A deep learning method, device and storage medium for identifying corn seedling stage plant number and leaf age

Abstract

The present application relates to the technical field of crop image recognition processing, in particular to a deep learning method, device and storage medium for identifying corn seedling stage plant number and leaf age, the present application obtains the data of corn seedlings by using near-ground equipment and unmanned aerial vehicle cameras, and then processes the data, and then uses the improved recognition model YOLOv8n-LP to accurately predict the plant number and leaf age of corn seedlings, and then uses a variety of evaluation indexes combined with a variety of model hybrid comparison to verify the high precision, high performance, small model space ratio and high adaptability of the present application, realizing large-scale standardized monitoring of the growth status of corn seedling stage, which not only can significantly improve the efficiency of farmland management, but also can provide scientific basis for crop production, and help the intelligentization and sustainable development of agriculture.

Description

Technical Field

[0001] This invention relates to the field of crop image recognition and processing technology, specifically a deep learning method, device, and storage medium for recognizing the number of corn seedlings and leaf age. Background Technology

[0002] Maize is one of the most widely planted crops globally. Although the planted area has decreased in recent years, consumption has increased significantly over the past decade. With limited arable land resources, increasing maize yields to meet the food demands of a growing population has become an important goal of global agricultural development.

[0003] Against this backdrop, we found that the seedling stage is the key stage that determines yield potential in the entire growth cycle of maize. Among them, the emergence rate and leaf development status are important indicators of early growth, which have a direct impact on subsequent growth and development and final yield.

[0004] With the development of precision agriculture, the demand for dynamic monitoring of crop growth is increasing, especially during the critical seedling stage of corn, where precise monitoring of plant number and leaf age is particularly important. This data is not only the basis for assessing crop growth status but also a crucial foundation for optimizing farmland management decisions and predicting crop yields.

[0005] However, current monitoring of corn seedlings still mainly relies on manual surveys. This traditional method is time-consuming, labor-intensive, and inefficient, and is greatly affected by subjective human factors, making it difficult to meet the needs of rapid and accurate monitoring in large-scale farmland.

[0006] Besides relying on manual surveys, remote sensing technology offers new possibilities for agricultural monitoring. For example, drones can be used to collect farmland information, providing a new technical means for monitoring crop growth. However, in complex field environments (such as weed cover, shadow interference, and large differences in soil background), the monitoring accuracy of existing remote sensing technologies is limited and cannot effectively adapt to complex scenarios. Furthermore, general image processing algorithms lack optimization for the specific characteristics of maize seedlings, making it difficult to accurately identify the number of maize plants and leaf age. At the same time, the processing speed of existing technologies is slow, making it difficult to meet the needs of real-time monitoring and rapid decision-making in large-scale farmland, especially in precision agricultural management. Therefore, the application of remote sensing technology to maize seedling monitoring still faces many limitations.

[0007] Based on the above reasons, this invention designs a deep learning method, device and storage medium for identifying the number of corn seedlings and leaf age, and constructs a lightweight and intelligent automatic identification system to realize large-scale standardized monitoring of the growth status of corn seedlings. This can not only significantly improve the efficiency of farmland management, but also provide scientific basis for crop production and help promote the intelligent and sustainable development of agriculture. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a deep learning method, device and storage medium for identifying the number of corn seedlings and leaf age. By constructing a lightweight and intelligent automatic identification system, it is possible to achieve large-scale standardized monitoring of the growth status of corn seedlings. This can not only significantly improve the efficiency of farmland management, but also provide a scientific basis for crop production and help promote the intelligent and sustainable development of agriculture.

[0009] This invention provides a deep learning method for identifying the number of corn seedlings and leaf age, comprising the following steps:

[0010] S1, Original Image Acquisition:

[0011] S1-1, data on corn seedlings were acquired using near-ground equipment and drone cameras. The data consisted of near-ground images and drone footage from stages V2 to V5.

[0012] V2 stage is the two-leaf stage, which means that two corn leaves are fully unfolded; V5 stage is the five-leaf stage, which means that five corn leaves are fully unfolded.

[0013] S1-2, After stitching the drone images, the stitched images are then cropped to obtain the images of the plots where the corn seedling stage needs to be identified;

[0014] S2, Training the recognition model:

[0015] S2-1, the processed images are organized into two datasets: near-ground NG and UAV;

[0016] S2-2, Draw a bounding box around each corn seedling and manually label it;

[0017] S2-3, In order to increase the diversity and quantity of the drone impact data training set, the drone imagery data is augmented.

