Rice canopy disease identification and positioning method based on unmanned aerial vehicle

Abstract

The invention relates to a rice canopy disease identification and positioning method based on an unmanned aerial vehicle, and belongs to the field of image processing, and the method comprises the steps: a ground station plans a flight path of the unmanned aerial vehicle according to the shape of a rice field, and transmits the flight path to an unmanned aerial vehicle control device; the unmanned aerial vehicle flies according to the planned flight path and carries out image data acquisition on rice in real time, RTK positioning data is read and stored through an airborne computer, and the RTK positioning data and the image data are bound; and the unmanned aerial vehicle uploads the RTK positioning data and the image data to a back-end server, the server inputs the image data as input into a trained disease recognition model for recognition, and generates a disease distribution diagram by using the RTK positioning data. Compared with a multispectral unmanned aerial vehicle detection method, the method has the advantages that the cost is reduced, an original yov8-seg network structure is optimized, the detection precision is improved, and low-cost and large-area rapid detection can be realized.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a method for identifying and locating rice canopy diseases based on an unmanned aerial vehicle. Background Art

[0002] Rice disease is one of the important factors affecting rice yield and quality. The occurrence of diseases often brings huge economic losses to agricultural production, which not only affects farmers' income, but also poses a threat to national food security. At present, the detection methods for rice diseases are mainly through manual detection and hyperspectral drone detection. However, the efficiency and detection accuracy of manual detection are low, which cannot meet the requirements of large-scale rice disease detection. On the other hand, although the detection effect of hyperspectral drones is significant, their high purchase cost makes it difficult for ordinary farmers to afford it. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for identifying and locating rice canopy diseases based on an unmanned aerial vehicle, thereby solving the deficiencies of the prior art.

[0004] The object of the present invention is achieved through the following technical solution: a method for identifying and locating rice canopy diseases based on an unmanned aerial vehicle, the method comprising: Step 1: The ground station plans the flight path of the drone according to the shape of the rice field and transmits the flight path to the drone control device; Step 2: The drone flies according to the planned flight path and collects image data of the rice in real time. The onboard computer reads and saves the RTK positioning data, and binds the RTK positioning data with the image data. Step 3: The drone uploads the RTK positioning data and image data to the backend server. The server uses the image data as input and inputs it into the trained disease recognition model for identification, and uses the RTK positioning data to generate a disease distribution map.

[0005] The disease recognition model includes an input end, a backbone feature extraction network, a neck fusion network and a detection head; The input end is responsible for introducing the image data into the disease recognition model; The backbone feature extraction network initially extracts image features through two initial convolutional layers, then enriches feature representation using residual blocks, and captures feature information of different scales through three rounds of alternating combinations of convolutional layers and residual blocks. It then outputs a feature vector of fixed size through a spatial pyramid pooling layer and feeds it into the MSCA attention mechanism to capture key feature information. The neck fusion network realizes the efficient integration of the features extracted by the backbone feature extraction network. The feature vector output by the MSCA attention mechanism is first processed by upsampling to ensure the consistency of the sizes of each part during feature fusion. Then, upsampling and two convolutional layers, feature fusion and residual block are applied again to promote the deep fusion of high-dimensional and low-dimensional features, and connected to the detection head in the C2f-DWR module. The detection head is used to predict the location and category information of the target.

[0006] The disease recognition model includes three detection heads, which respectively detect and segment targets larger than 8×8 pixels, larger than 16×16 pixels, and larger than 32×32 pixels.

[0007] The detection head includes two parallel links; the first link includes two convolutional layers, a two-dimensional convolutional layer and a bounding box loss module, which is used to achieve accurate prediction of the bounding box; The second link includes two convolutional layers, a two-dimensional convolutional layer and a classification loss module, which is used to realize the category determination of the target.

[0008] After refining the features through two convolutional layers, the first link uses a two-dimensional convolutional layer to output a segmentation probability map, and optimizes the segmentation accuracy through a segmentation loss module; The second link extracts auxiliary feature maps through two convolutional layers and one two-dimensional convolutional layer. The output auxiliary feature maps are optimized by the loss function in the classification loss module to provide supplementary information for the segmentation task.

