Tea garden drought real-time early warning method and system

Abstract

The invention provides a tea garden drought real-time early warning method and system in the technical field of intelligent agriculture. The method comprises the steps that S1, a large amount of historical monitoring data is collected to construct a data set, and an established tea garden drought prediction model is trained; s2, compressing the tea garden drought prediction model through a knowledge distillation technology, collecting actual monitoring data from the tea garden through a sensor array to perform data drift compensation training on the tea garden drought prediction model, and deploying the data to each edge computing device arranged in the tea garden; s3, the edge computing device inputs the real-time monitoring data collected by the sensor array into the deployed tea garden drought prediction model based on the load balancing technology and the multi-thread technology, and a tea garden drought prediction result is obtained; and S4, when the tea garden drought prediction result is abnormal, the tea garden drought prediction result is pushed to the management terminal in real time for early warning. The tea garden drought early warning system has the advantage that the timeliness and accuracy of tea garden drought early warning are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of smart agricultural technology, and in particular to a real-time early warning method and system for drought conditions in tea gardens. Background Art

[0002] Drought has a significant impact on tea plant growth, physiological metabolism, tea quality, and yield. Drought stress can have at least the following effects on tea plants: 1. It causes tea leaves to become smaller and thinner, with shortened leaf length and width, fewer leaves, and dwarfed plants. 2. It increases the number of lateral roots, but restricts overall root development, causing root hairs to dry out and die, and reducing absorptive capacity. 3. Long-term drought can stagnate tea plant growth, leading to slow shoot growth, bud and leaf shrinkage, and even death. 4. It inhibits photosynthesis, causing a decrease in photosynthetic rate. 5. It can lead to changes in the content of quality components in tea leaves, such as caffeine, amino acids, and tea polyphenols. Therefore, it is necessary to predict drought conditions in tea plantations and, based on these predictions, issue early warnings to better manage them.

[0003] However, the current tea garden drought early warning system faces two technical bottlenecks: First, the hierarchical architecture based on "sensors-cloud servers" has inherent transmission delays. Experimental tests have shown (referring to the GB / T33778-2017 communication standard) that the end-to-end delay of a typical IoT system is 8-12 seconds, resulting in insufficient real-time response capabilities for key parameters such as soil moisture. Second, existing early warning models mostly use a single machine learning algorithm (such as isolated random forests or BP neural networks), which has limited ability to integrate and process multi-dimensional heterogeneous data (soil conductivity, soil pH, TBQ total radiation, etc.). Cross-regional verification shows that its drought classification accuracy is only 72-78% (confidence interval 95%), which is difficult to meet the needs of precision agriculture decision-making.

[0004] Therefore, how to provide a real-time early warning method and system for tea garden drought to improve the timeliness and accuracy of tea garden drought warning has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a real-time early warning method and system for drought in tea gardens, so as to improve the timeliness and accuracy of early warning of drought in tea gardens.

[0006] In a first aspect, the present invention provides a real-time early warning method for drought in a tea garden, comprising the following steps:

[0007] Step S1: creating a tea garden drought prediction model based on the multimodal feature extraction layer, the heterogeneous model fusion layer, the soil entropy prediction layer, and the drought mapping output layer, and setting a loss function of the tea garden drought prediction model;

[0008] The multimodal feature extraction layer is used to extract features from the monitoring data to obtain splicing features; the heterogeneous model fusion layer is used to perform feature mining and fusion on the splicing features to obtain fusion features; the soil entropy prediction layer is used to calculate the soil entropy value based on the fusion features; the drought mapping output layer is used to map the soil entropy value according to the set drought classification rules, and output the tea garden drought prediction result with the drought period, drought level and drought location;

[0009] Step S2: Collect a large amount of historical monitoring data, pre-process and annotate each of the historical monitoring data to construct a data set;

[0010] Step S3, dividing the data set into a training set, a validation set, and a test set, training the tea garden drought prediction model using the training set and the loss function, dynamically pruning the tea garden drought prediction model during the training process, validating the trained tea garden drought prediction model using the validation set, and testing the verified tea garden drought prediction model using the test set;

[0011] Step S4: compressing the tested tea garden drought prediction model using knowledge distillation technology, collecting actual monitoring data from the tea garden through a sensor array, training the tea garden drought prediction model after knowledge distillation with data drift compensation using the actual monitoring data, and then deploying the tea garden drought prediction model to various edge computing devices located in the tea garden;

[0012] Step S5: The edge computing device inputs the real-time monitoring data collected by the sensor array into the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology to obtain the tea garden drought prediction result;

[0013] Step S6: The edge computing device monitors the tea garden drought prediction results for abnormalities. When an abnormality occurs, the tea garden drought prediction results are pushed to a pre-associated management terminal in real time for tea garden drought warning.

[0014] Furthermore, in step S1, the multimodal feature extraction layer is constructed based on a structured parameter processing module, an image processing module, and a feature splicing module; the structured parameter processing module is used to extract structured high-order features from the structured data of the monitoring data through a LASSO sub-model and a KNN sub-model; the image processing module is used to extract multi-scale image texture features from the remote sensing image of the monitoring data through a CNN sub-model; and the feature splicing module is used to splice the structured high-order features and the multi-scale image texture features to obtain spliced ​​features;

[0015] The heterogeneous model fusion layer is constructed based on a tree model channel, a kernel method channel, a linear channel, a decision channel and a dynamic feature fusion module; the tree model channel is used to extract nonlinear interaction sub-features from each splicing feature through a LightGBM sub-model, an XGBoost sub-model and an RF sub-model; the kernel method channel is used to extract high-dimensional space sub-features from each splicing feature through the kernel function of the LSVM sub-model; the linear channel is used to extract linear sub-features from each splicing feature through a linear regression sub-model; the decision channel is used to extract segmentation sub-features from each splicing feature through a decision tree sub-model; the dynamic feature fusion module is used to dynamically fuse nonlinear interaction sub-features, high-dimensional space sub-features, linear sub-features and segmentation sub-features through gated weighting to obtain fused features;

[0016] The soil entropy prediction layer is used to dynamically adjust the weights of each sub-feature in the fusion feature through the Bayesian optimizer, and output the soil entropy value through the deep residual network;

[0017] The specific drought classification rules are as follows:

[0018] The quantile of soil entropy value is (75,100], which corresponds to the first level of drought; the quantile is (25,75], which corresponds to the second level of drought; the quantile is (5,25], which corresponds to the third level of drought; the quantile is (0,5], which corresponds to the fourth level of drought.

[0019] Furthermore, the step S2 is specifically as follows:

[0020] Acquire a large amount of historical monitoring data including structured data and remote sensing images, wherein the structured data at least includes maximum light intensity, average light intensity, sunshine duration, maximum temperature, minimum temperature, average temperature, average daytime temperature, average nighttime temperature, daily temperature difference, maximum air humidity, minimum air humidity, average relative humidity, average daytime air humidity, average nighttime air humidity, maximum carbon dioxide concentration, minimum carbon dioxide concentration, average carbon dioxide concentration, average daytime carbon dioxide concentration, average nighttime carbon dioxide concentration, maximum air pressure value, minimum air pressure value value, average air pressure value, daytime average air pressure value, nighttime average air pressure value, air pressure difference, daily average PM2.5, daily average PM10, maximum TBQ total radiation, minimum TBQ total radiation, daily average TBQ total radiation, maximum soil temperature, minimum soil temperature, daily average soil temperature, soil temperature difference, maximum soil pH, minimum soil pH, average soil pH, maximum soil salinity value, minimum soil salinity value, average soil salinity value, maximum soil conductivity, minimum soil conductivity, average soil conductivity, rainfall, wind direction, wind speed, sampling location and sampling time;

[0021] The structured data in each of the historical monitoring data are preprocessed by at least filling missing values ​​and repairing outliers, and the remote sensing images in the historical monitoring data are preprocessed by at least noise reduction, contrast enhancement, morphological adjustment, scaling, and color space conversion; the missing value filling is based on the K-nearest neighbor filling method; and the outlier repair is based on the K-means clustering algorithm;

[0022] The pre-processed historical monitoring data are labeled with soil entropy value, drought level and time segment label to construct a data set; the time segment label is the month of the sampling time.

[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 K-fold cross-validation method, and a tea garden drought prediction model is trained using the training set. During the training process, the hyperparameters of the tea garden drought prediction model, including at least a learning rate, a batch size, and a number of training rounds, are continuously optimized. The tea garden drought prediction model is dynamically pruned to remove weights or channels with importance scores lower than a preset score until the loss function converges or reaches a preset convergence condition;

[0025] The trained tea garden drought prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if so, the verification passes, and the ROC curve and calibration curve are drawn using the test set to test the tea garden drought prediction model.

[0026] Furthermore, the step S4 is specifically as follows:

[0027] The tea garden drought prediction model that has passed the test is used as the teacher model. The number of network layers and the number of neurons in the teacher model are simplified to generate a student model. The distillation loss function of the student model is set based on the cross entropy loss and the KL divergence loss. The teacher model is called to train the student model through the knowledge distillation technology until the distillation loss function converges. The student model is used as the tea garden drought prediction model after knowledge distillation;

[0028] Actual monitoring data is collected from the tea garden through a sensor array including a humidity sensor, a light intensity sensor, a temperature sensor, a carbon dioxide sensor, an air pressure sensor, a particulate matter sensor, a global solar radiation sensor, a pH sensor, a salinity sensor, a conductivity sensor, a rainfall sensor, a wind speed and direction sensor, a locator, and a wireless communication module; the wireless communication module is used to obtain remote sensing images;

[0029] After data drift compensation training is performed on the tea garden drought prediction model after knowledge distillation using the actual monitoring data, the tea garden drought prediction model is deployed to each edge computing device located in the tea garden through containerization technology;

[0030] The step S5 is specifically as follows:

[0031] Each edge device maintains a heartbeat connection and exchanges load monitoring data with each other. The load monitoring data is parsed using load balancing technology. The real-time monitoring data collected by the sensor array is input into the tea garden drought prediction model with the smallest load through multi-threading technology for inference to obtain the tea garden drought prediction result.

