each key observation plant, at least 10 images from different angles and spectral data from 5 locations are collected. All high-resolution ground images of the plant are accompanied by records of environmental parameters such as acquisition time, light intensity, temperature, and humidity. If obstacles are encountered during the data collection process, the robot automatically plans an obstacle avoidance path to ensure continuous observation.

[0024] In this embodiment, multi-source sensing data of plants are fused under a unified spatiotemporal reference. All data are transformed to the same coordinate system, and the time synchronization accuracy is controlled within 1 second. For high-altitude UAV data, regional plant characteristics are mainly extracted; for forest UAV data, the focus is on extracting three-dimensional morphological structure and spatial distribution features; and for ground robot data, the focus is on microscopic morphology and fine spectral features. A data quality evaluation system is established, including indicators such as image sharpness (based on Laplacian variance calculation), spectral signal-to-noise ratio (greater than 30dB), and point cloud density (greater than 1000 points / square meter). For data that does not meet the quality standards, it is automatically marked and a secondary acquisition process is initiated. All qualified data is packaged according to standard data format to generate a standardized data package containing metadata, raw data, and preprocessed data.

[0025] In one optional embodiment, the plant-ground-air collaborative sensing device employs a hierarchical scheduling strategy. The high-altitude UAV first completes a large-scale survey, and upon discovering anomalies, immediately sends reconnaissance commands to the forest-based UAV. After completing detailed mapping, the forest-based UAV sends the coordinates of target points requiring ground verification to the ground robot. Simultaneously, dynamic task priorities are set, and for high-confidence suspected intrusion areas, a three-level platform synchronous observation mode is activated. At the end of each operational cycle, a sensing coverage report and a data quality assessment report are automatically generated, providing a basis for task planning in the next cycle. Through this multi-level, multi-scale collaborative sensing mechanism, full-scale observation is achieved.

[0026] In an optional embodiment, an automatic monitoring data filling module is established. This module, through high-precision positioning modules, environmental sensors, and image recognition units integrated on UAVs and ground robots, automatically collects and generates structured monitoring data during sensing operations. This data includes precise location coordinates, collection time, phenological stages based on image recognition (such as the vegetative growth stage and flowering stage), and estimated plant height and crown width. The collected data is transmitted to a data processing center in real time, automatically parsed, and mapped to the corresponding fields of a preset customs monitoring report template. This achieves automatic filling and generation of monitoring forms, greatly reducing the burden of manual data entry and improving the accuracy and timeliness of the data.

[0027] Meanwhile, the ground robot platform can be integrated as a mobile base station for unmanned aerial vehicles (UAVs). This platform features autonomous deployment and retrieval, precise positioning, and wireless charging capabilities, enabling the UAVs to automatically return to the platform for charging, data transmission, and storage after completing their data collection tasks. This forms a ground-air collaborative mobile operation unit integrating mobile deployment, autonomous collaboration, energy replenishment, and data relay, making it particularly suitable for scenarios requiring long-term, large-scale patrol monitoring, such as customs jurisdictions.

[0028] S2. Perform preliminary identification and output screening confidence scores on the multi-source sensing data of the plants, and screen out suspected invasive weed areas based on the screening confidence scores.

[0029] In this embodiment, an improved YOLOv5 architecture is used as the core detection model for large-scale semantic segmentation of aerial image data of plants. A coordinate attention mechanism is introduced into the backbone network based on the YOLOv5 architecture to enhance the model's ability to perceive the morphological details of weeds. The input images are preprocessed and uniformly scaled to 640×640 resolution, and data augmentation techniques are used to fuse samples under different lighting conditions, scales, and backgrounds to improve the model's generalization ability in complex environments. During the training phase, an adaptive gradient pruning strategy is adopted, with an initial learning rate set to 0.01. A cosine annealing scheduler is used to optimize the training process, and a focus loss function is used to address the class imbalance problem. The model first generates candidate detection boxes, and then performs pixel-level segmentation of the region within each detection box through a semantic segmentation branch. This branch is based on a lightweight neural network structure, embedding depthwise separable convolutions in the encoder to reduce computational cost, and fusing multi-level features through skip connections in the decoder. Finally, a class probability map for each pixel is output, and connected regions with probability values ​​higher than 0.75 are identified as preliminary abnormal plant regions, and their boundary geometric parameters and morphological features are recorded.

