A method for preventing and controlling canada goldenrod based on leaf vein feature fine-grained identification

Abstract

The application belongs to the field of agricultural information technology, and particularly relates to a Canada goldenrod prevention and control method based on leaf vein feature fine-grained identification. A high-resolution image dataset containing four types of seedling leaves of species is constructed, and the characteristics of tri-foliate veins of Canada goldenrod and the characteristics of pinnate veins of other species are highlighted and labeled. A deep learning model with a leaf vein structure attention module is designed, multi-scale feature extraction and triple comparison learning are used to force the model to focus on the vein direction, bifurcation mode and topological structure difference. Real-time images are collected by a drone and are divided into grids, and the species category, seedling confidence and spatial distribution are output in grid units. Based on the identification result, a differentiated and accurate prevention and control scheme is automatically matched. The seedling identification accuracy of the Canada goldenrod reaches 96.8%, the false alarm rates of the sumacweed, the small bentgrass and the annual meadowgrass are as low as 3.2%, 3.5% and 2.8% respectively, the pesticide saving rate reaches 52%-68%, and the other plant injury rate is reduced from 18.5% of the traditional method to 2.1%.

Description

Technical Field

[0001] This invention belongs to the field of agricultural information technology and invasive alien plant control technology, specifically relating to a method for controlling Canadian goldenrod based on fine-grained identification of leaf vein features. Background Technology

[0002] Canadian goldenrod (Solidago canadensis L.) is an important invasive alien plant in my country, seriously threatening the country's ecological environment and economic security. In response to the severe impact of Canadian goldenrod, the Ministry of Agriculture and Rural Affairs has included it in the "List of Key Invasive Alien Species under Management," and many regions have established monitoring and early warning systems and launched special control campaigns.

[0003] Existing technologies for controlling such invasive weeds include, for example, a Chinese patent that provides a method and system for detecting and applying pesticides in farmland using visible light. This method involves acquiring visible light images of the farmland; inputting these images into an improved YOLOv7 model to obtain weed bounding boxes and a weed density matrix; the weed density matrix including weed density values ​​for multiple regions; establishing a nonlinear mapping relationship between weed density values ​​and a pre-set pesticide application level set; and applying pesticides using a variable method based on the weed density matrix and the nonlinear mapping relationship. This improves the accuracy of weed identification and pesticide application during the seedling stage, thereby increasing pesticide utilization.

[0004] However, existing technologies for treating Canadian goldenrod still have the following problems: First, the morphology of Canadian goldenrod varies significantly at different growth stages (seedling stage, vegetative growth stage, reproductive growth stage), and it is similar in appearance to other Asteraceae plants, making it prone to misjudgment using traditional image recognition. Second, the seedling stage is the optimal window for chemical control of Canadian goldenrod, as the plants are small, their root systems are not yet fully developed, and they are sensitive to pesticides, achieving the best control effect at the lowest cost. However, the seedlings of Canadian goldenrod at this stage are extremely similar in morphology to the seedlings of other Asteraceae plants that are abundant in my country, such as *Symplocos sumatra*, as shown in the attached image. Figure 1 and 2 As shown, all of them have basal leaves growing in clusters, with subtle differences in leaf shape and color, making them highly susceptible to misjudgment using traditional manual identification or general image recognition methods.

[0005] Therefore, developing a control method and approach for Goldenrod in the seedling stage that can be applied to the prevention and control of Goldenrod is an urgent technical problem to be solved. Summary of the Invention

[0006] To address the technical problems of large identification errors and high false alarm rates when existing invasive weed control systems and methods are applied to goldenrod seedlings, this invention provides a goldenrod control method based on fine-grained identification of leaf vein features. By using a fine-grained identification method based on leaf vein structural features, the identification success rate can be significantly improved.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition, characterized by the following steps:

[0009] Step S1: Construct a high-resolution fine-grained image dataset, which contains leaf images of seedling Canadian goldenrod and easily confused plants, and the leaf images are labeled with the leaf vein structure of seedling Canadian goldenrod and easily confused plants.

[0010] Step S2: Based on the dataset, train a deep learning recognition model that integrates a leaf vein structure attention module. The leaf vein structure attention module is used to drive the model to focus on and extract leaf vein discriminative features to distinguish between Canadian goldenrod and easily confused plants. The leaf vein discriminative features include the number of main veins, the direction of the main veins, the branching pattern of the lateral veins, the leaf vein density, and the topological relationship between the leaf margin serrations and the leaf veins.

[0011] Step S3: Collect real-time images of the target area, divide the images into grids, and input them into the trained deep learning recognition model. Output the species classification probability and fine-grained confidence score in each grid, and give the recognition result.

