Planting hole detection method based on unmanned aerial vehicle
By constructing the YOLO-PH model, combining high-resolution drone sensors and feature extraction optimization technology, the problem of insufficient accuracy of planting hole detection in complex scenarios is solved, efficient and intelligent detection is achieved, labor costs are reduced, and the application value of drone remote sensing images is improved.
Patent Information
- Application Number
- CN202510490951.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-18
- Publication Date
- 2025-08-12
AI Technical Summary
Existing object detection algorithms are insufficient in the accuracy of plant hole detection based on edge features, especially in complex scenarios, resulting in high labor costs and inefficiency in manual counting.
Using the YOLO-PH model, the initial model consisting of the backbone network, neck network and SD-Head detection head was constructed, combined with high-resolution drone sensors to obtain data, perform multiple aerial shots, perform geometric correction and radiation correction, build a high-quality data set, and enhance multi-scale feature extraction through the C2f_DyGhostConv module, and use ATSS strategy and SD-Head detection head for planting hole detection, optimize model parameters to improve accuracy.
The intelligent level of planting hole detection has been improved, the cost of manual detection has been reduced, and the application value of drone remote sensing images in agricultural monitoring and forestry management has been improved, with the accuracy increased by 1.3-1.1%, the detection speed increased by 13.8%, and the missed detection rate has been reduced to 6.5%.
Smart Images
Figure CN120472308A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of planting hole detection, and more particularly, to a planting hole detection method based on a drone. Background Art
[0002] The integration of drones and planting hole detection technology offers an effective solution to addressing the challenges of manual counting, high labor costs, and low efficiency in large-scale planting operations. However, existing object detection algorithms still suffer from insufficient accuracy when detecting plant holes based on edge features, especially in complex scenes. To address this issue, a drone-based planting hole detection method is proposed. Summary of the Invention
[0003] The purpose of the present invention is to provide a planting hole detection method based on drone to solve the above problems.
[0004] The above technical objectives of the present invention are achieved through the following technical solutions:
[0005] A first aspect of the present invention provides a planting hole detection method based on a drone, comprising the following steps:
[0006] S1. Obtain the processed planting hole dataset;
[0007] S2. Build the initial YOLO-PH model consisting of the backbone network, neck network, and SD-Head detection head;
[0008] S3. Training and evaluating the initial planting hole automatic detection model to obtain an optimized YOLO-PH model;
[0009] S4. Use the optimized YOLO-PH model for automatic detection and visually annotate the planting holes in the form of bounding boxes.
[0010] In combination with the first aspect, the present invention is further configured as follows: in step S1, under good weather conditions, a high-resolution drone sensor is used to perform multiple aerial photography missions in a designated research area to obtain a planting hole dataset;
[0011] The planting hole dataset is sequentially subjected to geometric correction, radiation correction, and orthorectification preprocessing steps to ultimately generate an orthophoto. The orthophoto is then cropped to generate multiple sub-images of fixed size. An intelligent screening algorithm is then used to automatically remove areas with unclear boundaries and most images that do not contain planting holes, thereby obtaining a processed planting hole dataset.
[0012] In combination with the first aspect, the present invention is further configured as follows: in step S2, the backbone network adopts the CSPDarknet53 framework, the original C2f module is replaced by the C2f_DyGhostConv module, the DyGhostConv layer of the C2f_DyGhostConv module replaces the standard convolution with a combination of dynamic convolution and Ghost convolution, and the backbone network extracts multi-scale features through C2f_DyGhostConv and outputs P3, P4 and P5 feature maps.
[0013] In combination with the first aspect, the present invention is further configured as follows: the neck network fuses multi-scale features to enhance small target semantic information.
[0014] In combination with the first aspect, the present invention is further configured as follows: the SD-Head detection head adopts convolution parameter sharing and group normalization to reduce the number of parameters and improve the field of view perception ability. The SD-Head detection head adopts a single convolution module to process P3, P4 and P5 multi-scale features, performs shared convolution calculations, outputs the bounding box coordinates, confidence and category probability of each grid, filters the final detection results through non-maximum suppression, and outputs the location and number of planting holes.