[0032] The step S6 is specifically as follows:

[0033] The edge computing device monitors the tea garden drought prediction results for anomalies through preset monitoring rules. When an anomaly occurs, the tea garden drought prediction results and the corresponding real-time monitoring data are compressed into a compressed warning data packet, and the compressed warning data packet is pushed in real time to a pre-associated management terminal for tea garden drought warning.

[0034] In a second aspect, the present invention provides a real-time early warning system for drought in tea gardens, comprising the following modules:

[0035] A tea garden drought prediction model creation module is used to create a tea garden drought prediction model based on the multimodal feature extraction layer, the heterogeneous model fusion layer, the soil entropy prediction layer, and the drought mapping output layer, and set the loss function of the tea garden drought prediction model;

[0036] The multimodal feature extraction layer is used to extract features from the monitoring data to obtain splicing features; the heterogeneous model fusion layer is used to perform feature mining and fusion on the splicing features to obtain fusion features; the soil entropy prediction layer is used to calculate the soil entropy value based on the fusion features; the drought mapping output layer is used to map the soil entropy value according to the set drought classification rules, and output the tea garden drought prediction result with the drought period, drought level and drought location;

[0037] A data set construction module is used to collect a large amount of historical monitoring data, pre-process the historical monitoring data, and construct a data set after annotating the data;

[0038] a tea garden drought prediction model training module, configured to divide the data set into a training set, a validation set, and a test set, train the tea garden drought prediction model using the training set and a loss function, dynamically prune the tea garden drought prediction model during the training process, validate the trained tea garden drought prediction model using the validation set, and test the validated tea garden drought prediction model using the test set;

[0039] A tea garden drought prediction model deployment module is used to compress the tested tea garden drought prediction model using knowledge distillation technology, collect actual monitoring data from the tea garden through a sensor array, perform data drift compensation training on the knowledge distilled tea garden drought prediction model using the actual monitoring data, and then deploy the tea garden drought prediction model to various edge computing devices located in the tea garden;

[0040] The tea garden drought prediction module is used to input the real-time monitoring data collected by the sensor array into the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology to obtain the tea garden drought prediction results;

[0041] The real-time drought warning module is used for edge computing equipment to monitor the tea garden drought prediction results for abnormalities, and when an abnormality occurs, the tea garden drought prediction results are pushed to the pre-associated management terminal in real time to issue a tea garden drought warning.

[0042] Furthermore, in the tea garden drought prediction model creation module, the multimodal feature extraction layer is constructed based on the structured parameter processing module, the image processing module and the feature splicing module; the structured parameter processing module is used to extract structured high-order features from the structured data of the monitoring data through the LASSO sub-model and the KNN sub-model; the image processing module is used to extract multi-scale image texture features from the remote sensing images of the monitoring data through the CNN sub-model; the feature splicing module is used to splice the structured high-order features and the multi-scale image texture features to obtain splicing features;

[0043] The heterogeneous model fusion layer is constructed based on a tree model channel, a kernel method channel, a linear channel, a decision channel and a dynamic feature fusion module; the tree model channel is used to extract nonlinear interaction sub-features from each splicing feature through a LightGBM sub-model, an XGBoost sub-model and an RF sub-model; the kernel method channel is used to extract high-dimensional space sub-features from each splicing feature through the kernel function of the LSVM sub-model; the linear channel is used to extract linear sub-features from each splicing feature through a linear regression sub-model; the decision channel is used to extract segmentation sub-features from each splicing feature through a decision tree sub-model; the dynamic feature fusion module is used to dynamically fuse nonlinear interaction sub-features, high-dimensional space sub-features, linear sub-features and segmentation sub-features through gated weighting to obtain fused features;

[0044] The soil entropy prediction layer is used to dynamically adjust the weights of each sub-feature in the fusion feature through the Bayesian optimizer, and output the soil entropy value through the deep residual network;

[0045] The specific drought classification rules are as follows:

[0046] The quantile of soil entropy value is (75,100], which corresponds to the first level of drought; the quantile is (25,75], which corresponds to the second level of drought; the quantile is (5,25], which corresponds to the third level of drought; the quantile is (0,5], which corresponds to the fourth level of drought.

[0047] Furthermore, the dataset construction module is specifically used to:

[0048] Acquire a large amount of historical monitoring data including structured data and remote sensing images, wherein the structured data at least includes maximum light intensity, average light intensity, sunshine duration, maximum temperature, minimum temperature, average temperature, average daytime temperature, average nighttime temperature, daily temperature difference, maximum air humidity, minimum air humidity, average relative humidity, average daytime air humidity, average nighttime air humidity, maximum carbon dioxide concentration, minimum carbon dioxide concentration, average carbon dioxide concentration, average daytime carbon dioxide concentration, average nighttime carbon dioxide concentration, maximum air pressure value, minimum air pressure value value, average air pressure value, daytime average air pressure value, nighttime average air pressure value, air pressure difference, daily average PM2.5, daily average PM10, maximum TBQ total radiation, minimum TBQ total radiation, daily average TBQ total radiation, maximum soil temperature, minimum soil temperature, daily average soil temperature, soil temperature difference, maximum soil pH, minimum soil pH, average soil pH, maximum soil salinity value, minimum soil salinity value, average soil salinity value, maximum soil conductivity, minimum soil conductivity, average soil conductivity, rainfall, wind direction, wind speed, sampling location and sampling time;

[0049] The structured data in each of the historical monitoring data are preprocessed by at least filling missing values ​​and repairing outliers, and the remote sensing images in the historical monitoring data are preprocessed by at least noise reduction, contrast enhancement, morphological adjustment, scaling, and color space conversion; the missing value filling is based on the K-nearest neighbor filling method; and the outlier repair is based on the K-means clustering algorithm;

[0050] The pre-processed historical monitoring data are labeled with soil entropy value, drought level and time segment label to construct a data set; the time segment label is the month of the sampling time.

[0051] Furthermore, the tea garden drought prediction model training module is specifically used to:

[0052] The data set is divided into a training set, a validation set, and a test set based on a K-fold cross-validation method, and a tea garden drought prediction model is trained using the training set. During the training process, the hyperparameters of the tea garden drought prediction model, including at least a learning rate, a batch size, and a number of training rounds, are continuously optimized. The tea garden drought prediction model is dynamically pruned to remove weights or channels with importance scores lower than a preset score until the loss function converges or reaches a preset convergence condition;

[0053] The trained tea garden drought prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if so, the verification passes, and the ROC curve and calibration curve are drawn using the test set to test the tea garden drought prediction model.

[0054] Furthermore, the tea garden drought prediction model deployment module is specifically used to:

[0055] The tea garden drought prediction model that has passed the test is used as the teacher model. The number of network layers and the number of neurons in the teacher model are simplified to generate a student model. The distillation loss function of the student model is set based on the cross entropy loss and the KL divergence loss. The teacher model is called to train the student model through the knowledge distillation technology until the distillation loss function converges. The student model is used as the tea garden drought prediction model after knowledge distillation;

[0056] Actual monitoring data is collected from the tea garden through a sensor array including a humidity sensor, a light intensity sensor, a temperature sensor, a carbon dioxide sensor, an air pressure sensor, a particulate matter sensor, a global solar radiation sensor, a pH sensor, a salinity sensor, a conductivity sensor, a rainfall sensor, a wind speed and direction sensor, a locator, and a wireless communication module; the wireless communication module is used to obtain remote sensing images;

[0057] After data drift compensation training is performed on the tea garden drought prediction model after knowledge distillation using the actual monitoring data, the tea garden drought prediction model is deployed to each edge computing device located in the tea garden through containerization technology;

[0058] The tea garden drought prediction module is specifically used for:

[0059] Each edge device maintains a heartbeat connection and exchanges load monitoring data with each other. The load monitoring data is parsed using load balancing technology. The real-time monitoring data collected by the sensor array is input into the tea garden drought prediction model with the smallest load through multi-threading technology for inference to obtain the tea garden drought prediction result.

[0060] The drought real-time early warning module is specifically used for:

[0061] The edge computing device monitors the tea garden drought prediction results for anomalies through preset monitoring rules. When an anomaly occurs, the tea garden drought prediction results and the corresponding real-time monitoring data are compressed into a compressed warning data packet, and the compressed warning data packet is pushed in real time to a pre-associated management terminal for tea garden drought warning.