[0012] Step S4: Automatically match differentiated and precise prevention and control solutions based on the identification and judgment results.

[0013] Furthermore, in step S1, the easily confused plants include one or more of the following: *Erigeron simonii*, *Erigeron canadensis*, and *Erigeron annuus*.

[0014] Furthermore, in step S1, the leaf vein structure marking method is a three-level annotation, including the following steps:

[0015] Level 1 labeling: Mark the number, location, and direction of the main veins of *Solidago canadensis* and easily confused plants during the seedling stage;

[0016] Secondary labeling: Label the lateral veins of seedlings of Canadian goldenrod and easily confused plants, including the angle of attachment, number of forks, and termination position. The accuracy of the angle of attachment is required to be ±5°, the number of forks is 0-3, and the termination position is the leaf margin, leaf tip, or leaf middle.

[0017] Level 3 annotation: Annotate the topological relationship between leaf veins and leaf margin serrations of seedlings of Canadian goldenrod and easily confused plants. The topological relationship includes the connection relationship between serrations and the ends of lateral veins, and the proportion of serrations connected to the ends of lateral veins.

[0018] Furthermore, in step S2, the leaf vein structure attention module employs a multi-scale leaf vein feature extraction structure. This extraction structure is used to extract leaf image markers for seedling-stage Canadian goldenrod and easily confused plants. The extraction structure includes:

[0019] Low-level feature extraction branch: Set 3×3 and 5×5 convolution kernels, receptive field 7×7 pixels, used to extract low-level texture features of main vein edges, corners and directions in the first-level annotation;

[0020] Intermediate feature extraction branch: Using 7×7 and 11×11 convolutional kernels with a receptive field of 31×31 pixels, it is used to extract intermediate structural features such as lateral vein attachment angle, number of bifurcations and termination position in the secondary annotation;

[0021] High-level feature extraction branch: Using 15×15 and 21×21 convolution kernels with a receptive field of 63×63 pixels or more, it is used to extract the topological relationship between leaf veins and leaf margin serrations in the third-level annotation, as well as the global features of the spatial distribution pattern of main and lateral veins.

[0022] Furthermore, in step S2, the training employs a leaf vein feature metric learning strategy based on contrastive learning, including the following steps:

[0023] Construct triplet training samples, each triplet containing an anchor sample, a positive sample, and a negative sample. Then, use the triplet loss function to force the anchor sample and the positive sample to be closer in the feature space, while pushing the anchor sample and the negative sample further apart.

[0024] The anchor sample is a leaf image of Canadian goldenrod, the positive sample is a leaf image of a different plant of the same species of Canadian goldenrod, and the negative sample is a leaf image of a plant that is easily confused with another plant.

[0025] The triplet loss function is defined as follows:

[0026]

[0027] in For Euclidean distance, Set to 1.0.

[0028] Further, in step S3, the species classification probabilities include P(Canada Goldenrod), P(Sumatran White Grass), P(Erigeron davidii), P(Erigeron annuus), and P(background). P(Canada Goldenrod) represents the probability that the grid image is Canada Goldenrod; P(Sumatran White Grass) represents the probability that the grid image is Sumatran White Grass; P(Erigeron davidii) represents the probability that the grid image is Erigeron davidii; P(Erigeron annuus) represents the probability that the grid image is Erigeron annuus; and P(background) represents the probability that the grid image is the background image.

[0029] The fine-grained confidence level ranges from 0 to 1.

[0030] Furthermore, in step S3, the identification results include the Canadian goldenrod region, the region to be verified, and the non-target species or background region;

[0031] The criteria for identifying the Canadian goldenrod region are P(Canadian goldenrod) > 0.95 and fine-grained confidence level > 0.90;

[0032] The criteria for judging the region to be reviewed are 0.85 < P (Canada Goldenrod) ≤ 0.95 and fine-grained confidence level > 0.90;

[0033] The criteria for determining non-target species or background areas are P(Canada goldenrod) ≤ 0.85 and fine-grained confidence level > 0.90.

