A Field Weed Infestation Early Warning Method Based on Weed Density Detection

Abstract

This invention relates to a field weed infestation early warning method based on weed density detection: First, the operating coordinates of a rice transplanter are obtained, and the transplanter's operating trajectory is constructed. Then, the mechanical characteristics of the transplanter's implements are used to locate the theoretical seedling coordinates. Next, the spacing of the actual seedling geographical coordinates is used to construct a theoretical seedling circle region with the theoretical seedling geographical coordinates as the origin and a preset distance as the radius. Then, images of the farmland after transplanting are collected using a drone, and a deep learning target detection model is constructed. A computational mapping model is constructed to map the theoretical seedling circle region onto the farmland image. A plant quantity statistics algorithm model is constructed to perform quantity statistics calculations on the plants detected by the target detection algorithm in the farmland image. Target detection processing and plant quantity statistics are performed on the farmland images outside the theoretical seedling circle region. A weed infestation early warning threshold is set; exceeding this threshold indicates severe weed infestation, requiring an early warning.

Description

Technical Field

[0001] This invention relates to the agricultural field, specifically to a field weed infestation early warning method based on weed density detection. Background Technology

[0002] During the growth of crops such as tea, peanuts, and rice, growers provide suitable nutrients and water to create a suitable growing environment. However, this environment also easily breeds weeds. Weeds block sunlight, steal nutrients from crops, and encroach on above-ground and underground space, leading to reduced yields and ultimately impacting economic benefits. Therefore, weeding is often necessary during crop growth, and weed control begins with weed detection. Traditional weed detection methods mainly involve manual inspection; however, this method is time-consuming, labor-intensive, inefficient, and has a high rate of missed detections. Therefore, there is an urgent need for a new, highly automated, efficient, and accurate weed detection method.

[0003] Regarding the automated detection of weeds, patent application CN115546639A discloses "A method for detecting forest weeds based on an improved YOLOv5 model," which acquires forest vegetation image data and inputs the acquired vegetation image data into an improved YOLOv5 model to obtain the detection results of weeds in the forest. Another example is patent application CN114818909A, which discloses "A method and device for detecting weeds based on crop growth characteristics." This method acquires images of various plants in the field and the growth stages of the crops to obtain a database of images to be detected. The database is then input into a trained first classifier to perform content recognition on the images, and weed images are obtained based on the content recognition results. The weed images are uploaded to the detection database and manually labeled. Based on the manual labeling results, the images are classified and stored in a first sample database and a third sample database. Finally, the weed images are input into a trained second classifier to perform feature recognition on the images. Based on the feature recognition results, crop images are removed from the weed images to obtain the weed detection results.

[0004] However, the aforementioned existing technologies have the following problems:

[0005] The forest weed detection method based on the improved YOLOv5 model uses computer vision as its core and deep learning, image processing and other technologies to identify and locate specific objects in images or videos. However, the number and types of weeds around crops are numerous and the weeds have extremely high similarity in features. There is also overlap and mutual occlusion between weeds, which makes the recognition accuracy and speed not meet expectations. In addition, it is difficult to build a sample library and label the data. The workload of labeling samples is large and the cost is high. Moreover, the labeling is prone to errors.

[0006] The weed detection method based on crop growth characteristics detects weeds based on crop growth characteristics, which to some extent avoids the problem of the difficulty in constructing a total weed sample library due to the large variety of weeds and the large differences in weed characteristics under different growth environments. However, it requires identification and labeling of different crop stages and different crop organs, which is a lot of work and the identification accuracy is not high. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a field weed infestation early warning method based on weed density detection. The field weed infestation early warning method has the advantages of wide applicability, high accuracy and simple detection process.

[0008] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0009] A field weed infestation early warning method based on weed density detection includes the following steps:

[0010] S1. Obtain the working positioning coordinates of the rice transplanter and construct the working trajectory line of the rice transplanter;

[0011] S2. Based on the mechanical characteristics of the rice transplanter, the theoretical coordinate position of the seedlings is determined.

