Tea tree tender shoot freeze injury detection and early warning method, system and equipment based on scene migration

Abstract

The invention discloses a tea tree tender shoot freeze injury detection and early warning method, system and device based on scene migration, a model can be effectively migrated to a field environment by introducing a CycleGAN style migration and domain adaptive technology, and accurate detection and trend prediction of freeze injury are realized in combination with Mask R-CNN and BiLSTM models. According to the method, the problem of generalization when a laboratory training model migrates to a field environment is effectively solved, the cross-scene adaptive capacity of the model is improved, accurate detection and freeze injury trend prediction of a tea tree tender shoot freeze injury area are realized, and the accuracy and reliability are high.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and agricultural intelligent detection technology, and specifically to a method, system and equipment for detecting and warning frost damage to tea tree buds based on scene migration. Background Art

[0002] Traditional methods for detecting frost damage to tea buds rely primarily on manual inspections and empirical judgment. This requires not only advanced technical expertise and extensive practical experience, but also suffers from poor accuracy, slow speed, and the inability to perform quantitative analysis, making it difficult to meet the practical needs of large-scale, real-time frost damage monitoring. However, there is a lack of systematic research and published results on the automated identification of frost damage to tea buds. Especially given the complex conditions under which frost damage occurs and the difficulty in obtaining samples, there is no mature technology for highly accurate, automated frost damage level identification.

[0003] Currently, although the available data mainly comes from laboratory simulations of freezing conditions, these images differ from the actual field environment and have the following main problems:

[0004] 1. Inadequate cross-scenario detection capabilities: In laboratory environments, tea bud images have stable lighting conditions, background information, and camera parameters, allowing the detection model to learn highly accurate frost damage signatures. However, in field environments, factors such as fluctuating light intensity, leaf occlusion, and complex backgrounds can significantly degrade the performance of laboratory-trained models, directly impacting the accuracy and generalization of frost damage detection.

[0005] 2. Lack of high-quality field data: To improve detection accuracy, image data of frost damage in field environments is difficult to obtain due to the seasonal nature of frost damage. Furthermore, the uncertainty of changes in frost damage severity leads to high labeling costs for field data. Furthermore, existing methods rely on laboratory data for training, but the distribution of this data differs significantly from field data, making it difficult to directly transfer models between the laboratory and the field. Summary of the Invention

[0006] The purpose of the present invention is to provide a method, system and equipment for detecting and warning frost damage to tea tree buds based on scene migration to solve the above defects.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A method for detecting and warning frost damage to tea tree buds based on scene migration includes the following steps:

[0009] S1. Collect image information data of tea tree buds to be tested in laboratory and field environments;

[0010] S2. Cross-scene data generation: using the CycleGAN network for image style transfer, converting laboratory labeled data into data with the style of a field environment to eliminate data distribution differences between scenes;

[0011] S3. Input the data after style transfer and domain adaptation processing into the Mask R-CNN model to detect and segment the target area of ​​the tea tree bud frost damage area and extract the ROI features of the frost damage area;

[0012] S4. Use the BiLSTM model to perform time series modeling on the frost damage ROI features at multiple consecutive time points, analyze the dynamic evolution trend of frost damage, and thus achieve online accurate detection of frost damage to tea tree buds, frost damage trend prediction, and early warning.

[0013] Preferably, in step S1, the image information data of tea tree buds to be detected in the laboratory environment and the field environment are collected, including: the image dataset D annotated by the laboratory lab and Daejeon Unlabeled Image Dataset D field , as follows:

[0014]

[0015]

[0016] Where D lab represents the set of frost damage image samples labeled by the laboratory, x i lab represents the image sample collected by the i-th laboratory, y i cls Represents image x i lab Corresponding frost damage level label, y i mask Represents image x i lab The corresponding pixel-level frost damage area mask, D field represents unlabeled image data in a field environment, represents the image sample collected from the i-th field, H*W represents the length and width of the image, N lab represents the total number of samples labeled in the laboratory, and N field represents the total number of unlabeled samples in the field.