[0015] In combination with the first aspect, the present invention is further configured as follows: in step S3, the YOLO-PH model is trained using the processed training set, and the network parameters are optimized in combination with the target loss function, and the optimization parameters are dynamically adjusted based on the adaptive learning rate strategy. In the model optimization stage, the trained YOLO-PH model is hyperparameter tuned using the validation set, and mAP50, mAP75, mAP50:95, Parameter, GFLOPs and Detection Speed are used for quantitative evaluation to obtain the optimized YOLO-PH model.
[0016] The second aspect of the present invention also provides a device / equipment / system for drone-based planting hole detection, comprising a memory, a processor, and a computer program stored on the memory, characterized in that the processor executes the computer program to implement the steps of any of the above methods.
[0017] The third aspect of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of any of the above methods when executed by a processor.
[0018] A fourth aspect of the present invention further provides a computer program product, comprising a computer program / instruction, which implements the steps of any of the above methods when executed by a processor.
[0019] In summary, the present invention has the following beneficial effects:
[0020] The YOLO-PH model proposed in this paper not only improves the intelligence level of planting hole detection, but also effectively reduces the cost of manual detection and improves the application value of UAV remote sensing images in agricultural monitoring and forestry management.
[0021] The quantitative data are as follows:
[0022] Accuracy improvement: mAP50 reaches 96% (+1.3%), mAP50:95 reaches 51.6% (+1.1%).
[0023] Efficiency optimization: FLOPs dropped to 4.2G (-48.8%), and detection speed increased to 2207FPS (+13.8%).
[0024] Adaptability to complex scenes: In complex areas, the F1-score reaches 0.95 and the missed detection rate is only 6.5%. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 This is a flow chart of a planting hole detection method based on a drone in an embodiment of the present invention;
[0026] Figure 2 This is a diagram of the overall architecture of YOLO-PH in an embodiment of the present invention, noting the connection relationship between the backbone, neck, and SD-Head modules;
[0027] Figure 3 This is the dynamic convolution and Ghost operation process within the C2f_DyGhostConv module in an embodiment of the present invention;
[0028] Figure 4 1 is a schematic diagram of the calculation of the dynamic threshold value of the ATSS strategy in an embodiment of the present invention;
[0029] Figure 5 This is the implementation details of SD-Head shared convolution and group normalization in an embodiment of the present invention;
[0030] Figure 6 This is the complete technical route from drone image input to detection result output in the embodiment of the present invention. DETAILED DESCRIPTION
[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0032] Example 1:
[0033] A planting hole detection method based on drone includes the following steps:
[0034] S1. Obtain a processed implant hole dataset and divide the processed implant hole dataset into a training set, a validation set, and a test set;
[0035] S2. Build the initial YOLO-PH model consisting of the backbone network, neck network, and SD-Head detection head;
[0036] S3. Training and evaluating the initial planting hole automatic detection model to obtain an optimized YOLO-PH model;
[0037] S4. Use the optimized YOLO-PH model to automatically detect the target image and visually annotate the planting holes in the form of bounding boxes.
[0038] Step S1 is specifically as follows: Under good weather conditions, high-resolution drone sensors are used to perform multiple aerial photography missions in the designated study area to obtain full-coverage, high-precision aerial survey images. The collected raw data undergoes preprocessing steps such as geometric correction, radiation correction, and orthorectification to finally generate orthophotos. In order to improve the robustness of model training and data utilization efficiency, the image data is further cropped to generate multiple sub-images of a fixed size (512×512 pixels), and the intelligent screening algorithm is used to automatically remove areas with unclear boundaries and most images that do not contain targets (planting holes), thereby constructing a high-quality planting hole dataset. To ensure the generalization ability of the model, the dataset is divided into training set, validation set, and test set in a ratio of 7:2:1, which are used for model training, hyperparameter optimization, and final performance evaluation, respectively.
[0039] In step S2,
[0040] Backbone network: Adopting the CSPDarknet53 framework, the original C2f module is replaced with the C2f_DyGhostConv module to enhance the multi-scale feature expression capability.
