Tea seedling drought detection method and system combined with chlorophyll fluorescence imaging

Abstract

The invention provides a tea seedling drought detection method and system combined with chlorophyll fluorescence imaging in the technical field of tea seedling planting management, and the method comprises the steps: S1, obtaining chlorophyll fluorescence historical images of a large number of tea seedlings under different drought grades, performing preprocessing including data annotation and data expansion on each historical image and then constructing a data set; s2, creating a tea seedling drought detection model based on a YOLOv8n network, and setting a loss function of the tea seedling drought detection model as an MPD IoU function; s3, dividing the data set into a training set, a verification set and a test set based on a preset proportion, and respectively training, verifying and testing the tea seedling drought detection model through the training set, the verification set and the test set; and S4, deploying the tea seedling drought detection model passing the test, and performing tea seedling drought detection by using the deployed tea seedling drought detection model. The method has the advantage that the precision of tea seedling drought detection is greatly improved.

Description

Technical Field

[0001] The invention relates to the technical field of tea seedling planting management, and in particular to a tea seedling drought detection method and system combined with chlorophyll fluorescence imaging. Background Art

[0002] Water is an indispensable condition for plant growth and photosynthesis, and plays a vital role in the absorption and transportation of plant nutrients. Water shortage will reduce the efficiency of plant photosynthesis and hinder transpiration. In severe cases, plants will wither and die. With the volatile global climate and frequent natural disasters, especially the increase in drought disasters, agricultural production has been severely negatively affected. In this context, studying the response mechanism of plants to drought stress is of great significance for improving the drought resistance of crops and ensuring the sustainable development of agriculture.

[0003] Drought stress causes the net photosynthetic rate, transpiration rate, and stomatal conductance of tea seedlings to decrease, while the intercellular CO2 concentration increases, malondialdehyde (MDA) and peroxidase (POD) increase, causing varying degrees of damage to the yield and quality of tea. In view of this, in-depth research on the physiological and biochemical responses of tea seedlings to drought stress is beneficial to the cultivation and management of tea seedlings and lays the foundation for the timely prevention and control of drought.

[0004] With the rapid development of machine learning and deep learning technologies, the application of related technologies in the agricultural field is becoming more and more extensive. High-throughput phenotyping technology has the ability to quickly, non-destructively, and multi-dimensionally analyze plant characteristics, effectively improving the efficiency of crop breeding and environmental response research. As an emerging non-destructive detection method, chlorophyll fluorescence imaging technology has been proven to be of great value in monitoring plant physiological status. Dong et al. used chlorophyll fluorescence parameters, texture features, etc. to describe image features, and calculated the Pearson correlation of features. The selected features were used as inputs to the BP neural network model to determine the chilling injury level of tomato seedlings, with a detection accuracy of more than 90%. Zhang et al. successfully identified the state of rice plants under herbicide and heavy metal stress through hyperspectral imaging and chlorophyll fluorescence imaging technology, providing an effective new method for plant stress phenotyping analysis. By establishing a stress detection model, a deep fusion model combined with a convolutional neural network (CNN) was achieved to achieve a detection accuracy of 97.7%. The chlorophyll fluorescence imaging (CFI) system constructed by Muhammad Akbar Andi Arief et al. effectively measures the maximum photochemical quantum efficiency (Fv / Fm) of strawberry plants under different abiotic stresses, accurately captures the plant's response to stress, and demonstrates its potential for application in plant health monitoring. Kaihua Wei et al. used fluorescence imaging combined with convolutional neural networks (CNN) to classify five types of tea. By preprocessing tea samples and training with VGG16 and ResNet-34 models, they found that fluorescence images significantly surpassed traditional white light images in classification accuracy, achieving a high accuracy rate of 97.5%. However, in the field of tea science, the application of chlorophyll fluorescence imaging technology is limited, and existing research has certain problems in the accuracy of drought stress identification.

[0005] Therefore, how to provide a tea seedling drought detection method and system combined with chlorophyll fluorescence imaging to improve the accuracy of tea seedling drought detection has become a technical problem that needs to be solved urgently. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a tea seedling drought detection method and system combined with chlorophyll fluorescence imaging, so as to improve the accuracy of tea seedling drought detection.

[0007] In a first aspect, the present invention provides a method for detecting drought in tea seedlings combined with chlorophyll fluorescence imaging, comprising the following steps:

[0008] Step S1, obtaining a large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels, and constructing a data set after preprocessing each of the chlorophyll fluorescence historical images including data annotation and data expansion;

[0009] Step S2: creating a tea seedling drought detection model based on the YOLOv8n network, and setting the loss function of the tea seedling drought detection model to the MPDIoU function;

[0010] Step S3, dividing the data set into a training set, a validation set and a test set based on a preset ratio, and respectively training, validating and testing the tea seedling drought detection model through the training set, the validation set and the test set;

[0011] Step S4: deploying the tea seedling drought detection model that has passed the test, and performing tea seedling drought detection using the deployed tea seedling drought detection model.

[0012] Furthermore, the step S1 is specifically as follows:

[0013] Acquire a large number of historical chlorophyll fluorescence images of tea seedlings under different drought levels, wherein the drought level is no drought, mild stress or severe stress;

[0014] Each of the chlorophyll fluorescence historical images is labeled with drought levels and drought areas, and data expansion is performed on the labeled chlorophyll fluorescence historical images, including at least brightness adjustment, contrast adjustment, image flipping, image translation and image rotation, so as to complete preprocessing of each of the chlorophyll fluorescence historical images, and a data set is constructed based on the preprocessed chlorophyll fluorescence historical images.