[0034] Furthermore, in step S4, the differentiated and precise prevention and control plan includes:

[0035] When the identification result is "Canadian goldenrod region", spot spraying is carried out at 50%-60% of the baseline amount, and the spraying agent is Canadian goldenrod fermentation liquid; no spraying instruction is generated for areas identified as non-target species or background areas; for areas identified as needing verification, manual verification is carried out.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. This invention innovatively proposes a deep learning recognition model based on fine-grained leaf vein features, which can significantly improve the recognition accuracy of *Solidago canadensis* seedlings. Addressing the technical challenge of distinguishing easily confused *Solidago canadensis* seedlings from easily confused Asteraceae plants (*Erigeron canadensis*, *Erigeron simonii*, and *Erigeron annuus*) using traditional image recognition methods, this invention proposes for the first time to use leaf vein structure as the core discriminative feature and designs a deep learning recognition model that integrates a leaf vein structure attention module (VSAM). This model, through multi-scale feature extraction branches, can simultaneously capture fine-grained features such as the midrib, lateral veins, and leaf vein-leaf margin topological relationships. Combined with a triplet contrastive learning strategy, it forces the model to focus on the essential differences in leaf vein structure between species. Experiments have shown that the present invention achieves an accuracy rate of 96.8% in identifying seedlings of Canadian goldenrod, and the false alarm rates for Sumatran white wine grass, Erigeron canadensis, and Erigeron annuus are as low as 3.2%, 3.5%, and 2.8%, respectively. Compared with the existing identification model with an accuracy rate of 85.6%, the present invention improves the accuracy rate by 11.2 percentage points and significantly reduces the risk of misjudgment.

[0038] 2. This invention constructs a three-level fine-grained annotation system during the data annotation stage, providing structured, multi-layered supervisory signals for leaf vein feature learning. Existing technologies for labeling plant leaves often remain at the species level or with simple leaf outlines, lacking a refined description of the vein topology. This invention proposes for the first time a first-level (main vein), second-level (lateral vein), and third-level (vein-leaf margin topology) annotation system for identifying Canadian goldenrod. It quantitatively defines key discriminative indicators such as the number and orientation of main veins, the angle of lateral vein attachment (accuracy ±5°), the number of bifurcations (levels 0-3), termination positions, and the connection relationship between serrations and the ends of lateral veins. This annotation system not only provides a clear feature learning objective for deep learning models, enabling them to focus on discriminative regions of leaf veins layer by layer from low-level texture and intermediate structure to global topology (the IoU between Grad-CAM heatmap and botanical annotation reaches 0.87), but also possesses good interpretability and transferability, and can be extended to fine-grained identification tasks of other invasive plants.

[0039] 3. This invention designs a Leaf Vein Structure Attention Module (VSAM) to achieve efficient extraction and adaptive fusion of multi-scale leaf vein features. Unlike general channel attention (such as SE) or hybrid attention (such as CBAM), this invention specifically designs parallel multi-scale extraction branches for the scale characteristics of leaf vein structures. The low-level branch captures low-level textures such as the edges and corners of the main veins using small convolutional kernels, sensitive to subtle texture changes such as the "three main veins protruding on the underside of the leaf." The mid-level branch uses medium-sized convolutional kernels to extract intermediate structural features such as the direction and bifurcation angle of lateral veins, accurately distinguishing the direction differences between "three-veined veins" and "pinnate veins." The high-level branch extracts global features such as the number of main veins and the topological relationship between leaf veins and leaf margins using large convolutional kernels, stably distinguishing the structural differences between "three main veins" and "one main vein." Ablation experiments show that the introduction of the VSAM module improves the model accuracy by 7.3%, verifying the rationality and effectiveness of the module design.

[0040] 4. Achieving an integrated closed loop of "identification-decision-control" significantly reduces pesticide dosage and ecological side effects. This invention deeply integrates a high-precision identification model with a differentiated and precise control strategy. By dividing drone aerial images into grids, it outputs species classification probabilities, fine-grained confidence scores, and leaf vein feature activation heatmaps on a grid-by-grid basis, dynamically dividing the region into "Canada Goldenrod region," "region awaiting verification," and "non-target region." For confirmed areas, only 50%-60% of the baseline amount is applied for precise spot spraying. This method achieves a pesticide saving rate of 52%-68%, while maintaining a 30-day plant control efficacy of 93.5%. Simultaneously, the accidental damage rate to other plants is significantly reduced from 18.5% in existing technologies to 2.1%, significantly reducing the impact of chemical pesticides on the ecological environment while ensuring control effectiveness.