[0030] Furthermore, based on the coordinate information of the initial abnormal plant areas, forest drones and ground robots are coordinated to collect multi-angle data. A multi-level feature fusion pipeline is constructed based on the low-altitude multi-angle image data acquired by the forest drones and the high-resolution image data from the ground robots. For the low-altitude multi-angle images, a dense 3D point cloud is first generated using the structure-of-motion (SOMO) algorithm, and then three levels of features are extracted: the spatial structure of the plant is represented based on the point cloud; and directional gradient histograms and local binary pattern features are extracted from the multi-angle images to describe the macroscopic morphology of the plant. For the high-resolution ground images, the focus is on extracting microscopic morphological features: wavelet transform is used to extract multi-directional texture features, and superpixel segmentation algorithms are used to extract morphological features such as leaf margin serration and leaf vein distribution patterns.

[0031] In the feature fusion stage, a feature weighting module based on an attention mechanism is established to automatically learn the importance weights of different feature dimensions. Low-altitude features and ground features are spliced ​​together at the feature level, and then dimensionality is reduced through a fully connected layer to finally generate plant morphology and texture fusion features.

[0032] In this embodiment, a hierarchical clustering and spatial constraint-based partitioning strategy is adopted based on the plant morphology and texture fusion features. First, based on the Euclidean distance of feature vectors, a density-based spatial clustering of applications with noise (DBSCAN) algorithm is used to aggregate pixels with similar features into primary clusters. Then, a region adjacency graph is constructed, where nodes represent primary clusters, edge weights are calculated jointly by feature similarity and spatial distance, and a normalized cut algorithm is used for graph partitioning. During the partitioning process, the number of clusters is dynamically optimized to ensure that each sub-region satisfies feature consistency (intra-class variance less than 0.1) and spatial continuity (minimum area threshold of 100 pixels). The final output of the abnormal plant sub-region includes its outline polygon, area, feature mean, and spatial index, providing structured input for subsequent confidence calculations.

[0033] In this embodiment, a Lightweight Gradient Boosting Machine (LightGBM) is used as the classifier to calculate the initial confidence score based on the plant morphology and texture fusion features. The feature vectors are input into a LightGBM model containing 200 decision trees. This model optimizes training efficiency through gradient one-sided sampling and mutually exclusive feature binding techniques. The training data comes from labeled samples, covering weed samples at different growth stages, under different lighting conditions, and with varying background complexity. During inference, the model outputs the probability value of each abnormal sub-region of a plant belonging to the invasive weed category as the initial confidence score, satisfying the following relationship: in, As the initial confidence level, For the Sigmoid function, For the number of decision trees, For the index variable of the decision tree, For the first The weight coefficients of each decision tree, For the first The prediction output of a decision tree, The feature vector representing the fusion features of plant morphology and texture. This is the bias vector.

[0034] In this embodiment, the local plant prior knowledge base is stored using a relational database and includes a phenological phase module and a growth characteristic module. The phenological phase module records monthly phenological data of dominant plants in the target area over the past 5 years (such as budding stage, leaf expansion stage, and flowering stage), and is updated through a combination of satellite remote sensing and ground observation. The growth characteristic module stores morphological parameters (such as leaf length-to-width ratio and stem branching pattern) and spectral reflectance curves of typical invasive weeds. The knowledge base enables fast retrieval through spatiotemporal indexing and supports querying matching prior knowledge by geographic coordinates, date, and environmental conditions.

[0035] Furthermore, a multi-level matching strategy is designed when matching plant morphology and texture features with a prior knowledge base. First, phenological data is retrieved based on the current date and geographical location to quantify the phenological matching degree. If the features of the target sub-region conflict with the phenological features of the dominant plant at the current time (e.g., typical summer plant morphology appearing in winter), a phenological anomaly marker is triggered. Subsequently, a dynamic time warping algorithm is used to compare the spectral curves of the abnormal plant sub-regions with the standard curves of invasive weeds in the knowledge base to calculate spectral similarity. Finally, cosine similarity is used to measure the degree of morphological parameter matching. A prior conformity score is generated by weighting the phenological matching degree, spectral similarity, and morphological matching degree, satisfying the following relationship: in, This represents the prior agreement score. These are the weighting coefficients. For phenological period matching degree, For spectral similarity, For morphological matching degree.

[0036] The above phenological period matching degree satisfies the following relationship: in, For phenological period matching degree, It is a natural exponential function. The sequence number within the year of the current date. This represents the sequence number of expected phenological dates within the year. This refers to the time window of phenological periods.