[0015] Further, in step S2, the YOLOv8n network is composed of an input layer, a backbone network layer, a neck network layer and a detection head, and the input layer, the backbone network layer, the neck network layer and the detection head are connected in sequence;

[0016] The input layer is used to input the chlorophyll fluorescence historical image;

[0017] The backbone network layer is used to extract image features from chlorophyll fluorescence historical images and is constructed based on the MobileViT network;

[0018] The neck network layer is used to upsample the image features to obtain an upsampled feature map, using a PANet structure;

[0019] The detection head is used to output the drought level of tea seedling drought and the predicted probability of drought area according to the upsampled feature map.

[0020] Furthermore, in step S2, the formula of the MPDIoU function is:

[0021]

[0022] in, represents the loss value of the MPDIoU function; MPDIoU represents the intersection-over-union ratio of the minimum point distance; A represents the predicted box; B represents the true box; w represents the width of the chlorophyll fluorescence history image; h represents the height of the chlorophyll fluorescence history image; Indicates the upper left corner of the prediction box; Indicates the lower right corner of the prediction box; Indicates the upper left corner of the real box; Indicates the lower right corner of the ground-truth box.

[0023] Furthermore, the step S3 is specifically as follows:

[0024] The data set is divided into a training set, a validation set and a test set based on a ratio of 7:2:1, and the tea seedling drought detection model is trained by the training set. During the training process, the tea seedling drought detection model is continuously optimized, including at least a learning rate, a learning rate factor, a weight decay, a learning rate warm-up phase, a learning rate warm-up momentum, a warm-up deviation learning rate, a batch size, and a training cycle hyperparameter, until the loss value of the loss function is less than a preset loss threshold;

[0025] The trained tea seedling drought detection model is verified by the verification set to determine whether the recognition accuracy of the tea seedling drought detection model is greater than a preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if yes, the verification passes, and:

[0026] The accuracy, recall rate, average precision mean and balance score of the tea seedling drought detection model are calculated through the test set to test the tea seedling drought detection model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0027] In a second aspect, the present invention provides a tea seedling drought detection system combined with chlorophyll fluorescence imaging, comprising the following modules:

[0028] A data set construction module is used to obtain a large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels, and to construct a data set after preprocessing each of the chlorophyll fluorescence historical images including data annotation and data expansion;

[0029] A tea seedling drought detection model creation module is used to create a tea seedling drought detection model based on the YOLOv8n network, and set the loss function of the tea seedling drought detection model to be the MPDIoU function;

[0030] A tea seedling drought detection model training module is used to divide the data set into a training set, a validation set and a test set based on a preset ratio, and train, validate and test the tea seedling drought detection model through the training set, the validation set and the test set respectively;

[0031] The tea seedling drought detection module is used to deploy the tea seedling drought detection model that has passed the test, and use the deployed tea seedling drought detection model to perform tea seedling drought detection.

[0032] Furthermore, the data set construction module is specifically used for:

[0033] Acquire a large number of historical chlorophyll fluorescence images of tea seedlings under different drought levels, wherein the drought level is no drought, mild stress or severe stress;

[0034] Each of the chlorophyll fluorescence historical images is labeled with drought levels and drought areas, and data expansion is performed on the labeled chlorophyll fluorescence historical images, including at least brightness adjustment, contrast adjustment, image flipping, image translation and image rotation, so as to complete preprocessing of each of the chlorophyll fluorescence historical images, and a data set is constructed based on the preprocessed chlorophyll fluorescence historical images.

[0035] Furthermore, in the tea seedling drought detection model creation module, the YOLOv8n network consists of an input layer, a backbone network layer, a neck network layer and a detection head, and the input layer, the backbone network layer, the neck network layer and the detection head are connected in sequence;

[0036] The input layer is used to input the chlorophyll fluorescence historical image;

[0037] The backbone network layer is used to extract image features from chlorophyll fluorescence historical images and is constructed based on the MobileViT network;

[0038] The neck network layer is used to upsample the image features to obtain an upsampled feature map, using a PANet structure;

[0039] The detection head is used to output the drought level of tea seedling drought and the predicted probability of drought area according to the upsampled feature map.

[0040] Furthermore, in the tea seedling drought detection model creation module, the formula of the MPDIoU function is:

[0041]

[0042] in, represents the loss value of the MPDIoU function; MPDIoU represents the intersection-over-union ratio of the minimum point distance; A represents the predicted box; B represents the true box; w represents the width of the chlorophyll fluorescence history image; h represents the height of the chlorophyll fluorescence history image; Indicates the upper left corner of the prediction box; Indicates the lower right corner of the prediction box; Indicates the upper left corner of the real box; Indicates the lower right corner of the ground-truth box.