[0041] Neck network: Based on the FPN structure, it also uses the C2f_DyGhostConv module for feature fusion and outputs three scale feature maps (P3, P4, P5).
[0042] SD-Head detection head: The shared convolution module processes multi-scale features, introduces the Scale layer to unify the feature scale, and improves the perception of small targets through GN.
[0043] Among them, the core module design
[0044] (1) C2f_DyGhostConv module (see attached Figure 2 )
[0045] Structural composition:
[0046] DyGhostConv layer: replaces the standard convolution with a combination of dynamic convolution and ghost convolution.
[0047] Dynamic convolution kernel: Adaptively generates four sets of convolution kernel weights based on input features to enhance feature diversity.
[0048] Ghost convolution: Generates partial features through 1×1 convolution, and then expands the feature channel through 3×3 depth-wise separable convolution, reducing the number of parameters by 50%.
[0049] Cross-stage connection: retain the multi-branch structure of the C2f module and fuse shallow details with deep semantic features through skip connections.
[0050] The C2f_DyGhostConv module works as follows: After the input feature map is processed by DyGhostConv to extract dynamic features, it is processed in two ways: one retains the original resolution, and the other generates multi-scale features through downsampling. The two features are concatenated (concat) and fused with a 1×1 convolution to output the enhanced feature map.
[0051] (2) ATSS label allocation strategy (refer to the attached Figure 3 )
[0052] Adaptive positive sample selection:
[0053] For each ground-truth box (GT), the IoU between the candidate anchor box and the GT is calculated at each level of the feature pyramid, and the IoU mean + standard deviation is selected as the dynamic threshold.
[0054] High-quality positive samples are screened according to the threshold to avoid sample distribution bias caused by a fixed IoU threshold.
[0055] Advantages: Based on the uniformity of planting hole scale, the ratio of positive and negative samples is adaptively adjusted to improve training efficiency.
[0056] (3) Siblings Detection Head (Compare to the attached Figure 4 )
[0057] Shared convolution design:
[0058] A single convolution module is used to process the P3-P5 multi-scale features, and the feature map is normalized through the Scale layer to eliminate scale differences.
[0059] 3×3 CBS (Conv+BN+SiLU) and 1×1 convolution are used to generate detection results, and the number of parameters is reduced to 1.54M.
[0060] Group Normalization (GN): Batch Normalization (BN) in the convolutional layers of the detection head is replaced by Group Normalization (GN).
[0061] The channels are divided into 32 groups, and the mean and variance are calculated independently for each group to solve the problem of BN performance degradation during small batch training.
[0062] The robustness of the model in complex lighting and occlusion scenarios is improved, and the false detection rate is reduced to 3.2% (253).
[0063] The SD-Head detection head uses convolution parameter sharing and group normalization to reduce the number of parameters and improve visual perception capabilities. By performing shared convolution calculations, it outputs the bounding box coordinates, confidence level, and category probability of each grid. The final detection results are filtered through non-maximum suppression to output the location and number of planting holes.
[0064] In step S3 of the present invention, the YOLO-PH model is trained using the processed training data, and network parameters are optimized using a combination of various loss functions, such as L1, L2, and L4, to improve the model's recognition accuracy and robustness for implant hole targets. During training, the loss function converges. In the sample experiment for implant hole detection, convergence was achieved after 200 epochs in most cases. Dynamically adjusting the optimization parameters based on an adaptive learning rate strategy can further improve convergence speed and reduce the risk of overfitting. During the model optimization phase, hyperparameters are tuned using an independent validation set to ensure that the model can maintain high-precision detection capabilities under varying lighting conditions, resolutions, and complex backgrounds. To comprehensively evaluate the detection performance of the model, multiple indicators such as mAP50 (average precision @IoU=50%), mAP75 (average precision @IoU=75%), mAP50:95 (mean average precision under different IoU thresholds), Parameter (model parameter number), GFLOPs (floating point operations) and DetectionSpeed (detection speed) are used for quantitative evaluation to ensure that the model achieves the best balance between accuracy, computational complexity and real-time detection performance.