[0018] S2-4, the near-ground NG dataset and the UAV dataset are independent of each other, and both are divided into training set, validation set and test set in an 8:1:1 ratio for training the model;

[0019] S3, recognition model YOLOv8n-LP:

[0020] The YOLOv8n-LP recognition model is based on YOLOv8n and adjusts the backbone network, network neck and network head components to reduce the number of parameters without affecting accuracy.

[0021] S4, Corn seedling plant count:

[0022] S4-1: Use the trained model to predict the plot images of corn seedlings that need to be identified, generate images with individual plant bounding boxes, and save the corresponding box position information to a txt file.

[0023] S4-2, each line in the txt file records the confidence score and position information of a box. By counting the number of lines in the txt file, the predicted number of corn seedlings can be obtained.

[0024] The actual number of plants in the bounding box images of a single plant in S4-3 and S4-1 was obtained by labeling the images and generating bounding box information before training, and the same method as the prediction results was used for statistical counting.

[0025] S4-4 is used to assess the accuracy of maize seedling count results by comparing predicted results with actual plant counts.

[0026] S5, estimation of leaf age during corn seedling stage:

[0027] Using near-ground NG platform and UAV dataset, the leaf tip of maize was identified based on the YOLOv8n-LP recognition model to estimate the leaf age of individual maize plants and plots.

[0028] S6, Measurement result verification:

[0029] The performance of the YOLOv8n-LP recognition model was evaluated using precision (p), recall (r), and average precision (AP) metrics, and the coefficient of determination R was used as the evaluation metric. 2 The root mean square error (RMSE) and relative root mean square error (rRMSE) are used to evaluate the accuracy of the model recognition results. The specific formulas are as follows:

[0030] Precision: Formula 1: ;

[0031] Recall: Formula 2: ;

[0032] Average Precision: Formula 3: ;

[0033] in It is the differential unit in the integral. It is an algebraic symbol;

[0034] Coefficient of determination R 2 Formula 4: ;

[0035] Root Mean Square Error (RMSE): Formula 5:

[0036] Relative root mean square error rRMSE: Formula 6: ;

[0037] Where TP represents the number of samples correctly predicted as positive by the model, FP represents the number of samples that were actually negative but were incorrectly predicted as positive by the model, and FN represents the number of samples that were actually positive but were incorrectly predicted as negative by the model. This represents the actual value of the i-th observation. This represents the predicted value of the i-th observation. This represents the average value of the observed values.

[0038] S3 specifically includes:

[0039] S3-1, the backbone network uses the C2f_DAttention module to enhance the dynamic attention mechanism in the feature extraction process, thereby improving the ability to capture details of corn leaves;

[0040] S3-2, the network neck is a bidirectional feature pyramid network BiFPN, and a 3×3 convolutional layer is added before each multi-scale feature fusion to reduce computational cost;

[0041] S3-3 uses the EfficientHead module in the network header, which reduces computational load and improves detection accuracy through shared convolution operations and multi-scale information fusion.

[0042] S3-4 describes the implementation of Layer Adaptive Model Pruning (LAMP) on the YOLOv8n-LP identification model. This dynamically adjusts the pruning rate of each layer, removes unimportant layers, and retains the parameters of key layers. This approach enables high-precision and high-efficiency detection of maize plants and further reduces computational requirements.

[0043] S5 specifically includes:

[0044] S5-1, when counting seedlings, firstly, cropped images of individual corn plants are generated, and leaf tips are labeled using a tagging method, thereby constructing a new dataset related to leaf age;

[0045] S5-2, a new dataset related to leaf age, includes a near-ground NG platform dataset and a UAV dataset, which are independent of each other and are divided into training, validation and test sets in an 8:1:1 ratio for training the model.

[0046] S5-3 first predicts the sum of leaf ages of all individual corn plants, then estimates the average leaf age of the plot by dividing the sum of leaf ages by the predicted number of plants. The true leaf age of the plot is obtained by dividing the sum of the true leaf ages at the time of labeling by the true number of plants.

[0047] The near-ground equipment consists of digital cameras fixed on high poles.

[0048] Stages V2 through V5 represent the stages from two fully unfolded leaves to five fully unfolded leaves.

[0049] Data augmentation processing includes brightness adjustment, histogram equalization, white balance, and color setting adjustments.

[0050] An electronic device includes a processor and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform a deep learning method for recognizing the number of corn seedlings and leaf age.

[0051] A non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute a deep learning method for identifying the number of corn seedlings and leaf age.