[0043] Furthermore, the tea seedling drought detection model training module is specifically used for:

[0044] The data set is divided into a training set, a validation set and a test set based on a ratio of 7:2:1, and the tea seedling drought detection model is trained by the training set. During the training process, the tea seedling drought detection model is continuously optimized, including at least a learning rate, a learning rate factor, a weight decay, a learning rate warm-up phase, a learning rate warm-up momentum, a warm-up deviation learning rate, a batch size, and a training cycle hyperparameter, until the loss value of the loss function is less than a preset loss threshold;

[0045] The trained tea seedling drought detection model is verified by the verification set to determine whether the recognition accuracy of the tea seedling drought detection model is greater than a preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if yes, the verification passes, and:

[0046] The accuracy, recall rate, average precision mean and balance score of the tea seedling drought detection model are calculated through the test set to test the tea seedling drought detection model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0047] The advantages of the present invention are:

[0048] A large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels are obtained, and each chlorophyll fluorescence historical image is preprocessed including data annotation and data expansion to construct a data set; then a tea seedling drought detection model is created based on the YOLOv8n network, and the loss function of the tea seedling drought detection model is set to the MPDIoU function; then the data set is divided into a training set, a validation set and a test set based on a preset ratio, and the tea seedling drought detection model is trained, validated and tested by the training set, validation set and test set respectively; finally, the tea seedling drought detection model that passes the test is deployed, and the tea seedling drought detection is performed using the deployed tea seedling drought detection model; the YOLOv8n network consists of an input layer, a backbone network layer, a neck network layer and a detection head. The input layer is used to input chlorophyll fluorescence historical images; the backbone network layer is used to extract image features from chlorophyll fluorescence historical images, and is constructed based on the MobileViT network; the neck network layer is used to upsample the image features to obtain an upsampled feature map, and the upsampled feature map is obtained by using the upsampled feature map. PANet structure; the detection head is used to output the drought level of tea seedlings and the predicted probability of drought areas based on the upsampled feature map; that is, the backbone network layer of the YOLOv8n network is improved based on the MobileViT network. The MobileViT network cleverly combines the advantages of convolutional neural networks (CNN) and transformers, balances the needs of local feature extraction and global feature capture, and significantly reduces the number of model parameters and computational complexity to adapt to the visual recognition performance on resource-constrained mobile devices; the loss function of the tea seedling drought detection model is set to the MPDIoU function instead of the traditional CIoU function to improve the speed of model convergence and stability performance, and the MPDIoU function reduces the dependence on secondary factors, so that the model can focus more on the main positioning task, which not only helps the model converge faster during training, but also effectively improves the detection accuracy and robustness of the model, and ultimately greatly improves the accuracy of tea seedling drought detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0050] Figure 1 The present invention is a flow chart of a method for detecting drought in tea seedlings combined with chlorophyll fluorescence imaging.

[0051] Figure 2 The present invention is a schematic structural diagram of a tea seedling drought detection system combined with chlorophyll fluorescence imaging.

[0052] Figure 3 It is a structural schematic diagram of the tea seedling drought detection model of the present invention.

[0053] Figure 4 It is a schematic diagram of the structure of the MobileViT network of the present invention. DETAILED DESCRIPTION

[0054] The technical solution in the embodiment of the present application has the following overall idea: tea seedling drought detection is performed based on a tea seedling drought detection model created based on the YOLOv8n network, and the backbone network layer of the YOLOv8n network is improved based on the MobileViT network. The MobileViT network cleverly combines the advantages of convolutional neural networks (CNN) and transformers, balances the needs of local feature extraction and global feature capture, and significantly reduces the number of parameters and computational complexity of the model to adapt to the visual recognition performance on resource-constrained mobile devices; the loss function of the tea seedling drought detection model is set to the MPDIoU function instead of the traditional CIoU function to improve the speed of model convergence and stability performance, and the MPDIoU function reduces the dependence on secondary factors, so that the model can focus more on the main positioning task, which not only helps the model converge faster during training, but also effectively improves the detection accuracy and robustness of the model, thereby improving the accuracy of tea seedling drought detection.

[0055] Please refer to Figures 1 to 4 As shown, a preferred embodiment of a tea seedling drought detection method combined with chlorophyll fluorescence imaging of the present invention comprises the following steps:

[0056] Step S1, obtaining a large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels, and constructing a data set after preprocessing each of the chlorophyll fluorescence historical images including data annotation and data expansion;

[0057] Step S2: creating a tea seedling drought detection model based on the YOLOv8n network, and setting the loss function of the tea seedling drought detection model to the MPDIoU function;

[0058] Step S3, dividing the data set into a training set, a validation set and a test set based on a preset ratio, and respectively training, validating and testing the tea seedling drought detection model through the training set, the validation set and the test set;

[0059] Step S4: deploying the tea seedling drought detection model that has passed the test, and performing tea seedling drought detection using the deployed tea seedling drought detection model.

[0060] The step S1 is specifically as follows:

[0061] Acquire a large number of historical chlorophyll fluorescence images of tea seedlings under different drought levels, wherein the drought level is no drought, mild stress or severe stress;

[0062] Each of the chlorophyll fluorescence historical images is labeled with drought levels and drought areas, and each of the labeled chlorophyll fluorescence historical images is subjected to data expansion including at least brightness adjustment, contrast adjustment, image flipping, image translation, and image rotation to complete preprocessing of each of the chlorophyll fluorescence historical images, and a data set is constructed based on each of the preprocessed chlorophyll fluorescence historical images. By performing the data expansion operation, the sample size of the data set is effectively expanded, thereby effectively improving the learning effect and generalization ability of the tea seedling drought detection model.