[0012] S3. Combining the actual spacing of the seedling geographical coordinates, construct a theoretical seedling circular domain with the theoretical seedling geographical coordinates as the origin and the preset distance as the radius.

[0013] S4. Use drones to collect images of farmland after rice transplanting;

[0014] S5. Construct a deep learning target detection model for identifying and locating plants using the collected farmland images;

[0015] S6. Construct a computational mapping model to map the theoretical seedling circle domain to farmland images collected by UAV;

[0016] S7. Construct a plant quantity counting algorithm model to perform quantity counting calculations on the plants detected by the target detection algorithm in farmland images;

[0017] S8. Perform target detection processing and plant count on the farmland images outside the theoretical seedling circle area collected, and set a threshold for the number of weeds to be warned. If the threshold is exceeded, it means that the weed infestation is serious and an early warning is required.

[0018] Preferably, in step S1, an RTK antenna is installed on the top of the rice transplanter and an RTK receiver is installed inside it to record the real-time positioning coordinates of the rice transplanter during operation, and the recorded real-time positioning coordinates of the rice transplanter are connected by a smooth curve to construct the operation trajectory line of the rice transplanter.

[0019] Preferably, in step S2, the mechanical characteristics of the rice transplanter during operation are determined, wherein the mechanical characteristics are the planting spacing between seedlings; and the position coordinates of the theoretical seedlings are determined by determining the seedling spacing between two adjacent theoretical seedlings.

[0020] Preferably, in step S3, a theoretical seedling position coordinate is used as the origin, and a preset distance of no more than 20 centimeters is set as the radius to construct a seedling theoretical circle domain. All plants within this seedling theoretical circle domain are considered seedlings.

[0021] Preferably, in step S4, a drone is used to collect images of farmland after rice transplanting. The drone used should be a high-precision drone with RTK positioning function. The drone's shooting pixels and shooting height need to meet the shooting requirements and be able to record the coordinate position information of the drone when shooting farmland images.

[0022] Preferably, in step S5, the construction steps of the deep learning object detection model are as follows:

[0023] S51. Utilize drones to collect sufficient farmland images and preprocess the farmland images;

[0024] S52. Set up the Anaconda environment and download the YOLO model;

[0025] S53. Divide the collected farmland image samples into training and test sets at a ratio of 7:3, and use labeling tools to label the plants in the farmland images taken by the drone.

[0026] S54. Train the model and test the training results. When the accuracy meets the requirements, the deep learning target detection model is obtained.

[0027] Preferably, in step S6, the construction step of the computational mapping model is as follows:

[0028] By calculating the latitude and longitude range of each pixel in each farmland image taken by the drone, the operation trajectory of the rice transplanter that conforms to the latitude and longitude range of the image and the theoretical seedling circle domain are mapped onto the farmland image, so that the corresponding positions in the farmland image taken by the drone have the theoretical seedling circle domain, which corresponds to the actual situation.

[0029] Preferably, in step S7, the construction steps of the plant number counting algorithm model are as follows:

[0030] The farmland image is processed by a deep learning object detection model to detect all the plants in the farmland image. Code for counting the number of plants is added to build an algorithm model for counting the number of plants in the farmland image. This algorithm model is used to count the number of plants detected in the farmland image.

[0031] Preferably, in step S8, the latitude and longitude range of each pixel in the farmland image captured by the UAV is calculated, the working trajectory of the rice transplanter within this latitude and longitude range is found, a theoretical seedling circle domain is constructed based on the mechanical characteristics of the rice transplanter, and the theoretical seedling circle domain is mapped to the farmland image captured by the UAV through a calculation mapping model. The farmland image outside the theoretical seedling circle domain is processed by a deep learning object detection model to find all plants in the farmland image outside the theoretical seedling circle domain, and the number of detected plants is counted by a plant count algorithm. By combining the actual farmland environment and planting conditions, a warning threshold is set, and the counted number of plants is compared with the warning threshold. If the threshold is exceeded, a weed infestation warning is issued.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] (1) The field weed early warning method based on weed density detection of the present invention does not require labeling and identifying specific weeds, but identifies and locates all plants, which to a certain extent avoids the problems of low accuracy and difficulty in labeling weeds.