[0037] Furthermore, the initial confidence score and prior conformity score are fused using the Dempster-Shafer Evidence Theory (DS Theory). The identification framework is defined to include invasive weeds and non-invasive weeds. A first basic probability is assigned to the initial confidence score, and a second basic probability is assigned to the prior conformity score. The fused confidence function is calculated using Dempster's combination rule, where a conflict factor is used to adjust the degree of contradiction between pieces of evidence. Finally, the confidence score corresponding to "invasive weeds" in the confidence function is used as the screening confidence score, which satisfies the following relationship: in, To screen for confidence levels, Indicates invasive weeds, Assign a value to the first basic probability. As the initial confidence level, This represents Dempster's combinatorial rule. Assign a value to the second basic probability. This represents the prior agreement score. It is a conflict factor.

[0038] In this embodiment, a dynamic threshold strategy is employed to compare the screening confidence level with a preset threshold. The base threshold is set to 0.65 and is adaptively adjusted based on environmental factors: when the image sharpness is higher than 0.8, the threshold is lowered to 0.6 to improve sensitivity; in areas with insufficient light or dense vegetation, the threshold is raised to 0.7 to reduce false alarms. During the comparison, a triple (coordinates, screening confidence level, feature) is generated for each abnormal sub-region of a plant, retaining only sub-regions with a confidence level higher than the threshold. Subsequently, a non-maximum suppression algorithm is used to merge adjacent regions with a spatial overlap greater than 0.5 to output suspected invasive weed regions, accompanied by confidence labels and spatial boundary information.

[0039] S3. Use the plant-ground-air collaborative sensing device to perform fine sensing of the suspected invasive weed area to obtain multi-dimensional feature data, and identify the invasive weed plants through classification.

[0040] In this embodiment, after receiving the coordinates of a suspected invasive weed area, a forest drone is dispatched to collect low-altitude, multi-angle image data of the plants. First, a scale-invariant feature transform algorithm is used to extract feature points, and a random sampling consensus algorithm is used to eliminate mismatched points. Then, multi-view stereo technology is used to generate dense point cloud data, and a three-dimensional point cloud model is built based on the point cloud data to extract three types of three-dimensional structural features: the three-dimensional volume and span ratio of the plant canopy are calculated through principal component analysis; the entropy value of the spatial orientation of branches and leaves is calculated based on the point cloud normal distribution; and the plant surface mesh is reconstructed using the Alpha Shape Algorithm (α-shape) to calculate the curvature distribution characteristics. Simultaneously, areas with abnormal morphology in the point cloud are automatically labeled as key plants to be identified based on the three-dimensional structural features.

[0041] It should be noted that the labeling criteria include point cloud density anomalies (more than twice the standard deviation of the mean), height abrupt changes (height difference between adjacent points > 0.5 meters), or structural irregularities (anisotropy index > 0.7).

[0042] Furthermore, after receiving the coordinates of key plants to be identified from the forest drone, the ground robot collects data through autonomous path planning. First, it obtains its initial position, then uses LiDAR SLAM technology to construct a centimeter-level accurate environmental map. Path planning employs an improved A* algorithm, comprehensively considering terrain slope, vegetation cover, and passage safety. Upon reaching the vicinity of the target point, the ground robot uses a binocular stereo vision system to accurately measure distances and adjusts its robotic arm to the optimal observation point. Multimodal sensing systems are used to collect detailed stem features: a macro camera captures stem epidermal texture, a hyperspectral imager captures stem spectral features, and a structured light 3D scanner acquires stem cross-sectional morphology. All collected data simultaneously records environmental parameters (light intensity, temperature, and humidity) and transmits them back to the data processing center in real time.

[0043] In this embodiment, three-dimensional structural features and stem detail features are used as multi-dimensional feature data, and a cascaded classification model is used to distinguish the multi-dimensional feature data. The cascaded classification model adopts a two-level architecture: the first-level classifier is a random forest model, which takes three-dimensional structural features as input and outputs coarse screening recognition results; the second-level classifier is a deep neural network, which takes coarse screening recognition results and stem detail features as input and finally outputs fine recognition results.

[0044] It should be noted that both classifiers were trained on a training set containing diverse samples, covering typical invasive weeds and their similar species.

[0045] In this embodiment, a decision-making mechanism based on category consistency and a multi-level review process are adopted to jointly analyze the coarse screening and fine screening results, and finally determine the invasive weed plants.

