Rapid detection method and system for plant number of high-oil corn in seedling stage

Abstract

The invention belongs to the technical field of agricultural information, and discloses a rapid detection method and system for the plant number of high-oil corn in the seedling stage, and the method comprises the following steps: selecting a high-oil corn monitoring field; the multi-dimensional data of the high-oil corn field is obtained through carrying image acquisition equipment by an unmanned aerial vehicle; performing multi-dimensional data preprocessing; seedling segmentation and identification; performing plant number statistics and verification; and storing and analyzing data. According to the rapid detection method and system for the plant number of the high-oil corn in the seedling stage, data support is provided for analysis and prediction of the growth condition of the corn, field management measures are optimized, and the yield and quality of the corn are improved; the plant number of the high-oil corn in the seedling stage can be rapidly and accurately detected, and the efficiency and accuracy of field management are improved; the method has wide applicability and can be used for corn field management in different growth stages and growth environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural information technology, and in particular to a rapid detection method and system for high-oil corn seedling stage plant number. BACKGROUND

[0002] Traditional corn seedling stage plant number detection relies on manual counting, which has the problems of low efficiency and large error. Although current unmanned aerial vehicle remote sensing and computer vision technology has made progress in crop monitoring, such as the threshold segmentation method based on RGB image with an accuracy of 89%-96%, and the YOLOv8 model with an accuracy of 98.22% in ordinary corn seedling identification, it still faces significant limitations for high-oil corn:

[0003] 1. Weak environmental adaptability, the bright reflective characteristics of high-oil corn leaves cause the accuracy of traditional image segmentation method to drop by 6%-10% on cloudy days, and the low contrast between seedlings and weeds and residues further increases the false detection rate;

[0004] 2. Insufficient model generalization, existing deep learning models are mostly based on ordinary corn training, and lack optimization for high-oil corn morphological differences (narrow leaf sheath, dark green leaf color), and need to adjust parameters according to growth stages;

[0005] 3. Hardware deployment bottleneck, high-precision models (such as FasterR-CNN) rely on GPU computing power, and lightweight models (such as YOLO-FE2) have a miss detection rate of >15% under high-density planting;

[0006] 4. Lack of agronomic integration, existing systems do not associate key agronomic parameters such as uniformity of emergence and oil content, which restricts decision-making value.

[0007] Therefore, it is urgent to develop a high-oil corn seedling stage plant number rapid detection system with environmental robustness, special model optimization and embedded deployment capability, to fill the technical gap. SUMMARY

[0008] The purpose of the present application is to provide a rapid detection method and system for high-oil corn seedling stage plant number, which combines modern sensor technology and image processing technology to realize rapid and accurate detection of high-oil corn seedling stage plant number.

[0009] To achieve the above purpose, the present application provides a rapid detection method for high-oil corn seedling stage plant number, comprising the following steps:

[0010] S1, selecting a high-oil corn monitoring field;

[0011] S2, acquiring multi-dimensional data of high-oil corn monitoring field by unmanned aerial vehicle carrying image acquisition equipment;

[0012] S3, multi-dimensional data preprocessing;

[0013] S4, seedling segmentation and recognition;

[0014] S5, plant number statistics and verification;

[0015] S6, data storage and analysis.

[0016] Preferably, in S1, a high-oil corn monitoring field is selected, and a representative field covering different soil types, a planting density of 35-65 thousand plants per hectare, and terrain conditions is selected as the monitoring field.

[0017] Preferably, in S2, multi-dimensional data of the high-oil corn monitoring field is obtained by a UAV carrying an image acquisition device, and the specific content is as follows:

[0018] S21, a UAV carrying an RGB camera and a satellite positioning system is used to collect data of the high-oil corn monitoring field;

[0019] S22, the multi-dimensional image data obtained in the high-oil corn monitoring field, and the specific content is as follows:

[0020] (1) Spatial data: orthographic images of the monitoring field are obtained by a UAV, with a resolution of ≤2 cm / pixel;

[0021] (2) Spectral data: the RGB channel reflectivity of the seedling canopy of the monitoring field is collected;

[0022] (3) Structural data: based on the structure from motion SfM method, a digital surface model DSM containing seedling projection leaf area and plant height point cloud is constructed.

[0023] Preferably, in S3, multi-dimensional data preprocessing, and the specific content is as follows:

[0024] S31, denoising processing: eliminate flight vibration noise points by Non-Local Means algorithm;

[0025] S32, illumination correction: compensate for leaf reflection based on the Retinex theory model;

[0026] S33, contrast enhancement: use CLAHE algorithm to improve the separation degree between seedlings and background.