[0041] 5. The model exhibits strong generalization ability and maintains high stability across different seedling stages and cross-regional scenarios. Cross-regional generalization experiments show that the model still achieves an average recognition accuracy of 94.8% for Canadian goldenrod on independent test sets in regions not involved in training, with a performance difference of only -2.0% compared to the training set. Seedling stage experiments show that from early to late seedling stages, the model's recognition accuracy remains stable between 95.2% and 97.8%, maintaining excellent performance even for the smallest and least distinctive early-stage plants. This demonstrates that the recognition strategy based on leaf vein structure features possesses species specificity and growth stage invariance, unaffected by geographical environment and plant development status, and has broad application value. Attached Figure Description

[0042] Figure 1 Images showing the characteristics of seedlings of Canadian goldenrod (A), Sumatran white wine grass (B), Erigeron candelilla (C), and Erigeron annuus (D);

[0043] Figure 2These are actual images of the leaf characteristics of Canadian goldenrod, Sumatran white wine grass, Erigeron candelilla, and Erigeron annuus (from left to right). Detailed Implementation

[0044] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings, examples and experimental examples.

[0045] Example

[0046] This embodiment provides a method for controlling Canadian goldenrod based on fine-grained identification of leaf vein features, characterized by the following steps:

[0047] Step S1: Construct a high-resolution, fine-grained image dataset containing leaf images of seedling *Solidago canadensis* and easily confused plants. The leaf images are labeled with the vein structures of seedling *Solidago canadensis* and easily confused plants. Specifically, leaf images of *Solidago canadensis* were collected during its seedling stage. The collection period was from March 15th to April 20th, 2025, in wild habitats in Jurong City and Nanjing City, Jiangsu Province, and Chuzhou City, Anhui Province. Three sampling points were selected in each region, and at each sampling point, at least 500 images of seedling *Solidago canadensis* and various easily confused plants were collected. Easily confused plants in the Asteraceae family include *Erigeron sumatrensis*, *Erigeron canadensis*, and *Erigeron annuus*. Images of the collected plants are attached. Figure 1 and 2 As shown. To ensure high-resolution leaf images, a DJI Mavic 3E drone equipped with a 20-megapixel mechanical shutter camera was used to acquire leaf images of various target plants at a flight altitude of 1.5-2.0 meters, with an image resolution of 0.08 mm / pixel. A DSLR camera with a 100mm macro lens was also used for supplementary ground-based image acquisition.

[0048] When marking the leaf vein structure of *Solidago canadensis* and easily confused plants in the seedling stage, the LabelMe annotation tool was used. Three botanical professionals independently annotated the leaves, and the intersection of the annotation results was taken; any points of disagreement were decided by an expert panel. The leaf vein structure features included the basal tri-veining characteristic of *Solidago canadensis*, the pinnate venation characteristic of *Solanum lyratum*, the pinnate venation with sparsely toothed margins of *Erigeron annuus*, and the pinnate venation with basal leaf clusters of *Erigeron annuus*.

[0049] A three-level annotation system is adopted, and the annotation content includes:

[0050] Level 1 annotation (main vein annotation): Annotates the number, location, edges, corners, and direction of the main veins. For example, *Solidago canadensis* is annotated with 3 main veins (tri-veined from the base), while *Erigeron suma*, *Erigeron canadensis*, and *Erigeron annuus* are annotated with 1 main vein (pinnate vein). Annotations use Bézier curves, with keypoint spacing ≤ 10 pixels.

[0051] Secondary labeling (lateral vein labeling): Label the angle of lateral vein (accuracy ±5°), number of branches (0-3 levels), and termination position (leaf margin / leaf tip / leaf middle).

[0052] Level 3 annotation (leaf vein-leaf margin topology annotation): Annotate the connection relationship between leaf veins and leaf margin serrations, and record the proportion of serrations connected to the ends of lateral veins.

[0053] The final dataset contains 6320 images of Canadian goldenrod, 5840 images of Sumatran white wine grass, 6010 images of Erigeron candelilla, 5950 images of Erigeron annuus, and 4000 background images, for a total of 28120 images. At least 4000 images are needed for each species for training, 1000 for validation, and the remaining images for testing.

[0054] Step S2: Based on the dataset, train a deep learning recognition model that integrates a leaf vein structure attention module. The leaf vein structure attention module is used to drive the model to focus on and extract leaf vein discriminative features to distinguish between Canadian goldenrod and easily confused plants. The leaf vein discriminative features include the number of main veins, the direction of the main veins, the branching pattern of the lateral veins, the leaf vein density, and the topological relationship between the leaf margin serrations and the leaf veins.

[0055] Specifically, in this embodiment, the overall architecture of the deep learning recognition model integrating the vein structure attention module adopts an improved multi-branch convolutional neural network, with the backbone network using ResNet-50, fine-tuned based on ImageNet pre-trained weights. The vein structure attention module (VSAM) is embedded after the 3rd and 4th residual blocks of the backbone network. This module contains three parallel multi-scale feature extraction branches:

[0056] ① Low-level texture extraction branch: Using 3×3 and 5×5 convolution kernels with a receptive field of 7×7 pixels, this branch extracts low-level texture features such as the edges, corners, and orientation of the main veins from the primary annotation. This branch is sensitive to the texture gradient changes formed by the "three main vein protrusions on the underside of the leaf".

