Toxic plant euphorbia fischeriana identification and spatial positioning method based on improved YOLOv8

Abstract

The invention relates to the technical field of grassland ecological remote sensing intelligent detection, in particular to a toxic plant euphorbia fischeriana identification and spatial positioning method based on YOLOv8. The method comprises the following steps: optimizing a trunk feature extraction network, a neck network and a convolution module of a YOLOv8 network model to form an improved RFC-YOLOv8 network model; compared with the prior art, the method not only improves the detection precision of YOLOv8, but also enables the network model to be lighter, and improves the reasoning speed, thereby achieving the quick and accurate recognition and positioning of euphorbia fischeriana in a natural grassland, and finally achieving the spatial distribution drawing of euphorbia fischeriana.

Description

Technical Field

[0001] The present invention relates to the technical field of grassland ecological remote sensing intelligent detection, and in particular to a YOLOv8-based method for identifying and spatially locating a poisonous plant Euphorbia chamaejasme. Background Art

[0002] Euphorbia fischeriana Steud. is a perennial herb that is widely distributed in the northeastern part of China's grasslands. It is not eaten by livestock because of its toxicity. Euphorbia fischeriana Steud. has strong resistance to adversity and strong reproductive capacity. Once it invades natural grasslands, it can form a dominant population in a short period of time, which can inhibit the growth of other high-quality forage grasses, reduce forage quality and grassland carrying capacity. Therefore, timely and accurate monitoring of the number and spatial distribution pattern of poisonous plants in natural grasslands is of great significance to grassland ecological protection and production. At present, the investigation of poisonous plants in natural grasslands mainly relies on manual field monitoring by professionals. Due to the vast grassland area and complex terrain, traditional methods have limitations in coverage and timeliness.

[0003] In recent years, the rapid development of UAV remote sensing has provided new solutions for grassland monitoring. With its flexibility, high spatial resolution and other advantages, UAVs have been widely used in grassland vegetation monitoring. At the same time, deep learning and computer vision technology have brought more possibilities for vegetation monitoring, especially in target detection based on deep learning. Due to its strong learning ability and wide adaptability, it has been widely studied and developed in the field of plant monitoring. Among them, the YOLO series of models has become an important tool in the field of target detection due to its efficient real-time processing capabilities and accuracy.

[0004] Existing research has improved the accuracy and efficiency of herbaceous plant detection by constructing dedicated data sets and improving the YOLO model. However, monitoring of poisonous plants in natural grasslands still faces challenges. First, the accuracy of target recognition in the complex background of natural grasslands needs to be improved; second, how to accurately obtain the geographic coordinate information of poisonous plants in a larger range, thereby providing data support for natural grassland ecological monitoring and effective management. Summary of the invention

[0005] Based on the above purpose, the purpose of the present invention is to overcome the shortcomings of the prior art and propose a method for identifying and spatially locating the poisonous plant Euphorbia chamaejasme based on improved YOLOv8, comprising the following steps:

[0006] S1. Dataset preparation: Use drones to collect images of the target Euphorbia chamaejasme growth area, perform image preprocessing and period annotation as the original dataset, of which 80% is the training set, 10% is the validation set, and 10% is the test set;

[0007] S2. Extract image features to optimize the YOLOv8 network model: According to the performance characteristics of Euphorbia chamaejasme at different stages, the backbone feature extraction network, neck network and convolution module of the original YOLOv8 network model are optimized to form an improved RFC-YOLOv8 network model;

[0008] S3. Model training and result prediction: Use the training set in S1 to train the constructed RFC-YOLOv8 network model, obtain the optimal network model for detecting Euphorbia chamaejasme, and output the RFC-YOLOv8 network recognition drone jpg format image of the training set;

[0009] S4. Spatial positioning and accuracy evaluation: The RFC-YOLOv8 network of the training set recognizes the UAV jpg format image, matches the pixel coordinates of the recognition frame with the tif format image with geographic coordinates, obtains the actual geographic location information of the wolfberry and the location information of the wolfberry in the UAV image, and performs accuracy evaluation based on the ground survey points;

[0010] S5. Generate a schematic diagram of the spatial distribution pattern of Stellera chamaejasme in natural grasslands to detect the spatiotemporal distribution of Stellera chamaejasme.