[0017] Preferably, in step S2, a CycleGAN network is used to perform image style transfer, specifically: tea tree bud frost damage data under different environments is generated through unsupervised learning, and used to enhance the training set to improve the model's cross-scene adaptability.

[0018] Preferably, in step S3, the domain adaptation processing is specifically: using domain adaptation technology, including adversarial training and feature alignment, to introduce a domain discriminator in the feature extraction layer of the Mask R-CNN target detection network to reduce the domain bias of different tea garden environments and improve the generalization ability of the model.

[0019] Preferably, in step S3, the Mask R-CNN model is used to detect and segment the target area of ​​the tea tree bud frost damage area, and outputs a bounding box bbox of the frost damage area, a mask mask and the extracted frost damage area feature ROI; an indicator or score of the severity of the frost damage in the ROI area is I frost , and use the Sigmoid function to get the weight w of the frost damage area frost Then, it is multiplied with the original ROI feature to enhance the recognition ability of the frost-damaged area and provide feature information for subsequent time series analysis;

[0020] During the training process, the Mask R-CNN model jointly optimizes L det Perform end-to-end optimization, the loss function includes the focal loss L Focal , bounding box regression loss L SmoothL1 and mask segmentation loss L Dice , improving the accuracy of frost-damaged area detection, bounding box positioning accuracy, and instance mask segmentation quality;

[0021]

[0022]

[0023] Where w frost is the frost damage area weight, which is used to indicate the severity of frost damage in the ROI area. Its value range is between 0 and 1 and is obtained using the Sigmoid function. frost Generally, the value is between [0,1]; the coefficient 10 and the bias 0.5 are used to adjust the steepness of the Sigmoid curve so that when I frost When I > 0.5, the weight quickly approaches 1, indicating that the characteristics of the frost damage area are significant; when I frost When <0.5, the weight rapidly approaches 0, indicating that the freezing damage characteristics in the area are not obvious;

[0024] Where, f ROI The calculation method is as follows: first, perform ROIAlign operation on the input feature map F and the detected target box bbox to obtain the initial regional features, and then compare the obtained ROI features with the frost damage area weight w frost Multiply them together to highlight the characteristic information of the frost-damaged area;

[0025] Where, Ldet Jointly optimize the loss function for multiple tasks, used for joint optimization of target detection and instance segmentation; L Focal L is the focal loss, which is used for target classification and is suitable for handling situations where positive and negative samples are extremely unbalanced, focusing on the classification of difficult samples. SmoothL1 Mainly used for regression optimization of the bounding box bbox, calculating the error between the predicted bounding box and the true bounding box; L Dice Mainly used for instance segmentation task loss to measure the degree of overlap between the predicted mask and the true mask.

[0026] Preferably, in step S4, the BiLSTM model is used to process the ROI features of the frozen area extracted by Mask R-CNN and output the feature h at each time point. t , and then use the attention mechanism to calculate the corresponding attention weight a for each time feature t , and then fuse to obtain the feature vector Z after the attention weight is weighted, input to the fully connected classification layer, and output the frost damage level p(y|c) through the softmax function;

[0027]

[0028]

[0029] Where a t represents the attention weight of the feature at the tth moment, that is, the importance of the feature at each moment; T represents the number of time steps of the input sequence, such as 3 time points in the morning, noon and afternoon; h t is the hidden feature output by BiLSTM at the tth time step; W and V are the weight parameters to be learned; h K represents the features of BiLSTM at the kth time point and the information before and after it, Z represents the feature vector weighted by the attention weight, which is the weighted sum of the time series features; W c and b c are the weights and biases of the classification layer to be learned; p(y|c) is the predicted probability output of the classification task, such as the prediction of frost damage level; the softmax function is used to convert the output features into probability values ​​for multi-classification decision making.

[0030] Preferably, in step S4, the frost damage level of the tea tree buds includes four levels: no frost damage, mild frost damage, moderate frost damage and severe frost damage; combined with the temporal information of BiLSTM to optimize the classification accuracy, it can be integrated into a drone or ground monitoring equipment to achieve real-time image acquisition and online frost damage detection, thereby improving the frost damage monitoring capability of tea tree planting areas.