[0065] Specifically, in step S4, after model training and optimization, this method uses the optimal trained model to automatically detect the target image and visually annotate the implant holes using bounding boxes to ensure clear and intuitive detection results. To further verify detection accuracy, the model is inferred using different test samples, and error analysis and performance evaluation are performed using metrics such as precision, recall, F1-score, and intersection over union (IoU).
[0066] Furthermore, this invention uses visualization techniques to display the spatial distribution and accuracy of detection results using histograms, scatter plots, and heat maps, providing more intuitive data support for automated planting hole detection technology. This method not only enhances the intelligent level of planting hole detection, but also effectively reduces manual inspection costs, increasing the application value of drone remote sensing imagery in agricultural monitoring and forestry management.
[0067] Test Example 1: Performance comparison of different detection models
[0068] In order to verify the superiority of the present invention, the performance of the YOLO-PH model proposed in the present invention is compared with various existing improved Yolo models. The experimental results are shown in Table 1:
[0069] Table 1 Performance comparison of various existing improved Yolo models
[0070]
[0071] Test Example 2: Comparative Experiment on Implant Hole Detection Using Different Detection Models
[0072] In order to verify the superiority of the present invention, the accuracy index of the planting hole detection results of the YOLO-PH model proposed in the present invention is compared with various existing improved Yolo models. The experimental results are shown in Table 2:
[0073] Table 2 Precision index of implant hole detection results
[0074]
[0075] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.
Claims
1. A planting hole detection method based on drone, characterized by: The steps include: S1. Obtain the processed planting hole dataset; S2. Build the initial YOLO-PH model consisting of the backbone network, neck network, and SD-Head detection head; S3. Training and evaluating the initial planting hole automatic detection model to obtain an optimized YOLO-PH model; S4. Use the optimized YOLO-PH model for automatic detection and visually annotate the planting holes in the form of bounding boxes.
2. The method for detecting planting holes based on drone according to claim 1 is characterized in that: In step S1, under good weather conditions, a high-resolution UAV sensor is used to perform multiple aerial photography missions in the designated study area to obtain a planting hole dataset; The planting hole dataset is sequentially subjected to geometric correction, radiation correction, and orthorectification preprocessing steps to ultimately generate an orthophoto. The orthophoto is then cropped to generate multiple sub-images of fixed size. An intelligent screening algorithm is then used to automatically remove areas with unclear boundaries and most images that do not contain planting holes, thereby obtaining a processed planting hole dataset.
3. The method for detecting planting holes based on drone according to claim 1 is characterized in that: In step S2, the backbone network adopts the CSPDarknet53 framework, replaces the original C2f module with the C2f_DyGhostConv module, and replaces the standard convolution with a combination of dynamic convolution and Ghost convolution in the DyGhostConv layer of the C2f_DyGhostConv module. The backbone network extracts multi-scale features through C2f_DyGhostConv and outputs P3, P4 and P5 feature maps.
4. The method for detecting planting holes using a drone according to claim 3, wherein: The neck network fuses multi-scale features to enhance the semantic information of small objects.
5. The method for detecting planting holes using a drone according to claim 3, wherein: The SD-Head detection head uses convolution parameter sharing and group normalization to reduce the number of parameters and improve visual field perception capabilities; The SD-Head detection head uses a single convolution module to process P3, P4 and P5 multi-scale features, performs shared convolution calculations, outputs the bounding box coordinates, confidence level and category probability of each grid, filters the final detection results through non-maximum suppression, and outputs the location and number of planting holes.
6. The method for detecting planting holes based on drones according to claim 1 is characterized in that: In step S3, the processed dataset is used to train the YOLO-PH model, and the network parameters are optimized in combination with the target loss function. The optimization parameters are dynamically adjusted based on the adaptive learning rate strategy. In the model optimization stage, the validation set is used to tune the hyperparameters of the trained YOLO-PH model. The mAP50, mAP75, mAP50:95, Parameter, GFLOPs and Detection Speed are used for quantitative evaluation to obtain the optimized YOLO-PH model.
7. A device / equipment / system for detecting planting holes based on drones, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.
8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that When the computer program / instructions are executed by a processor, the steps of the method described in any one of claims 1 to 6 are implemented.
9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.