[0011] Furthermore, the image preprocessing and period labeling in S1 specifically includes splicing and sliding cropping of the collected images, the size of the cropped images is 640×640 pixels, and the cropped images are labeled to divide the wolfberry into three growth periods, namely, early, middle and late stages.

[0012] Furthermore, the optimization of the backbone feature extraction network described in S2 is as follows: using the RevCol structure, the input features are divided into two groups, and the Forward and Inverse operations are implemented in each Block. The formula is as follows:

[0013] Forward:x t =F t (x t-1 ,x t-2 ,…,x t-m+1 )+γx t-m

[0014] Inverse:x t-m =γ -1 [x t-1 ,x t-2 ,…,x t-m+1 ]

[0015] Here, m represents the recursive order, allowing tensors of different shapes to be used to represent features at different semantic levels or resolutions.

[0016] Furthermore, the neck network described in S2 is optimized as follows: the focused diffusion pyramid network is used as the neck structure of YOLOv8, and the FocusFeature is used as the core part of the focused diffusion pyramid network.

[0017] Furthermore, in the YOLOv8 network model optimization described in S2, the C3Ghost module is used to replace the C2f module in the original YOLOv8.

[0018] Furthermore, S5 describes generating a schematic diagram of the spatial distribution pattern of wolfsbane in natural grasslands as follows: using 10 meters × 10 meters as a statistical unit, classifying the wolfsbane Euphorbia identified in the drone image, setting every 5 wolfsbane Euphorbia as one class, and generating a schematic diagram of the spatial distribution pattern of wolfsbane in natural grasslands based on the classification information.

[0019] The beneficial effects of the present application are as follows: compared with the prior art, the present application improves YOLOv8, not only improving the detection accuracy of YOLOv8, but also improving its backbone network architecture, making the network model more lightweight, while improving the reasoning speed, thereby achieving fast and accurate identification and positioning of wolfsbane in natural grasslands, and ultimately realizing the spatial distribution mapping of wolfsbane. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the method for identifying and locating Euphorbia chamaejasme based on improved YOLOv8 of the present invention;

[0021] Figure 2 These are the field sampling pictures of Euphorbia chamaejasme at different periods and the corresponding drone images;

[0022] Figure 3 This is a schematic diagram of image enhancement of Euphorbia chamaejasme;

[0023] Figure 4 A structural diagram of a Euphorbia chamaejasme detection model based on improved YOLOv8 provided by the present invention;

[0024] Figure 5 Schematic diagram of the reversible module in the Revcol structure;

[0025] Figure 6 This is a schematic diagram of the focused diffusion pyramid network module proposed by the present invention;

[0026] Figure 7 Matching map for identification and geolocation of Euphorbia chamaejasme;

[0027] Figure 8 This is a schematic diagram of the spatial distribution pattern of Euphorbia chamaejasme in natural grasslands generated by the present invention. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with a specific implementation case. This implementation case is implemented based on the technology of the present invention, and a detailed implementation method and specific operation process are now given to illustrate that the present invention is creative, but the protection scope of the present invention is not limited to the following examples.

[0029] Taking the identification and spatial location of the poisonous plant Euphorbia chamaejasme in the natural grassland research area of ​​Hailar District, Hulunbuir City, Inner Mongolia Autonomous Region, China as an example, the method flow chart is as follows Figure 1 As shown, specifically:

[0030] S1. Dataset creation: Use drones to collect images of the growth area of ​​the target Euphorbia chamaejasme, perform image preprocessing and period annotation as the original dataset, of which 80% is the training set, 10% is the validation set, and 10% is the test set.