[0031] Preferably, a tea tree bud frost damage detection and early warning system based on scene migration is applied to the tea tree bud frost damage detection and early warning method based on scene migration, comprising:

[0032] An image acquisition module, including an RGB camera and a colloidal gold reagent detection device, is used to collect image information to be analyzed of the agricultural products to be tested;

[0033] Data enhancement module, which performs style conversion through CycleGAN to improve data diversity;

[0034] The target detection module detects frost-damaged areas based on Mask R-CNN and outputs ROI features;

[0035] Time series analysis module, based on BiLSTM, for frost damage trend prediction and risk warning;

[0036] The user interaction module is used to display the detection results and provide suggestions for preventing frost damage in tea gardens.

[0037] Preferably, an electronic device comprises a processor and a memory storing a computer program, wherein the processor is used to execute the computer program stored in the memory and implement the steps in the scene migration-based tea tree bud frost damage detection and early warning method.

[0038] The beneficial effects of the present invention are:

[0039] This paper uses CycleGAN style transfer and domain adaptation technology to effectively solve the generalization problem when migrating laboratory-trained models to field environments, thereby improving the model's cross-scenario adaptability. By combining Mask R-CNN with the BiLSTM time series model, it achieves accurate detection of frost-damaged areas of tea tree buds and prediction of frost damage trends, making up for the deficiency of traditional single-frame detection in being unable to perform frost damage trend analysis. It also provides intelligent classification and real-time early warning of frost damage levels, helping tea farmers to make precise management and prevention and control decisions, and has high practical application value and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the method of the present invention;

[0041] Figure 2 Schematic diagram comparing the actual detection effects of the method of the present invention in laboratory and field environments. DETAILED DESCRIPTION

[0042] The present invention is further described below with reference to the embodiments. It should be noted that these are merely examples and illustrations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should be deemed to fall within the scope of protection of the present invention.

[0043] Example 1:

[0044] like Figure 1 As shown, the present invention proposes a method for detecting and warning frost damage to tea tree buds based on scene migration, comprising the following steps:

[0045] S1. Collect image information data of tea tree buds to be tested in laboratory environment and field environment.

[0046] Image information of tea tree buds was collected in laboratory and field environments. The laboratory data (source domain) was collected using a 20-megapixel RGB camera, and the annotation followed the "GB / T35810-2018" standard to generate an annotated dataset with pixel-level frost damage masks. The field data (target domain) was collected using a drone-mounted multispectral camera (RGB+near-infrared) to collect unlabeled images covering different lighting conditions (morning frost and midday strong light).

[0047] Collect the image information data of tea tree buds to be tested in the laboratory environment and field environment, including: laboratory labeled image dataset D lab and Daejeon Unlabeled Image Dataset D field , as follows:

[0048]

[0049]

[0050] Where D lab represents the set of frost damage image samples labeled by the laboratory, x i lab represents the image sample collected by the i-th laboratory, y i cls Represents image x i lab Corresponding frost damage level label, y i mask Represents image x i lab The corresponding pixel-level frost damage area mask, D field represents unlabeled image data in a field environment, represents the image sample collected from the i-th field, H*W represents the length and width of the image, N labrepresents the total number of samples labeled in the laboratory, and N field represents the total number of unlabeled samples in the field.

[0051] S2. Cross-scene data generation: Use the CycleGAN network to perform image style transfer processing, converting laboratory labeled data into data with the style of a field environment to eliminate data distribution differences between scenes.

[0052] The CycleGAN network is used for image style transfer. Specifically, unsupervised learning is used to generate tea tree bud frost damage data in different environments, and this data is used to enhance the training set and improve the model's cross-scenario adaptability.

[0053] The specific steps in style transfer include calculating the importance of the attention weight matrix for each layer and pruning the network by combining global attention information with local convolutional features. The pruning process uses group sparsity regularization to prune nodes with smaller weights in groups, improving pruning efficiency while preserving the model's key features.

