Efficient leaf area estimation method

Abstract

The invention discloses an efficient leaf area estimation method based on deep learning, and provides a new solution for solving the problems that destructive sampling is time-consuming and a traditional computer vision technology is easily interfered by the environment in a traditional leaf area measurement method. According to the method, a self-supervised pre-training framework based on a DINOv2 architecture is constructed, and efficient model adaptation is realized in combination with a two-stage transfer learning strategy; developing a Canopy-Mix token-level hybrid enhancement technology, and improving the robustness of a model to a complex shielding scene by dynamically fusing a plurality of canopy images; and designing a hybrid optimization strategy of smooth L1 loss and Log-Cosh loss, and balancing prediction deviation and abnormal value sensitivity. According to the method, the single plant treatment time is shortened, batch treatment can be completed, and the correlation between the prediction result and the fresh weight and dry weight of the leaves is high. The invention provides an efficient and accurate non-destructive detection tool for seedling stage crop breeding phenotype analysis, nutrition early diagnosis and cultivation management.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of agricultural artificial intelligence, and particularly relates to a high-efficiency leaf area estimation method. BACKGROUND

[0002] In the field of agricultural artificial intelligence, the accuracy and efficiency of crop phenotype analysis directly affect the breeding process and cultivation management decision. The leaf area estimation method is the technical basis for crop phenotype analysis, and its quantitative results directly serve the analysis of phenotype characteristics, the construction of growth models and the optimization of agricultural decision-making. The current leaf area estimation method has significant technical defects: (1) Destructive measurement methods, such as LAI punch weighing method and AGB drying method, are time-consuming and labor-intensive and cannot be scaled up.

[0003] (2) Traditional computer vision technology relies on manual feature extraction and is easily affected by environmental light interference, with a relative error of more than 35%.

[0004] (3) Existing AI solutions face two major bottlenecks: first, supervised learning models require massive pixel-level annotation data, which is costly and has weak cross-species generalization ability; second, leaf overlap and occlusion lead to systematic underestimation of predicted values, and point cloud reconstruction solutions have high equipment costs and are difficult to promote.

[0005] Therefore, a new high-efficiency leaf area estimation method is needed. SUMMARY

[0006] The present application proposes an innovative solution to the problems in the prior art, using a self-supervised pre-training framework (DINOv2) combined with a two-stage transfer learning framework, introducing Canopy-Mix token-level mixed enhancement technology, and using a hybrid strategy of smooth L1 and Log-Cosh loss to break through the limitations of traditional methods in sample size, generalization ability and cost-effectiveness.

[0007] The present application proposes a high-efficiency leaf area estimation method, comprising the following steps: S1, collecting data for model training and evaluation, taking a bird's-eye view of the plants in the field using a camera, and then using a transparent glass plate to cover the plant leaves and calibration plate on a black background plate for shooting, calculating the calibration conversion coefficient, and segmenting the leaves to obtain the actual leaf area value; S2, model self-supervised pre-training, constructing a visual Transformer teacher-student network based on the DINOv2 architecture, integrating multiple source plant datasets, generating global and local views, using a cross-entropy loss function, and completing multiple rounds of large-scale self-supervised training; S3, model fine-tuning and leaf area regression, adding a multi-layer perceptron regression head at the end of the visual Transformer teacher-student network, unfreezing the backbone network parameters in stages, using the Canopy-Mix token mixing enhancement technology, and the smooth L1+Log-Cosh hybrid loss, completing the model parameter fine-tuning; S4, model evaluation and verification, using cross-validation to divide the data set, calculating the error index, and verifying the generalization ability and batch processing stability of the model; S5, model deployment and application, real-time calculation of output leaf area value, correlation of fresh weight and dry weight data, support for cultivation management and variety selection.

[0008] Preferably, the S1 comprises: S11, selecting the main camera, setting the main camera parameters, using a vertical bamboo pole with scales as height assistance, controlling the image acquisition environment, and taking a bird's eye view of the plants in the field under natural light environment; S12, using a transparent glass plate to cover the plant leaves and calibration plate on the black background plate for shooting; S13, using visual and machine learning software library to calculate the calibration conversion coefficient, which is the ratio of the actual size of the calibration plate to the number of pixels in the image; S14, segmenting the leaf area based on hue, saturation and lightness; S15, counting the total number of leaf pixels after segmentation, multiplying the calibration conversion coefficient to get the actual leaf area value.