[0031] In this embodiment, a DJI Mavic 3M drone was used to photograph the target area. The shooting height was set to 30 meters and the ground resolution was 1.3 cm / pixel. The shooting time was: June 23, 2023 (the aboveground biomass of Euphorbia chamaejasme was relatively higher than that of other grassland plants, in the early growth stage) and July 15 (the leaves of Euphorbia chamaejasme turned red, and the aboveground parts of a few plants began to dry up, in the middle growth stage). The shooting results are as follows: Figure 2 (Field sampling pictures of Euphorbia chamaejasme at different periods and corresponding drone images) Figure 2 (a) is the ground image taken at different growth stages. Figure 2 (b) Images taken by drone at different times.

[0032] The collected images were preprocessed by sliding the stitched images, and the size of the cropped images was 640 × 640 pixels. The cropped images were annotated. At the same time, in order to simulate the image of Wolfberry in the natural environment and improve the robustness of the detection model in the recognition process, this study performed data enhancement on the image of Wolfberry by rotating (horizontally and vertically), brightness transformation and Gaussian blur processing (such as Figure 3 As shown in the figure, it is a schematic diagram of image enhancement of Euphorbia chamaejasme, where (a) is the original image, (b) is the rotated image, (c) is the Gaussian blur processing (Gaussian noise) image; (d) is the light and dark change processing image), which constitutes the original data set for training the improved YOLOv8 network model, of which 80% is the training set, 10% is the validation set, and 10% is the test set.

[0033] S2.YOLOv8 network model optimization: According to the performance characteristics of Euphorbia chamaejasme at different stages, the backbone feature extraction network, neck network and convolution module of the original YOLOv8 network model were optimized to form an improved RFC-YOLOv8 network model.

[0034] In this embodiment, if Figure 4 As shown in (Structure diagram of wolfsbane detection model based on improved YOLOv8), the optimization is divided into the optimization of the backbone feature extraction network, the neck network and the convolution module. Specifically, the optimization of the backbone feature extraction network is as follows: the input features are divided into two groups, and the Forward and Inverse operations are implemented in each Block. The formula is as follows:

[0035] Forward:x t =F t (x t-1 ,x t-2 ,…,x t-m+1 )+γx t-m

[0036] Inverse:x t-m =γ -1 [x t-1 ,x t-2 ,…,x t-m+1 ]

[0037] Here, m represents the recursive order, allowing tensors of different shapes to be used to represent features at different semantic levels or resolutions.

[0038] In this embodiment, the input feature is the result of feature extraction of the image through a neural network, which is an image feature map. Figure 5 As shown in FIG. 1 (a schematic diagram of a reversible module in a Revcol structure), in this application, after the image collected in S1 is subjected to feature extraction by a neural network, it passes through a reversible module in Revcol. The Revcol module first divides the input feature X into two groups X0 and X1. For the subsequent Block, such as the tth Block, it uses the output X of the two previous blocks as t-1 and X t-2 As input, generate output X t , the formula is as follows:

[0039] Forward:x t =F t (x t-1 ,x t-2 ,…,x t-m+1 )+γx t-m

[0040] Inverse:x t-m =γ -1 [x t-1 ,x t-2 ,…,x t-m+1 ]

[0041] Here, m represents the recursive order, allowing tensors of different shapes to be used to represent features of different semantic levels or resolutions. t represents any nonlinear operation on the residual function, and γ represents a simple reversible operation.