[0034] (2) The field weed damage early warning method based on weed density detection of the present invention combines rice transplanter operation information. By determining the theoretical seedling circle domain, when processing the farmland image captured by the UAV, only the part outside the theoretical seedling circle domain needs to be processed, which can reduce the workload of image processing and improve detection efficiency to a certain extent.

[0035] (3) The field weed damage early warning method based on weed density detection of the present invention can increase the accuracy and efficiency of detection by reducing the image processing area and reducing the detection targeting. Attached Figure Description

[0036] Figure 1 This is a diagram of the rice transplanter's operating trajectory; in the diagram, 1 (black circle) represents the rice transplanter's positioning coordinates, and 2 (transplanting trajectory line) represents the rice transplanter's operating trajectory.

[0037] Figure 2 This is a theoretical diagram of seedlings.

[0038] Figure 3 The theoretical circle domain diagram for rice seedlings.

[0039] Figure 4 This is a flowchart of the field weed infestation early warning method based on weed density detection according to the present invention.

[0040] In the diagram: 1-Positioning coordinates of the rice transplanter; 2-Trajectory line of the rice transplanting operation; 3-Theoretical seedlings (the spacing between theoretical seedlings is 40 cm); 4-Circular area of ​​theoretical seedlings (the radius of the circular area of ​​theoretical seedlings does not exceed 20 cm). Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0042] See Figures 1-4 The field weed infestation early warning method based on weed density detection of the present invention includes the following steps:

[0043] S1. Obtain the working positioning coordinates of the rice transplanter and construct the working trajectory line of the rice transplanter;

[0044] S2. Based on the mechanical characteristics of the rice transplanter, the theoretical coordinate position of the seedlings is determined.

[0045] S3. Combining the actual spacing of the seedling geographical coordinates, construct a theoretical seedling circular domain with the theoretical seedling geographical coordinates as the origin and the preset distance as the radius.

[0046] S4. Use drones to collect images of farmland after rice transplanting;

[0047] S5. Construct a deep learning target detection model for identifying and locating plants using the collected farmland images;

[0048] S6. Construct a computational mapping model to map the theoretical seedling circle domain to farmland images collected by UAV;

[0049] S7. Construct a plant quantity counting algorithm model to perform quantity counting calculations on the plants detected by the target detection algorithm in farmland images;

[0050] S8. Perform target detection processing and plant count on the farmland images outside the theoretical seedling circle area collected, and set a threshold for the number of weeds to be warned. If the threshold is exceeded, it means that the weed infestation is serious and an early warning is required.

[0051] See Figures 1-4 In step S1, an RTK antenna is installed on top of the rice transplanter and an RTK receiver is installed inside it to record the real-time positioning coordinates of the rice transplanter during operation. The recorded real-time positioning coordinates of the rice transplanter are then connected by a smooth curve to construct the rice transplanter's operating trajectory line (see [link]). Figure 1 ).

[0052] See Figures 1-4In step S2, the mechanical characteristics of the rice transplanter during operation are determined, wherein the mechanical characteristics are the planting spacing between seedlings. In this embodiment, the planting spacing is 40 centimeters. The position coordinates of the theoretical seedlings are determined by determining the spacing between two adjacent theoretical seedlings; that is, one theoretical seedling is determined every 40 centimeters along the rice transplanter's operating trajectory (see...). Figure 2 ).

[0053] See Figures 1-4 In step S3, the theoretical seedling position coordinates are used as the origin, and a preset distance of no more than 20 centimeters is set as the radius to construct a theoretical seedling circle domain. All plants within this theoretical seedling circle domain are considered seedlings.