[0009] Preferably, the S2 comprises: S21, selecting DINOv2 self-supervised framework as the technical base, and constructing teacher-student network based on visual Transformer architecture; S22, integrating multi-source heterogeneous plant data set for pre-training, establishing the cross-species generalization ability of the model; S23, performing large-scale self-supervised training, configuring multiple iteration cycles and batch processing image quantity, and outputting a pre-trained model with universal plant shape perception ability.

[0010] Preferably, the S3 comprises: S31, adding two layers of multi-layer perceptron regression head at the end of the pre-trained visual Transformer teacher-student network, freezing the backbone network parameters, using AdamW optimizer, configuring hierarchical learning rate, combining cosine annealing scheduler, and realizing the initial mapping of leaf area and model output; S32, the last four layers of the visual Transformer teacher-student network are thawed, and the regression head is trained end-to-end, using the Canopy-Mix token mixing enhancement technology, and the smooth L1 and Log-Cosh mixed loss weight ratio is 0.7:0.3; The Canopy-Mix token mixing enhancement technology comprises: (1) Randomly mixing the Transformer token sequences of global and local images by Beta distribution to generate synthetic samples with multi-image features; (2) Performing weighted averaging on leaf area label values using the same mixing ratio as the token.

[0011] Preferably, the S4 comprises: S41, randomly dividing the samples collected in the S1 step into multiple groups, using cross-validation, and performing multiple rounds of iteration; S42, calculating error indicators during each round of iteration to form an evaluation report and evaluate the model effect; The error indicators include the coefficient of determination, the mean absolute error, the root mean square error, and the relative root mean square error.

[0012] Preferably, the S5 comprises: S51, using a mobile phone or a drone device to take aerial photos of plants in the field; S52, loading the fine-tuned parameter prediction model, receiving image input, and outputting leaf area values to the terminal or cloud in real time; S53, analyzing and applying the prediction output results, correlating leaf fresh weight and dry weight data, and supporting cultivation management and variety selection.

[0013] The efficient leaf area estimation method has the following beneficial effects: (1) Learning the general morphological features of non-oilseed rape plants through a self-supervised pre-training framework (DINOv2), combined with Canopy-Mix token-level mixing enhancement, the model's cross-species / environment generalization ability is significantly improved, solving the domain shift problem of traditional ImageNet pre-training models.

[0014] (2) The mixed strategy of smooth L1 and Log-Cosh loss effectively reduces the influence of outliers and maintains high robustness in complex field backgrounds (such as soil noise and leaf overlap).

[0015] (3) Only a small amount of labeled samples are needed to complete model fine-tuning, greatly reducing labeling costs. Using a smartphone or a drone platform instead of traditional multi-view shooting devices significantly reduces equipment investment.

[0016] (4) The measurement time of single plant leaf area is shortened from 5 minutes of traditional destructive method to 5 seconds, and the work efficiency is significantly improved. The batch processing capacity reaches 1000 plants / hour, which meets the demand of large-scale breeding phenotype analysis.

[0017] The high-efficiency leaf area estimation method provided by the present application provides an efficient, accurate and intelligent tool for plant breeding phenotype analysis. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A high-efficiency leaf area estimation method according to an embodiment of the present application is shown in the flowchart. Figure 2 A leaf flat diagram of a high-efficiency leaf area estimation method according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0019] In order to further understand the present application, the embodiments of the present application are described below in combination with examples, but it should be understood that these descriptions are only for further illustrating the features and advantages of the present application, and are not limitations on the claims of the present application.

[0020] The following abbreviations or terms are used in the embodiments of the present application: LAI: Leaf Area Index, a key physiological indicator reflecting the total area of plant leaves per unit of land area.

[0021] AGB: Above Ground Biomass, a core parameter for crop growth monitoring.

