Vineyard growth state real-time monitoring and intelligent regulation and control method based on deep learning

Abstract

The invention discloses a vineyard growth state real-time monitoring and intelligent regulation and control method based on deep learning, and relates to the technical field of agricultural intelligent management. According to the invention, environment, image and root system data are fused through a multi-source data acquisition system, a time-space fusion deep learning network is constructed to realize accurate prediction, and an automatic execution mechanism is linked to carry out intelligent regulation and control. Edge computing and cloud platform cooperative processing are adopted, the real-time performance and the long-term analysis capability are improved, the problems that in the prior art, monitoring is not comprehensive, the real-time performance is poor, and intelligent regulation and control are lacked are solved, and the grape yield and quality are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of agricultural intelligent management, and particularly relates to a method for real-time monitoring and intelligent control of the growth status of vineyards based on deep learning. Background Art

[0002] Under the traditional vineyard management mode, it mainly relies on manual monitoring and simple technical means. Manual monitoring usually depends on experience, using a ruler to measure the size of grape clusters and the naked eye to observe the color of leaves to judge the growth status of grapes. This method is not only highly subjective but also inefficient, making it difficult to conduct high-frequency and comprehensive monitoring of large-scale vineyards. There are great limitations in manual judgment, and the initial symptoms of diseases and pests are extremely easy to be ignored. Once diseases and pests break out on a large scale, it will have a serious impact on grape yield and quality.

[0003] The monitoring system based on sensors has improved the monitoring situation to a certain extent. By deploying temperature and humidity sensors, light sensors, soil humidity sensors, etc. in the vineyard, it can collect some environmental parameters and transmit the data to the data center. However, this system has obvious deficiencies. The sensors can only detect specific physical quantities and cannot directly reflect the health status of grape plants themselves, such as whether they are infected with diseases and whether the nutrients are sufficient. There is a lack of effective integration of data from different sensors, making it difficult to evaluate the growth status of vineyards as a whole. In a complex environment, such as in bad weather with strong winds and heavy rains, the sensor data is prone to deviation, and even the equipment will be damaged.

[0004] The monitoring technology based on remote sensing images uses satellites or drones to obtain remote sensing images of vineyards and monitors grape growth by analyzing image features. However, the resolution of this technology is limited, and it is difficult to accurately identify details such as lesions on single grape leaves and small patches of diseases and pests. The image interpretation algorithms are mostly based on traditional image processing and pattern recognition methods, unable to fully mine the complex information in the images, resulting in low accuracy of the monitoring results. The data processing cost is relatively high and the real-time performance is poor, unable to meet the requirements of real-time monitoring of the growth status of vineyards.

[0005] In summary, the existing vineyard management technologies have obvious defects in terms of the comprehensiveness, accuracy, real-time performance, and intelligent control ability of monitoring, and there is an urgent need for an innovative technical solution to solve these problems. Summary of the Invention

[0006] The present invention aims to provide a method for real-time monitoring and intelligent control of the growth status of vineyards based on deep learning, so as to solve the problems of incomplete, inaccurate, poor real-time performance, and lack of intelligent control ability in the prior art, realize the intelligent and refined management of vineyards, and improve the yield and quality of grapes.

[0007] To solve the above technical problems, the present invention is realized through the following technical solutions: The present invention is a method for real-time monitoring and intelligent regulation of the growth status of vineyards based on deep learning, including the following steps: S1. Construct a multi-source data acquisition system, including: An environmental data acquisition module, which is used to collect the micro-environment data of the vineyard by deploying temperature and humidity sensors, light sensors, and CO2 sensors, and obtain the soil nutrient content and distribution through soil hyperspectral analysis technology; An image data acquisition module, which is used to obtain the above-ground part images of grape plants, the overall images of the vineyard, and root images through multi-angle high-definition cameras, drones, and root image acquisition devices respectively; A root data acquisition module, which is used to regulate the root growth environment by optimizing the fog cultivation device and monitor the root growth amount in real time based on image processing algorithms; S2. Build a spatio-temporal fusion deep learning network and train it, including: Construct a deep learning model that combines a long short-term memory network (LSTM) and a convolutional neural network (CNN), and input the historical environmental data, real-time image data, and root growth amount data collected by the multi-source data acquisition system; Through transfer learning technology, fine-tune the deep learning model based on a pre-trained model to adapt to different grape varieties and climate conditions; Train the deep learning model so that it can predict the pest and disease risks, water requirements, and nutrient requirements of the vineyard within the next 72 hours; S3. According to the output result of step S2, judge whether regulation is needed. If so, trigger the intelligent regulation system to perform the following operations: Link the supplementary light, drip irrigation system, and drone spraying equipment to automatically adjust the environmental parameters according to the predicted insufficient light, water requirements, and pest and disease risks; Set multiple alarm thresholds. When the monitored data or prediction results exceed the thresholds, notify the farmers through a mobile terminal and trigger emergency measures.