[0052] Compared with existing technologies, this invention acquires and processes data on corn seedlings using near-ground equipment and drone cameras, and then uses the improved YOLOv8n-LP recognition model to accurately predict the number of corn seedlings and leaf age. Subsequently, a combination of multiple evaluation indicators and multiple models was used for comparative analysis to verify the high precision, high performance, small model space ratio, and high adaptability of this invention. This enables large-scale standardized monitoring of corn seedling growth, which can not only significantly improve farmland management efficiency, but also provide scientific basis for crop production, and contribute to the intelligent and sustainable development of agriculture. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of near-ground data acquisition according to the present invention.

[0054] Figure 2 This is a schematic diagram of the UAV data acquisition and processing of the present invention.

[0055] Figure 3 This is a schematic diagram of the UAV data augmentation process of the present invention.

[0056] Figure 4 This is a schematic diagram of the YOLOv8n-LP recognition model of the present invention.

[0057] Figure 5 This is a schematic diagram of the near-ground image detection results for plant counting according to the present invention.

[0058] Figure 6 This is a schematic diagram of the image detection results of the plant counting drone of the present invention.

[0059] Figure 7 This is a schematic diagram of the leaf age estimation and detection process of the present invention.

[0060] Figure 8Comparison of evaluation results of various maize seedling detection models of the present invention Figure 1 .

[0061] Figure 9 Comparison of evaluation results of various maize seedling detection models of the present invention Figure 2 .

[0062] Figure 10 The following are schematic diagrams comparing the corn seedling counts of the model of the present invention under various conditions on different platforms: Figure (a) shows the YOLOv8n-L model applied to near-ground data, Figure (b) shows the YOLOv8n-LP model applied to near-ground data, Figure (c) shows the YOLOv8n-L model applied to UAV data, and Figure (d) shows the YOLOv8n-LP model applied to UAV data.

[0063] Figure 11 The confusion matrix for leaf age estimation using the model of this invention is shown in Figure (a), where Figure (b) shows the YOLOv8n-L model applied to near-ground data, Figure (c) shows the YOLOv8n-LP model applied to near-ground data, and Figure (d) shows the YOLOv8n-L model applied to UAV data.

[0064] Figure 12 This is a schematic diagram of the Precision-Recall curve of the present invention. Detailed Implementation

[0065] The present invention will now be further described with reference to the accompanying drawings.

[0066] See Figures 1-12 This invention provides a deep learning method for identifying the number of corn seedlings and leaf age:

[0067] I. Acquiring the original image:

[0068] like Figures 1-2 As shown, to ensure the accuracy and applicability of the analysis results, near-ground equipment (a Daheng MER2-302-56U3M / C digital camera fixed on a 4.1-meter-high pole) and a DJI M600-Pro drone equipped with a Sony α7II digital camera were used to acquire data on corn seedlings.

[0069] Data acquisition progressed from stage V2 (two fully unfolded leaves) to stage V5 (five fully unfolded leaves), collecting a total of 1918 near-ground images and 744 UAV images. The original UAV images were stitched together using Agisoft Metashape software, and then cropped using ArcGIS 10.8 software to obtain the cell images.

[0070] II. Training of the corn seedling stage identification model:

[0071] The processed images were organized into two datasets: near-ground (NG) and unmanned aerial vehicle (UAV). These two datasets were independent of each other and were divided into training, validation and test sets in an 8:1:1 ratio to train the model.

[0072] The LabelImg tool was used to draw bounding boxes around each corn seedling for manual annotation. The NG dataset contains 1918 images (annotating 74887 seedlings), and the drone dataset contains 744 images (annotating 90176 seedlings). Due to the relatively small amount of drone data, brightness adjustments were made to increase the diversity and quantity of the training set, enabling the model to learn more powerful features. Figure 3 a) Histogram equalization ( Figure 3 b) White balance ( Figure 3 c) and color settings adjustment ( Figure 3 d) Data augmentation and processing.

[0073] III. Recognition Model: YOLOv8n-Light-Pruned (YOLOv8n-LP)

[0074] YOLOv8n-LP is an optimized variant of YOLOv8n that reduces the number of parameters without compromising accuracy by adjusting the backbone, neck, and head components.

[0075] In the backbone network, the C2f_DAttention module was adopted to enhance the dynamic attention mechanism in the feature extraction process and improve the ability to capture details of corn leaves.

[0076] The network neck was replaced with a bidirectional feature pyramid network (BiFPN), and a 3x3 convolutional layer was added before each multi-scale feature fusion to reduce computational cost.