[0027] Preferably, in S4, seedling segmentation and recognition, and the specific content is as follows:

[0028] S41, seedling segmentation is performed by a double-branch convolutional network DC-Net to generate a binary segmentation mask;

[0029] S42, the missing targets of the segmentation mask are supplemented by the optimized YOLOv10 model.

[0030] S43, fuse S41, S42 results, perform morphological opening operation to eliminate adhesion; mark independent single plants through connected domain marking algorithm.

[0031] Preferably, in S41, the seedling segmentation is performed through a double-branch convolutional network DC-Net to generate a binary segmentation mask, and the specific content is as follows:

[0032] (1) Branch 1 extracts a seedling pixel-level mask through a U-Net architecture;

[0033] (2) Branch 2 fuses to generate a prior segmentation map through a color model HSV and an index ExG;

[0034] (3) Fuse the output results of the double branches to generate a binary segmentation mask; wherein the seedling area = 1 and the background = 0.

[0035] Preferably, in S42, the YOLOv10 model is optimized to supplement the missing targets of the segmentation mask, and the specific optimization is as follows:

[0036] (1) Replace the YOLOv10 backbone network with GhostNetV2 to adapt to embedded deployment;

[0037] (2) Increase a channel-spatial dual attention reflection perception module GAM at the end of the Backbone to suppress the interference of high-oil corn leaf reflection;

[0038] (3) Introduce a loss function that fuses an enhanced intersection over union loss function EIoU Loss and a focal loss function FocalLoss to reduce the dense seedling miss detection rate;

[0039] Preferably, in S5, the plant number is counted and verified, and the specific content is as follows:

[0040] S51, automatic counting: generate a plant number statistics graph based on the connected domain centroid coordinates to count and monitor the number of single plants in the field;

[0041] S52, precision verification:

[0042] (1) Randomly extract 5% of the area for manual review;

[0043] (2) Calculate the F1-score to evaluate the counting accuracy, as shown below:

[0044]

[0045] Wherein, Precision is the precision; Recall is the recall.

[0046] Preferably, in S6, data storage and analysis, the specific content is as follows:

[0047] S61, integrate the multi-dimensional data generated by S1-S5;

[0048] S62, based on the stored data, perform agronomic analysis;

[0049] First, calculate the density per mu, as follows:

[0050]

[0051] Second, calculate the seedling uniformity index, as follows:

[0052]

[0053] Wherein, ECI is the seedling uniformity index; N_actual is the actual emergence number; N_expected is the expected emergence number; CV_height is the coefficient of variation in height.

[0054] A rapid detection system for high-oil corn seedling stage plant number, comprising an image acquisition module, an image preprocessing module, a target segmentation module, a counting and statistical module, a data storage and analysis module;

[0055] The image acquisition module includes a UAV, an RGB camera, a satellite positioning system, and an image transmission device; the module is used for collecting image data of the monitored field, and transmitting the data into the image preprocessing module through the image transmission device;

[0056] The image preprocessing module includes an embedded GPU, Retinex algorithm and CLAHE algorithm running in sequence to form an algorithm chain, which eliminates environmental interference and enhances target features, and transmits the processed data into the target segmentation module;

[0057] The target segmentation module includes a double-branch convolutional network DC-Net model and an optimized YOLOv10 model, which is used to separate seedlings from the background and transmit the segmentation results to the counting and statistical module;

[0058] The counting and statistical module outputs the plant number detection results and spatial distribution through connected domain analysis;

[0059] The data storage and analysis module stores the detection results of the counting and statistical module and provides data analysis functions to help users understand the growth status of corn;

[0060] The image preprocessing module, the target segmentation module and the counting and statistical module run on the embedded GPU, and the model is quantized and accelerated by FP16 through the TensorRT tool, which is suitable for field mobile deployment and real-time generation of field plant number heat map and seedling shortage area early warning coordinates.

[0061] Therefore, the present application adopts the above-mentioned rapid detection method and system for high-oil corn seedling stage plant number, which has the following beneficial effects:

[0062] (1) The method and system of the present application can quickly and accurately detect the plant number of high-oil corn at the seedling stage, improving the efficiency and accuracy of field management.

[0063] (2) The method and system of the present application have wide applicability and can be used for corn field management at different growth stages and growth environments.

[0064] (3) The method and system of the present application provide data support for the analysis and prediction of corn growth conditions, which helps to optimize field management measures and improve corn yield and quality.

[0065] (4) The present application uses image processing technology to realize automatic recognition and counting, avoiding the inaccuracy and time-consuming nature of manual counting.