[0022] SSL: Self-Supervised Learning, a machine learning paradigm that can learn visual representations without labeled data.

[0023] ViT: Vision Transformer, an image processing model architecture based on self-attention mechanism.

[0024] DINOv2: Self-DIstillation with NO labels v2, an algorithm for self-supervised pre-training.

[0025] RRMSE: Relative Root Mean Square Error, the percentage of the prediction error relative to the mean of the true value.

[0026] Canopy-Mix: This invention employs a token-level hybrid enhancement technique that generates enhanced samples with multi-image features by mixing the Transformer token sequences of two images according to a Beta distribution, thereby improving the robustness of the model to complex occlusion scenarios.

[0027] An embodiment of the present invention provides an efficient leaf area estimation method, such as... Figure 1 As shown, it includes the following steps: S1. Collect data for model training and evaluation. This includes the following sub-steps: S11. Select the main camera, set its parameters, and use a vertical bamboo pole with markings as a height aid to control the image acquisition environment. Perform an overhead shot of the field under natural light conditions.

[0028] In this embodiment of the invention, a commercial mobile phone is selected as the image acquisition device. A main camera is used, configured with a large aperture of f / 1.42, a shutter speed of 1 / 954s, and an ISO sensitivity of 50 to ensure image clarity. During shooting, a bamboo pole with a 60cm graduation is used to mark the shooting height to avoid image distortion caused by drastic changes in height.

[0029] rape( Brassica napus The L. (rapeseed) was sown on October 20, 2024, and the seedlings were transplanted to the field on November 15, 2024, at a planting density of 225,000 plants / hectare. The soil type was paddy soil, and the fertilizer application rate was 750 kg / hectare of rapeseed-specific slow-release fertilizer (N-P2O5-K2O=25:7:8). Field management practices for all plants were consistent throughout the growing season.

[0030] This embodiment of the invention was photographed from above on January 1, 2025, collecting 833 rapeseed plant samples from seedlings to the 5-leaf stage. The main differences between the samples were leaf area, leaf shape, and growth status. A 4cm×3cm calibration board was placed next to each sample to establish a conversion benchmark between pixels and actual size, ensuring the accuracy of subsequent area calculations. The overhead photography was conducted under natural light in the field, using a bamboo pole with a 60cm graduation to mark the shooting height, and keeping the mobile phone parallel to the ground during shooting.

[0031] S12. Use a transparent glass plate to cover the plant's leaves and calibration plate on a black background for photography.

[0032] After filming in the field, such as Figure 2 As shown, the leaves of the plant were picked and laid flat on a black background along with the calibration board to ensure high contrast. The leaves were then pressed down with a transparent glass plate to keep them spread out before taking the picture.

[0033] S13, Use visual and machine learning software library to identify calibration board and calculate proportion.

[0034] Use OpenCV library to detect the edge of the calibration board, and establish the conversion coefficient from pixels to centimeters by calculating the proportion of the actual size of the calibration board to the number of pixels in the image. OpenCV is an open-source cross-platform computer vision and machine learning software library, mainly used for real-time image processing and visual application development.

[0035] S14, Leaf area segmentation.

[0036] Based on the HSV color space model (hue 35-75, saturation 80-255, lightness 50-255), extract the leaf pixels, and use threshold segmentation method to filter out the soil background. HSV color space is a three-dimensional color model based on hue (Hue), saturation (Saturation) and lightness (Value). Morphological operations can also be used to optimize the segmentation results and ensure the integrity of the leaf contour.

[0037] S15, Pixel area conversion.

[0038] Count the total number of pixels of the segmented leaf, and multiply it by the calibration conversion coefficient to get the actual leaf area value. This step realizes non-contact measurement through calibration board calibration, replacing traditional punch method for leaf area measurement and other manual measurement methods.

[0039] Step S1 provides high-precision, structured input data for model training through standardized acquisition process and automation processing, which is the basis for building an efficient leaf area estimation system.

[0040] S2, Self-supervised pre-training.