[0008] As a preferred technical solution of the present invention, in step S1: The soil hyperspectral analysis technology uses a portable ground object spectrometer to collect soil reflectance spectral data, and inversely calculates the soil nutrient content through partial least squares method or principal component regression algorithm; The root image acquisition module includes: Optimize the fog cultivation device to provide a controllable growth environment for the roots by adjusting the spray interval and dynamically adapting the nutrient solution formula; The built-in high-definition camera periodically collects root images, and uses the improved U-Net or MaskR-CNN algorithm for image segmentation to calculate root length, surface area, and volume metrics.

[0009] As a preferred technical solution of the present invention, in step S2: The spatio-temporal fusion deep learning network includes ResNet50 as the CNN backbone network, combined with a spatial pyramid pooling (SPP) module to adapt to input images of different sizes; The transfer learning technique is specifically: freeze some convolutional layers of the pre-trained model, fine-tune the fully connected layer based on specific vineyard data, and gradually unfreeze the convolutional layers to improve the generalization ability of the model.

[0010] As a preferred technical solution of the present invention, in the transfer learning technique, select a model pre-trained on a public plant dataset and vineyard data from different regions. When fine-tuning for a specific vineyard, the fine-tuning dataset accounts for 15%-20% of the total data, and the parameters of the last few convolutional layers and fully connected layers of the fine-tuning pre-trained model are adjusted.

[0011] As a preferred technical solution of the present invention, the training process in step S2 includes: Data preprocessing uses normalization and Z-score standardization, and the ratio of the training set, validation set, and test set is 7:2:1 or 6:2:2; The loss function is the weighted sum of mean squared error and cross entropy, with a weight ratio of 0.5:0.5. The optimizer uses Adam or AdamW, and the learning rate dynamic decay strategy is to halve every 10 epochs or trigger decay based on the validation loss.

[0012] As a preferred technical solution of the present invention, in step S3: The multi-level alarm threshold includes at least three levels, corresponding to different degrees of abnormality. The triggering methods include mobile application push, SMS notification, and automatic activation of emergency devices; The emergency measures include a spray cooling system, unfolding of sunshade nets, or increasing the frequency of drone spraying, and the parameter adjustment range is dynamically adjusted according to the prediction results.

[0013] As a preferred technical solution of the present invention, it also includes the collaboration between edge computing and the cloud platform: Deploy edge computing nodes at the vineyard site to receive and process data collected by sensors in real time, perform preliminary screening, cleaning, and analysis to reduce the amount and frequency of data transmitted to the cloud; The cloud platform receives the data uploaded by the edge nodes for in-depth big data analysis, mines the long-term trends of the vineyard growth status, provides planting suggestions for farmers, and stores historical data.

[0014] As a preferred technical solution of the present invention, the method supports collaborative management of multiple vineyards, specifically as follows: The cloud platform integrates data of multiple vineyards and generates regionalized planting suggestions through cluster analysis; The edge computing nodes are interconnected through a Mesh network to achieve data sharing and redundant backup.

[0015] As a preferred technical solution of the present invention, it further includes a user interaction module, specifically as follows: The mobile terminal displays the vineyard monitoring data, model prediction results, and execution status of control measures in real time; Farmers can manually adjust the alarm threshold or intervene in the parameters of the control equipment through the interface, and the system records the operation logs and synchronizes them to the cloud platform.

[0016] As a preferred technical solution of the present invention, the drone spraying equipment in the intelligent control system includes: A dynamic path planning algorithm generates an optimal flight path according to the terrain of the vineyard and the distribution of pests and diseases; the spraying dose and frequency are adjusted based on the pest and disease risk level predicted by the model, with the dose range of 100 - 800 ml / mu and the spraying height of 1 - 3 meters.