[0063] In step S2, the YOLOv8n network consists of an input layer, a backbone network layer, a neck network layer and a detection head, and the input layer, the backbone network layer, the neck network layer and the detection head are connected in sequence;

[0064] The input layer is used to input the chlorophyll fluorescence historical image;

[0065] The backbone network layer is used to extract image features from chlorophyll fluorescence historical images and is constructed based on the MobileViT network;

[0066] The neck network layer is used to upsample the image features to obtain an upsampled feature map, using a PANet structure;

[0067] The detection head is used to output the drought level of tea seedling drought and the predicted probability of drought area according to the upsampled feature map.

[0068] The YOLO series of algorithms are typical representatives in the field of single-stage target detection. The core idea is to simplify the target detection task into a regression problem, predict the category and spatial position of the target, and compared with traditional target detection methods, the characteristics are fast and efficient training. The YOLOv8 network optimizes the computational efficiency and real-time problems in traditional target detection methods, and significantly improves the accuracy while ensuring the detection speed. The YOLOv8 network consists of an input layer (Input), a backbone network layer (Backbone), a neck network layer (Neck), and a detection head (Head); the size of the feature map (chlorophyll fluorescence history image) input by the input layer is 640×640; the neck network layer adopts the PANet structure. Compared with the top-down FPN, the PANet structure introduces a bottom-up fusion mechanism, which helps to effectively transfer the color, shape and spatial features of the low layer to the high layer, thereby enhancing the representation ability of the model; the detection head adopts a decoupled design, and each scale of the detection head is independently configured, abandoning the preset anchor frame, which not only improves the detection accuracy, but also solves the inherent conflict between classification and regression tasks, and speeds up the detection speed.

[0069] The YOLOv8 network can flexibly adjust the parameter scale of the backbone network layer by adjusting the three factors of depth, width and scale to adapt to different application scenarios and computing resource requirements. The present invention selects the YOLOv8n network with the smallest number of parameters, and achieves high-precision and fast detection by optimizing the loss function as weighted intersection-over-union (MPDIoU), which is more suitable for resource-constrained scenarios such as embedded devices.

[0070] Traditional convolutional neural networks excel in extracting local features, but are insufficient in obtaining global feature information; on the contrary, the Vision Transformer model is good at capturing global features, but its large parameter size places too high a demand on computing resources, making it difficult to deploy on mobile devices. The MobileViT network cleverly combines the advantages of convolutional neural networks (CNNs) and transformers, balancing the needs of local feature extraction and global feature capture, while significantly reducing the number of model parameters and computational complexity to adapt to visual recognition performance on resource-constrained mobile devices. The MobileViT network is mainly composed of convolutional layers, MV2, MobileViT Block, global pooling layers, and fully connected layers.

[0071] As the pre-processing part of the MobileViT network, the convolution layer is responsible for extracting preliminary local features from the input chlorophyll fluorescence historical image, and using the local perception ability of the convolution operation to lay the foundation for subsequent complex feature extraction; MV2 is derived from MobileNetV2 and adopts an inverted residual structure. By first expanding and then reducing the number of channels, it improves recognition accuracy while maintaining computational efficiency. MV2 uses a combination of 1x1 convolution and 3x3 convolution, combined with the RELU6 activation function, which improves the recognition accuracy and robustness of the model; the core of the MobileViT network is its unique MobileViT Block, which consists of a local information encoding module, a global information decoding module, and a feature information fusion module.

[0072] In order to learn a global representation with spatial inductive bias, MobileViT Block first expands X into N non-overlapping patches X u =R P*N*D , where P = wh, is the number of patches, h≤n,w≤n is the height and width of the patch, and each pixel p∈{1,2,…,P} in the patch is modeled by global features through Transformer to obtain X G(p) = Transformer(X U (p)),1≤p≤P.

[0073] The local information encoding module is responsible for further refining and extracting local features of the chlorophyll fluorescence history image. The global information decoding module uses the Unfold-Transformer-fold structure to globally model the local features to capture the global context information of the chlorophyll fluorescence history image and obtain the global features. The feature information fusion module effectively integrates local features and global features to extract richer image feature representations. Through this design, the MobileViT network not only reduces the parameter scale of the model while maintaining high accuracy, but also improves the network's inference speed, providing a new, efficient and lightweight option for mobile visual recognition tasks.

[0074] The MobileViT network aims to achieve high-precision, low-computation fast reasoning by reducing model parameters and computational complexity, effectively utilizing limited computing resources, and thus making the tea seedling drought detection model have a wide range of application value on mobile devices. Through the synergistic extraction of local and global features, the MobileViT network achieves comprehensive extraction of image features, thereby improving the performance of the tea seedling drought detection model on complex visual tasks.

[0075] In step S2, the formula of the MPDIoU function is:

[0076]

[0077] in, represents the loss value of the MPDIoU function; MPDIoU represents the intersection-over-union ratio of the minimum point distance; A represents the predicted box; B represents the true box; w represents the width of the chlorophyll fluorescence history image; h represents the height of the chlorophyll fluorescence history image; Indicates the upper left corner of the prediction box; Indicates the lower right corner of the prediction box; Indicates the upper left corner of the real box; Indicates the lower right corner of the ground-truth box.