[0041] Step S2 completes the function of building basic feature learning capability in the whole leaf area estimation process. This step pre-trains non-target crop images to enable the model to understand plant morphology across species, providing initialization parameter support for subsequent accurate regression of rape leaf area. The specific implementation includes the following sub-steps: S21, Basic framework construction and parameter initialization.

[0042] Select DINOv2 self-supervised framework as the technical base, which is based on a 12-layer 768-dimensional visual Transformer architecture (ViT-Base) to build a teacher-student network. The teacher network weight is updated through the exponential moving average mechanism, and the momentum coefficient is gradually linearly increased from the initial 0.996 to 1, ensuring the stability of feature representation in the knowledge distillation process. This stage needs to complete the model architecture code implementation, GPU resource allocation and initial weight randomization settings.

[0043] S22, Multi-source heterogeneous data integration and enhancement.

[0044] To construct a robust and general feature extractor, the embodiments of the present application integrate five public plant datasets images. These datasets do not contain oilseed rape, aiming to let the model learn diverse plant morphological features, forming a pre-training data pool containing 25535 non-oilseed rape images. The public plant datasets include: (1) CVPPP dataset: 810 Arabidopsis and tobacco rosette leaf overhead views, showing basic overlapping patterns.

[0045] (2) Flavia dataset: 1,907 single leaf scans of 32 species, showing fine leaf shape and vein details.

[0046] (3) Plant Pathology 2021 dataset: 18,635 field apple leaf images, covering different health conditions.

[0047] (4) Plant Seedlings dataset: 408 overhead views of 12 crop / weed seedlings.

[0048] (5) VegAnn dataset: 3,775 images of complex field scenes of 26+ vegetable crops.

[0049] Two spatial transformations are performed on each original image: a global view of 224x224 pixels is generated to maintain overall structural awareness, and a local view of 96x96 pixels is generated simultaneously to enhance detail sensitivity. By constraining the consistency of probability distribution between different views through cross-entropy loss function, a feature learning scenario with spatial invariance is constructed.

[0050] S23, perform large-scale self-supervised training.

[0051] Configure 300 iteration cycles, process 256 images per batch, set the initial learning rate to 1e-4, and use the cosine annealing strategy. The temperature parameter τ is controlled at 0.07 to balance feature discrimination and uniformity. During the training process, the KL divergence loss of the teacher network and the student network is monitored in real time, and when the loss decreases by less than 0.1% for 5 consecutive rounds, the early stopping mechanism is automatically triggered. Finally, a pre-trained model with general plant morphological perception ability is output, which solves the domain bias problem of traditional ImageNet (a large visual database for visual object recognition software research) pre-trained model in leaf edge detection, texture feature extraction, etc. intermediate layer representation, and improves the domain adaptation ability.

[0052] S3: Model fine-tuning and leaf area regression.

[0053] Step S3 completes the function of converting general plant morphological features into specific leaf area prediction capability. This step realizes efficient adaptation of the pre-trained model on the rapeseed data through phased parameter unfreezing and joint optimization strategy, and finally constructs a leaf area estimation model with centimeter-level precision. The specific implementation includes the following sub-steps: S31, regression head initialization and feature migration.

[0054] Two layers of multilayer perceptron (MLP) are added at the end of the ViT-Base backbone network as a specific regression head, with a hidden layer dimension of 512 and a ReLU activation function. After loading the pre-trained visual Transformer weights in step S2, all backbone network parameters are frozen, and only the regression head is trained.

[0055] The optimizer is AdamW (a deep learning adaptive optimizer), which is configured with a hierarchical learning rate strategy: the learning rate of the backbone network is set to 5e-5, and the learning rate of the regression head is increased to 5e-4, with a cosine annealing scheduler to realize learning rate smooth decay. This stage focuses on regression head training, and through the first 50 iterations, the mapping relationship between leaf area and model output is optimized to ensure effective use of pre-trained features.

[0056] S32, perform full-scale fine-tuning and joint optimization.

[0057] Unfreeze the last four layers of the ViT-Base network parameters, and perform end-to-end joint optimization with the trained regression head. Keep the AdamW optimizer configuration, and dynamically monitor the validation set performance. When the validation loss does not decrease for five consecutive rounds, trigger the early stopping mechanism.