[0046] The plant is identified as the corresponding invasive weed only if the classification label of the coarse screening result is completely consistent with the fine screening result, and the result is marked as "verified" for subsequent early warning decision-making.

[0047] When two identification results have inconsistent classification labels, a multi-level review process is initiated: First, a trigger feature dimension difference analysis is performed, comparing the key morphological feature vectors relied upon by the two-level classifiers to identify the feature difference patterns that lead to the discrepancy; second, the corresponding arbitration classifier is invoked based on the feature difference pattern. This classifier is independently trained based on specific feature conflict scenarios in historical disputed samples and provides an arbitration result by analyzing the feature space distribution pattern; if the arbitration classifier still cannot reach a clear conclusion, the sample is automatically included in the manual review sequence, and a ground robot is simultaneously instructed to relocate and collect high-resolution images of multiple parts of the plant, such as the back of the leaves and the base of the stem. The final manual judgment conclusions of all disputed samples will be used as standard data to drive the incremental learning and optimization iteration of the cascade classification model, especially strengthening the ability to distinguish similar species that are prone to discrepancies.

[0048] It should be noted that the arbitration classifier is constructed by coupling support vector machines and random forests. Its core function is to resolve identification conflicts by building a feature difference space. The input to this classifier is the feature difference vector between the coarse-screening and fine-screening results, including parameters such as morphological feature deviation, texture feature distance, and spectral response difference. By establishing an optimal separating hyperplane in the high-dimensional feature difference space and combining decision tree ensemble learning to perform multi-granularity analysis of complex conflict patterns, accurate arbitration of disputed samples is achieved. This classifier uses kernel functions to handle non-linearly separable feature conflict scenarios and dynamically evaluates the contribution of different feature dimensions to the arbitration decision through an adaptive weight adjustment mechanism to output the arbitration result, ensuring consistency between the classification labels of the coarse-screening and fine-screening results.

[0049] S4. Construct a weed invasion risk diffusion prediction model, and obtain the normalized invasion risk index of the alien weed plants based on the weed invasion risk diffusion prediction model.

[0050] Specifically, S4 includes the following steps: S41. Construct a prediction model for the risk of weed invasion and spread.

[0051] In this embodiment, a set of invasion risk assessment factors is acquired and quantified. Invasion source intensity factors are obtained through spatial analysis methods, including three assessment dimensions: plant distribution density, biomass index, and reproductive potential index. Environmental driving force factors integrate multi-source environmental data, including elements such as climate suitability, soil compatibility, and hydrological conditions, and principal component analysis is used to obtain principal component factors from multiple environmental variables. Land type resistance factors are constructed based on landscape ecology principles, comprehensively considering factors such as land use type, vegetation cover, and intensity of human activities. Different land types are assigned differentiated basic resistance coefficients; for example, natural forests are low-resistance areas (coefficient 0.1-0.3), farmland is a medium-resistance area (coefficient 0.4-0.6), and built-up areas are high-resistance areas (coefficient 0.7-1.0). All factor data are spatially registered and normalized to form a standardized set of invasion risk assessment factors.

[0052] The above intrusion source strength factors satisfy the following relationship: in, Invasion source intensity factor, Plant distribution density, Biomass index As a reproductive potential index, These are the weight coefficients for the corresponding dimensions.

[0053] It should be noted that the weighting coefficients are determined by expert scoring or historical data regression to measure the relative contribution of different dimensions to the intensity of the intrusion source, and the weighting coefficients add up to 1.

[0054] The above environmental driving factors satisfy the following relationship: in, As an environmental driving force factor, The number of principal component factors, The index variable of the principal component factor. For the first The contribution weights of each principal component factor. For the first Principal component factor scores.

[0055] The resistance factors for the above land types satisfy the following relationship: in, For land type resistance factors, Based on the basic drag coefficient, To adjust the coefficient, To normalize vegetation cover, To normalize the intensity of human activities.

[0056] The above set of intrusion risk assessment factors can be represented as a three-dimensional feature vector, satisfying the following relationship: in, For the set of intrusion risk assessment factors, Invasion source intensity factor, As an environmental driving force factor, This represents the land type resistance factor.