[0078] IoU (Intersection over Union) is an important indicator for measuring the degree of overlap between the predicted box and the true box in the field of target detection. The loss function of the traditional YOLOv8 network adopts the CIoU function that only considers the intersection area of ​​the bounding box (prediction box and true box), the center point distance and the aspect ratio of the bounding box. The CIoU function will be affected by the model convergence speed due to excessive reliance on the measurement of the aspect ratio of the bounding box. In order to improve the convergence speed and stability performance, the present invention adopts the MPDIoU function for optimization. The core idea is to introduce the concept of minimum point distance, that is, by minimizing the distance between the upper left corner and the lower right corner of the predicted box and the true box, to more accurately evaluate the similarity of the bounding box, which is suitable for processing the regression problem in the case of overlapping and non-overlapping bounding boxes, while taking into account the difference in center point distance and width and height. The MPDIoU function reduces the dependence on secondary factors, allowing the model to focus more on the main positioning task, helping the model to converge faster during training, while improving the detection accuracy and robustness of the model.

[0079] The step S3 is specifically as follows:

[0080] The data set is divided into a training set, a validation set and a test set based on a ratio of 7:2:1, and the tea seedling drought detection model is trained by the training set. During the training process, the tea seedling drought detection model is continuously optimized, including at least a learning rate, a learning rate factor, a weight decay, a learning rate warm-up phase, a learning rate warm-up momentum, a warm-up deviation learning rate, a batch size, and a training cycle hyperparameter, until the loss value of the loss function is less than a preset loss threshold;

[0081] The trained tea seedling drought detection model is verified by the verification set to determine whether the recognition accuracy of the tea seedling drought detection model is greater than a preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if yes, the verification passes, and:

[0082] The precision (Precision), recall (Recall), mean Average Precision (mAP) and balance score (F1) of the tea seedling drought detection model are calculated through the test set to test the tea seedling drought detection model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0083] Precision measures the proportion of samples predicted to be positive that actually belong to the positive category, thus reflecting the accuracy of the classification results; recall rate indicates the proportion of all positive category samples that the model can correctly identify, reflecting the comprehensiveness of the model; average precision reflects the average accuracy of a category under different thresholds, which is a comprehensive indicator for evaluating positioning accuracy and prediction accuracy, calculated as the area under the PR curve; mean average precision is the average of various average precisions, which is used to evaluate the overall performance of the model on multiple categories; the balance score is a comprehensive measure of precision and recall rate, calculated by the harmonic mean, and the formula is:

[0084]

[0085] A preferred embodiment of a tea seedling drought detection system combined with chlorophyll fluorescence imaging of the present invention includes the following modules:

[0086] A data set construction module is used to obtain a large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels, and to construct a data set after preprocessing each of the chlorophyll fluorescence historical images including data annotation and data expansion;

[0087] A tea seedling drought detection model creation module is used to create a tea seedling drought detection model based on the YOLOv8n network, and set the loss function of the tea seedling drought detection model to be the MPDIoU function;

[0088] A tea seedling drought detection model training module is used to divide the data set into a training set, a validation set and a test set based on a preset ratio, and train, validate and test the tea seedling drought detection model through the training set, the validation set and the test set respectively;

[0089] The tea seedling drought detection module is used to deploy the tea seedling drought detection model that has passed the test, and use the deployed tea seedling drought detection model to perform tea seedling drought detection.

[0090] The data set construction module is specifically used for:

[0091] Acquire a large number of historical chlorophyll fluorescence images of tea seedlings under different drought levels, wherein the drought level is no drought, mild stress or severe stress;

[0092] Each of the chlorophyll fluorescence historical images is labeled with drought levels and drought areas, and each of the labeled chlorophyll fluorescence historical images is subjected to data expansion including at least brightness adjustment, contrast adjustment, image flipping, image translation, and image rotation to complete preprocessing of each of the chlorophyll fluorescence historical images, and a data set is constructed based on each of the preprocessed chlorophyll fluorescence historical images. By performing the data expansion operation, the sample size of the data set is effectively expanded, thereby effectively improving the learning effect and generalization ability of the tea seedling drought detection model.

[0093] In the tea seedling drought detection model creation module, the YOLOv8n network consists of an input layer, a backbone network layer, a neck network layer and a detection head, and the input layer, the backbone network layer, the neck network layer and the detection head are connected in sequence;

[0094] The input layer is used to input the chlorophyll fluorescence historical image;

[0095] The backbone network layer is used to extract image features from chlorophyll fluorescence historical images and is constructed based on the MobileViT network;

[0096] The neck network layer is used to upsample the image features to obtain an upsampled feature map, using a PANet structure;

[0097] The detection head is used to output the drought level of tea seedling drought and the predicted probability of drought area according to the upsampled feature map.

[0098] The YOLO series of algorithms are typical representatives in the field of single-stage target detection. The core idea is to simplify the target detection task into a regression problem, predict the category and spatial position of the target, and compared with traditional target detection methods, the characteristics are fast and efficient training. The YOLOv8 network optimizes the computational efficiency and real-time problems in traditional target detection methods, and significantly improves the accuracy while ensuring the detection speed. The YOLOv8 network consists of an input layer (Input), a backbone network layer (Backbone), a neck network layer (Neck), and a detection head (Head); the size of the feature map (chlorophyll fluorescence history image) input by the input layer is 640×640; the neck network layer adopts the PANet structure. Compared with the top-down FPN, the PANet structure introduces a bottom-up fusion mechanism, which helps to effectively transfer the color, shape and spatial features of the low layer to the high layer, thereby enhancing the representation ability of the model; the detection head adopts a decoupled design, and each scale of the detection head is independently configured, abandoning the preset anchor frame, which not only improves the detection accuracy, but also solves the inherent conflict between classification and regression tasks, and speeds up the detection speed.