[0058] The Canopy-Mix token-level mixing enhancement technology is introduced, which mixes the Transformer token sequences of the global and local two images according to the Beta distribution (α=0.5, β=0.5), generates enhanced samples with multi-image features, and synchronously mixes the leaf area label values. In the embodiment of the present invention, the Canopy-Mix technology generates synthetic samples with multi-image features by randomly combining token sequences from different images through token-level mixing. This mixing method not only retains part of the original image features, but also introduces new combinations, forcing the model to learn more general and abstract feature representations.

[0059] The loss function adopts a mixed strategy of smooth L1 (handling outliers) and Log-Cosh (improving gradient stability), with a weight distribution of 0.7:0.3, effectively suppressing the interference of outliers on gradient updates. After 150 rounds of subsequent iterations, the model reaches a precision level of R²=0.805 and MAE=22.207 cm² in five-fold cross-validation.

[0060] S4, model evaluation and verification.

[0061] Step S4 completes the function of quantifying model performance and ensuring the reliability of actual application. This step verifies the precision and generalization ability of the model in the rapeseed leaf area estimation task through five-fold cross-validation and multi-dimensional performance statistics, providing quantitative basis for subsequent deployment. The specific implementation includes the following sub-steps: S41, division and configuration of data set.

[0062] In the embodiment of the application, the 833 rapeseed samples collected are randomly divided into 5 groups, each containing plants of different leaf stages and growth states. In each evaluation, 1 group is selected as the test set (20%), and the remaining 4 groups are combined as the training set (80%). Through 5 iterations, each sample is tested once to form a complete cross-validation cycle. This process is realized through a custom data loader, with a built-in random seed control to ensure repeatability.

[0063] S42, perform multi-index performance statistics.

[0064] In each test, the model predicts the leaf area of the test set images, and records the true value and predicted value simultaneously. Four core indicators are calculated: the coefficient of determination (R²) measures the degree of linear correlation, the mean absolute error (MAE) reflects the absolute error level, the root mean square error (RMSE) assesses the sensitivity to outliers, and the relative root mean square error (RRMSE) standardizes the error proportion. After five rounds of cross-validation, the average values of each indicator are summarized to form the final evaluation report. In a typical experiment, this method improves the prediction correlation by 26.6% compared to traditional methods.

[0065] S5, model deployment and application.

[0066] This step realizes the real-time application of the leaf area estimation model in the field environment through standardized device configuration, precise calibration system and efficient inference process. The specific implementation includes the following sub-steps: S51, use a mobile phone or a drone to take pictures of the plants from above in the field, which is simple and easy to operate.

[0067] S52, model inference process and prediction output.

[0068] The established model outputs the leaf area value in cm² for the photographed pictures. The prediction results can be displayed in real time on the mobile phone screen and transmitted to the cloud through Bluetooth or Wi-Fi, supporting a throughput of 1000 plants / hour.

[0069] S53, analysis and application of prediction output results.

[0070] The output result can be associated with fresh weight and dry weight, and can also include plant number, predicted value and timestamp, facilitating subsequent phenotype analysis. The result directly supports field dynamic cultivation management and breeding screening, and accelerates the breeding process of high-yield and stress-resistant varieties.

[0071] The embodiment of the present application completes the establishment and application of the leaf area prediction model through the above S1-S5 steps. The method in the embodiment is superior to other traditional methods. As shown in Table 1, the determination coefficient R of the predicted leaf area and the true value of the leaf area is 0.805, the correlation coefficient r of the predicted leaf area and the leaf fresh weight and dry weight is 0.900 and 0.885 respectively. 2

[0072] Table 1: Comparison of the effects of the method of the embodiment and the traditional method

[0073] The efficient leaf area estimation method of the embodiment of the present application has high precision and large processing capacity, and can provide efficient and accurate assistance for plant breeding phenotype analysis.