[0054] The feature enhancement structure formula of the residual connection is as follows:

[0055] FAM(F)=σ{Conv 1×1 (F)}⊙F+F,

[0056] In the formula, FAM(F) represents the residual connection feature enhancement map, and F represents the original feature map. The formula shows that: first, a 1×1 convolution is performed on the feature map F, and then it is activated with the sigmoid function (σ) to obtain a weighted feature map of the same size as F. This is then multiplied element-wise with the original feature map F to achieve weighted attention on the channel or spatial dimension of the original feature map F. Finally, the result is added to the original feature map F to form a residual connection feature enhancement structure. This structure can highlight the target area in cross-scene images, suppress interference information, and improve generalization ability.

[0057] The correspondence formula for characteristic losses of freezing damage is as follows:

[0058]

[0059] Lt otal =L adv +10Lc ycle +5L frost

[0060] Where, L frost represents the “freeze damage feature loss”, which aims to distinguish the difference between the generated style transfer image in the frost damage area and the actual field image; frost is the weight coefficient of the loss term, which is used to control the contribution of the loss to the overall loss function; E represents the mathematical expectation, which represents the average loss over all samples; Mfrost represents a mask matrix, which is used to indicate the frost-related regions in the image. It is usually a binary matrix (1 represents the frost-damaged area and 0 represents the non-frost-damaged area). frost (x lab ) indicates that the generator of CycleGAN is used to transform the laboratory image x lab Convert to an image with Ota style;x field Represents an unlabeled frost-damaged image collected in a field environment, which is used as the target image for style transfer comparison.

[0061] Where, L adv is the adversarial loss, used to ensure that the generated image has the real image style; L cycle To ensure the stability of image style conversion; L frost The style transfer effect is enhanced for frost-damaged areas, ensuring that the frost-damaged features are not lost during the conversion process, and ensuring that the generated images can truly reflect the visual characteristics of frost-damaged areas in field environments.

[0062] Through the above-mentioned residual connection feature enhancement structure formula and frost damage feature loss correspondence formula, the semantic information of the channel is enhanced while retaining local details, thereby converting laboratory annotated data into data with field environment style to eliminate the data distribution differences between scenes.

[0063] S3. Input the data after style transfer and domain adaptation processing into the Mask R-CNN model to perform target detection and segmentation of the frost-damaged area of ​​tea tree buds, and extract the ROI features of the frost-damaged area.

[0064] Domain adaptation processing, specifically: using domain adaptation techniques, including adversarial training and feature alignment (such as MMD and CORAL), to introduce a domain discriminator in the feature extraction layer of the Mask R-CNN object detection network to reduce domain bias in different tea garden environments and improve the generalization ability of the model.

[0065] The Mask R-CNN model is used to detect and segment targets in the frost-damaged area of ​​tea tree buds. The output includes the bounding box bbox of the frost-damaged area, the mask mask, and the extracted frost-damaged area feature ROI; an indicator or score of the severity of the frost damage in the ROI area is I frost , and use the Sigmoid function to get the weight w of the frost damage area frost Then, it is multiplied with the original ROI feature to enhance the recognition ability of the frost-damaged area and provide feature information for subsequent time series analysis;

[0066] During the training process, the Mask R-CNN model jointly optimizes L detPerform end-to-end optimization, the loss function includes the focal loss L Focal , bounding box regression loss L SmoothL1 and mask segmentation loss L Dice , improving the accuracy of frost-damaged area detection, bounding box positioning accuracy, and instance mask segmentation quality;

[0067]

[0068]

[0069] Where w frost is the frost damage area weight, which is used to indicate the severity of frost damage in the ROI area. Its value range is between 0 and 1 and is obtained using the Sigmoid function. frost Generally, the value is between [0,1]; the coefficient 10 and the bias 0.5 are used to adjust the steepness of the Sigmoid curve so that when I frost When I > 0.5, the weight quickly approaches 1, indicating that the characteristics of the frost damage area are significant; when I frost When <0.5, the weight rapidly approaches 0, indicating that the freezing damage characteristics in the area are not obvious;