[0062] The advantages of the present invention are:

[0063] 1. A tea garden drought prediction model is created through the multimodal feature extraction layer, heterogeneous model fusion layer, soil entropy prediction layer and drought mapping output layer, and the loss function of the tea garden drought prediction model is set; a large amount of historical monitoring data is collected to construct a data set, and the data set is divided into a training set, a validation set and a test set. The tea garden drought prediction model is trained through the training set and the loss function, and the tea garden drought prediction model is dynamically pruned during the training process. The trained tea garden drought prediction model is verified through the validation set, and the verified tea garden drought prediction model is tested through the test set; then the tested tea garden drought prediction model is compressed through the knowledge distillation technology, actual monitoring data is collected from the tea garden through the sensor array, and the tea garden drought prediction model after knowledge distillation is trained with data drift compensation through each actual monitoring data, and then the tea garden drought prediction model is deployed to each edge computing device located in the tea garden; the edge computing device uses the real-time monitoring data collected by the sensor array to input the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology. The method uses a pre-trained tea garden drought prediction model to predict the tea garden drought, and performs abnormal monitoring on the tea garden drought prediction results. When an abnormality occurs, the tea garden drought prediction results are pushed to the pre-associated management terminal in real time for tea garden drought warning. That is, the tea garden drought warning is performed through the pre-trained tea garden drought prediction model. Since the tea garden drought prediction model combines the LASSO sub-model, KNN sub-model, CNN sub-model, LightGBM sub-model, XGBoost sub-model, RF sub-model, LSVM sub-model, linear regression sub-model and decision tree sub-model, it integrates the advantages of each model and greatly improves the feature extraction ability, generalization ability and robustness. In addition, the tea garden drought prediction model is dynamically pruned during training and knowledge distillation is performed before deployment, which effectively compresses the model size of the tea garden drought prediction model and facilitates deployment on edge computing devices with limited computing power. The edge computing device can locally process real-time monitoring data without uploading it to the server. In the prediction process, load balancing technology and multi-threading technology are combined to effectively improve the inference speed, thereby greatly improving the timeliness and accuracy of the tea garden drought warning.

[0064] 2. By extracting structured high-order features through LASSO / KNN and combining it with CNN to extract multi-scale image texture features, we can achieve multimodal feature fusion of tea garden environmental parameters (such as soil temperature and pH value) and remote sensing images (such as vegetation coverage), effectively improving the comprehensiveness of data representation.

[0065] 3. By setting up a heterogeneous fusion layer to integrate differentiated models such as LightGBM (efficient gradient boosting), XGBoost (precise segmentation), LSVM (high-dimensional kernel mapping), and decision tree (rule segmentation), and dynamically fusing nonlinear, linear, and segmentation sub-features through gated weighting, the model's adaptability to different data distributions is significantly improved, avoiding the bias of a single model.

[0066] 4. By setting up the soil entropy prediction layer and using the Bayesian optimizer to dynamically adjust the feature weights, combined with the deep residual network to solve the gradient vanishing problem, we can ensure the accurate modeling of complex nonlinear relationships and improve the accuracy of soil entropy prediction.

[0067] 5. By introducing dynamic pruning (removing low-importance weights / channels) and hyperparameter optimization (learning rate, batch size) in the training process of the tea garden drought prediction model, combined with K-fold cross-validation, overfitting can be effectively prevented, model complexity can be reduced, and generalization performance can be improved.

[0068] 6. Compress the teacher model into a student model through knowledge distillation, reducing computing resource requirements while retaining accuracy; combined with data drift compensation training (such as seasonal changes in environmental parameters), ensure the long-term stability of the tea garden drought prediction model on edge devices.

[0069] 7. Through the heartbeat connection and load balancing strategy between edge computing devices, combined with multi-threading technology to distribute real-time monitoring data, maximize the use of edge computing resources, and meet the real-time warning needs of the wide-area distribution of tea gardens.

[0070] 8. Drought levels (level 1 to 4) are divided by soil entropy value quantiles to objectively reflect data distribution characteristics and avoid subjectivity in threshold setting; abnormal monitoring rules are combined with data compression and real-time push to ensure rapid response to sudden droughts.

[0071] 9. By integrating more than 40 structured parameters such as light, temperature, humidity, and soil conductivity, as well as remote sensing images, it covers multi-dimensional factors such as meteorology, soil, and the environment; K-nearest neighbor filling and K-means anomaly repair improve data integrity and provide high-quality input for the model.

[0072] 10. By integrating technologies from multiple fields such as agricultural meteorology (drought classification rules), the Internet of Things (sensor arrays), edge computing (containerized deployment) and machine learning (heterogeneous models), an interdisciplinary collaborative solution is formed to break through the limitations of traditional single-disciplinary methods in tea garden monitoring.

[0073] 11. By automating the entire process from data acquisition (sensor array), preprocessing (KNN / K-means repair), model training (dynamic pruning) to warning push (compressed data packets), human intervention is reduced and system response efficiency is improved (for example, the delay of abnormal warning is reduced to seconds).

[0074] 12. By dividing the time labels by month, the model can automatically learn the seasonal patterns of drought in tea gardens (such as the difference between rainy season and dry season), improving the temporal matching degree between the prediction results and the actual environmental changes.

[0075] 13. The tea garden drought prediction model is independently run by each edge computing device, and distributed computing is achieved through heartbeat connection and load balancing to avoid the risk of single point failure (for example, the failure of an edge computing device does not affect the overall system) and improve the system's disaster recovery capabilities.

[0076] 14. By setting the tea garden drought prediction results to include drought period, drought level and drought location, combined with the GIS map integration of the management terminal, the spatial distribution of drought can be visualized, assisting decision makers to quickly locate problem areas and trace historical data.

[0077] 15. Through multimodal feature fusion (co-processing of structured data and remote sensing images) and dynamic integration of heterogeneous models (gated weighted fusion of differentiated models such as LightGBM, LSVM, and decision trees), combined with Bayesian optimization and dynamic pruning, high-precision soil entropy prediction is achieved; knowledge distillation and edge computing container deployment are adopted to reduce model complexity while ensuring real-time performance, and to improve the long-term stability of the system through data drift compensation and load balancing multi-threaded processing; the quantile drought classification rules and abnormal monitoring compression warning mechanism are innovatively introduced to objectively adapt to different tea garden environments, combined with the multi-dimensional data coverage of the sensor array and the end-to-end automated process, significantly improving the accuracy and response efficiency of drought warnings, while having low power consumption, easy scalability and cross-scenario generalization capabilities, providing a set of efficient, reliable and reusable edge intelligence solutions for smart agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0079] Figure 1 The present invention is a flow chart of a real-time early warning method for drought in a tea garden.

[0080] Figure 2 The present invention is a structural diagram of a real-time early warning system for drought in a tea garden. DETAILED DESCRIPTION

[0081] The technical solution in the embodiments of the present application has the following overall idea: tea garden drought prediction model is used to provide tea garden drought warning. Since the tea garden drought prediction model combines the LASSO sub-model, KNN sub-model, CNN sub-model, LightGBM sub-model, XGBoost sub-model, RF sub-model, LSVM sub-model, linear regression sub-model and decision tree sub-model, it integrates the advantages of each model and greatly improves the feature extraction ability, generalization ability and robustness; and the tea garden drought prediction model is dynamically pruned during training and knowledge distilled before deployment, which effectively compresses the model size of the tea garden drought prediction model and facilitates deployment on edge computing devices with limited computing power. The edge computing device can locally process real-time monitoring data without uploading it to the server. In the prediction process, load balancing technology and multi-threading technology are combined to effectively improve the inference speed, so as to improve the timeliness and accuracy of tea garden drought warning.

[0082] Please refer to Figures 1 to 2 As shown, a preferred embodiment of a real-time early warning method for drought in a tea garden according to the present invention comprises the following steps:

[0083] Step S1: creating a tea garden drought prediction model based on the multimodal feature extraction layer, the heterogeneous model fusion layer, the soil entropy prediction layer, and the drought mapping output layer, and setting a loss function of the tea garden drought prediction model;

[0084] The multimodal feature extraction layer is used to extract features from the monitoring data to obtain splicing features; the heterogeneous model fusion layer is used to perform feature mining and fusion on the splicing features to obtain fusion features; the soil entropy prediction layer is used to calculate the soil entropy value based on the fusion features; the drought mapping output layer is used to map the soil entropy value according to the set drought classification rules, and output the tea garden drought prediction result with the drought period, drought level and drought location; the drought mapping output layer outputs the probability of the drought level (level 1 drought / level 2 drought / level 3 drought / level 4 drought), the probability of the drought period, and the probability of the drought location through the fully connected layer + Softmax, and introduces the temperature scaling technology to calibrate the prediction confidence;

[0085] By setting the tea garden drought prediction results to include drought period, drought level and drought location, and combining them with GIS map integration on the management terminal, the spatial distribution of drought can be visualized, helping decision makers to quickly locate problem areas and trace historical data.

[0086] Step S2: Collect a large amount of historical monitoring data, pre-process and annotate each of the historical monitoring data to construct a data set;

[0087] Step S3, dividing the data set into a training set, a validation set, and a test set, training the tea garden drought prediction model using the training set and the loss function, dynamically pruning the tea garden drought prediction model during the training process, verifying the trained tea garden drought prediction model using the validation set, and testing the verified tea garden drought prediction model using the test set; using the AdamW optimizer in the training process, combined with cosine annealing learning rate scheduling;

[0088] Step S4: compressing the tested tea garden drought prediction model using knowledge distillation technology, collecting actual monitoring data from the tea garden through a sensor array, training the tea garden drought prediction model after knowledge distillation with data drift compensation using the actual monitoring data, and then deploying the tea garden drought prediction model to various edge computing devices located in the tea garden;

[0089] Step S5: The edge computing device inputs the real-time monitoring data collected by the sensor array into the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology to obtain the tea garden drought prediction result;

[0090] Step S6: The edge computing device monitors the tea garden drought prediction results for abnormalities. When an abnormality occurs, the tea garden drought prediction results are pushed to a pre-associated management terminal in real time for tea garden drought warning.