[0066] The technical solutions of the present application will be further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 is a technical roadmap of a rapid detection method for high-oil corn seedling stage plant number;

[0068] Figure 2 is a structural diagram of a rapid detection system for high-oil corn seedling stage plant number. DETAILED DESCRIPTION

[0069] The technical solutions of the present application will be further described in detail below through the accompanying drawings and examples.

[0070] As shown in Figure 1 , a rapid detection method for high-oil corn seedling stage plant number includes the following steps:

[0071] S1, selecting a high-oil corn monitoring field;

[0072] S2, acquiring multi-dimensional data of high-oil corn monitoring field by unmanned aerial vehicle carrying image acquisition equipment;

[0073] S3, preprocessing multi-dimensional data;

[0074] S4, seedling segmentation and recognition;

[0075] S5, plant number statistics and verification;

[0076] S6, data storage and analysis.

[0077] EMBODIMENT

[0078] A rapid detection method for high-oil corn seedling stage plant number, comprising the following steps:

[0079] S1, select high-oil corn monitoring field.

[0080] Select representative fields covering different soil types, planting densities of 35-65 thousand plants per hectare, and terrain conditions as monitoring fields to ensure model generalization ability.

[0081] S2, obtain multi-dimensional data of high-oil corn monitoring field by unmanned aerial vehicle carrying image acquisition equipment, the specific content is as follows:

[0082] S21, use RGB unmanned aerial vehicle to collect data of high-oil corn monitoring field, the specific content is as follows:

[0083] The unmanned aerial vehicle model is DJI M300 RTK, equipped with 20 million pixel RGB camera and PPK positioning system.

[0084] Flight conditions are as follows:

[0085] (1) Choose the flight time when the sky is clear, the light intensity is >80000 lux, and the sun elevation angle is >30° to reduce shadow interference.

[0086] (2) Choose the corn at 4-8 leaf stage, the seedling height is 15-40 cm, and the crown structure is clear.

[0087] (3) The flight height is 30 m, the ground resolution is 1.2 cm, the heading overlap rate is 80%, and the lateral overlap rate is 70%.

[0088] S22, collect more than 10,000 multi-dimensional image data of high-oil corn monitoring field covering sunny, weeds, and stubble scenes, and use Labelme tool to label "smallest circumscribed rectangle of seedling", wherein the multi-dimensional image data has the following specific content:

[0089] (1) Spatial data: obtain the orthographic image of the monitoring field by unmanned aerial vehicle, the resolution is ≤2 cm / pixel.

[0090] (2) Spectral data: collect the seedling canopy R, G, B channel reflectivity of the monitoring field.

[0091] (3) Structure data: based on the structure from motion SfM method, construct a digital surface model DSM containing seedling projection leaf area and plant height point cloud.

[0092] S3, pre-process the multi-dimensional data, the specific steps are as follows:

[0093] S31, denoising processing: eliminate flight vibration noise points through non-local mean filtering (Non-Local Means) algorithm.

[0094] S32, light correction: compensate for leaf reflection based on Retinex theory model, and solve the problem of high-oil corn reflection.

[0095] S33, contrast enhancement: use CLAHE algorithm to enhance the separation degree between seedlings and background.

[0096] S4, seedling segmentation and recognition.

[0097] S41, through the double-branch convolution network (DC-Net), the seedling segmentation is carried out, and the binary segmentation mask is generated, and the specific content is as follows:

[0098] (1) Branch 1 extracts the seedling pixel-level mask through the U-Net architecture.

[0099] (2) Branch 2 fuses to generate a prior segmentation map through the color model HSV and the exponential ExG.

[0100] (3) The output results of the double-branch fusion are generated, and the binary segmentation mask is generated, wherein the seedling area = 1 and the background = 0.

[0101] S42, through the optimized YOLOv10 model, the missing target of the segmentation mask is supplemented, and the specific optimization is as follows:

[0102] (1) Replace the YOLOv10 backbone network with GhostNetV2, reduce the parameter quantity by 58%, and adapt to embedded deployment.

[0103] (2) Increase the channel-space double attention reflection perception module (GAM) at the end of the Backbone to suppress the interference of high-oil corn leaf reflection.

[0104] (3) Introduce the enhanced intersection over union loss function EIoU Loss and the focal loss function Focal Loss to reduce the dense seedling miss rate.

[0105] S43, fuse the results of S41 and S42, perform morphological opening operation, eliminate adhesion, and label independent single plants through connected domain labeling algorithm.