[0042] The neck network is optimized as follows: Figure 6 As shown in the (Schematic diagram of the proposed Focus Diffusion Pyramid Network module), the Focus Diffusion Pyramid Network (FDPN) is used as the neck structure of YOLOv8, where FocusFeature is the core component of FDPN, focusing on extracting and focusing on multi-scale features. The specific process is as follows:

[0043] (1) High-resolution feature maps are downsampled through a custom ADown module to ensure the preservation of important information and reduce the computational burden;

[0044] (2) One-dimensional convolution is used to adjust the number of channels in the medium-resolution feature map to ensure the consistency and integration of features;

[0045] (3) The low-resolution feature map is converted into a high-resolution feature map through upsampling and one-dimensional convolution to enhance the extraction and retention of detail information;

[0046] (4) The strategy of introducing the multi-core perception network (PKINet) module extracts features of different scales through non-expanded multi-scale convolution kernels, captures local context information, and connects the three-scale feature maps in the channel dimension to form a comprehensive feature map, thereby improving the overall detection effect.

[0047] The convolution module is optimized as follows: the C3Ghost module is used to replace the C2f module in the original YOLOv8 to improve the efficiency and practicality of the model.

[0048] S3. Model training and result prediction: Use the training set in S1 to train the constructed RFC-YOLOv8 network model, obtain the optimal network model for detecting Euphorbia chamaejasme, and output the RFC-YOLOv8 network recognition drone jpg format image of the training set as the prediction result of the model training;

[0049] In a further implementation, the method of obtaining the optimal network model for detecting wolfberry is specifically: using the training set in S1 to train the constructed RFC-YOLOv8 network model, obtaining various parameter indicators of the model, and using the validation set to verify and optimize the trained model, and finally evaluating the accuracy of the test results to obtain the optimal network model for detecting wolfberry; the evaluation indicators include precision, recall, mAP, GFLOPs, Parameters and Euclidean distance.

[0050] S4. Spatial positioning and accuracy evaluation: The RFC-YOLOv8 network of the training set recognizes the jpg format image of the drone, matches the pixel coordinates of the recognition frame with the tif format image with geographic coordinates, obtains the actual geographic location information of the wolfsbane euphorbia and the location information of the wolfsbane euphorbia in the drone image, and performs accuracy evaluation based on the ground survey points; Figure 7 (Euphorbia chamaejasme identification and geographic location matching map) as shown.

[0051] S5. Generate a schematic diagram of the spatial distribution pattern of Euphorbia chamaejasme in natural grasslands.

[0052] In a further implementation, S5 is specifically: using 10 meters × 10 meters as a statistical unit, classifying the wolfberry trees identified in the drone image, setting every 5 wolfberry trees as one category, and generating a schematic diagram of the spatial distribution pattern of wolfberry trees in natural grasslands based on the classification information.

[0053] Figure 8 This is a schematic diagram of the spatial distribution pattern of Euphorbia chamaejasme in natural grasslands of a certain region generated according to the method of this application (wherein, Figure 8 (a) is a schematic diagram of the spatial distribution of Euphorbia chamaejasme in the region in June. Figure 8 (b) is a schematic diagram of the spatial distribution of Euphorbia chamaejasme in the same area in July, as shown in the figure: Figure 8 (a) and Figure 8 (b) Using the same color coding method, the color depth intuitively reflects the changes in the number and distribution density of Stellera chamaejasme in the study area. The comparison between June and July allows users to intuitively observe the differences in the spatial distribution of Stellera chamaejasme in different time periods, especially the changes in dense areas. The change in color can quickly identify which areas are the key areas of Stellera chamaejasme expansion or reduction, which is helpful for formulating targeted grassland management plans.

[0054] Spatiotemporal dynamic monitoring: The spatiotemporal dynamic change characteristics of Euphorbia chamaejasme are shown through grid division and quantitative statistics of images in June and July 2023. The present invention provides refined quantitative information of Euphorbia chamaejasme in the study area through a hierarchical display of 5 plants per order of magnitude, which can effectively track its expansion trend and provide a scientific basis for grassland managers to formulate reasonable management and governance measures. With this information, the growth dynamics of Euphorbia chamaejasme in different time periods can be accurately judged, and the time window and key areas for governance can be clarified.