[0091] By integrating technologies from multiple fields, including agricultural meteorology (drought classification rules), the Internet of Things (sensor arrays), edge computing (containerized deployment), and machine learning (heterogeneous models), an interdisciplinary collaborative solution is formed, breaking through the limitations of traditional single-disciplinary methods in tea garden monitoring.

[0092] By automating the entire process from data acquisition (sensor array), preprocessing (KNN / K-means repair), model training (dynamic pruning) to warning push (compressed data packets), manual intervention is reduced and system response efficiency is improved (for example, the delay in abnormal warning is reduced to seconds).

[0093] Through multimodal feature fusion (co-processing of structured data and remote sensing images) and dynamic integration of heterogeneous models (gated weighted fusion of differentiated models such as LightGBM, LSVM, and decision trees), combined with Bayesian optimization and dynamic pruning, high-precision soil entropy prediction is achieved; knowledge distillation and edge computing container deployment are adopted to reduce model complexity while ensuring real-time performance, and the long-term stability of the system is improved through data drift compensation and load balancing multi-threaded processing; the quantile drought classification rules and abnormal monitoring compression and early warning mechanism are innovatively introduced to objectively adapt to different tea garden environments, and combined with the multi-dimensional data coverage of the sensor array and the end-to-end automated process, the accuracy and response efficiency of drought warning are significantly improved. At the same time, it has low power consumption, easy scalability and cross-scenario generalization capabilities, providing a set of efficient, reliable and reusable edge intelligence solutions for smart agriculture.

[0094] In step S1, the multimodal feature extraction layer is constructed based on a structured parameter processing module, an image processing module, and a feature splicing module; the structured parameter processing module is used to extract structured high-order features from the structured data of the monitoring data through a LASSO sub-model and a KNN sub-model; the image processing module is used to extract multi-scale image texture features from the remote sensing image of the monitoring data through a CNN sub-model; and the feature splicing module is used to splice the structured high-order features and the multi-scale image texture features to obtain spliced ​​features;

[0095] By extracting structured high-order features through LASSO / KNN and combining it with CNN to extract multi-scale image texture features, we can achieve multimodal feature fusion of tea garden environmental parameters (such as soil temperature and pH value) and remote sensing images (such as vegetation coverage), effectively improving the comprehensiveness of data representation.

[0096] The heterogeneous model fusion layer is constructed based on a tree model channel, a kernel method channel, a linear channel, a decision channel and a dynamic feature fusion module; the tree model channel is used to extract nonlinear interaction sub-features from each splicing feature through a LightGBM sub-model, an XGBoost sub-model and an RF sub-model; the kernel method channel is used to extract high-dimensional space sub-features from each splicing feature through the kernel function of the LSVM sub-model; the linear channel is used to extract linear sub-features from each splicing feature through a linear regression sub-model; the decision channel is used to extract segmentation sub-features from each splicing feature through a decision tree sub-model; the dynamic feature fusion module is used to dynamically fuse nonlinear interaction sub-features, high-dimensional space sub-features, linear sub-features and segmentation sub-features through gated weighting to obtain fused features;

[0097] By setting up a heterogeneous fusion layer to integrate differentiated models such as LightGBM (efficient gradient boosting), XGBoost (precise segmentation), LSVM (high-dimensional kernel mapping), and decision tree (rule segmentation), and dynamically fusing nonlinear, linear, and segmentation sub-features through gated weighting, the model's adaptability to different data distributions is significantly improved, avoiding the bias of a single model.

[0098] The soil entropy prediction layer is used to dynamically adjust the weights of each sub-feature in the fusion feature through the Bayesian optimizer, and output the soil entropy value through the deep residual network;

[0099] By setting up the soil entropy prediction layer and using the Bayesian optimizer to dynamically adjust the feature weights, combined with the deep residual network to solve the gradient vanishing problem, we can ensure the accurate modeling of complex nonlinear relationships and improve the accuracy of soil entropy prediction.

[0100] Numerical features are extracted through the structured parameter processing module (LASSO+KNN), and the multi-scale texture features of remote sensing images are extracted by combining the image processing module (CNN). The meteorological, soil physical and chemical parameters and spatial image information are comprehensively utilized to comprehensively capture the multi-dimensional factors affecting soil entropy. By integrating models based on different principles such as tree models (LightGBM / XGBoost / RF), kernel methods (LSVM), linear regression, and decision trees, nonlinear interactions, high-dimensional spatial mapping, linear relationships, and decision boundary features are captured respectively, breaking through the limitations of a single model and greatly improving the accuracy of soil entropy prediction.

[0101] The specific drought classification rules are as follows:

[0102] The quantile of soil entropy value is (75,100], which corresponds to the first level of drought; the quantile is (25,75], which corresponds to the second level of drought; the quantile is (5,25], which corresponds to the third level of drought; the quantile is (0,5], which corresponds to the fourth level of drought.

[0103] The step S2 is specifically as follows:

[0104] Acquire a large amount of historical monitoring data including structured data and remote sensing images, wherein the structured data at least includes maximum light intensity, average light intensity, sunshine duration, maximum temperature, minimum temperature, average temperature, average daytime temperature, average nighttime temperature, daily temperature difference, maximum air humidity, minimum air humidity, average relative humidity, average daytime air humidity, average nighttime air humidity, maximum carbon dioxide concentration, minimum carbon dioxide concentration, average carbon dioxide concentration, average daytime carbon dioxide concentration, average nighttime carbon dioxide concentration, maximum air pressure value, minimum air pressure value value, average air pressure value, daytime average air pressure value, nighttime average air pressure value, air pressure difference, daily average PM2.5, daily average PM10, maximum TBQ total radiation, minimum TBQ total radiation, daily average TBQ total radiation, maximum soil temperature, minimum soil temperature, daily average soil temperature, soil temperature difference, maximum soil pH, minimum soil pH, average soil pH, maximum soil salinity value, minimum soil salinity value, average soil salinity value, maximum soil conductivity, minimum soil conductivity, average soil conductivity, rainfall, wind direction, wind speed, sampling location and sampling time;

[0105] By integrating more than 40 structured parameters and remote sensing images such as light, temperature and humidity, and soil conductivity, it covers multi-dimensional factors such as meteorology, soil, and the environment; K-nearest neighbor filling and K-means anomaly repair improve data integrity and provide high-quality input for the model.

[0106] The structured data in each of the historical monitoring data are preprocessed by at least filling missing values ​​and repairing outliers, and the remote sensing images in the historical monitoring data are preprocessed by at least noise reduction, contrast enhancement, morphological adjustment, scaling, and color space conversion; the missing value filling is based on the K-nearest neighbor filling method; and the outlier repair is based on the K-means clustering algorithm;

[0107] The pre-processed historical monitoring data are labeled with soil entropy value, drought level and time segment label to construct a data set; the time segment label is the month of the sampling time.

[0108] By dividing the time labels by month, the model can automatically learn the seasonal patterns of drought in tea gardens (such as the difference between rainy season and dry season), thereby improving the temporal matching between the prediction results and the actual environmental changes.

[0109] The step S3 is specifically as follows:

[0110] The data set is divided into a training set, a validation set, and a test set based on a K-fold cross-validation method, and a tea garden drought prediction model is trained using the training set. During the training process, the hyperparameters of the tea garden drought prediction model, including at least a learning rate, a batch size, and a number of training rounds, are continuously optimized. The tea garden drought prediction model is dynamically pruned to remove weights or channels with importance scores lower than a preset score until the loss function converges or reaches a preset convergence condition;

[0111] The trained tea garden drought prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if so, the verification passes, and the ROC curve and calibration curve are drawn using the test set to test the tea garden drought prediction model.

[0112] By introducing dynamic pruning (removing low-importance weights / channels) and hyperparameter optimization (learning rate, batch size) in the training process of the tea garden drought prediction model, combined with K-fold cross-validation, overfitting can be effectively prevented, model complexity can be reduced, and generalization performance can be improved.

[0113] By K-folding the data set, we can maximize the use of limited data, reduce the deviation caused by random division, and make the verification results more statistically significant. After training, we conduct double tests on the validation set (accuracy threshold) and the test set (confidence threshold) to ensure the reliability and robustness of the model and avoid overfitting or underfitting.

[0114] The step S4 is specifically as follows:

[0115] The tea garden drought prediction model that has passed the test is used as the teacher model. The number of network layers and the number of neurons in the teacher model are simplified to generate a student model. The distillation loss function of the student model is set based on the cross entropy loss and the KL divergence loss. The teacher model is called to train the student model through the knowledge distillation technology until the distillation loss function converges. The student model is used as the tea garden drought prediction model after knowledge distillation;

[0116] Actual monitoring data is collected from the tea garden through a sensor array including a humidity sensor, a light intensity sensor, a temperature sensor, a carbon dioxide sensor, an air pressure sensor, a particulate matter sensor, a global solar radiation sensor, a pH sensor, a salinity sensor, a conductivity sensor, a rainfall sensor, a wind speed and direction sensor, a locator, and a wireless communication module; the wireless communication module is used to obtain remote sensing images;

[0117] By collecting sampling location data through locators and dynamically binding it to GIS maps, drought forecasts can be accurately mapped to spatial coordinates. Compared to traditional methods that only mark plot numbers, this technology can accurately determine longitude and latitude coordinates, providing centimeter-level positioning support for precision agriculture operations such as drone irrigation.