[0070] Where, f ROI The calculation method is as follows: first, perform ROIAlign operation on the input feature map F and the detected target box bbox to obtain the initial regional features, and then compare the obtained ROI features with the frost damage area weight w frost Multiply them together to highlight the characteristic information of the frost-damaged area;

[0071] Where, L det Jointly optimize the loss function for multiple tasks, used for joint optimization of target detection and instance segmentation; L Focal L is the focal loss, which is used for target classification and is suitable for handling situations where positive and negative samples are extremely unbalanced, focusing on the classification of difficult samples. SmoothL1 Mainly used for regression optimization of the bounding box bbox, calculating the error between the predicted bounding box and the true bounding box; L Dice It is mainly used for instance segmentation task (mask prediction) loss to measure the overlap between the predicted mask and the true mask.

[0072] S4. Use the BiLSTM model to perform time series modeling on the frost damage ROI features at multiple consecutive time points, analyze the dynamic evolution trend of frost damage, and thus achieve online accurate detection of frost damage to tea tree buds and frost damage trend prediction and early warning.

[0073] The BiLSTM model is used to process the ROI features of the frozen area extracted by Mask R-CNN and output the feature h at each time point.t , and then use the attention mechanism to calculate the corresponding attention weight a for each time feature t , and then fuse to obtain the feature vector Z after the attention weight is weighted, input to the fully connected classification layer, and output the frost damage level p(y|c) through the softmax function;

[0074]

[0075]

[0076] Where a t represents the attention weight of the feature at the tth moment, that is, the importance of the feature at each moment; T represents the number of time steps of the input sequence, such as 3 time points in the morning, noon and afternoon; h t is the hidden feature output by BiLSTM at the tth time step; W and V are the weight parameters to be learned; h K represents the features of BiLSTM at the kth time point and the information before and after it, Z represents the feature vector weighted by the attention weight, which is the weighted sum of the time series features; W c and b c are the weights and biases of the classification layer to be learned; p(y|c) is the predicted probability output of the classification task, such as the prediction of frost damage level; the softmax function is used to convert the output features into probability values ​​for multi-classification decision making.

[0077] Based on the analysis results from the above steps, the corresponding tea bud frost damage classification results are output. These frost damage levels are classified into four categories: no frost damage, mild frost damage, moderate frost damage, and severe frost damage. Combined with BiLSTM's time series information to optimize classification accuracy, the system can be integrated into drones or ground-based monitoring equipment to analyze the dynamic evolution of frost damage, enabling real-time image acquisition, online frost damage detection, and frost damage trend prediction and early warning, improving frost damage monitoring capabilities in tea-growing areas.

[0078] Figure 2 The figure is a comparison diagram of the actual detection effect of the method of the present invention in the laboratory and the field environment. Figure 2 As can be seen, the image compares the image characteristics of laboratory and real-world field conditions, validating the adaptability and generalization performance of the proposed cross-scenario frost damage detection method in different environments. By introducing strategies such as style transfer, domain alignment, and residual enhancement, the model maintains high detection performance even in complex natural environments, effectively improving the practicality and reliability of the tea bud frost damage detection system.

[0079] Example 2:

[0080] The present invention proposes a tea tree bud frost damage detection system based on scene migration, which is applied to the tea tree bud frost damage detection and early warning method based on scene migration described in Example 1, comprising:

[0081] An image acquisition module, including an RGB camera and a colloidal gold reagent detection device, is used to collect image information to be analyzed of the agricultural products to be tested;

[0082] Data enhancement module, which performs style conversion through CycleGAN to improve data diversity;

[0083] The target detection module detects frost-damaged areas based on Mask R-CNN and outputs ROI features;

[0084] Time series analysis module, based on BiLSTM, for frost damage trend prediction and risk warning;

[0085] The user interaction module is used to display the detection results and provide suggestions for preventing frost damage in tea gardens.

[0086] Example 3:

[0087] An electronic device comprising a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other via the communication bus. The processor is configured to execute a computer program stored in the memory and implement the steps of the scenario-migration-based tea bud frost damage detection and early warning method described in Example 1.