[0009] The training of the disease recognition model includes the following contents: A1. Construct a dataset containing five diseases, namely, wilt, rice blast, leaf spot, brown spot and leafhopper, by using web crawlers. A2. Label the diseased areas of the dataset using LabelImg, and randomly divide the dataset into training set, test set, and validation set in a ratio of 7:1.5:1.5; A3. Optimize the original yolov8-seg network structure to obtain an optimized disease recognition model; A4. Use the labeled data to train the disease recognition model; A5. Verify the trained disease recognition model.

[0010] The optimization of the original yolov8-seg network structure includes: replacing the Conv module after the C2f module of the backbone feature extraction network part in the original yolov8-seg network structure with a HWD_Adown module, replacing the C2f module of the neck fusion network part with a C2f_DWRseg module, and introducing the MSCA attention mechanism and the Wise-IoU loss function.

[0011] The method of generating a disease distribution map using RTK positioning data includes: using the RTK positioning data of the upper left corner and the lower right corner of the rice field, and the RTK positioning data bound to the disease image, to calculate the relative position of the disease and the rice field, and mark it on the rice field image.

[0012] The method of calculating the relative position of the disease and the paddy field and marking the paddy field image includes: Based on the longitude of the upper left corner of the field and the longitude of the lower right corner , disease longitude And the pixel coordinates of the upper left corner of the field image and the pixel coordinates of the lower right corner Calculate the pixel points of the disease in the field image x The coordinates are ; Similarly, according to the latitude of the upper left corner of the field and the latitude of the lower right corner , disease latitude And the pixel coordinates of the upper left corner of the field image and the pixel coordinates of the lower right corner Calculate the pixel points of the disease in the field image y The coordinates are ; according to and Marking disease locations on rice field images.

[0013] The present invention has the following advantages: a method for identifying and locating rice canopy diseases based on drones, which reduces costs compared to multi-spectral drone detection methods, improves efficiency compared to manual detection methods, optimizes the original yolov8-seg network structure, improves detection accuracy, and enables low-cost, large-area, and rapid detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0016] like Figure 1 As shown, the present invention specifically relates to a method for identifying and locating rice canopy diseases based on drones, which realizes the identification and locating of rice canopy diseases by drones, and specifically includes the following contents: Step 1: Use the ground station to manually plan the flight path of the drone according to the shape of the rice field, and transmit the planned path to the drone control device through the data transmission module; Step 2: Take off the drone, and the drone flies according to the planned path. During the flight, the drone controls the image acquisition module to collect images of rice. The onboard computer saves the image information. At the same time, the onboard computer saves and reads the RTK positioning data, and binds the positioning data with the image data. Step 3: After the flight is completed, the drone automatically returns and lands, and uploads the image information and positioning information to the server; Step 4: After receiving the image information, the server uses it as input and inputs it into the trained disease recognition model. After completing the recognition, the positioning information is used to generate a disease distribution map.

[0017] Furthermore, the training of the disease recognition model includes the following: A1. Use web crawlers to build a dataset of five diseases: Fusarium wilt, rice blast, leaf spot, brown spot, and leafhopper, with 1,000 images for each disease. A2. Label the diseased areas of the dataset using LabelImg, and randomly divide the dataset into training set, test set, and validation set in a ratio of 7:1.5:1.5; A3. Optimize the original yolov8-seg network structure to obtain an optimized disease recognition model; A4. Use the labeled data to train the disease recognition model; A5. Verify the trained disease recognition model.