[0118] After data drift compensation training is performed on the tea garden drought prediction model after knowledge distillation using the actual monitoring data, the tea garden drought prediction model is deployed to each edge computing device located in the tea garden through containerization technology;

[0119] Through knowledge distillation, the teacher model is compressed into a student model, reducing computing resource requirements while retaining accuracy; combined with data drift compensation training (such as seasonal changes in environmental parameters), the long-term stability of the tea garden drought prediction model on edge devices is ensured.

[0120] The step S5 is specifically as follows:

[0121] Each edge device maintains a heartbeat connection and exchanges load monitoring data with each other. The load monitoring data is parsed using load balancing technology. The real-time monitoring data collected by the sensor array is input into the tea garden drought prediction model with the smallest load through multi-threading technology for inference to obtain the tea garden drought prediction result.

[0122] Through the heartbeat connection and load balancing strategy between edge computing devices, combined with multi-threading technology to distribute real-time monitoring data, the edge computing resources are maximized to meet the real-time warning needs of the wide-area distribution of tea gardens.

[0123] By independently running the tea garden drought prediction model on each edge computing device and implementing distributed computing through heartbeat connection and load balancing, the risk of single point failure can be avoided (for example, the failure of an edge computing device does not affect the overall system) and the system's disaster recovery capability can be improved.

[0124] The step S6 is specifically as follows:

[0125] The edge computing device monitors the tea garden drought prediction results for anomalies through preset monitoring rules. When an anomaly occurs, the tea garden drought prediction results and the corresponding real-time monitoring data are compressed into a compressed warning data packet, and the compressed warning data packet is pushed in real time to a pre-associated management terminal for tea garden drought warning.

[0126] The drought level (level 1 to level 4) is divided by the quantile of soil entropy value to objectively reflect the data distribution characteristics and avoid the subjectivity of threshold setting; the abnormal monitoring rules are combined with data compression and real-time push to ensure rapid response to sudden droughts.

[0127] A preferred embodiment of a real-time early warning system for drought in a tea garden according to the present invention includes the following modules:

[0128] A tea garden drought prediction model creation module is used to create a tea garden drought prediction model based on the multimodal feature extraction layer, the heterogeneous model fusion layer, the soil entropy prediction layer, and the drought mapping output layer, and set the loss function of the tea garden drought prediction model;

[0129] The multimodal feature extraction layer is used to extract features from the monitoring data to obtain splicing features; the heterogeneous model fusion layer is used to perform feature mining and fusion on the splicing features to obtain fusion features; the soil entropy prediction layer is used to calculate the soil entropy value based on the fusion features; the drought mapping output layer is used to map the soil entropy value according to the set drought classification rules, and output the tea garden drought prediction result with the drought period, drought level and drought location; the drought mapping output layer outputs the probability of the drought level (level 1 drought / level 2 drought / level 3 drought / level 4 drought), the probability of the drought period, and the probability of the drought location through the fully connected layer + Softmax, and introduces the temperature scaling technology to calibrate the prediction confidence;

[0130] By setting the tea garden drought prediction results to include drought period, drought level and drought location, and combining them with GIS map integration on the management terminal, the spatial distribution of drought can be visualized, helping decision makers to quickly locate problem areas and trace historical data.

[0131] A data set construction module is used to collect a large amount of historical monitoring data, pre-process the historical monitoring data, and construct a data set after annotating the data;

[0132] A tea garden drought prediction model training module is used to divide the data set into a training set, a validation set, and a test set, train the tea garden drought prediction model using the training set and a loss function, dynamically prune the tea garden drought prediction model during the training process, verify the trained tea garden drought prediction model using the validation set, and test the verified tea garden drought prediction model using the test set; the AdamW optimizer is used during the training process, combined with cosine annealing learning rate scheduling;

[0133] A tea garden drought prediction model deployment module is used to compress the tested tea garden drought prediction model using knowledge distillation technology, collect actual monitoring data from the tea garden through a sensor array, perform data drift compensation training on the knowledge distilled tea garden drought prediction model using the actual monitoring data, and then deploy the tea garden drought prediction model to various edge computing devices located in the tea garden;

[0134] The tea garden drought prediction module is used to input the real-time monitoring data collected by the sensor array into the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology to obtain the tea garden drought prediction results;

[0135] The real-time drought warning module is used for edge computing equipment to monitor the tea garden drought prediction results for abnormalities, and when an abnormality occurs, the tea garden drought prediction results are pushed to the pre-associated management terminal in real time to issue a tea garden drought warning.

[0136] By integrating technologies from multiple fields, including agricultural meteorology (drought classification rules), the Internet of Things (sensor arrays), edge computing (containerized deployment), and machine learning (heterogeneous models), an interdisciplinary collaborative solution is formed, breaking through the limitations of traditional single-disciplinary methods in tea garden monitoring.

[0137] By automating the entire process from data acquisition (sensor array), preprocessing (KNN / K-means repair), model training (dynamic pruning) to warning push (compressed data packets), manual intervention is reduced and system response efficiency is improved (for example, the delay in abnormal warning is reduced to seconds).

[0138] Through multimodal feature fusion (co-processing of structured data and remote sensing images) and dynamic integration of heterogeneous models (gated weighted fusion of differentiated models such as LightGBM, LSVM, and decision trees), combined with Bayesian optimization and dynamic pruning, high-precision soil entropy prediction is achieved; knowledge distillation and edge computing container deployment are adopted to reduce model complexity while ensuring real-time performance, and the long-term stability of the system is improved through data drift compensation and load balancing multi-threaded processing; the quantile drought classification rules and abnormal monitoring compression and early warning mechanism are innovatively introduced to objectively adapt to different tea garden environments, and combined with the multi-dimensional data coverage of the sensor array and the end-to-end automated process, the accuracy and response efficiency of drought warning are significantly improved. At the same time, it has low power consumption, easy scalability and cross-scenario generalization capabilities, providing a set of efficient, reliable and reusable edge intelligence solutions for smart agriculture.

[0139] In the tea garden drought prediction model creation module, the multimodal feature extraction layer is constructed based on a structured parameter processing module, an image processing module, and a feature splicing module; the structured parameter processing module is used to extract structured high-order features from the structured data of the monitoring data through a LASSO sub-model and a KNN sub-model; the image processing module is used to extract multi-scale image texture features from the remote sensing images of the monitoring data through a CNN sub-model; the feature splicing module is used to splice the structured high-order features and the multi-scale image texture features to obtain splicing features;

[0140] By extracting structured high-order features through LASSO / KNN and combining it with CNN to extract multi-scale image texture features, we can achieve multimodal feature fusion of tea garden environmental parameters (such as soil temperature and pH value) and remote sensing images (such as vegetation coverage), effectively improving the comprehensiveness of data representation.

[0141] The heterogeneous model fusion layer is constructed based on a tree model channel, a kernel method channel, a linear channel, a decision channel and a dynamic feature fusion module; the tree model channel is used to extract nonlinear interaction sub-features from each splicing feature through a LightGBM sub-model, an XGBoost sub-model and an RF sub-model; the kernel method channel is used to extract high-dimensional space sub-features from each splicing feature through the kernel function of the LSVM sub-model; the linear channel is used to extract linear sub-features from each splicing feature through a linear regression sub-model; the decision channel is used to extract segmentation sub-features from each splicing feature through a decision tree sub-model; the dynamic feature fusion module is used to dynamically fuse nonlinear interaction sub-features, high-dimensional space sub-features, linear sub-features and segmentation sub-features through gated weighting to obtain fused features;

[0142] By setting up a heterogeneous fusion layer to integrate differentiated models such as LightGBM (efficient gradient boosting), XGBoost (precise segmentation), LSVM (high-dimensional kernel mapping), and decision tree (rule segmentation), and dynamically fusing nonlinear, linear, and segmentation sub-features through gated weighting, the model's adaptability to different data distributions is significantly improved, avoiding the bias of a single model.

[0143] The soil entropy prediction layer is used to dynamically adjust the weights of each sub-feature in the fusion feature through the Bayesian optimizer, and output the soil entropy value through the deep residual network;

[0144] By setting up the soil entropy prediction layer and using the Bayesian optimizer to dynamically adjust the feature weights, combined with the deep residual network to solve the gradient vanishing problem, we can ensure the accurate modeling of complex nonlinear relationships and improve the accuracy of soil entropy prediction.

[0145] Numerical features are extracted through the structured parameter processing module (LASSO+KNN), and the multi-scale texture features of remote sensing images are extracted by combining the image processing module (CNN). The meteorological, soil physical and chemical parameters and spatial image information are comprehensively utilized to comprehensively capture the multi-dimensional factors affecting soil entropy. By integrating models based on different principles such as tree models (LightGBM / XGBoost / RF), kernel methods (LSVM), linear regression, and decision trees, nonlinear interactions, high-dimensional spatial mapping, linear relationships, and decision boundary features are captured respectively, breaking through the limitations of a single model and greatly improving the accuracy of soil entropy prediction.

[0146] The specific drought classification rules are as follows:

[0147] The quantile of soil entropy value is (75,100], which corresponds to the first level of drought; the quantile is (25,75], which corresponds to the second level of drought; the quantile is (5,25], which corresponds to the third level of drought; the quantile is (0,5], which corresponds to the fourth level of drought.