[0057] Furthermore, based on the set of invasion risk assessment factors, a hybrid architecture coupling cellular automata (CA) and multilayer perceptron (MLP) is used to establish a weed invasion risk diffusion prediction model. The cellular automata defines a spatial grid of rules, with each cell state including three categories: invaded, uninvaded, and potentially invaded. The MLP, as the core of the intelligent state transition rules, receives feature data in its input layer: the neighborhood state vector of the cellular automata, invasion source intensity factors, environmental driving force factors, and land type resistance factors. The network structure contains three hidden layers with 128, 64, and 32 nodes respectively, using modified linear units to introduce nonlinear transformations. The output layer outputs the cell state transition probabilities through a sigmoid function. Model training employs a spatiotemporal cross-validation strategy, using historical invasion sequence data as a supervision signal. Optimizers such as adaptive moment estimation algorithms are used to optimize network parameters, and early stopping is employed to prevent overfitting. During the prediction process, CA provides a framework for spatial dynamic evolution, while MLP provides intelligent state transition rules. The two work together through an iterative feedback mechanism to achieve accurate simulation of the invasion process.

[0058] The above-mentioned weed invasion risk diffusion prediction model satisfies the following relationship: in, Represents a cell In the next moment State transition The probability, Indicates the grid position. This indicates that the system has been hacked. For the Sigmoid function, Here is the weight matrix of the multilayer perceptron. To correct the linear unit, For cells The neighborhood state vector, Set of intrusion risk assessment factors in cells The feature vector at that location, This is the bias vector of the multilayer perceptron.

[0059] In an optional embodiment, the loss function used to train the weed invasion risk spread prediction model is a weighted cross-entropy form, satisfying the following relationship: in, For loss function, The total number of training samples. For the index variable of the training samples, For hyperparameters, For the first The true label of each sample Represents the natural logarithm. For the first The predicted probability of a sample.

[0060] It should be noted that hyperparameters are used to increase the contribution of positive samples to the total loss, and are usually set as the proportion of positive samples in the training samples.

[0061] S42. Obtain the normalized invasion risk index of the invasive weed plants based on the weed invasion risk diffusion prediction model.

[0062] In this embodiment, multiple iterative simulations are conducted within a predetermined future time period using a pre-established weed invasion risk diffusion prediction model. Using the currently validated distribution of invasive weeds as the initial seed point, a simulation time step is set, and multiple simulation instances are run simultaneously using a parallel computing architecture. Within each time step, the invasion probability of each spatial unit is calculated based on a CA-MLP coupled model, and a Monte Carlo random sampling method is used to determine state transitions. To fully assess the uncertainty of the prediction, multiple independent simulations are performed, each using a different random number seed, and the invasion state of each spatial unit in each simulation is recorded. Finally, the frequency of invasion of each spatial unit within the predetermined time period is statistically analyzed to generate a spatiotemporal invasion risk probability map. This probability map adopts a hierarchical storage structure, containing multiple time slices, where each pixel value represents the statistical invasion probability at that location at the corresponding time point, forming a complete spatiotemporal risk assessment data foundation.

[0063] Furthermore, multi-dimensional statistical analysis was performed on the spatiotemporal invasion risk probability map to extract core weed invasion indicators. The potential maximum impact area was calculated by setting a risk threshold. First, the probability map of each time slice was binarized, and high-risk areas were extracted separately. Then, the total area of ​​connected patches in the region was calculated, and the maximum value over the entire simulation period was used as the indicator. The average spatial diffusion rate was calculated based on the dynamic changes of invasion. The boundary contours of invading patches each year were accurately extracted using image processing algorithms. The displacement distance of the centroid between adjacent years was calculated, and then divided by the time interval to obtain the diffusion rate. Finally, the average value over the entire period was taken. The overlap of ecologically sensitive areas was achieved through spatial overlay analysis. The spatial intersection of the distribution maps of high-risk invasion areas and ecologically sensitive areas was calculated, and the ratio of the overlapping area to the total area of ​​ecologically sensitive areas was calculated to assess the potential threat level of invasion to key ecological areas.

[0064] In this embodiment, the dynamic weight coefficients of the core intrusion indicators are objectively calculated using the entropy weight method. First, an indicator data matrix containing all evaluation units is constructed, and the values ​​of each indicator are standardized to eliminate dimensional differences. Then, the entropy value of each indicator is calculated based on the information entropy theory. The degree of difference between the indicators is determined according to the magnitude of the entropy value. The smaller the entropy value, the stronger the distinguishing ability of the indicator in the evaluation, and the higher the weight should be assigned.