[0018] Furthermore, the disease recognition model is obtained by optimizing the original yolov8-seg network structure, which includes the input end, the backbone feature extraction network, the neck fusion network and the detection head; Among them, the image enters from the input end, which is responsible for introducing the image data into the disease recognition model; The backbone feature extraction network initially extracts image features through the initial two convolutional layers, then uses residual blocks to enrich feature representation, and undergoes three rounds of alternating combinations of convolutional layers and residual blocks to capture feature information of different scales. It then outputs a fixed-size feature vector through a spatial pyramid pooling layer and feeds it into the MSCA attention mechanism to capture key feature information. The neck fusion network realizes the efficient integration of the features extracted by the backbone feature extraction network. The feature vector output by the MSCA attention mechanism is first processed by upsampling to ensure the consistency of the sizes of each part during feature fusion. Then, upsampling and two convolutional layers, feature fusion and residual blocks are applied again to promote the deep integration of high-dimensional and low-dimensional features, and connected to the detection head in the C2f-DWR module. The detection head is the final output stage of the model and is used to predict the location and category information of the target.

[0019] Furthermore, the disease recognition model includes three detection heads, which respectively detect and segment targets larger than 8×8 pixels, larger than 16×16 pixels, and larger than 32×32 pixels.

[0020] Each detection head includes two parallel links. The first link includes two convolutional layers, a two-dimensional convolutional layer and a bounding box loss module, which are used to achieve accurate prediction of the bounding box. The second link also includes two convolutional layers, a two-dimensional convolutional layer and a classification loss module, which are used to achieve target category determination.

[0021] The two links work independently, respectively calculating and optimizing the bounding box loss and classification loss, and jointly improving the detection performance; the first link focuses on the precise segmentation of foreground and background pixels, and after refining the features through two convolutional layers, uses a two-dimensional convolutional layer to output a segmentation probability map, and optimizes the segmentation accuracy through the segmentation loss module; the second link extracts auxiliary feature maps through two convolutional layers and a two-dimensional convolutional layer, and the output auxiliary feature maps are optimized through the loss function in the classification loss module to provide supplementary information for the segmentation task. The two links are processed in parallel, optimizing their respective losses, and jointly improving the segmentation performance, ensuring accurate and efficient instance segmentation in complex scenarios.

[0022] Furthermore, the original yolov8-seg network structure is optimized, including replacing the Conv module after the C2f module in the backbone feature extraction network part of the original yolov8-seg network structure with the HWD_Adown module, replacing the C2f module in the neck fusion network part with the C2f_DWRseg module, and introducing the MSCA attention mechanism and Wise-IoU loss function.

[0023] Furthermore, the specific method of using positioning information to generate a disease distribution map is: using the RTK positioning data of the upper left corner and the lower right corner of the rice field, as well as the RTK positioning data bound to the disease image, the relative position of the disease and the rice field is calculated, and marked on the rice field image.

[0024] Furthermore, according to the longitude of the upper left corner of the field and the longitude of the lower right corner , disease longitude And the pixel coordinates of the upper left corner of the field image and the pixel coordinates of the lower right corner Calculate the pixel points of the disease in the field image x The coordinates are ; Similarly, according to the latitude of the upper left corner of the field and the latitude of the lower right corner , disease latitude And the pixel coordinates of the upper left corner of the field image and the pixel coordinates of the lower right corner Calculate the pixel points of the disease in the field image y The coordinates are ; according to and Marking disease locations on rice field images.

[0025] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for identifying and locating rice canopy diseases based on drones, characterized in that: The method comprises: Step 1: The ground station plans the flight path of the drone according to the shape of the rice field and transmits the flight path to the drone control device; Step 2: The drone flies according to the planned flight path and collects image data of the rice in real time. The onboard computer reads and saves the RTK positioning data, and binds the RTK positioning data with the image data. Step 3: The drone uploads the RTK positioning data and image data to the backend server. The server uses the image data as input and inputs it into the trained disease recognition model for identification, and uses the RTK positioning data to generate a disease distribution map.