[0099] The YOLOv8 network can flexibly adjust the parameter scale of the backbone network layer by adjusting the three factors of depth, width and scale to adapt to different application scenarios and computing resource requirements. The present invention selects the YOLOv8n network with the smallest number of parameters, and achieves high-precision and fast detection by optimizing the loss function as weighted intersection-over-union (MPDIoU), which is more suitable for resource-constrained scenarios such as embedded devices.

[0100] Traditional convolutional neural networks excel in extracting local features, but are insufficient in obtaining global feature information; on the contrary, the Vision Transformer model is good at capturing global features, but its large parameter size places too high a demand on computing resources, making it difficult to deploy on mobile devices. The MobileViT network cleverly combines the advantages of convolutional neural networks (CNNs) and transformers, balancing the needs of local feature extraction and global feature capture, while significantly reducing the number of model parameters and computational complexity to adapt to visual recognition performance on resource-constrained mobile devices. The MobileViT network is mainly composed of convolutional layers, MV2, MobileViT Block, global pooling layers, and fully connected layers.

[0101] As the pre-processing part of the MobileViT network, the convolution layer is responsible for extracting preliminary local features from the input chlorophyll fluorescence historical image, and using the local perception ability of the convolution operation to lay the foundation for subsequent complex feature extraction; MV2 is derived from MobileNetV2 and adopts an inverted residual structure. By first expanding and then reducing the number of channels, it improves recognition accuracy while maintaining computational efficiency. MV2 uses a combination of 1x1 convolution and 3x3 convolution, combined with the RELU6 activation function, which improves the recognition accuracy and robustness of the model; the core of the MobileViT network is its unique MobileViT Block, which consists of a local information encoding module, a global information decoding module, and a feature information fusion module.

[0102] In order to learn a global representation with spatial inductive bias, MobileViT Block first expands X into N non-overlapping paths cX u =R P*N*D , where O = wh, is the number of patches, h≤n,w≤n is the height and width of the patch, and each pixel p∈{1,2,…,P} in the patch is modeled by global features through Transformer to obtain X G(p) = Transformer(X U (p)),1≤p≤P.

[0103] The local information encoding module is responsible for further refining and extracting local features of the chlorophyll fluorescence history image. The global information decoding module uses the Unfold-Transformer-fold structure to globally model the local features to capture the global context information of the chlorophyll fluorescence history image and obtain the global features. The feature information fusion module effectively integrates local features and global features to extract richer image feature representations. Through this design, the MobileViT network not only reduces the parameter scale of the model while maintaining high accuracy, but also improves the network's inference speed, providing a new, efficient and lightweight option for mobile visual recognition tasks.

[0104] The MobileViT network aims to achieve high-precision, low-computation fast reasoning by reducing model parameters and computational complexity, effectively utilizing limited computing resources, and thus making the tea seedling drought detection model have a wide range of application value on mobile devices. Through the synergistic extraction of local and global features, the MobileViT network achieves comprehensive extraction of image features, thereby improving the performance of the tea seedling drought detection model on complex visual tasks.

[0105] In the tea seedling drought detection model creation module, the formula of the MPDIoU function is:

[0106]

[0107] in, represents the loss value of the MPDIoU function; MPDIoU represents the intersection-over-union ratio of the minimum point distance; A represents the predicted box; B represents the true box; w represents the width of the chlorophyll fluorescence history image; h represents the height of the chlorophyll fluorescence history image; Indicates the upper left corner of the prediction box; Indicates the lower right corner of the prediction box; Indicates the upper left corner of the real box; Indicates the lower right corner of the ground-truth box.

[0108] IoU (Intersection over Union) is an important indicator for measuring the degree of overlap between the predicted box and the true box in the field of target detection. The loss function of the traditional YOLOv8 network adopts the CIoU function that only considers the intersection area of ​​the bounding box (prediction box and true box), the center point distance and the aspect ratio of the bounding box. The CIoU function will be affected by the model convergence speed due to excessive reliance on the measurement of the aspect ratio of the bounding box. In order to improve the convergence speed and stability performance, the present invention adopts the MPDIoU function for optimization. The core idea is to introduce the concept of minimum point distance, that is, by minimizing the distance between the upper left corner and the lower right corner of the predicted box and the true box, to more accurately evaluate the similarity of the bounding box, which is suitable for processing the regression problem in the case of overlapping and non-overlapping bounding boxes, while taking into account the difference in center point distance and width and height. The MPDIoU function reduces the dependence on secondary factors, allowing the model to focus more on the main positioning task, helping the model to converge faster during training, while improving the detection accuracy and robustness of the model.