[0106] S5, plant number statistics and verification.

[0107] S51, automatic counting: generate plant number statistics chart based on connected domain centroid coordinates to count and monitor the number of single plants in the field.

[0108] S52, accuracy verification:

[0109] (1) Manually review by randomly extracting 5% of the area.

[0110] (2) Calculate the F1-score to evaluate the counting accuracy, requiring F1≥0.92 in complex field scenarios, as follows:

[0111]

[0112] where Precision is the precision rate; Recall is the recall rate.

[0113] S6, Data storage and analysis.

[0114] S61, Integrate the multi-dimensional data generated by S1-S5.

[0115] S62, Based on the stored data, perform agronomic analysis, as follows:

[0116] Calculate the density per mu, as follows:

[0117]

[0118] Calculate the uniformity index of emergence, as follows:

[0119]

[0120] where ECI is the uniformity index of emergence; N_actual is the actual emergence plant number; N_expected is the expected emergence plant number; CV_height is the coefficient of variation in height.

[0121] As shown in the following figure, a rapid detection system for high-oil corn seedling stage plant number includes an image acquisition module, an image preprocessing module, a target segmentation module, a counting and statistical module, and a data storage and analysis module. Figure 2

[0122] (1) The image acquisition module includes a DJI M300 RTK unmanned aerial vehicle equipped with a 20 million pixel RGB camera, a PPK positioning system, and an image transmission device, which is used to collect image data of the monitoring field and transmit the data into the image preprocessing module through the image transmission device.

[0123] (2) The image preprocessing module includes an NVIDIA Jetson AGX Orin embedded GPU running a Retinex+CLAHE algorithm chain, which eliminates environmental interference and enhances target features, and transmits the processed data into the target segmentation module.

[0124] ​(3) The target segmentation module includes a double-branch convolutional network (DC-Net) model and an optimized YOLOv10 model, which can accurately separate seedlings from the background. The parameter quantity of the DC-Net model is 4.8M, and the inference speed is 23fps@1080p.

[0125] (4) The counting and statistical module outputs the seedling number detection results and spatial distribution based on the output results of the target segmentation module through connected component analysis.

[0126] (5) The data storage and analysis module stores the detection results of the counting and statistical module and provides data analysis functions to help users understand the growth status of corn.

[0127] (6) The image preprocessing module, the target segmentation module and the counting and statistical module run on an embedded GPU, and the models are quantized and accelerated by FP16 through the TensorRT tool, which is suitable for field mobile deployment and can generate field seedling number heat maps and seedling shortage area warning coordinates in real time.

[0128] Therefore, the application adopts the above-mentioned rapid detection method and system for high-oil corn seedling number, which realizes automatic recognition and counting by using image processing technology, avoids the inaccuracy and time-consuming nature of manual counting, can quickly and accurately detect the seedling number of high-oil corn at the seedling stage, and improves the efficiency and accuracy of field management. The method and system of the application provide data support for the analysis and prediction of corn growth status, which helps to optimize field management measures, improve corn yield and quality, and can be used for corn field management at different growth stages and growth environments, and has wide applicability.

[0129] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the application and not to limit them, although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the application.

Claims

1. A rapid detection method for the number of seedlings in high-oil maize, characterized in that, Includes the following steps: S1. Select high-oil corn monitoring plots; S2. Acquire multi-dimensional field data for high-oil corn monitoring by using drones equipped with image acquisition devices; S3, Multi-dimensional data preprocessing; S4. Seedling segmentation and identification; S5. Plant count statistics and verification; S6. Data storage and analysis.

2. The rapid detection method for the number of seedlings in high-oil maize according to claim 1, characterized in that, In S1, high-oil maize monitoring plots were selected, specifically representative plots covering different soil types, planting densities of 35,000-65,000 plants / hectare, and topographical conditions, as monitoring plots.

3. The rapid detection method for the number of seedlings in high-oil maize according to claim 1, characterized in that, In S2, drones equipped with image acquisition devices are used to obtain multi-dimensional data from high-oil corn monitoring fields. The specific details are as follows: S21. Data collection for high-oil corn monitoring fields is carried out using drones equipped with RGB cameras and satellite positioning systems. S22. The acquired multi-dimensional image data of high-oil corn monitoring fields are as follows: (1) Spatial data: Orthophotos of the monitored fields were acquired by drones with a resolution of ≤2cm / pixel; (2) Spectral data: Collect the RGB channel reflectance of the seedling canopy in the monitoring field; (3) Structural data: Based on the structure of motion recovery (SfM) method, a digital surface model (DSM) containing the seedling projected leaf area and plant height point cloud was constructed.