[0074] The above embodiments are only used to help understand the method of the present application and its core idea. It should be pointed out that for ordinary skilled persons in the technical field, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.​

Claims

1. An efficient method for estimating leaf area, characterized in that, Includes the following steps: S1. Collect data for model training and evaluation. Take photos of the plants in the field from above using a camera. Then, use a transparent glass plate to cover the plant leaves and calibration plate on a black background for taking photos. Calculate the calibration conversion coefficient and segment the leaves to obtain the actual leaf area value. S2. Self-supervised pre-training of the model: Based on the DINOv2 architecture, a visual Transformer teacher-student network is built, integrating multi-source plant datasets to generate global and local views. The cross-entropy loss function is used to complete multiple rounds of large-scale self-supervised training. S3. Model fine-tuning and leaf area regression: Add a multilayer perceptron regression head to the end of the visual Transformer teacher-student network, unfreeze the backbone network parameters in stages, use Canopy-Mix token fusion enhancement technology and smooth L1+Log-Cosh fusion loss to complete the model parameter fine-tuning; S4. Model evaluation and validation: Cross-validation is used to divide the dataset, and statistical error indicators are used to validate the model's generalization ability and batch processing stability. S5, Model Deployment and Application: Real-time calculation and output of leaf area value, correlation of fresh weight and dry weight data, supporting cultivation management and variety breeding.

2. The efficient leaf area estimation method according to claim 1, characterized in that, S1 includes: S11. Select the main camera, set the main camera parameters, use a vertical bamboo pole with scales as a height aid, control the image acquisition environment, and take a bird's-eye view of the plants in the field under natural light. S12. Use a transparent glass plate to cover the plant leaves and the calibration plate on a black background for photography. S13. Using a vision and machine learning software library, calculate the calibration conversion coefficient; the calibration conversion coefficient is the ratio of the actual size of the calibration board to the number of pixels in the image. S14. Segmenting leaf regions based on hue, saturation, and brightness; S15. Count the total number of leaf pixels after segmentation, multiply by the calibration conversion coefficient to obtain the actual leaf area value.

3. The efficient leaf area estimation method according to claim 1, characterized in that, S2 includes: S21. The DINOv2 self-supervised framework was selected as the technical foundation, and a teacher-student network was built based on the visual Transformer architecture. S22. Integrate multi-source heterogeneous plant datasets for pre-training to establish the model's cross-species generalization ability; S23. Perform large-scale self-supervised training, configure multiple iteration cycles and batch processing image quantity, and output a pre-trained model with general plant morphology perception capabilities.

4. The efficient leaf area estimation method according to claim 1, characterized in that, S3 includes: S31. Add two multilayer perceptron regression heads to the end of the pre-trained visual Transformer teacher-student network, freeze the backbone network parameters, use the AdamW optimizer, configure the hierarchical learning rate, and combine the cosine annealing scheduler to achieve the initial mapping between leaf area and model output. S32. Unfreeze the parameters of the last four layers of the visual Transformer teacher-student network, perform end-to-end joint training with the regression head, use the Canopy-Mix token hybrid augmentation technique, and set the weight ratio of smoothed L1 and Log-Cosh hybrid loss to 0.7:0.

3. The Canopy-Mix token hybrid enhancement technology includes: (1) By randomly mixing the Transformer token sequences of global and local images using the Beta distribution, a synthetic sample with multi-image features is generated. (2) Perform a weighted average of the leaf area label values ​​using the same mixing ratio as the token.

5. The efficient leaf area estimation method according to claim 1, characterized in that, S4 includes: S41. The samples collected in step S1 are randomly divided into multiple groups, and cross-validation is used to perform multiple rounds of iteration. S42. Calculate the error index during each iteration, generate an evaluation report, and evaluate the model's performance. The error indices include the coefficient of determination, mean absolute error, root mean square error, and relative root mean square error.

6. The efficient leaf area estimation method according to any one of claims 1-5, characterized in that, S5 includes: S51. Use a mobile phone or drone to take aerial photos of the plants in the field. S52: Load the prediction model with fine-tuned parameters, receive image input, and output the leaf area value to the terminal or cloud in real time. S53. Analyze and apply the predicted output results, correlate leaf fresh weight and dry weight data, and support cultivation management and variety breeding.

Images