[0057] ② Mid-level structure extraction branch: Using 7×7 and 11×11 convolutional kernels with a receptive field of 31×31 pixels, this branch extracts mid-level structural features from the secondary annotations, such as the lateral vein attachment angle, number of bifurcations, and termination position. This branch is sensitive to the difference in orientation between "tri-veins from the base" and "feathered veins".

[0058] ③ High-level topology extraction branch: Using 15×15 and 21×21 convolutional kernels with a receptive field of 63×63 pixels or more, this branch is used to extract the topological relationships between leaf veins and leaf margin serrations in the third-level annotation, as well as global features such as the spatial distribution pattern of main and lateral veins. This branch is sensitive to the structural differences between "three main veins" and "one main vein".

[0059] The output feature maps of the three branches are adaptively fused through a channel attention mechanism to generate an enhanced leaf vein structure representation vector.

[0060] During training, a triplet loss function was introduced, and a leaf vein feature metric learning strategy based on contrastive learning was adopted. The optimizer used was AdamW, with an initial learning rate of 0.0001, weight decay of 0.01, a batch size of 64, and a training cycle of 150 epochs. The learning rate adopted a cosine annealing decay strategy, and the weight ratio between the contrastive learning branch and the cross-entropy branch was 1:1.

[0061] First, triplet training samples are constructed. Each triplet contains an anchor sample (a leaf image of Canadian goldenrod), positive samples (leaf images of different plants of the same species of Canadian goldenrod), and negative samples (leaf images of easily confused plants). A triplet loss function is used to force the anchor sample and positive sample to be closer in the feature space, while simultaneously increasing the distance between the anchor sample and negative sample.

[0062] In this embodiment, the triplet loss function is defined as:

[0063]

[0064] in For Euclidean distance, Set to 1.0 to ensure that in the feature space, the inter-class distance threshold for Canada goldenrod, Sumatran white wine grass, Erigeron canadensis, and Erigeron annuus is set to at least twice the intra-class distance.

[0065] Step S3: Use a drone equipped with a high-resolution camera to collect real-time images of the target area. After dividing the images into grids, input them into the trained deep learning recognition model. Output the species classification probability, leaf vein feature activation heatmap and fine-grained confidence score in each grid, and give the recognition judgment result.

[0066] Specifically, a drone equipped with a high-resolution camera is used to collect real-time images of the target area. After normalization, noise reduction, and illumination correction of the real-time images, the images are divided into a 64×64 pixel grid, with each grid corresponding to an area of ​​approximately 10cm×10cm on the ground, thus matching the size of the spraying unit.

[0067] The divided grid image is input into a trained deep learning recognition model. The deep learning recognition model outputs the species classification probability, leaf vein feature activation heatmap, and fine-grained confidence score for each grid cell. The species classification probability includes P(Canada Goldenrod), P(Sumatran White Grass), P(Erigeron davidii), P(Erigeron annuus), and P(Background), corresponding to the probability of occurrence of different species classifications. The fine-grained confidence score is the certainty score of the deep learning recognition model for the current judgment, ranging from 0 to 1, with a higher score indicating higher confidence.

[0068] The identification results include the Canadian goldenrod region, the region to be verified, and the non-target species or background region. The criteria for identifying the Canadian goldenrod region are P(Canadian goldenrod) > 0.95 and fine-grained confidence level > 0.90; the criteria for identifying the region to be verified are 0.85 < P(Canadian goldenrod) ≤ 0.95 and fine-grained confidence level > 0.90.

[0069] The criteria for determining non-target species or background areas are P(Canada goldenrod) ≤ 0.85 and fine-grained confidence level > 0.90.

[0070] Step S4: Automatically match differentiated and precise prevention and control solutions based on the identification and judgment results.

[0071] Specifically, when the identification result is "Canada Goldenrod area", a precise control instruction is generated for the geographic coordinates of that grid. This instruction includes a recommended spraying dosage and spraying method. For grids confirmed as seedling-stage Canada Goldenrod, a point-based precise spray is applied at 50%-60% of the baseline dosage. The spraying agent is Canada Goldenrod fermented liquid, which is consistent with the biological control fermented liquid disclosed in prior art publication CN120283790A. The specific preparation method is also consistent with existing technologies and will not be elaborated here. The corresponding baseline dosage is also consistent with existing technologies. No spraying instruction is generated for areas identified as non-target species or background areas; areas identified as requiring verification are manually verified.