[0109] The tea seedling drought detection model training module is specifically used for:

[0110] The data set is divided into a training set, a validation set and a test set based on a ratio of 7:2:1, and the tea seedling drought detection model is trained by the training set. During the training process, the tea seedling drought detection model is continuously optimized, including at least a learning rate, a learning rate factor, a weight decay, a learning rate warm-up phase, a learning rate warm-up momentum, a warm-up deviation learning rate, a batch size, and a training cycle hyperparameter, until the loss value of the loss function is less than a preset loss threshold;

[0111] The trained tea seedling drought detection model is verified by the verification set to determine whether the recognition accuracy of the tea seedling drought detection model is greater than a preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if yes, the verification passes, and:

[0112] The precision (Precision), recall (Recall), mean Average Precision (mAP) and balance score (F1) of the tea seedling drought detection model are calculated through the test set to test the tea seedling drought detection model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

[0113] Precision measures the proportion of samples predicted to be positive that actually belong to the positive category, thus reflecting the accuracy of the classification results; recall rate indicates the proportion of all positive category samples that the model can correctly identify, reflecting the comprehensiveness of the model; average precision reflects the average accuracy of a category under different thresholds, which is a comprehensive indicator for evaluating positioning accuracy and prediction accuracy, calculated as the area under the PR curve; mean average precision is the average of various average precisions, which is used to evaluate the overall performance of the model on multiple categories; the balance score is a comprehensive measure of precision and recall rate, calculated by the harmonic mean, and the formula is:

[0114]

[0115] In summary, the advantages of the present invention are:

[0116] A large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels are obtained, and each chlorophyll fluorescence historical image is preprocessed including data annotation and data expansion to construct a data set; then a tea seedling drought detection model is created based on the YOLOv8n network, and the loss function of the tea seedling drought detection model is set to the MPDIoU function; then the data set is divided into a training set, a validation set and a test set based on a preset ratio, and the tea seedling drought detection model is trained, validated and tested by the training set, validation set and test set respectively; finally, the tea seedling drought detection model that passes the test is deployed, and the tea seedling drought detection is performed using the deployed tea seedling drought detection model; the YOLOv8n network consists of an input layer, a backbone network layer, a neck network layer and a detection head. The input layer is used to input chlorophyll fluorescence historical images; the backbone network layer is used to extract image features from chlorophyll fluorescence historical images, and is constructed based on the MobileViT network; the neck network layer is used to upsample the image features to obtain an upsampled feature map, and the upsampled feature map is obtained by using the upsampled feature map. PANet structure; the detection head is used to output the drought level of tea seedlings and the predicted probability of drought areas based on the upsampled feature map; that is, the backbone network layer of the YOLOv8n network is improved based on the MobileViT network. The MobileViT network cleverly combines the advantages of convolutional neural networks (CNN) and transformers, balances the needs of local feature extraction and global feature capture, and significantly reduces the number of model parameters and computational complexity to adapt to the visual recognition performance on resource-constrained mobile devices; the loss function of the tea seedling drought detection model is set to the MPDIoU function instead of the traditional CIoU function to improve the speed of model convergence and stability performance, and the MPDIoU function reduces the dependence on secondary factors, so that the model can focus more on the main positioning task, which not only helps the model converge faster during training, but also effectively improves the detection accuracy and robustness of the model, and ultimately greatly improves the accuracy of tea seedling drought detection.

[0117] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A tea seedling drought detection method combined with chlorophyll fluorescence imaging, characterized in that: The steps include: Step S1, obtaining a large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels, and constructing a data set after preprocessing each of the chlorophyll fluorescence historical images including data annotation and data expansion; Step S2: creating a tea seedling drought detection model based on the YOLOv8n network, and setting the loss function of the tea seedling drought detection model to the MPDIoU function; Step S3, dividing the data set into a training set, a validation set and a test set based on a preset ratio, and respectively training, validating and testing the tea seedling drought detection model through the training set, the validation set and the test set; Step S4: deploying the tea seedling drought detection model that has passed the test, and performing tea seedling drought detection using the deployed tea seedling drought detection model.

2. The tea seedling drought detection method combined with chlorophyll fluorescence imaging as claimed in claim 1, characterized in that: The step S1 is specifically as follows: Acquire a large number of historical chlorophyll fluorescence images of tea seedlings under different drought levels, wherein the drought level is no drought, mild stress or severe stress; Each of the chlorophyll fluorescence historical images is labeled with drought levels and drought areas, and data expansion is performed on the labeled chlorophyll fluorescence historical images, including at least brightness adjustment, contrast adjustment, image flipping, image translation and image rotation, so as to complete preprocessing of each of the chlorophyll fluorescence historical images, and a data set is constructed based on the preprocessed chlorophyll fluorescence historical images.

3. The tea seedling drought detection method combined with chlorophyll fluorescence imaging as claimed in claim 1, characterized in that: In step S2, the YOLOv8n network consists of an input layer, a backbone network layer, a neck network layer and a detection head, and the input layer, the backbone network layer, the neck network layer and the detection head are connected in sequence; The input layer is used to input the chlorophyll fluorescence historical image; The backbone network layer is used to extract image features from chlorophyll fluorescence historical images and is constructed based on the MobileViT network; The neck network layer is used to upsample the image features to obtain an upsampled feature map, using a PANet structure; The detection head is used to output the drought level of tea seedling drought and the predicted probability of drought area according to the upsampled feature map.