2. The method for identifying and locating rice canopy diseases based on an unmanned aerial vehicle according to claim 1, characterized in that: The disease recognition model includes an input end, a backbone feature extraction network, a neck fusion network and a detection head; The input end is responsible for introducing the image data into the disease recognition model; The backbone feature extraction network initially extracts image features through two initial convolutional layers, then enriches feature representation using residual blocks, and captures feature information of different scales through three rounds of alternating combinations of convolutional layers and residual blocks. It then outputs a feature vector of fixed size through a spatial pyramid pooling layer and feeds it into the MSCA attention mechanism to capture key feature information. The neck fusion network realizes the efficient integration of the features extracted by the backbone feature extraction network. The feature vector output by the MSCA attention mechanism is first processed by upsampling to ensure the consistency of the sizes of each part during feature fusion. Then, upsampling and two convolutional layers, feature fusion and residual block are applied again to promote the deep fusion of high-dimensional and low-dimensional features, and connected to the detection head in the C2f-DWR module. The detection head is used to predict the location and category information of the target.

3. The method for identifying and locating rice canopy diseases based on an unmanned aerial vehicle according to claim 2, characterized in that: The disease recognition model includes three detection heads, which respectively detect and segment targets larger than 8×8 pixels, larger than 16×16 pixels, and larger than 32×32 pixels.

4. The method for identifying and locating rice canopy diseases based on drones according to claim 3, characterized in that: The detection head includes two parallel links; the first link includes two convolutional layers, a two-dimensional convolutional layer and a bounding box loss module, which is used to achieve accurate prediction of the bounding box; The second link includes two convolutional layers, a two-dimensional convolutional layer and a classification loss module, which is used to realize the category determination of the target.

5. The method for identifying and locating rice canopy diseases based on drones according to claim 4, characterized in that: After refining the features through two convolutional layers, the first link uses a two-dimensional convolutional layer to output a segmentation probability map, and optimizes the segmentation accuracy through a segmentation loss module; The second link extracts auxiliary feature maps through two convolutional layers and one two-dimensional convolutional layer. The output auxiliary feature maps are optimized by the loss function in the classification loss module to provide supplementary information for the segmentation task.

6. The method for identifying and locating rice canopy diseases based on drones according to claim 1, characterized in that: The training of the disease recognition model includes the following contents: A1. Construct a dataset containing five diseases, namely, wilt, rice blast, leaf spot, brown spot and leafhopper, by using web crawlers. A2. Label the diseased areas of the dataset using LabelImg, and randomly divide the dataset into training set, test set, and validation set in a ratio of 7:1.5:1.5; A3. Optimize the original yolov8-seg network structure to obtain an optimized disease recognition model; A4. Use the labeled data to train the disease recognition model; A5. Verify the trained disease recognition model.

7. The method for identifying and locating rice canopy diseases based on drones according to claim 6, characterized in that: The optimization of the original yolov8-seg network structure includes: replacing the Conv module after the C2f module of the backbone feature extraction network part in the original yolov8-seg network structure with a HWD_Adown module, replacing the C2f module of the neck fusion network part with a C2f_DWRseg module, and introducing the MSCA attention mechanism and the Wise-IoU loss function.

8. The method for identifying and locating rice canopy diseases based on drones according to claim 1, characterized in that: The method of generating a disease distribution map using RTK positioning data includes: using the RTK positioning data of the upper left corner and the lower right corner of the rice field, and the RTK positioning data bound to the disease image, to calculate the relative position of the disease and the rice field, and mark it on the rice field image.

9. The method for identifying and locating rice canopy diseases based on drones according to claim 8, characterized in that: The method of calculating the relative position of the disease and the paddy field and marking the paddy field image includes: Based on the longitude of the upper left corner of the field and the longitude of the lower right corner , disease longitude And the pixel coordinates of the upper left corner of the field image and the pixel coordinates of the lower right corner Calculate the pixel points of the disease in the field image x The coordinates are ; Similarly, according to the latitude of the upper left corner of the field and the latitude of the lower right corner , disease latitude And the pixel coordinates of the upper left corner of the field image and the pixel coordinates of the lower right corner Calculate the pixel points of the disease in the field image y The coordinates are ; according to and Marking disease locations on rice field images.

Images