[0072] Experiment Example 1: Performance of a Deep Learning Recognition Model Integrating a Leaf Vein Structure Attention Module

[0073] Experimental Methods: Following the control and experimental groups shown in Table 1 below, different model architectures, attention modules, and learning models were set up, using the same hardware configuration and training parameters. The accuracy of identifying seedlings of *Solidago canadensis* and easily confused plants was compared after training. Hardware Configuration: NVIDIA A100 GPU × 4, 80GB / card. Training Parameters: The optimizer used was AdamW, with an initial learning rate of 0.0001, weight decay of 0.01, batch size of 64, and 150 training cycles. The learning rate adopted a cosine annealing decay strategy. In the experimental group, the margin of the triplet loss function was set to 1.0, and the weight ratio of the contrast learning branch to the cross-entropy branch was 1:1. Experimental results are shown in Table 2 below:

[0074] Table 1. Composition of Deep Learning Recognition Models in the Control and Experimental Groups of Example 1

[0075]

[0076] Table 2. Experimental results after training for the control group and experimental group in Experiment Example 1.

[0077]

[0078] Experimental Results Analysis: The experimental results show that the deep learning recognition model based on fine-grained leaf vein features proposed in this invention significantly improves the recognition accuracy of *Solidago canadensis* seedlings, reaching 96.8%. The false positive rates for *Symplocos suma*, *Erigeron canadensis*, and *Erigeron annuus* are as low as 3.2%, 3.5%, and 2.8%, respectively, indicating that the model has the ability to successfully recognize *Solidago canadensis* seedlings. Control groups 1-3, using existing architectures and attention modules respectively, only achieved a maximum recognition accuracy of 85.6% after training on the same test set, far lower than the results of this invention. The false positive rates for *Symplocos suma*, *Erigeron canadensis*, and *Erigeron annuus* reached as high as 21%, 22.5%, and 19.8%, respectively, rendering control groups 1-3 unusable in practical applications.

[0079] Experiment Example 2: Ablation Experiment

[0080] Experimental Methods: An ablation experiment was conducted on the complete model, removing different modules one by one to test the contribution of each module in the identification model. The experimental results are shown in Table 3 below:

[0081] Table 3 Statistical table of experimental results in Experiment Example 2

[0082]

[0083] Experimental results analysis: The ablation experiment results show that both the VSAM module and the triplet loss function make significant contributions, and removing any module will lead to a performance drop of more than 5%. Among the three multi-scale branches, the mid-level structure extraction branch (responsible for leaf vein direction and bifurcation angle) contributes the most. The complete model has the best overall performance, and the best recognition effect is achieved through the collaborative work of each module.

[0084] Experiment Example 3: Comparison Experiment with Existing Technology Models

[0085] Experimental Methods: The deep learning recognition model trained by this invention and the existing technology model were run on the same test set. The test set was the test set used by the recognition model not trained by this invention, containing 500 images each of Canadian goldenrod, Sumatran white wine grass, Erigeron canadensis, and Erigeron annuus. The existing technology model adopted the deep learning-based method for precise weeding operation area detection in vegetable fields proposed by Li Weili et al. (Deep Learning-based Method for Precision Weeding Operation Area Detection in Vegetable Fields [J]. Fujian Journal of Agricultural Sciences, 2024, 39(2): 1-8). This method adopts the technical route of image gridding, deep learning classification (ShuffleNet / DenseNet / ResNet), and operation area labeling. The recognition accuracy and confusion matrix prediction results of Canadian goldenrod, Sumatran white wine grass, Erigeron canadensis, and Erigeron annuus were tested respectively. The experimental results are shown in Tables 4-6 below:

[0086] Table 4. Statistical table of recognition accuracy results in the comparative experiment of Experiment Example 3

[0087]

[0088] Table 5. Statistical table of prediction results of confusion matrix of existing technology model in Experiment Example 3 (comparison experiment)

[0089]

[0090] Table 6. Statistical table of confusion matrix prediction results of the present invention model in the comparative experiment of Example 3.

[0091]

[0092] Experimental results analysis: The results show that the method of the present invention is significantly superior to the existing technology in all four classification tasks. In particular, it reduces the confusion rate between Canadian goldenrod and Sumatran white wine grass from 12.5% ​​to 1.5%, which verifies the technical advantages of fine-grained identification based on leaf vein features combined with precise prevention and control.