[0077] In the network header, the EfficientHead module is used, which reduces the computational load and improves the detection accuracy by sharing convolution operations and fusing multi-scale information.

[0078] In addition, to further reduce computational requirements, Layer Adaptive Model Pruning (LAMP) was implemented on the model, dynamically adjusting the pruning rate of each layer, removing unimportant layers, and retaining the parameters of key layers to ensure high accuracy and efficiency in maize plant detection.

[0079] IV. Count of corn seedlings:

[0080] The trained model is used to predict the number of corn seedlings in the plot images that need to be counted, generating images with bounding boxes for individual plants and saving the corresponding box location information to a txt file. Each line records the confidence score and location information of a box. By counting the number of lines in the txt file, the predicted number of corn seedlings can be obtained.

[0081] The true number of plants per image was obtained by labeling the images and generating bounding box information before training, and then statistically counting was performed using the same method as the prediction results. By comparing the prediction results and the true number of plants, the accuracy of the corn seedling count results can be evaluated.

[0082] V. Leaf Age Estimation in Maize Seedlings: Using the NG platform and UAV dataset, the leaf tips of maize plants are identified based on the proposed model to estimate the leaf age of individual maize plants and plots.

[0083] The steps are similar to those for plant counting. When counting seedlings, a cropped image of a single corn plant is first generated, and the leaf tips are labeled using a tagging method, thereby constructing a new dataset related to leaf age.

[0084] The NG dataset contains 17,712 images of individual maize plants, with 15,940 in the training set and 1,772 in the test set. The UAV platform dataset contains 14,244 images, with 12,819 in the training set (25,639 after data augmentation) and 1,425 in the test set. Similarly, these two leaf age datasets were divided into training, validation, and test sets in an 8:1:1 ratio and trained using the YOLOv8n-LP model. This process first predicts the leaf age of individual maize plants, then estimates the average leaf age of the plot by dividing the leaf age by the predicted number of plants. The true leaf age of the plot is obtained by dividing the true leaf age at the time of annotation by the actual number of plants.

[0085] VI. Verification of Measurement Results:

[0086] The performance of the model is evaluated using metrics such as precision, average precision, and recall. The accuracy of the model's recognition results is evaluated using the coefficient of determination (R²), root mean square error (RMSE), and relative root mean square error (rRMSE).

[0087] Precision refers to the proportion of samples correctly predicted as positive (TP) out of all samples predicted as positive (TP+FP). A higher precision indicates a more accurate prediction of positive samples. The expression is as follows:

[0088] Accuracy: Formula 1: ;

[0089] Recall is the proportion of samples correctly predicted as positive (TP) to the total number of samples that were actually positive (TP+FN). A higher recall indicates a stronger ability of the model to correctly identify actual positive samples. The expression is as follows:

[0090] Recall rate: Formula 2: ;

[0091] Average precision (AP) is the average precision of a model across different recall thresholds, reflecting the overall performance of an object detection or classification model at different decision thresholds. A higher AP indicates that the model maintains high precision across different recall levels. Its calculation expression is as follows:

[0092] Average accuracy: Formula 3: ;

[0093] Formula 3 is an integral expression. It is the differential unit in an integral expression and has no special meaning;

[0094] It is an algebraic symbol, in Formula 3. As a whole, the Precision-Recall curve (e.g.) Figure 12 The smooth form of r' (as shown) is transformed into a monotonically decreasing function. Here, r' can be replaced by any symbol without affecting the calculation result.

[0095] The coefficient of determination (R²) measures the goodness of fit between the model's predicted and actual values, assessing the model's ability to interpret data. A R² value closer to 1 indicates that the model effectively explains data variations and provides better predictions. This represents the actual value of the i-th observation. This represents the predicted value of the i-th observation. This represents the average of the observed values. Its calculation expression is as follows:

[0096] Coefficient of determination R 2 Formula 4: ;

[0097] Root Mean Square Error (RMSE) is the root mean square of the error between the model's predicted values ​​and the actual values, measuring the magnitude of the model's prediction error. A smaller RMSE indicates a lower prediction error and a better fit. Its calculation expression is as follows:

[0098] Root mean square error: Formula 5: ;

[0099] The relative root mean square error (rRMSE) is the ratio of the root mean square error (RMSE) to the mean of the true values. It measures the relative magnitude of the model's prediction error relative to the actual observed values. The smaller the rRMSE, the smaller the model's prediction error relative to the fluctuations in the data, and the better the fit.