[0054] See Figures 1-4 In step S4, a drone is used to collect images of farmland after rice transplanting. The drone used should be a high-precision drone with RTK positioning function. The drone's shooting pixels and shooting height need to meet the shooting requirements and be able to record the drone's coordinate position information when shooting farmland images.

[0055] See Figures 1-4 In step S5, the construction steps of the deep learning object detection model are as follows:

[0056] S51. Utilize drones to collect sufficient farmland images and preprocess the farmland images;

[0057] S52. Set up the Anaconda environment and download the YOLO model;

[0058] S53. Divide the collected farmland image samples into training and test sets at a ratio of 7:3, and use labeling tools to label the plants in the farmland images taken by the drone.

[0059] S54. Train the model and test the training results. When the accuracy meets the requirements, the deep learning target detection model is obtained.

[0060] See Figures 1-4 In step S6, the construction steps of the computational mapping model are as follows:

[0061] By calculating the latitude and longitude range of each pixel in each farmland image taken by the drone, the operation trajectory of the rice transplanter that conforms to the latitude and longitude range of the image and the theoretical seedling circle domain are mapped to the farmland image. This ensures that the corresponding positions in the farmland image taken by the drone have the theoretical seedling circle domain, which corresponds to the actual situation. In other words, by constructing a corresponding computational mapping model, the theoretical seedling circle domain is mapped to the farmland image taken by the drone, ensuring that the corresponding positions in the farmland image taken by the drone have the theoretical seedling circle domain, which corresponds to the actual situation.

[0062] See Figures 1-4 In step S7, the construction steps of the plant number counting algorithm model are as follows:

[0063] The farmland image is processed by a deep learning object detection model to detect all the plants in the farmland image. Code for counting the number of plants is added to build an algorithm model for counting the number of plants in the farmland image. This algorithm model is used to count the number of plants detected in the farmland image.

[0064] See Figures 1-4 In step S8, the latitude and longitude range of each pixel in the farmland image captured by the UAV is calculated, the working trajectory of the rice transplanter within this latitude and longitude range is found, and a theoretical seedling circle domain is constructed based on the mechanical characteristics of the rice transplanter. The theoretical seedling circle domain is mapped to the farmland image captured by the UAV through a calculation mapping model. The farmland image outside the theoretical seedling circle domain is processed by a deep learning object detection model to find all plants in the farmland image outside the theoretical seedling circle domain. The number of detected plants is counted by a plant count algorithm. By combining the actual farmland environment and planting conditions, a warning threshold is set, and the counted number of plants is compared with the warning threshold. If the threshold is exceeded, a weed infestation warning is issued.

[0065] Finally, the field weed infestation early warning method based on weed density detection of the present invention has the advantages of higher detection accuracy and higher efficiency. This is because the field weed infestation early warning method based on weed density detection of the present invention does not only focus on improving the accuracy of weed detection, but also innovatively detects all plants for weed infestation early warning. The field weed infestation early warning method based on weed density detection of the present invention avoids distinguishing between weeds and seedlings, and selects the broad category of weeds and seedlings: plants, which greatly helps to improve the overall detection accuracy. In addition, the field weed infestation early warning method based on weed density detection of the present invention introduces rice transplanter operation information. The theoretical seedling circle domain is constructed using the rice transplanter operation information. In actual image processing, the area within the theoretical seedling circle domain is removed, which can reduce the detection area to a certain extent and improve the detection efficiency.

[0066] The above are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above content. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A field weed infestation early warning method based on weed density detection, characterized in that, Includes the following steps: S1. Obtain the working positioning coordinates of the rice transplanter and construct the working trajectory line of the rice transplanter; S2. Based on the mechanical characteristics of the rice transplanter, the theoretical coordinate position of the seedlings is determined. S3. Combining the actual spacing of the seedling geographical coordinates, construct a theoretical seedling circular domain with the theoretical seedling geographical coordinates as the origin and the preset distance as the radius. S4. Use drones to collect images of farmland after rice transplanting; S5. Construct a deep learning target detection model for identifying and locating plants using the collected farmland images; S6. Construct a computational mapping model to map the theoretical seedling circle domain to farmland images collected by UAV; S7. Construct a plant quantity counting algorithm model to perform quantity counting calculations on the plants detected by the target detection algorithm in farmland images; S8. Perform target detection processing and plant count on the farmland images outside the theoretical seedling circle area collected, and set a threshold for the number of weeds to be warned. If the threshold is exceeded, it means that the weed infestation is serious and an early warning is required.