[0093] Experiment Example 4: Field Precision Control Verification Experiment

[0094] Experimental Methods: A field verification experiment was conducted in an area invaded by *Solidago canadensis* in Jurong City, Jiangsu Province. Three experimental plots, each 2 mu in size, were selected, containing mixed growth areas of *Solidago canadensis*, *Sophora flavescens*, *Erigeron canadensis*, and *Erigeron annuus*. Aerial photography was conducted using a drone (flight altitude 2 meters, grid 64×64 pixels, corresponding to 10cm×10cm ground). Identification and control were compared between control group 1 (ResNet-50 + uniform spraying) and experimental group (the method of this invention: ResNet-50 + VSAM + ternary compound + precision spraying). In control group 1, the entire plot was uniformly sprayed after identification (450 L / hm²), while in experimental group, only the grids identified as *Solidago canadensis* were targeted with spot spraying (50% of the baseline amount, i.e., 225 L / hm²). The experimental results are shown in Table 7 below.

[0095] Table 7 Statistical table of experimental results for Example 4

[0096]

[0097] Analysis of experimental results:

[0098] As shown in Table 7, the method of this invention (experimental group) achieved a recognition accuracy of 95.8% for Canadian goldenrod, significantly higher than the 72.3% of the control group 1. Regarding pesticide dosage, the experimental group used only 216 L / hm², saving 52.0% compared to the control group 1, while maintaining a 30-day control efficacy of 93.5%, slightly higher than the 91.2% of the control group 1. Particularly noteworthy is the fact that the experimental group had a mis-damage rate of only 2.1% for other plants, far lower than the 18.5% of the control group 1. These results demonstrate that this invention, through precise identification and differentiated spraying strategies, significantly reduces pesticide dosage and ecological damage to non-target plants while ensuring control efficacy, achieving the goal of green and precise pest control.

[0099] Experiment Example 5: Verification of Cross-Regional Generalization Ability

[0100] Experimental Methods: Independent test sets were collected from Anhui Province (not a training region), Zhejiang Province, and Jiangxi Province. 200 images each of *Solidago canadensis*, *Sophora flavescens*, *Erigeron canadensis*, and *Erigeron annuus* were collected from each region, totaling 2400 images. These images were imported into the recognition model trained in the examples to test the model's ability to recognize seedling-stage *Solidago canadensis* across different regions. The experimental results are shown in Table 8 below:

[0101] Table 8 Statistical table of experimental results in Experiment Example 5

[0102]

[0103] The results show that the method of the present invention has good cross-regional generalization ability. Even on independent test sets in different provinces, the recognition accuracy remains above 94%, which proves that the recognition strategy based on leaf vein structure features is species-specific and is not affected by regional environment.

[0104] Experiment Example 6: Recognition Stability Experiment of Canadian Goldenrod at Different Seedling Stages

[0105] Experimental Methods: 400 images each of Canadian goldenrod seedlings at different seedling stages (early, middle, and late stages) were selected and imported into the recognition model provided in this invention. The recognition accuracy was tested, and the results are shown in Table 9 below. Early seedling stage: 10-15 days after emergence, plant height 5-10cm, 2-4 true leaves; Middle seedling stage: 20-25 days after emergence, plant height 10-20cm, 5-8 true leaves; Late seedling stage: 30-35 days after emergence, plant height 20-35cm, 8-12 true leaves.

[0106] Table 9 Statistical Table of Experimental Results for Example 6

[0107]

[0108] Experimental results analysis: The results show that the method of the present invention maintains an accuracy rate of over 95% at all stages of the seedling stage, and the recognition performance slightly improves as the plant grows and the leaf vein structure becomes clearer. Even in the early seedling stage (the stage when the plant is smallest and the features are least obvious), the accuracy rate still reaches 95.2%, meeting the needs of practical applications.

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for controlling Canadian goldenrod based on fine-grained identification of leaf vein features, characterized in that, Includes the following steps: Step S1: Construct a high-resolution fine-grained image dataset, which contains leaf images of seedling Canadian goldenrod and easily confused plants, and the leaf images are labeled with the leaf vein structure of seedling Canadian goldenrod and easily confused plants. Step S2: Based on the dataset, train a deep learning recognition model that integrates a leaf vein structure attention module. The leaf vein structure attention module is used to drive the model to focus on and extract leaf vein discriminative features to distinguish between Canadian goldenrod and easily confused plants. The leaf vein discriminative features include the number of main veins, the direction of the main veins, the branching pattern of the lateral veins, the leaf vein density, and the topological relationship between the leaf margin serrations and the leaf veins. Step S3: Collect real-time images of the target area, divide the images into grids, and input them into the trained deep learning recognition model. Output the species classification probability and fine-grained confidence score in each grid, and give the recognition result. Step S4: Automatically match differentiated and precise prevention and control solutions based on the identification and judgment results.