[0088] This paper uses CycleGAN style transfer and domain adaptation technology to effectively solve the generalization problem when migrating laboratory-trained models to field environments, thereby improving the model's cross-scenario adaptability. By combining Mask R-CNN with the BiLSTM time series model, it achieves accurate detection of frost-damaged areas of tea tree buds and prediction of frost damage trends, making up for the deficiency of traditional single-frame detection in being unable to perform frost damage trend analysis. It also provides intelligent classification and real-time early warning of frost damage levels, helping tea farmers to make precise management and prevention and control decisions, and has high practical application value and promotion potential.

[0089] The above is an exemplary description of the invention. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as such non-substantial improvements are made by adopting the method concept and technical solution of the present invention, or the concept and technical solution of the invention are directly applied to other occasions without improvement, they are all within the scope of protection of the present invention.

Claims

1. A method for detecting and warning frost damage to tea tree buds based on scene migration, characterized in that: The following steps are involved: S1. Collect image information data of tea tree buds to be tested in laboratory and field environments; S2. Cross-scene data generation: using the CycleGAN network for image style transfer, converting laboratory labeled data into data with the style of a field environment to eliminate data distribution differences between scenes; S3. Input the data after style transfer and domain adaptation processing into the Mask R-CNN model to detect and segment the target area of ​​the tea tree bud frost damage area and extract the ROI features of the frost damage area; S4. Use the BiLSTM model to perform time series modeling on the frost damage ROI features at multiple consecutive time points, analyze the dynamic evolution trend of frost damage, and thus achieve online accurate detection of frost damage to tea tree buds, frost damage trend prediction, and early warning.

2. The method for detecting and warning of frost damage to tea buds based on scene migration according to claim 1, characterized in that: In step S1, the image information data of tea tree buds to be detected in the laboratory environment and the field environment are collected, including: the image dataset D lab and Daejeon Unlabeled Image Dataset D field , as follows: Where D lab represents the set of frost damage image samples labeled by the laboratory, x i lab represents the image sample collected by the i-th laboratory, y i cls Represents image x i lab Corresponding frost damage level label, y i mask Represents image x i lab The corresponding pixel-level frost damage area mask, D field represents unlabeled image data in a field environment, represents the image sample collected from the i-th field, H*W represents the length and width of the image, N lab represents the total number of samples labeled in the laboratory, and N field represents the total number of unlabeled samples in the field.

3. The method for detecting and warning of frost damage to tea tree buds based on scene migration according to claim 1, characterized in that: In step S2, the CycleGAN network is used to perform image style transfer. Specifically, unsupervised learning is used to generate tea tree bud frost damage data in different environments, and the data is used to enhance the training set and improve the model's cross-scenario adaptability.

4. The method for detecting and warning of frost damage to tea tree buds based on scene migration according to claim 1, characterized in that: In step S3, the domain adaptation processing is specifically: using domain adaptation technology, including adversarial training and feature alignment, to introduce a domain discriminator in the feature extraction layer of the Mask R-CNN object detection network to reduce the domain bias of different tea garden environments and improve the generalization ability of the model.