[0148] The dataset construction module is specifically used for:

[0149] Acquire a large amount of historical monitoring data including structured data and remote sensing images, wherein the structured data at least includes maximum light intensity, average light intensity, sunshine duration, maximum temperature, minimum temperature, average temperature, average daytime temperature, average nighttime temperature, daily temperature difference, maximum air humidity, minimum air humidity, average relative humidity, average daytime air humidity, average nighttime air humidity, maximum carbon dioxide concentration, minimum carbon dioxide concentration, average carbon dioxide concentration, average daytime carbon dioxide concentration, average nighttime carbon dioxide concentration, maximum air pressure value, minimum air pressure value value, average air pressure value, daytime average air pressure value, nighttime average air pressure value, air pressure difference, daily average PM2.5, daily average PM10, maximum TBQ total radiation, minimum TBQ total radiation, daily average TBQ total radiation, maximum soil temperature, minimum soil temperature, daily average soil temperature, soil temperature difference, maximum soil pH, minimum soil pH, average soil pH, maximum soil salinity value, minimum soil salinity value, average soil salinity value, maximum soil conductivity, minimum soil conductivity, average soil conductivity, rainfall, wind direction, wind speed, sampling location and sampling time;

[0150] By integrating more than 40 structured parameters and remote sensing images such as light, temperature and humidity, and soil conductivity, it covers multi-dimensional factors such as meteorology, soil, and the environment; K-nearest neighbor filling and K-means anomaly repair improve data integrity and provide high-quality input for the model.

[0151] The structured data in each of the historical monitoring data are preprocessed by at least filling missing values ​​and repairing outliers, and the remote sensing images in the historical monitoring data are preprocessed by at least noise reduction, contrast enhancement, morphological adjustment, scaling, and color space conversion; the missing value filling is based on the K-nearest neighbor filling method; and the outlier repair is based on the K-means clustering algorithm;

[0152] The pre-processed historical monitoring data are labeled with soil entropy value, drought level and time segment label to construct a data set; the time segment label is the month of the sampling time.

[0153] By dividing the time labels by month, the model can automatically learn the seasonal patterns of drought in tea gardens (such as the difference between rainy season and dry season), thereby improving the temporal matching between the prediction results and the actual environmental changes.

[0154] The tea garden drought prediction model training module is specifically used for:

[0155] The data set is divided into a training set, a validation set, and a test set based on a K-fold cross-validation method, and a tea garden drought prediction model is trained using the training set. During the training process, the hyperparameters of the tea garden drought prediction model, including at least a learning rate, a batch size, and a number of training rounds, are continuously optimized. The tea garden drought prediction model is dynamically pruned to remove weights or channels with importance scores lower than a preset score until the loss function converges or reaches a preset convergence condition;

[0156] The trained tea garden drought prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if so, the verification passes, and the ROC curve and calibration curve are drawn using the test set to test the tea garden drought prediction model.

[0157] By introducing dynamic pruning (removing low-importance weights / channels) and hyperparameter optimization (learning rate, batch size) in the training process of the tea garden drought prediction model, combined with K-fold cross-validation, overfitting can be effectively prevented, model complexity can be reduced, and generalization performance can be improved.

[0158] By K-folding the data set, we can maximize the use of limited data, reduce the deviation caused by random division, and make the verification results more statistically significant. After training, we conduct double tests on the validation set (accuracy threshold) and the test set (confidence threshold) to ensure the reliability and robustness of the model and avoid overfitting or underfitting.

[0159] The tea garden drought prediction model deployment module is specifically used for:

[0160] The tea garden drought prediction model that has passed the test is used as the teacher model. The number of network layers and the number of neurons in the teacher model are simplified to generate a student model. The distillation loss function of the student model is set based on the cross entropy loss and the KL divergence loss. The teacher model is called to train the student model through the knowledge distillation technology until the distillation loss function converges. The student model is used as the tea garden drought prediction model after knowledge distillation;

[0161] Actual monitoring data is collected from the tea garden through a sensor array including a humidity sensor, a light intensity sensor, a temperature sensor, a carbon dioxide sensor, an air pressure sensor, a particulate matter sensor, a global solar radiation sensor, a pH sensor, a salinity sensor, a conductivity sensor, a rainfall sensor, a wind speed and direction sensor, a locator, and a wireless communication module; the wireless communication module is used to obtain remote sensing images;

[0162] By collecting sampling location data through locators and dynamically binding it to GIS maps, drought forecasts can be accurately mapped to spatial coordinates. Compared to traditional methods that only mark plot numbers, this technology can accurately determine longitude and latitude coordinates, providing centimeter-level positioning support for precision agriculture operations such as drone irrigation.

[0163] After data drift compensation training is performed on the tea garden drought prediction model after knowledge distillation using the actual monitoring data, the tea garden drought prediction model is deployed to each edge computing device located in the tea garden through containerization technology;

[0164] Through knowledge distillation, the teacher model is compressed into a student model, reducing computing resource requirements while retaining accuracy; combined with data drift compensation training (such as seasonal changes in environmental parameters), the long-term stability of the tea garden drought prediction model on edge devices is ensured.

[0165] The tea garden drought prediction module is specifically used for:

[0166] Each edge device maintains a heartbeat connection and exchanges load monitoring data with each other. The load monitoring data is parsed using load balancing technology. The real-time monitoring data collected by the sensor array is input into the tea garden drought prediction model with the smallest load through multi-threading technology for inference to obtain the tea garden drought prediction result.

[0167] Through the heartbeat connection and load balancing strategy between edge computing devices, combined with multi-threading technology to distribute real-time monitoring data, the edge computing resources are maximized to meet the real-time warning needs of the wide-area distribution of tea gardens.

[0168] By independently running the tea garden drought prediction model on each edge computing device and implementing distributed computing through heartbeat connection and load balancing, the risk of single point failure can be avoided (for example, the failure of an edge computing device does not affect the overall system) and the system's disaster recovery capability can be improved.

[0169] The drought real-time early warning module is specifically used for:

[0170] The edge computing device monitors the tea garden drought prediction results for anomalies through preset monitoring rules. When an anomaly occurs, the tea garden drought prediction results and the corresponding real-time monitoring data are compressed into a compressed warning data packet, and the compressed warning data packet is pushed in real time to a pre-associated management terminal for tea garden drought warning.

[0171] The drought level (level 1 to level 4) is divided by the quantile of soil entropy value to objectively reflect the data distribution characteristics and avoid the subjectivity of threshold setting; the abnormal monitoring rules are combined with data compression and real-time push to ensure rapid response to sudden droughts.

[0172] In summary, the advantages of the present invention are:

[0173] 1. A tea garden drought prediction model is created through the multimodal feature extraction layer, heterogeneous model fusion layer, soil entropy prediction layer and drought mapping output layer, and the loss function of the tea garden drought prediction model is set; a large amount of historical monitoring data is collected to construct a data set, and the data set is divided into a training set, a validation set and a test set. The tea garden drought prediction model is trained through the training set and the loss function, and the tea garden drought prediction model is dynamically pruned during the training process. The trained tea garden drought prediction model is verified through the validation set, and the verified tea garden drought prediction model is tested through the test set; then the tested tea garden drought prediction model is compressed through the knowledge distillation technology, actual monitoring data is collected from the tea garden through the sensor array, and the tea garden drought prediction model after knowledge distillation is trained with data drift compensation through each actual monitoring data, and then the tea garden drought prediction model is deployed to each edge computing device located in the tea garden; the edge computing device uses the real-time monitoring data collected by the sensor array to input the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology. The method uses a pre-trained tea garden drought prediction model to predict the tea garden drought, and performs abnormal monitoring on the tea garden drought prediction results. When an abnormality occurs, the tea garden drought prediction results are pushed to the pre-associated management terminal in real time for tea garden drought warning. That is, the tea garden drought warning is performed through the pre-trained tea garden drought prediction model. Since the tea garden drought prediction model combines the LASSO sub-model, KNN sub-model, CNN sub-model, LightGBM sub-model, XGBoost sub-model, RF sub-model, LSVM sub-model, linear regression sub-model and decision tree sub-model, it integrates the advantages of each model and greatly improves the feature extraction ability, generalization ability and robustness. In addition, the tea garden drought prediction model is dynamically pruned during training and knowledge distillation is performed before deployment, which effectively compresses the model size of the tea garden drought prediction model and facilitates deployment on edge computing devices with limited computing power. The edge computing device can locally process real-time monitoring data without uploading it to the server. In the prediction process, load balancing technology and multi-threading technology are combined to effectively improve the inference speed, thereby greatly improving the timeliness and accuracy of the tea garden drought warning.

[0174] 2. By extracting structured high-order features through LASSO / KNN and combining it with CNN to extract multi-scale image texture features, we can achieve multimodal feature fusion of tea garden environmental parameters (such as soil temperature and pH value) and remote sensing images (such as vegetation coverage), effectively improving the comprehensiveness of data representation.

[0175] 3. By setting up a heterogeneous fusion layer to integrate differentiated models such as LightGBM (efficient gradient boosting), XGBoost (precise segmentation), LSVM (high-dimensional kernel mapping), and decision tree (rule segmentation), and dynamically fusing nonlinear, linear, and segmentation sub-features through gated weighting, the model's adaptability to different data distributions is significantly improved, avoiding the bias of a single model.

[0176] 4. By setting up the soil entropy prediction layer and using the Bayesian optimizer to dynamically adjust the feature weights, combined with the deep residual network to solve the gradient vanishing problem, we can ensure the accurate modeling of complex nonlinear relationships and improve the accuracy of soil entropy prediction.

[0177] 5. By introducing dynamic pruning (removing low-importance weights / channels) and hyperparameter optimization (learning rate, batch size) in the training process of the tea garden drought prediction model, combined with K-fold cross-validation, overfitting can be effectively prevented, model complexity can be reduced, and generalization performance can be improved.