[0100] Relative root mean square error: Formula 6: ;

[0101] VII. Model Comparison Statistics:

[0102] 1. Model Comparison:

[0103] Comparing YOLOv8n-LP with other classic recognition models, such as... Figures 8-9 The following table shows the evaluation results of various maize seedling detection models:

[0104] Among existing models, YOLOv8n achieves the highest accuracy at 0.945, with an R² of 0.90 and an rRMSE of 8.53%. The proposed YOLOv8n-LP model maintains high performance (p=0.953, rRMSE=6.25%) while significantly reducing the model size (1.8MB) and proving highly efficient in practical applications.

[0105] 2. Count of corn seedlings:

[0106] The proposed YOLOv8n-L and YOLOv8n-LP models were observed on different platforms ( Figure 7 The maize seedling counting performance was excellent under various conditions, including leaf age, image resolution, seedling distribution, and planting density. Table 1 summarizes the R², RMSE, and rRMSE indices for each experiment.

[0107]

[0108] The results show that the YOLOv8n-L and YOLOv8n-LP models exhibit consistent performance trends under different factors, and their rRMSE is below 10%.

[0109] 3. Estimation of leaf age during the corn seedling stage:

[0110] Considering that field crop management is usually carried out at the plot scale, this embodiment further evaluated the accuracy of estimating leaf age in plot images based on single-plant leaf age data. Experimental results show that both models exhibit good adaptability on different data acquisition platforms. Table 2 summarizes the R², RMSE, and rRMSE indices for each experiment.

[0111]

[0112] The results show that YOLOv8n-LP performs well on both the NG and UAV datasets, achieving accurate estimation results with rRMSE of 5.73% and 9.24%, respectively.

[0113] Confusion matrix ( Figure 11 The data shows that the actual values ​​are highly consistent with the predicted values, and most values ​​are clustered on the diagonal. This clustering indicates that the model's predictions are generally accurate. Figure 11 a and Figure 11 b represents the performance of the YOLOv8n-L model on the NG and UAV datasets, respectively. Figure 11 c and Figure 11 d represents the performance of the YOLOv8n-LP model on the NG and UAV datasets, respectively.

[0114] The confusion matrix is ​​used to demonstrate the accuracy of the classification model. In this confusion matrix, the horizontal axis is the true leaf age, the vertical axis is the leaf age predicted by the model, and the number is the number of samples belonging to that category. For example, in (a), the top left corner 217 corresponds to V2 on both the horizontal and vertical axes, indicating that there are 217 samples that actually belong to the V2 leaf age and are also detected as V2 leaf age by the model.

[0115] The above are merely preferred embodiments of the present invention, intended only to aid in understanding the method and core ideas of this application. The scope of protection of the present invention is not limited to the above embodiments; all technical solutions falling within the scope of the present invention's concept are within its protection. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

[0116] This invention comprehensively addresses the problems of insufficient accuracy and low efficiency caused by the reliance on manual methods for monitoring maize seedlings in existing technologies, as well as the inability of traditional remote sensing methods to adapt to complex field environments, resulting in insufficient accuracy, slow precision identification and processing due to the lack of specific feature optimization, and inability to adapt to large-scale monitoring and decision-making. By constructing a lightweight, intelligent automatic identification system, it achieves large-scale standardized monitoring of maize seedling growth. Furthermore, by promoting the widespread application of automation technology in the field of precision agriculture, this invention can not only significantly improve the efficiency of farmland management but also provide scientific basis for crop production, contributing to the intelligent and sustainable development of agriculture.