2. The field weed infestation early warning method based on weed density detection according to claim 1, characterized in that, In step S1, by installing an RTK antenna on the top of the rice transplanter and an RTK receiver inside it, the real-time positioning coordinates of the rice transplanter during operation are recorded, and the recorded real-time positioning coordinates of the rice transplanter are connected by a smooth curve to construct the operation trajectory line of the rice transplanter.

3. The field weed infestation early warning method based on weed density detection according to claim 2, characterized in that, In step S2, the mechanical characteristics of the rice transplanter during operation are determined, wherein the mechanical characteristics are the planting spacing between seedlings; the position coordinates of the theoretical seedlings are determined by determining the seedling spacing between two adjacent theoretical seedlings.

4. The field weed infestation early warning method based on weed density detection according to claim 3, characterized in that, In step S3, the theoretical seedling position coordinates are used as the origin, and a preset distance of no more than 20 centimeters is set as the radius to construct a theoretical seedling circle domain. All plants within this theoretical seedling circle domain are considered seedlings.

5. The field weed infestation early warning method based on weed density detection according to claim 4, characterized in that, In step S4, a drone is used to collect images of farmland after rice transplanting. The drone used should be a high-precision drone with RTK positioning function. The drone's shooting pixels and shooting height need to meet the shooting requirements and be able to record the drone's coordinate position information when shooting farmland images.

6. The field weed infestation early warning method based on weed density detection according to claim 5, characterized in that, In step S5, the construction steps of the deep learning object detection model are as follows: S51. Use drones to collect sufficient farmland images and preprocess the farmland images; S52. Set up the Anaconda environment and download the YOLO model; S53. Divide the collected farmland image samples into training and test sets at a ratio of 7:3, and use labeling tools to label the plants in the farmland images taken by the drone. S54. Train the model and test the training results. When the accuracy meets the requirements, the deep learning target detection model is obtained.

7. The field weed infestation early warning method based on weed density detection according to claim 6, characterized in that, In step S6, the construction steps of the computational mapping model are as follows: By calculating the latitude and longitude range of each pixel in each farmland image taken by the drone, the operation trajectory of the rice transplanter that conforms to the latitude and longitude range of the image and the theoretical seedling circle domain are mapped onto the farmland image, so that the corresponding positions in the farmland image taken by the drone have the theoretical seedling circle domain, which corresponds to the actual situation.

8. The field weed infestation early warning method based on weed density detection according to claim 7, characterized in that, In step S7, the construction steps of the plant number counting algorithm model are as follows: The farmland image is processed by a deep learning object detection model to detect all the plants in the farmland image. Code for counting the number of plants is added to build an algorithm model for counting the number of plants in the farmland image. This algorithm model is used to count the number of plants detected in the farmland image.

9. The field weed infestation early warning method based on weed density detection according to claim 8, characterized in that, In step S8, the latitude and longitude range of each pixel in the farmland image captured by the UAV is calculated, and the working trajectory of the rice transplanter within this latitude and longitude range is found. A theoretical seedling circle domain is constructed based on the mechanical characteristics of the rice transplanter's implements. The theoretical seedling circle domain is mapped onto the farmland image captured by the UAV using a calculation mapping model. The farmland image outside the theoretical seedling circle domain is processed using a deep learning object detection model to find all plants in the farmland image outside the theoretical seedling circle domain. The number of detected plants is counted using a plant count algorithm. By combining the actual farmland environment and planting conditions, a warning threshold is set. The counted number of plants is compared with the warning threshold. If the threshold is exceeded, a weed infestation warning is issued.

Images