4. The tea seedling drought detection method combined with chlorophyll fluorescence imaging as claimed in claim 1, characterized in that: In step S2, the formula of the MPDIoU function is: in, represents the loss value of the MPDIoU function; MPDIoU represents the intersection-over-union ratio of the minimum point distance; A represents the predicted box; B represents the true box; w represents the width of the chlorophyll fluorescence history image; h represents the height of the chlorophyll fluorescence history image; Indicates the upper left corner of the prediction box; Indicates the lower right corner of the prediction box; Indicates the upper left corner of the real box; Indicates the lower right corner of the ground-truth box.

5. The tea seedling drought detection method combined with chlorophyll fluorescence imaging as claimed in claim 1, characterized in that: The step S3 is specifically as follows: The data set is divided into a training set, a validation set and a test set based on a ratio of 7:2:1, and the tea seedling drought detection model is trained by the training set. During the training process, the tea seedling drought detection model is continuously optimized, including at least a learning rate, a learning rate factor, a weight decay, a learning rate warm-up phase, a learning rate warm-up momentum, a warm-up deviation learning rate, a batch size, and a training cycle hyperparameter, until the loss value of the loss function is less than a preset loss threshold; The trained tea seedling drought detection model is verified by the verification set to determine whether the recognition accuracy of the tea seedling drought detection model is greater than a preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if yes, the verification passes, and: The accuracy, recall rate, average precision mean and balance score of the tea seedling drought detection model are calculated through the test set to test the tea seedling drought detection model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

6. A tea seedling drought detection system combined with chlorophyll fluorescence imaging, characterized in that: Includes the following modules: A data set construction module is used to obtain a large number of chlorophyll fluorescence historical images of tea seedlings under different drought levels, and to construct a data set after preprocessing each of the chlorophyll fluorescence historical images including data annotation and data expansion; A tea seedling drought detection model creation module is used to create a tea seedling drought detection model based on the YOLOv8n network, and set the loss function of the tea seedling drought detection model to be the MPDIoU function; A tea seedling drought detection model training module is used to divide the data set into a training set, a validation set and a test set based on a preset ratio, and train, validate and test the tea seedling drought detection model through the training set, the validation set and the test set respectively; The tea seedling drought detection module is used to deploy the tea seedling drought detection model that has passed the test, and use the deployed tea seedling drought detection model to perform tea seedling drought detection.

7. A tea seedling drought detection system combined with chlorophyll fluorescence imaging as claimed in claim 6, characterized in that: The data set construction module is specifically used for: Acquire a large number of historical chlorophyll fluorescence images of tea seedlings under different drought levels, wherein the drought level is no drought, mild stress or severe stress; Each of the chlorophyll fluorescence historical images is labeled with drought levels and drought areas, and data expansion is performed on the labeled chlorophyll fluorescence historical images, including at least brightness adjustment, contrast adjustment, image flipping, image translation and image rotation, so as to complete preprocessing of each of the chlorophyll fluorescence historical images, and a data set is constructed based on the preprocessed chlorophyll fluorescence historical images.

8. The tea seedling drought detection system combined with chlorophyll fluorescence imaging as claimed in claim 6, characterized in that: In the tea seedling drought detection model creation module, the YOLOv8n network consists of an input layer, a backbone network layer, a neck network layer and a detection head, and the input layer, the backbone network layer, the neck network layer and the detection head are connected in sequence; The input layer is used to input the chlorophyll fluorescence historical image; The backbone network layer is used to extract image features from chlorophyll fluorescence historical images and is constructed based on the MobileViT network; The neck network layer is used to upsample the image features to obtain an upsampled feature map, using a PANet structure; The detection head is used to output the drought level of tea seedling drought and the predicted probability of drought area according to the upsampled feature map.

9. The tea seedling drought detection system combined with chlorophyll fluorescence imaging as claimed in claim 6, characterized in that: In the tea seedling drought detection model creation module, the formula of the MPDIoU function is: in, represents the loss value of the MPDIoU function; MPDIoU represents the intersection-over-union ratio of the minimum point distance; A represents the predicted box; B represents the true box; w represents the width of the chlorophyll fluorescence history image; h represents the height of the chlorophyll fluorescence history image; Indicates the upper left corner of the prediction box; Indicates the lower right corner of the prediction box; Indicates the upper left corner of the real box; Indicates the lower right corner of the ground-truth box.

10. The tea seedling drought detection system combined with chlorophyll fluorescence imaging as claimed in claim 6, characterized in that: The tea seedling drought detection model training module is specifically used for: The data set is divided into a training set, a validation set and a test set based on a ratio of 7:2:1, and the tea seedling drought detection model is trained by the training set. During the training process, the tea seedling drought detection model is continuously optimized, including at least a learning rate, a learning rate factor, a weight decay, a learning rate warm-up phase, a learning rate warm-up momentum, a warm-up deviation learning rate, a batch size, and a training cycle hyperparameter, until the loss value of the loss function is less than a preset loss threshold; The trained tea seedling drought detection model is verified by the verification set to determine whether the recognition accuracy of the tea seedling drought detection model is greater than a preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if yes, the verification passes, and: The accuracy, recall rate, average precision mean and balance score of the tea seedling drought detection model are calculated through the test set to test the tea seedling drought detection model. If the test fails, the training set is expanded to continue training; if the test passes, the training is terminated.

Images