4. The rapid detection method for the number of seedlings in high-oil maize according to claim 1, characterized in that, In S3, multi-dimensional data preprocessing includes the following details: S31. Noise Reduction Processing: The Non-LocalMeans algorithm is used to eliminate flight vibration noise. S32. Light correction: Compensating for leaf reflection based on the Retinex theoretical model; S33. Contrast Enhancement: The CLAHE algorithm, which limits contrast adaptive histogram equalization, is used to improve the separation between seedlings and the background.

5. The rapid detection method for the number of seedlings in high-oil maize according to claim 1, characterized in that, In S4, seedling segmentation and identification are detailed below: S41. Use the dual-branch convolutional network DC-Net to segment seedlings and generate a binarized segmentation mask; S42. Supplement the missing targets in the segmentation mask using the optimized YOLOv10 model; S43, combined with the results of S41 and S42, performs morphological opening operations to eliminate adhesion; and uses a connected component labeling algorithm to label independent individual plants.

6. The rapid detection method for the number of seedlings in high-oil maize according to claim 5, characterized in that, In S41, seedling segmentation is performed using the dual-branch convolutional network DC-Net to generate a binarized segmentation mask, as detailed below: (1) Branch 1 extracts the seedling pixel-level mask through the U-Net architecture; (2) Branch 2 uses the color model HSV and the exponent ExG to generate a prior segmentation map; (3) Combine the output results of the two branches to generate a binary segmentation mask; where seedling region = 1 and background = 0.

7. The rapid detection method for the number of seedlings in high-oil maize according to claim 5, characterized in that, In S42, the optimized YOLOv10 model is used to supplement the missing targets in the segmentation mask. The specific optimizations are as follows: (1) Replace the YOLOv10 backbone network with GhostNetV2 to adapt to embedded deployment; (2) Add a channel-space dual attention reflective sensing module (GAM) to the end of the backbone to suppress reflective interference from high-oil corn leaves; (3) Introduce a loss function that combines the enhanced cross-union ratio loss function EIoU Loss and the focal loss function FocalLoss to reduce the false negative rate of dense seedlings.

8. The rapid detection method for the number of seedlings in high-oil maize according to claim 1, characterized in that, In S5, the plant count and verification are detailed below: S51. Automatic counting: Generates a plant count chart based on the centroid coordinates of connected components to count the number of individual plants in the monitored field. S52, Accuracy Verification: (1) Manual verification was conducted by randomly selecting 5% of the area; (2) Calculate the F1-score to evaluate the counting accuracy, as shown below: Precision is the accuracy rate; Recall is the recall rate.

9. The rapid detection method for the number of seedlings in high-oil maize according to claim 1, characterized in that, In S6, data storage and analysis are detailed below: S61. Integrate and store multi-dimensional data generated by S1-S5; S62. Conduct agronomic analysis based on stored data; First, calculate the density per acre, as shown below: Secondly, the seedling uniformity index is calculated, as shown below: Wherein, ECI is the seedling uniformity index; N_actual is the actual number of seedlings that have emerged; N_expected is the expected number of seedlings that have emerged; and CV_height is the seedling height variation coefficient.

10. A rapid detection method for the number of seedlings in high-oil maize according to claims 1-9, applied to a rapid detection system for the number of seedlings in high-oil maize, characterized in that, The system includes an image acquisition module, an image preprocessing module, a target segmentation module, a counting and statistics module, and a data storage and analysis module; The image acquisition module includes a drone, an RGB camera, a satellite positioning system, and image transmission equipment; This module is used to collect image data of the monitored fields and transmit the data to the image preprocessing module through the image transmission device; The image preprocessing module includes an embedded GPU that sequentially runs the Retinex algorithm and the CLAHE algorithm to form an algorithm chain, eliminates environmental interference, enhances target features, and then inputs the processed data into the target segmentation module. The target segmentation module includes a dual-branch convolutional network DC-Net model and an optimized YOLOv10 model, used to separate seedlings from the background, and the segmentation results are fed into the counting and statistics module. The counting and statistics module outputs the plant count detection results and spatial distribution through connected component analysis; The data storage and analysis module stores the detection results from the counting and statistics module and provides data analysis functions to help users understand the corn growth status. The image preprocessing module, target segmentation module, and counting and statistics module run on an embedded GPU. The model is accelerated by FP16 quantization using the TensorRT tool, adapted for mobile deployment in the field, and generates field plant number heatmaps and early warning coordinates of missing seedling areas in real time.

Images