[0065] After obtaining the dynamic weighting coefficients, the three core indicators are linearly weighted and fused. Before fusion, each indicator is normalized to a uniform numerical range. The calculated normalized intrusion risk index comprehensively reflects the scale and speed of the intrusion. Its numerical range is defined between 0 and 1, with higher values ​​indicating more severe intrusion risks, providing accurate quantitative basis for subsequent graded early warning.

[0066] The above normalized intrusion risk index satisfies the following relationship: in, To normalize the intrusion risk index, As the index variable for core intrusion indicators, The dynamic weighting coefficients for core intrusion indicators This is the normalized core intrusion indicator value.

[0067] S5. Classify the weed invasion level according to the normalized invasion risk index, and obtain the warning confidence level of the weed invasion level to generate early warning information for invasive weeds.

[0068] In this embodiment, a time-series sliding window mechanism is used to dynamically set the warning level thresholds, with the window size set to the most recent six monitoring periods. At each update time, the normalized intrusion risk index sequence within the window is extracted, and trend analysis is performed first: the slope coefficient of the sequence is calculated through linear regression, and the significance of the trend is tested. If a significant upward trend is shown, the thresholds at each level are lowered according to the trend strength; if a downward trend is shown, the thresholds are raised accordingly. Simultaneously, the volatility of the sequence within the window (the ratio of standard deviation to mean) is calculated. When the volatility exceeds 1.5 times the historical benchmark, the threshold interval between levels is widened to improve stability. Finally, differentiated adjustment ranges are set based on seasonal characteristics (such as the growing season and non-growing season) to ensure that the thresholds reflect both recent risk changes and maintain long-term robustness. All threshold parameters are stored in a dynamic configuration library and are automatically updated periodically.

[0069] Furthermore, based on dynamically set warning level thresholds, the normalized intrusion risk index is mapped to four intrusion levels. Threshold ranges are set as follows: Attention Level [0, 0.3), Alert Level [0.3, 0.6), Action Level [0.6, 0.8], and Emergency Level [0.8, 1.0]. The mapping process includes: escalating immediately when the index rises above the threshold, and downgrading only when it falls below the threshold for two consecutive monitoring cycles. Each level corresponds to a standardized response plan: Attention Level only requires recording monitoring data; Alert Level requires periodic review; Action Level requires on-site verification and preparation of prevention and control resources; Emergency Level immediately initiates a cross-departmental emergency response. The level mapping results are accompanied by timestamps and trend indicators, and are visualized using color coding (blue-yellow-orange-red).

[0070] In this embodiment, the data quality uncertainty component is quantified through multi-dimensional evaluation indicators, including the spatial resolution of image data (with 30 meters as the baseline, uncertainty increases by 0.1 for every 10 meters decrease), spectral signal-to-noise ratio (uncertainty increases linearly when it is less than 30 dB), meteorological conditions at the time of acquisition (uncertainty increases by 0.2 in rainy / strong wind weather), and data coverage completeness (uncertainty increases proportionally when the missing rate is >10%). The weights of each indicator are determined using the analytic hierarchy process (AHP), and the data quality uncertainty component in the range of 0-1 is finally obtained by weighting. The calculation process of the model uncertainty component is as follows: First, 50 random forward propagations are performed using Monte Carlo to calculate the variance of the prediction results; then, historical error patterns are analyzed, and the root mean square error between the model prediction value and the actual value in the last 12 prediction periods is statistically analyzed; finally, the KL divergence of the distribution of real-time input data and training data is combined, and the three are fused in a proportional weight (e.g., 4:3:3) to generate the model uncertainty component, the value of which is also normalized to the range of 0-1.

[0071] Furthermore, referring to the Dempster-Shafer evidence theory in step S2, the identification framework is defined to include reliable and unreliable components. The data quality uncertainty component and the model uncertainty component are fused to obtain the warning confidence level, satisfying the following relationship: in, To determine the confidence level for early warning, Indicates reliability. For the data quality uncertainty component, This represents the uncertainty component of the model.

[0072] In this embodiment, the warning confidence level is divided into four intervals (low [0, 0.6], medium [0.6, 0.8], high [0.8, 0.9], and extremely high [0.9, 1.0]). A 4×4-dimensional warning decision matrix is ​​constructed by combining the weed invasion level and the warning confidence level. The matrix has a total of 16 combinations of weed invasion level and warning confidence level. Different combinations are preset with different disposal strategies to form differentiated warning strategies. For example, when the combination of "action level + medium confidence level" is used, the strategy combination of "on-site verification + preparation of prevention and control resources + submission of verification report within 24 hours" is automatically triggered. When the combination of "emergency level + high confidence level" is used, the response process of "issuing red warning + emergency handling" is executed.