5. The method for detecting and warning of frost damage to tea tree buds based on scene migration according to claim 1, characterized in that: In step S3, the Mask R-CNN model is used to detect and segment the target area of ​​the tea tree bud frost damage area, and outputs a bounding box bbox of the frost damage area, a mask mask and the extracted frost damage area feature ROI; an indicator or score of the severity of the frost damage in the ROI area is I frost , and use the Sigmoid function to get the weight w of the frost damage area frost Then, it is multiplied with the original ROI feature to enhance the recognition ability of the frost-damaged area and provide feature information for subsequent time series analysis; During the training process, the Mask R-CNN model jointly optimizes L det Perform end-to-end optimization, the loss function includes the focal loss L Focal , bounding box regression loss L SmoothL1 and mask segmentation loss L Dice , improving the accuracy of frost-damaged area detection, bounding box positioning accuracy, and instance mask segmentation quality; <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> det <h2 style=";text-align:left;direction:ltr"> =0.7L<h2 style=";text-align:left;direction:ltr"> Focal <h2 style=";text-align:left;direction:ltr"> +0.2L<h2 style=";text-align:left;direction:ltr"> SmoothL1 <h2 style=";text-align:left;direction:ltr"> +0.1LDice, Where w frost is the frost damage area weight, which is used to indicate the severity of frost damage in the ROI area. Its value range is between 0 and 1 and is obtained using the Sigmoid function. frost Generally, the value is between [0,1]; the coefficient 10 and the bias 0.5 are used to adjust the steepness of the Sigmoid curve so that when I frost When I > 0.5, the weight quickly approaches 1, indicating that the characteristics of the frost damage area are significant; when I frost When <0.5, the weight rapidly approaches 0, indicating that the freezing damage characteristics in the area are not obvious; Where, f ROI The calculation method is as follows: first, perform ROIAlign operation on the input feature map F and the detected target box bbox to obtain the initial regional features, and then compare the obtained ROI features with the frost damage area weight w frost Multiply them together to highlight the characteristic information of the frost-damaged area; Where, L det Jointly optimize the loss function for multiple tasks, used for joint optimization of target detection and instance segmentation; L Focal L is the focal loss, which is used for target classification and is suitable for handling situations where positive and negative samples are extremely unbalanced, focusing on the classification of difficult samples. SmoothL1 Mainly used for regression optimization of the bounding box bbox, calculating the error between the predicted bounding box and the true bounding box; L Dice Mainly used for instance segmentation task loss to measure the degree of overlap between the predicted mask and the true mask.

6. The method for detecting and warning of frost damage to tea tree buds based on scene migration according to claim 1, characterized in that: In step S4, the BiLSTM model is used to process the ROI features of the frozen area extracted by Mask R-CNN and output the feature h at each time point. t , and then use the attention mechanism to calculate the corresponding attention weight a for each time feature t , and then fuse to obtain the feature vector Z after the attention weight is weighted, input to the fully connected classification layer, and output the frost damage level p(y|c) through the softmax function; Where a t represents the attention weight of the feature at the tth moment, that is, the importance of the feature at each moment; T represents the number of time steps of the input sequence, such as 3 time points in the morning, noon and afternoon; h t is the hidden feature output by BiLSTM at the tth time step; W and V are the weight parameters to be learned; h K represents the features of BiLSTM at the kth time point and the information before and after it, Z represents the feature vector weighted by the attention weight, which is the weighted sum of the time series features; W c and b c are the classification layer weights and biases to be learned; p(y|c) is the predicted probability output of the classification task, such as the frost damage level prediction; The softmax function is used to convert output features into probability values ​​for multi-classification decision making.

7. The method for detecting and warning of frost damage to tea tree buds based on scene migration according to claim 1, characterized in that: In step S4, the frost damage levels of the tea tree buds include four levels: no frost damage, mild frost damage, moderate frost damage and severe frost damage; combined with the BiLSTM time series information to optimize the classification accuracy, it can be integrated into drones or ground monitoring equipment to achieve real-time image acquisition and online frost damage detection, thereby improving the frost damage monitoring capabilities in tea tree planting areas.

8. A tea tree bud frost damage detection and early warning system based on scene migration, applied to the tea tree bud frost damage detection and early warning method based on scene migration according to any one of claims 1 to 7, characterized in that: include: An image acquisition module, including an RGB camera and a colloidal gold reagent detection device, is used to collect image information to be analyzed of the agricultural products to be tested; Data enhancement module, which performs style conversion through CycleGAN to improve data diversity; The target detection module detects frost-damaged areas based on Mask R-CNN and outputs ROI features; Time series analysis module, based on BiLSTM, for frost damage trend prediction and risk warning; The user interaction module is used to display the detection results and provide suggestions for preventing frost damage in tea gardens.

9. An electronic device comprising a processor and a memory storing a computer program, characterized in that: The processor is used to execute the computer program stored in the memory and implement the steps in the tea tree bud frost damage detection and early warning method based on scene migration as described in any one of claims 1 to 7.

Images