Claims

1. A deep learning method for identifying the number of corn seedlings and leaf age, characterized in that, Includes the following steps: S1, Original Image Acquisition: S1-1, Data on corn seedlings is acquired using near-ground equipment and drone cameras. The data consists of near-ground images and drone footage from stages V2 to V5. The V2 stage is the two-leaf stage, which is the stage where two corn leaves are fully unfolded; the V5 stage is the five-leaf stage, which is the stage where five corn leaves are fully unfolded. S1-2, After stitching the UAV images, the stitched images are then cropped to obtain the plot images of the corn seedlings that need to be identified. S2, Training the recognition model: S2-1, the processed images are organized into two datasets: near-ground NG and UAV; S2-2, Draw a bounding box around each corn seedling and manually label it; S2-3, In order to increase the diversity and quantity of the UAV impact data training set, the UAV image data is subjected to data augmentation processing; S2-4, the near-ground NG dataset and the UAV dataset are independent of each other, and both are divided into training set, validation set and test set in an 8:1:1 ratio for training the model; S3, recognition model YOLOv8n-LP: The recognition model YOLOv8n-LP is based on YOLOv8n, which adjusts the backbone network, network neck and network head components to reduce the number of parameters without affecting accuracy. S4, Corn seedling plant count: S4-1: Use the trained model to predict the plot images of corn seedlings that need to be identified, generate images with individual plant bounding boxes, and save the corresponding box position information to a txt file. S4-2, Each line in the txt file records the confidence level and position information of a box. By counting the number of lines in the txt file, the predicted number of corn seedlings is obtained. S4-3, the actual number of plants in the image of a single plant bounding box in S4-1 is obtained by marking the image and generating box information before training, and the same method as the prediction result is used for statistical counting. S4-4 is used to assess the accuracy of maize seedling count results by comparing predicted results with actual plant counts. S5, estimation of leaf age during corn seedling stage: Using the near-ground NG platform and UAV dataset, the YOLOv8n-LP recognition model is used to identify corn leaf tips, which is then used to estimate the leaf age of individual corn plants and plots. S6, Measurement result verification: The performance of the YOLOv8n-LP recognition model was evaluated using precision, recall, and average precision metrics, and the coefficient of determination R was used as the evaluation metric. 2 The root mean square error (RMSE) and relative root mean square error (rRMSE) are used to evaluate the accuracy of the model recognition results. The specific formulas are as follows: Precision: Formula 1: ; Recall: Formula 2: ; Average Precision: Formula 3: ;in It is the differential unit in the integral. It is an algebraic symbol; Coefficient of determination R 2 Formula 4: ; Root Mean Square Error (RMSE): Formula 5: Relative root mean square error rRMSE: Formula 6: ; Where TP represents the number of samples correctly predicted as positive by the model, FP represents the number of samples that were actually negative but were incorrectly predicted as positive by the model, and FN represents the number of samples that were actually positive but were incorrectly predicted as negative by the model. This represents the actual value of the i-th observation. This represents the predicted value of the i-th observation. This represents the average value of the observed values.

2. The deep learning method for identifying the number of corn seedlings and leaf age according to claim 1, characterized in that, S3 specifically includes: S3-1, The backbone network adopts the C2f_DAttention module to enhance the dynamic attention mechanism in the feature extraction process, thereby improving the ability to capture details of corn leaves; S3-2, the network neck is a bidirectional feature pyramid network BiFPN, and a 3×3 convolutional layer is added before each multi-scale feature fusion to reduce computational cost; S3-3, the network head uses the EfficientHead module, which reduces computational load and improves detection accuracy through shared convolution operations and multi-scale information fusion; S3-4, Layer Adaptive Model Pruning (LAMP) is implemented on the YOLOv8n-LP identification model to dynamically adjust the pruning rate of each layer, remove unimportant layers, and retain the parameters of key layers. This is used for high-precision and high-efficiency detection of maize plants and to further reduce computational requirements.

3. The deep learning method for identifying the number of corn seedlings and leaf age according to claim 1, characterized in that, S5 specifically includes: S5-1, when counting seedlings, firstly, cropped images of individual corn plants are generated, and leaf tips are labeled using a tagging method, thereby constructing a new dataset related to leaf age; S5-2, a new dataset related to leaf age, includes a near-ground NG platform dataset and a UAV dataset, which are independent of each other and are divided into training, validation and test sets in an 8:1:1 ratio for training the model. S5-3 first predicts the sum of leaf ages of all individual corn plants, then estimates the average leaf age of the plot by dividing the sum of leaf ages by the predicted number of plants. The true leaf age of the plot is obtained by dividing the sum of the true leaf ages at the time of labeling by the true number of plants.

4. The deep learning method for identifying the number of corn seedlings and leaf age according to claim 1, characterized in that, The near-ground device is a digital camera fixed on a high pole.

5. The deep learning method for identifying the number of corn seedlings and leaf age according to claim 1, characterized in that, The V2 to V5 stages are the stages with two fully unfolded leaves to five fully unfolded leaves.

6. The deep learning method for identifying the number of corn seedlings and leaf age according to claim 1, characterized in that, The data augmentation process includes brightness adjustment, histogram equalization, white balance adjustment, and color setting adjustment.

7. An electronic device, characterized in that, The device includes a processor and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform a deep learning method for identifying the number of corn seedlings and leaf age as described in any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the deep learning method for identifying the number of corn seedlings and leaf age as described in any one of claims 1 to 6.

Images