[0073] Furthermore, by matching level-confidence coordinate points in real time, structured early warning information for invasive weeds is generated from the differentiated early warning strategy. This information includes five elements: early warning level identifier, spatial distribution of risk, list of key areas, specific response measures, and expected response time. Simultaneously, a strategy optimization mechanism is established to dynamically adjust the matrix content based on historical early warning response data (such as response effectiveness evaluation and feedback delay statistics). The effectiveness of the strategy is periodically verified through testing, and the differentiated early warning strategy is updated to ensure that early warning decisions always maintain optimal alignment with the actual situation.

[0074] Please see Figure 2 In an optional embodiment, the present invention provides a ground-air collaborative confidence-based graded early warning system for invasive alien weeds. The system includes an input device, an output device, a processor, and a memory, all interconnected. The memory stores a computer program comprising program instructions, and the processor is configured to invoke the program instructions to execute specific steps as described in the relevant embodiments of the ground-air collaborative confidence-based graded early warning method for invasive alien weeds provided by the present invention. The ground-air collaborative confidence-based graded early warning system for invasive alien weeds provided by the present invention has a complete structure, is objective and stable, and enhances the overall applicability and practical application capability of the present invention.

[0075] In summary, this invention provides a ground-air collaborative confidence-based hierarchical early warning method and system for invasive alien weeds, constructing a ground-air collaborative perception and decision-making closed loop: First, through the collaborative operation of high-altitude UAVs, forest UAVs, and ground robots, multi-level data collection from macro to micro levels is achieved; second, a two-level mechanism of "preliminary screening-refined identification" is adopted, combining prior knowledge to accurately locate invasive alien weed plants; subsequently, a dynamic diffusion prediction model coupled with cellular automata and multilayer perceptron is constructed, capable of simulating the invasion situation and generating a quantified normalized invasion risk index; finally, Dempster-Shafer evidence theory is introduced, integrating the uncertainty of data and models to assess the early warning confidence, and based on the "invasion level-early warning confidence" early warning decision matrix, differentiated early warning information is generated, realizing a leap in early warning capability from qualitative to quantitative, from static to dynamic, and from single conclusions to reliability assessment. This invention is easy to understand, computationally simple, requires less workload, and is convenient for engineering applications, providing a theoretical foundation and technical support for the further development of plant protection monitoring technology.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. An air-ground cooperative confidence grading early warning method for alien weed invasion, characterized in that, The method comprises the following steps: Collecting plant multi-source sensing data of a target area by a plant air-ground collaborative sensing device; Performing preliminary identification on the plant multi-source sensing data to output a screening confidence, and screening a suspected alien weed area based on the screening confidence; Performing fine sensing on the suspected alien weed area by the plant air-ground collaborative sensing device to obtain multi-dimensional feature data, and determining an alien weed plant through classification and identification; Constructing a weed invasion risk diffusion prediction model, and obtaining a normalized invasion risk index of the alien weed plant according to the weed invasion risk diffusion prediction model; Dividing a weed invasion level according to the normalized invasion risk index, obtaining a warning confidence of the weed invasion level, and generating alien weed invasion warning information.

2. The method of claim 1, wherein the method further comprises: The plant multi-source sensing data of the target area is collected by the plant air-ground collaborative sensing device, which comprises: The plant air-ground collaborative sensing device is composed of a high-altitude unmanned aerial vehicle, a forest unmanned aerial vehicle and a ground robot; The plant high-altitude image data of the target area is obtained by the high-altitude unmanned aerial vehicle; The plant low-altitude multi-angle image data of the target area is collected by the forest unmanned aerial vehicle; The plant ground high-definition image data of the target area is obtained based on the ground robot; The plant high-altitude image data, the plant low-altitude multi-angle image data and the plant ground high-definition image data are taken as the plant multi-source sensing data.