[0055] Finally, this application greatly improves the detection accuracy of Euphorbia chamaejasminoides by optimizing the YOLOv8 model. As shown in Table 1, the accuracy evaluation table of the optimized YOLOv8 model (RFC-YOLOv8) and the YOLOv8 before optimization (YOLOv8) is provided. It can be seen from the table that the optimized model has higher precision (P), recall rate (R), and average precision (APR).

[0056] @IoU=0.5(mAP@0.5), average precision mean @IoU=0.5:0.95(mAP@0.5:0.95), parameter quantity (Params) and floating-point operation number (Gflops) have all been improved to varying degrees.

[0057] Table 1 RFC-YOLOv8 and YOLOv8 accuracy evaluation table

[0058]

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying and spatially locating the poisonous plant Euphorbia chamaejasme based on improved YOLOv8, characterized in that: The steps include: S1. Dataset preparation: Use drones to collect images of the target Euphorbia chamaejasme growth area, perform image preprocessing and period annotation as the original dataset, of which 80% is the training set, 10% is the validation set, and 10% is the test set; S2. Extract image features to optimize the YOLOv8 network model: According to the performance characteristics of Euphorbia chamaejasme at different stages, the backbone feature extraction network, neck network and convolution module of the original YOLOv8 network model are optimized to form an improved RFC-YOLOv8 network model; S3. Model training and result prediction: Use the training set in S1 to train the constructed RFC-YOLOv8 network model, obtain the optimal network model for detecting Euphorbia chamaejasme, and output the RFC-YOLOv8 network recognition drone jpg format image of the training set; S4. Spatial positioning and accuracy evaluation: The RFC-YOLOv8 network of the training set recognizes the UAV jpg format image, matches the pixel coordinates of the recognition frame with the tif format image with geographic coordinates, obtains the actual geographic location information of the wolfberry and the location information of the wolfberry in the UAV image, and performs accuracy evaluation based on the ground survey points; S5. Generate a schematic diagram of the spatial distribution pattern of Stellera chamaejasme in natural grasslands to detect the spatiotemporal distribution of Stellera chamaejasme.

2. The identification and spatial positioning method according to claim 1, characterized in that: The image preprocessing and period labeling in S1 specifically includes splicing and sliding cropping the collected images, the size of the cropped images is 640×640 pixels, and the cropped images are labeled to divide the wolfberry into three growth periods, namely, the early stage, the middle stage and the late stage.

3. The identification and spatial positioning method according to claim 1, characterized in that: The optimization of the backbone feature extraction network described in S2 is as follows: using the RevCol structure, the input features are divided into two groups, and the Forward and Inverse operations are implemented in each Block. The formula is as follows: Forward:x t =F t (x t-1 ,x t-2 ,…,x t-m+1 )+γx t-m Inverse:x t-m =c -1 [x] t-1 ,x t-2 ,…,x t-m+1 ] Here, m represents the recursive order, allowing tensors of different shapes to be used to represent features at different semantic levels or resolutions.

4. The identification and spatial positioning method according to claim 1, characterized in that: The neck network optimization described in S2 is as follows: using the focused diffusion pyramid network as the neck structure of YOLOv8, and using FocusFeature as the core part of the focused diffusion pyramid network.

5. The identification and spatial positioning method according to claim 1, characterized in that: In the YOLOv8 network model optimization described in S2, the C3Ghost module is used to replace the C2f module in the original YOLOv8.

6. The identification and spatial positioning method according to claim 1, characterized in that: The specific method of generating a schematic diagram of the spatial distribution pattern of wolfsbane in natural grasslands as described in S5 is as follows: using 10 meters × 10 meters as a statistical unit, classifying the wolfsbane Euphorbia identified in the drone image, setting every 5 wolfsbane Euphorbia as one class, and generating a schematic diagram of the spatial distribution pattern of wolfsbane in natural grasslands based on the classification information.

Images