[0178] 6. Compress the teacher model into a student model through knowledge distillation, reducing computing resource requirements while retaining accuracy; combined with data drift compensation training (such as seasonal changes in environmental parameters), ensure the long-term stability of the tea garden drought prediction model on edge devices.

[0179] 7. Through the heartbeat connection and load balancing strategy between edge computing devices, combined with multi-threading technology to distribute real-time monitoring data, maximize the use of edge computing resources, and meet the real-time warning needs of the wide-area distribution of tea gardens.

[0180] 8. Drought levels (level 1 to 4) are divided by soil entropy value quantiles to objectively reflect data distribution characteristics and avoid subjectivity in threshold setting; abnormal monitoring rules are combined with data compression and real-time push to ensure rapid response to sudden droughts.

[0181] 9. By integrating more than 40 structured parameters such as light, temperature, humidity, and soil conductivity, as well as remote sensing images, it covers multi-dimensional factors such as meteorology, soil, and the environment; K-nearest neighbor filling and K-means anomaly repair improve data integrity and provide high-quality input for the model.

[0182] 10. By integrating technologies from multiple fields such as agricultural meteorology (drought classification rules), the Internet of Things (sensor arrays), edge computing (containerized deployment) and machine learning (heterogeneous models), an interdisciplinary collaborative solution is formed to break through the limitations of traditional single-disciplinary methods in tea garden monitoring.

[0183] 11. By automating the entire process from data acquisition (sensor array), preprocessing (KNN / K-means repair), model training (dynamic pruning) to warning push (compressed data packets), human intervention is reduced and system response efficiency is improved (for example, the delay of abnormal warning is reduced to seconds).

[0184] 12. By dividing the time labels by month, the model can automatically learn the seasonal patterns of drought in tea gardens (such as the difference between rainy season and dry season), improving the temporal matching degree between the prediction results and the actual environmental changes.

[0185] 13. The tea garden drought prediction model is independently run by each edge computing device, and distributed computing is achieved through heartbeat connection and load balancing to avoid the risk of single point failure (for example, the failure of an edge computing device does not affect the overall system) and improve the system's disaster recovery capabilities.

[0186] 14. By setting the tea garden drought prediction results to include drought period, drought level and drought location, combined with the GIS map integration of the management terminal, the spatial distribution of drought can be visualized, assisting decision makers to quickly locate problem areas and trace historical data.

[0187] 15. Through multimodal feature fusion (co-processing of structured data and remote sensing images) and dynamic integration of heterogeneous models (gated weighted fusion of differentiated models such as LightGBM, LSVM, and decision trees), combined with Bayesian optimization and dynamic pruning, high-precision soil entropy prediction is achieved; knowledge distillation and edge computing container deployment are adopted to reduce model complexity while ensuring real-time performance, and to improve the long-term stability of the system through data drift compensation and load balancing multi-threaded processing; the quantile drought classification rules and abnormal monitoring compression warning mechanism are innovatively introduced to objectively adapt to different tea garden environments, combined with the multi-dimensional data coverage of the sensor array and the end-to-end automated process, significantly improving the accuracy and response efficiency of drought warnings, while having low power consumption, easy scalability and cross-scenario generalization capabilities, providing a set of efficient, reliable and reusable edge intelligence solutions for smart agriculture.

[0188] Although the specific embodiments of the present invention are described above, those skilled in the art should understand that the specific embodiments described are merely 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 within the scope of protection of the claims of the present invention.

Claims

1. A real-time early warning method for drought in a tea garden, characterized by: The steps include: Step S1: creating a tea garden drought prediction model based on the multimodal feature extraction layer, the heterogeneous model fusion layer, the soil entropy prediction layer, and the drought mapping output layer, and setting a loss function of the tea garden drought prediction model; The multimodal feature extraction layer is used to extract features from the monitoring data to obtain splicing features; the heterogeneous model fusion layer is used to perform feature mining and fusion on the splicing features to obtain fusion features; the soil entropy prediction layer is used to calculate the soil entropy value based on the fusion features; the drought mapping output layer is used to map the soil entropy value according to the set drought classification rules, and output the tea garden drought prediction result with the drought period, drought level and drought location; Step S2: Collect a large amount of historical monitoring data, pre-process and annotate each of the historical monitoring data to construct a data set; Step S3, dividing the data set into a training set, a validation set, and a test set, training the tea garden drought prediction model using the training set and the loss function, dynamically pruning the tea garden drought prediction model during the training process, validating the trained tea garden drought prediction model using the validation set, and testing the verified tea garden drought prediction model using the test set; Step S4: compressing the tested tea garden drought prediction model using knowledge distillation technology, collecting actual monitoring data from the tea garden through a sensor array, training the tea garden drought prediction model after knowledge distillation with data drift compensation using the actual monitoring data, and then deploying the tea garden drought prediction model to various edge computing devices located in the tea garden; Step S5: The edge computing device inputs the real-time monitoring data collected by the sensor array into the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology to obtain the tea garden drought prediction result; Step S6: The edge computing device monitors the tea garden drought prediction results for abnormalities. When an abnormality occurs, the tea garden drought prediction results are pushed to a pre-associated management terminal in real time for tea garden drought warning.

2. The method for real-time early warning of drought in a tea garden according to claim 1, wherein: In step S1, the multimodal feature extraction layer is constructed based on a structured parameter processing module, an image processing module, and a feature splicing module; the structured parameter processing module is used to extract structured high-order features from the structured data of the monitoring data through a LASSO sub-model and a KNN sub-model; the image processing module is used to extract multi-scale image texture features from the remote sensing image of the monitoring data through a CNN sub-model; and the feature splicing module is used to splice the structured high-order features and the multi-scale image texture features to obtain spliced ​​features; The heterogeneous model fusion layer is constructed based on a tree model channel, a kernel method channel, a linear channel, a decision channel and a dynamic feature fusion module; the tree model channel is used to extract nonlinear interaction sub-features from each splicing feature through a LightGBM sub-model, an XGBoost sub-model and an RF sub-model; the kernel method channel is used to extract high-dimensional space sub-features from each splicing feature through the kernel function of the LSVM sub-model; the linear channel is used to extract linear sub-features from each splicing feature through a linear regression sub-model; the decision channel is used to extract segmentation sub-features from each splicing feature through a decision tree sub-model; the dynamic feature fusion module is used to dynamically fuse nonlinear interaction sub-features, high-dimensional space sub-features, linear sub-features and segmentation sub-features through gated weighting to obtain fused features; The soil entropy prediction layer is used to dynamically adjust the weights of each sub-feature in the fusion feature through the Bayesian optimizer, and output the soil entropy value through the deep residual network; The specific drought classification rules are as follows: The quantile of soil entropy value is (75,100], which corresponds to the first level of drought; the quantile is (25,75], which corresponds to the second level of drought; the quantile is (5,25], which corresponds to the third level of drought; the quantile is (0,5], which corresponds to the fourth level of drought.

3. The real-time early warning method for drought in a tea garden according to claim 1, characterized in that: The step S2 is specifically as follows: Acquire a large amount of historical monitoring data including structured data and remote sensing images, wherein the structured data at least includes maximum light intensity, average light intensity, sunshine duration, maximum temperature, minimum temperature, average temperature, average daytime temperature, average nighttime temperature, daily temperature difference, maximum air humidity, minimum air humidity, average relative humidity, average daytime air humidity, average nighttime air humidity, maximum carbon dioxide concentration, minimum carbon dioxide concentration, average carbon dioxide concentration, average daytime carbon dioxide concentration, average nighttime carbon dioxide concentration, maximum air pressure value, minimum air pressure value value, average air pressure value, daytime average air pressure value, nighttime average air pressure value, air pressure difference, daily average PM2.5, daily average PM10, maximum TBQ total radiation, minimum TBQ total radiation, daily average TBQ total radiation, maximum soil temperature, minimum soil temperature, daily average soil temperature, soil temperature difference, maximum soil pH, minimum soil pH, average soil pH, maximum soil salinity value, minimum soil salinity value, average soil salinity value, maximum soil conductivity, minimum soil conductivity, average soil conductivity, rainfall, wind direction, wind speed, sampling location and sampling time; The structured data in each of the historical monitoring data are preprocessed by at least filling missing values ​​and repairing outliers, and the remote sensing images in the historical monitoring data are preprocessed by at least noise reduction, contrast enhancement, morphological adjustment, scaling, and color space conversion; the missing value filling is based on the K-nearest neighbor filling method; and the outlier repair is based on the K-means clustering algorithm; The pre-processed historical monitoring data are labeled with soil entropy value, drought level and time segment label to construct a data set; the time segment label is the month of the sampling time.

4. The method for real-time early warning of drought in a tea garden according to claim 1, wherein: 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 K-fold cross-validation method, and a tea garden drought prediction model is trained using the training set. During the training process, the hyperparameters of the tea garden drought prediction model, including at least a learning rate, a batch size, and a number of training rounds, are continuously optimized. The tea garden drought prediction model is dynamically pruned to remove weights or channels with importance scores lower than a preset score until the loss function converges or reaches a preset convergence condition; The trained tea garden drought prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if so, the verification passes, and the ROC curve and calibration curve are drawn using the test set to test the tea garden drought prediction model.