3. The method of claim 2, wherein the method further comprises: The plant multi-source sensing data is preliminarily identified to output a screening confidence, and a suspected alien weed area is screened based on the screening confidence, which comprises: A large-range semantic segmentation is performed on the plant high-altitude image data to obtain a preliminary plant abnormal area; Based on the plant low-altitude multi-angle image data and the plant ground high-definition image data, multi-view feature fusion enhancement is performed on the preliminary plant abnormal area to extract plant morphological texture fusion features; The preliminary plant abnormal area is divided according to the plant morphological texture fusion features to obtain a plant abnormal sub-area; The screening confidence of the plant abnormal sub-area is calculated, and the screening confidence is compared with a preset threshold to determine the suspected alien weed area.

4. The method of claim 3, wherein the method further comprises: The screening confidence of the plant abnormal sub-area is calculated, which comprises: Based on the plant morphological texture fusion features, an initial confidence of the plant abnormal sub-area is obtained by a lightweight classifier; A local plant priori knowledge base of the target area is obtained, which comprises plant phenological phases and plant growth characteristics; The plant morphological texture fusion features are matched and compared with the priori knowledge base to obtain a priori coincidence score; The initial confidence and the priori coincidence score are combined to obtain the screening confidence.

5. The method of claim 2, wherein the method further comprises: The suspected alien weed area is finely sensed by the plant air-ground collaborative sensing device to obtain multi-dimensional feature data, and an alien weed plant is determined through classification and identification, which comprises: The forest unmanned aerial vehicle collects low-altitude multi-angle image data of the suspected alien weed area, generates a three-dimensional point cloud model based on the low-altitude multi-angle image data to obtain a three-dimensional structure feature, and labels a key plant to be identified; Under the guidance of the forest unmanned aerial vehicle, the ground robot autonomously plans a path to an optimal observation point to collect stem detail features of the key plant to be identified; The three-dimensional structure feature and the stem detail features are used as the multi-dimensional feature data; Based on the multi-dimensional feature data, a cascade classification model is used for identification, including a first-level classifier that obtains a coarse screening identification result based on the three-dimensional structure feature, and a second-level classifier that obtains a fine identification result based on the coarse screening identification result and the stem detail features; The coarse screening identification result and the fine identification result are jointly analyzed to determine the alien weed plant.

6. The method of claim 1, wherein the method further comprises: The construction of the weed invasion risk diffusion prediction model includes: An invasion risk assessment factor set of the alien weed plant is obtained, including an invasion source intensity factor, an environmental driving force factor, and a land type resistance factor; Based on the invasion risk assessment factor set, a cellular automaton and a multilayer perception machine are coupled to construct the weed invasion risk diffusion prediction model.

7. The method of claim 6, wherein the method further comprises: The normalized invasion risk index of the alien weed plant is obtained according to the weed invasion risk diffusion prediction model, including: In a future preset period, the weed invasion risk diffusion prediction model is used for iterative simulation to obtain a spatiotemporal invasion risk probability map of the alien weed plant; A multi-dimensional statistical analysis is performed on the spatiotemporal invasion risk probability map to obtain core weed invasion indicators, including a potential maximum impact area, an average spatial diffusion speed, and an ecological sensitive area overlap degree; A dynamic weight coefficient of the core weed invasion indicators is obtained, and the core weed invasion indicators are linearly weighted and fused based on the dynamic weight coefficient to obtain the normalized invasion risk index.

8. The method of claim 1, wherein the method further comprises: According to the normalized invasion risk index, a weed invasion level is divided, and a warning confidence of the weed invasion level is obtained to generate alien weed invasion warning information, including: A sliding window mechanism is used to dynamically set a warning level threshold by analyzing the change trend of the normalized invasion risk index; According to the warning level threshold, the normalized invasion risk index is mapped to obtain a weed invasion level, including a concern level, an alert level, an action level, and an emergency level; The warning confidence is obtained based on an uncertainty component; A warning decision matrix is constructed based on the weed invasion level and the warning confidence, and a differentiated warning strategy is obtained according to the warning decision matrix to generate the alien weed invasion warning information.

9. The method of claim 8, wherein the method further comprises: The warning confidence is obtained based on the uncertainty component, including: The uncertainty component is obtained, including a data quality uncertainty component and a model uncertainty component; The uncertainty component is fused by Dempster-Shafer evidence theory to obtain the warning confidence.

10. An aerial-ground cooperative confidence ranking early warning system for alien weed invasion, characterized in that, The system comprises an input device, an output device, a processor and a memory, which are connected with each other, wherein the memory is used for storing a computer program, the computer program comprises program instructions, and the processor is configured to invoke the program instructions to execute the method for air-ground cooperative confidence grading early warning of alien weed invasion according to any one of claims 1-9.

Images