2. The method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition according to claim 1, characterized in that, In step S1, the easily confused plants include one or more of the following: *Erigeron simonii*, *Erigeron canadensis*, and *Erigeron annuus*.

3. The method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition according to claim 2, characterized in that, In step S1, the leaf vein structure marking method is a three-level labeling method, which includes the following steps: Level 1 labeling: Label the number, location, edges, corners, and direction of the main veins of seedlings of Canadian goldenrod and easily confused plants; Secondary labeling: Label the lateral veins of seedlings of Canadian goldenrod and easily confused plants, including the angle of attachment, number of forks, and termination position. The accuracy of the angle of attachment is required to be ±5°, the number of forks is 0-3, and the termination position is the leaf margin, leaf tip, or leaf middle. Level 3 annotation: Annotate the topological relationship between leaf veins and leaf margin serrations of seedlings of Canadian goldenrod and easily confused plants. The topological relationship includes the connection relationship between serrations and the ends of lateral veins, and the proportion of serrations connected to the ends of lateral veins.

4. The method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition according to claim 3, characterized in that, In step S2, the leaf vein structure attention module employs a multi-scale leaf vein feature extraction structure. This extraction structure is used to extract leaf image markers for seedling-stage Canadian goldenrod and easily confused plants. The extraction structure includes: Low-level feature extraction branch: Set 3×3 and 5×5 convolution kernels, receptive field 7×7 pixels, used to extract low-level texture features of main vein edges, corners and directions in the first-level annotation; Intermediate feature extraction branch: Using 7×7 and 11×11 convolutional kernels with a receptive field of 31×31 pixels, it is used to extract intermediate structural features such as lateral vein attachment angle, number of bifurcations and termination position in the secondary annotation; High-level feature extraction branch: Using 15×15 and 21×21 convolution kernels with a receptive field of 63×63 pixels or more, it is used to extract the topological relationship between leaf veins and leaf margin serrations in the third-level annotation, as well as the global features of the spatial distribution pattern of main and lateral veins.

5. The method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition according to claim 4, characterized in that, In step S2, the training employs a contrastive learning-based leaf vein feature metric learning strategy, including the following steps: Construct triplet training samples, each triplet containing an anchor sample, a positive sample, and a negative sample. Then, use the triplet loss function to force the anchor sample and the positive sample to be closer in the feature space, while pushing the anchor sample and the negative sample further apart. The anchor sample is a leaf image of Canadian goldenrod, the positive sample is a leaf image of a different plant of the same species of Canadian goldenrod, and the negative sample is a leaf image of a plant that is easily confused with another plant. The triplet loss function is defined as follows: in For Euclidean distance, Set to 1.

0.

6. The method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition according to claim 4, characterized in that, In step S3, the species classification probabilities include P(Canada Goldenrod), P(Sumatran White Grass), P(Erigeron davidii), P(Erigeron annuus), and P(background). P(Canada Goldenrod) represents the probability that the grid image is Canada Goldenrod; P(Sumatran White Grass) represents the probability that the grid image is Sumatran White Grass; P(Erigeron davidii) represents the probability that the grid image is Erigeron davidii; P(Erigeron annuus) represents the probability that the grid image is Erigeron annuus; and P(background) represents the probability that the grid image is the background image. The fine-grained confidence level ranges from 0 to 1.

7. The method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition according to claim 6, characterized in that, In step S3, the identification results include the Canadian goldenrod region, the region to be verified, and the non-target species or background region; The criteria for identifying the Canadian goldenrod region are P(Canadian goldenrod) > 0.95 and fine-grained confidence level > 0.90; The criteria for judging the region to be reviewed are 0.85 < P (Canada Goldenrod) ≤ 0.95 and fine-grained confidence level > 0.90; The criteria for determining non-target species or background areas are P(Canada goldenrod) ≤ 0.85 and fine-grained confidence level > 0.

90.

8. The method for controlling Canadian goldenrod based on fine-grained leaf vein feature recognition according to claim 4, characterized in that, In step S4, the differentiated and precise prevention and control plan includes: when the identification result is "Canadian goldenrod area", spot precise spraying is carried out at 50%-60% of the baseline amount, and the spraying agent is Canadian goldenrod fermentation liquid; no spraying instruction is generated for areas identified as non-target species or background areas; and manual verification is carried out for areas identified as areas to be verified.

Images