5. The tea garden drought real-time early warning method according to claim 1, characterized in that: The step S4 is specifically as follows: The tea garden drought prediction model that has passed the test is used as the teacher model. The number of network layers and the number of neurons in the teacher model are simplified to generate a student model. The distillation loss function of the student model is set based on the cross entropy loss and the KL divergence loss. The teacher model is called to train the student model through the knowledge distillation technology until the distillation loss function converges. The student model is used as the tea garden drought prediction model after knowledge distillation; Actual monitoring data is collected from the tea garden through a sensor array including a humidity sensor, a light intensity sensor, a temperature sensor, a carbon dioxide sensor, an air pressure sensor, a particulate matter sensor, a global solar radiation sensor, a pH sensor, a salinity sensor, a conductivity sensor, a rainfall sensor, a wind speed and direction sensor, a locator, and a wireless communication module; the wireless communication module is used to obtain remote sensing images; After data drift compensation training is performed on the tea garden drought prediction model after knowledge distillation using the actual monitoring data, the tea garden drought prediction model is deployed to each edge computing device located in the tea garden through containerization technology; The step S5 is specifically as follows: Each edge device maintains a heartbeat connection and exchanges load monitoring data with each other. The load monitoring data is parsed using load balancing technology. The real-time monitoring data collected by the sensor array is input into the tea garden drought prediction model with the smallest load through multi-threading technology for inference to obtain the tea garden drought prediction result. The step S6 is specifically as follows: The edge computing device monitors the tea garden drought prediction results for anomalies through preset monitoring rules. When an anomaly occurs, the tea garden drought prediction results and the corresponding real-time monitoring data are compressed into a compressed warning data packet, and the compressed warning data packet is pushed in real time to a pre-associated management terminal for tea garden drought warning.

6. A real-time early warning system for tea garden drought, characterized by: Includes the following modules: A tea garden drought prediction model creation module is used to create a tea garden drought prediction model based on the multimodal feature extraction layer, the heterogeneous model fusion layer, the soil entropy prediction layer, and the drought mapping output layer, and set the loss function of the tea garden drought prediction model; The multimodal feature extraction layer is used to extract features from the monitoring data to obtain splicing features; the heterogeneous model fusion layer is used to perform feature mining and fusion on the splicing features to obtain fusion features; the soil entropy prediction layer is used to calculate the soil entropy value based on the fusion features; the drought mapping output layer is used to map the soil entropy value according to the set drought classification rules, and output the tea garden drought prediction result with the drought period, drought level and drought location; A data set construction module is used to collect a large amount of historical monitoring data, pre-process the historical monitoring data, and construct a data set after annotating the data; a tea garden drought prediction model training module, configured to divide the data set into a training set, a validation set, and a test set, train the tea garden drought prediction model using the training set and a loss function, dynamically prune the tea garden drought prediction model during the training process, validate the trained tea garden drought prediction model using the validation set, and test the validated tea garden drought prediction model using the test set; A tea garden drought prediction model deployment module is used to compress the tested tea garden drought prediction model using knowledge distillation technology, collect actual monitoring data from the tea garden through a sensor array, perform data drift compensation training on the knowledge distilled tea garden drought prediction model using the actual monitoring data, and then deploy the tea garden drought prediction model to various edge computing devices located in the tea garden; The tea garden drought prediction module is used to input the real-time monitoring data collected by the sensor array into the deployed tea garden drought prediction model based on load balancing technology and multi-threading technology to obtain the tea garden drought prediction results; The real-time drought warning module is used for edge computing equipment to monitor the tea garden drought prediction results for abnormalities, and when an abnormality occurs, the tea garden drought prediction results are pushed to the pre-associated management terminal in real time to issue a tea garden drought warning.

7. The tea garden drought real-time early warning system according to claim 6, characterized in that: In the tea garden drought prediction model creation module, the multimodal feature extraction layer is constructed based on a structured parameter processing module, an image processing module, and a feature splicing module; the structured parameter processing module is used to extract structured high-order features from the structured data of the monitoring data through a LASSO sub-model and a KNN sub-model; the image processing module is used to extract multi-scale image texture features from the remote sensing images of the monitoring data through a CNN sub-model; the feature splicing module is used to splice the structured high-order features and the multi-scale image texture features to obtain splicing features; The heterogeneous model fusion layer is constructed based on a tree model channel, a kernel method channel, a linear channel, a decision channel and a dynamic feature fusion module; the tree model channel is used to extract nonlinear interaction sub-features from each splicing feature through a LightGBM sub-model, an XGBoost sub-model and an RF sub-model; the kernel method channel is used to extract high-dimensional space sub-features from each splicing feature through the kernel function of the LSVM sub-model; the linear channel is used to extract linear sub-features from each splicing feature through a linear regression sub-model; the decision channel is used to extract segmentation sub-features from each splicing feature through a decision tree sub-model; the dynamic feature fusion module is used to dynamically fuse nonlinear interaction sub-features, high-dimensional space sub-features, linear sub-features and segmentation sub-features through gated weighting to obtain fused features; The soil entropy prediction layer is used to dynamically adjust the weights of each sub-feature in the fusion feature through the Bayesian optimizer, and output the soil entropy value through the deep residual network; The specific drought classification rules are as follows: The quantile of soil entropy value is (75,100], which corresponds to the first level of drought; the quantile is (25,75], which corresponds to the second level of drought; the quantile is (5,25], which corresponds to the third level of drought; the quantile is (0,5], which corresponds to the fourth level of drought.

8. The tea garden drought real-time early warning system according to claim 6, characterized in that: The dataset construction module is specifically used for: Acquire a large amount of historical monitoring data including structured data and remote sensing images, wherein the structured data at least includes maximum light intensity, average light intensity, sunshine duration, maximum temperature, minimum temperature, average temperature, average daytime temperature, average nighttime temperature, daily temperature difference, maximum air humidity, minimum air humidity, average relative humidity, average daytime air humidity, average nighttime air humidity, maximum carbon dioxide concentration, minimum carbon dioxide concentration, average carbon dioxide concentration, average daytime carbon dioxide concentration, average nighttime carbon dioxide concentration, maximum air pressure value, minimum air pressure value value, average air pressure value, daytime average air pressure value, nighttime average air pressure value, air pressure difference, daily average PM2.5, daily average PM10, maximum TBQ total radiation, minimum TBQ total radiation, daily average TBQ total radiation, maximum soil temperature, minimum soil temperature, daily average soil temperature, soil temperature difference, maximum soil pH, minimum soil pH, average soil pH, maximum soil salinity value, minimum soil salinity value, average soil salinity value, maximum soil conductivity, minimum soil conductivity, average soil conductivity, rainfall, wind direction, wind speed, sampling location and sampling time; The structured data in each of the historical monitoring data are preprocessed by at least filling missing values ​​and repairing outliers, and the remote sensing images in the historical monitoring data are preprocessed by at least noise reduction, contrast enhancement, morphological adjustment, scaling, and color space conversion; the missing value filling is based on the K-nearest neighbor filling method; and the outlier repair is based on the K-means clustering algorithm; The pre-processed historical monitoring data are labeled with soil entropy value, drought level and time segment label to construct a data set; the time segment label is the month of the sampling time.

9. The tea garden drought real-time early warning system according to claim 6, characterized in that: The tea garden drought prediction 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 K-fold cross-validation method, and a tea garden drought prediction model is trained using the training set. During the training process, the hyperparameters of the tea garden drought prediction model, including at least a learning rate, a batch size, and a number of training rounds, are continuously optimized. The tea garden drought prediction model is dynamically pruned to remove weights or channels with importance scores lower than a preset score until the loss function converges or reaches a preset convergence condition; The trained tea garden drought prediction model is verified by the validation set to determine whether the prediction accuracy is greater than the preset accuracy threshold. If not, the verification fails and the training set is expanded to continue training; if so, the verification passes, and the ROC curve and calibration curve are drawn using the test set to test the tea garden drought prediction model.

10. The tea garden drought real-time early warning system according to claim 6, characterized in that: The tea garden drought prediction model deployment module is specifically used for: The tea garden drought prediction model that has passed the test is used as the teacher model. The number of network layers and the number of neurons in the teacher model are simplified to generate a student model. The distillation loss function of the student model is set based on the cross entropy loss and the KL divergence loss. The teacher model is called to train the student model through the knowledge distillation technology until the distillation loss function converges. The student model is used as the tea garden drought prediction model after knowledge distillation; Actual monitoring data is collected from the tea garden through a sensor array including a humidity sensor, a light intensity sensor, a temperature sensor, a carbon dioxide sensor, an air pressure sensor, a particulate matter sensor, a global solar radiation sensor, a pH sensor, a salinity sensor, a conductivity sensor, a rainfall sensor, a wind speed and direction sensor, a locator, and a wireless communication module; the wireless communication module is used to obtain remote sensing images; After data drift compensation training is performed on the tea garden drought prediction model after knowledge distillation using the actual monitoring data, the tea garden drought prediction model is deployed to each edge computing device located in the tea garden through containerization technology; The tea garden drought prediction module is specifically used for: Each edge device maintains a heartbeat connection and exchanges load monitoring data with each other. The load monitoring data is parsed using load balancing technology. The real-time monitoring data collected by the sensor array is input into the tea garden drought prediction model with the smallest load through multi-threading technology for inference to obtain the tea garden drought prediction result. The drought real-time warning module is specifically used for: The edge computing device monitors the tea garden drought prediction results for anomalies through preset monitoring rules. When an anomaly occurs, the tea garden drought prediction results and the corresponding real-time monitoring data are compressed into a compressed warning data packet, and the compressed warning data packet is pushed in real time to a pre-associated management terminal for tea garden drought warning.

Images