[0017] The present invention has the following beneficial effects: By constructing a multi-source data acquisition system, comprehensive acquisition of vineyard environment, image, and root data is realized, providing rich input data for the deep learning model; Using a spatio-temporal fusion deep learning network, combining LSTM and CNN, accurate prediction of the growth status of the vineyard is achieved, including pest and disease risks, water requirements, and nutrient requirements; According to the prediction results, the intelligent control system is triggered to realize automatic adjustment of environmental parameters and triggering of emergency measures, improving the intelligent level of vineyard management; Through the collaboration of edge computing and the cloud platform, real-time processing and in-depth analysis of data are realized, providing planting suggestions and historical data storage for farmers; Supporting collaborative management of multiple vineyards and user interaction, improving the practicality of the system and the user experience.

[0018] Of course, it is not necessary for any product implementing the present invention to achieve all the above advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0020] Figure 1 is the system architecture diagram adopted in the present invention; Figure 2 is the structure diagram of the spatio-temporal fusion deep learning network in the present invention; Figure 3 is the logic diagram of the intelligent control system in the present invention. Specific Embodiments

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0022] Please refer to Figures 1-3 as shown, the present invention discloses a method for real-time monitoring and intelligent control of the growth status of vineyards based on deep learning, including: S1. Construct a multi-source data acquisition system; S2. Build a deep learning model and train it; S3. Determine whether regulation is required. If so, execute the intelligent control system.

[0023] A specific embodiment of the present invention is as follows: Among them I. Implementation of the multi-source data acquisition system Implementation of environmental data acquisition: Uniformly deploy temperature and humidity sensors, light sensors, and CO2 sensors in the vineyard at a certain interval to ensure that the micro-environment data of different regions of the vineyard can be comprehensively and accurately collected. For example, in a vineyard with an area of 50 hectares, sensors are deployed at a density of one monitoring point per 500 square meters. Regularly use a portable ground object spectrometer, such as ASD FieldSpec4, to sample and analyze the vineyard soil. Select multiple sampling points in different regions of the vineyard. For each sampling point, collect soil samples at a depth of 0-20 cm from the soil surface. Place the spectrometer probe vertically 2 cm above the soil sample to obtain the reflectance spectral data of the soil in the 350-2500 nm band, and use partial least squares method or principal component regression algorithm to analyze the soil nutrient content and distribution.

[0024] Implementation of Image Data Acquisition: Install high-definition cameras at different heights and angles in the vineyard. For example, install them at heights of 1.5 meters, 2 meters, and 3 meters respectively to ensure that images of all parts of the grape plants can be captured. The cameras are set to automatically take pictures every 15 minutes, and the shooting resolution is not less than 1920×1080 pixels. Use drones to fly according to the preset flight routes and time intervals to collect overall images of the vineyard. The flight height is set to 50 meters, the flight speed is 5 m / s, and the flight time interval is 2 hours. For root image acquisition, install an optimized aeroponic device near the roots of the grape plants. By adjusting the spraying interval (such as 10 - 30 seconds) and the nutrient solution formula (adjust the nitrogen, phosphorus, and potassium ratios according to the grape growth stage), while creating good conditions for root growth, use the built-in high-definition camera (such as OV2710) to collect root images every 24 hours.

[0025] Implementation of Root Data Acquisition: According to the growth stages (germination stage, flowering stage, fruit swelling stage, etc.) and requirements of the grape plants, precisely adjust the spraying interval and nutrient solution formula of the aeroponic device. Use improved U-Net or MaskR-CNN algorithms to analyze the collected root images, calculate root growth indicators such as root length, root surface area, root volume, etc., and transmit the data to the data processing center in real time.

[0026] II. Implementation of Spatiotemporal Fusion Deep Learning Network Construction and Training Implementation of Model Construction: In the data processing center, use the Python programming language and deep learning frameworks (such as TensorFlow or PyTorch) to build a deep learning model combining LSTM and CNN. Taking the PyTorch framework as an example, the LSTM network contains 3 hidden layers, with 128 neurons in each layer; the CNN network uses ResNet50 as the backbone network, removes the last fully connected layer of the original model, retains the previous convolutional layers and pooling layers, and connects a spatial pyramid pooling (SPP) module behind it. Concatenate the time series features output by LSTM and the image features output by CNN+SPP, and then perform classification and prediction through several fully connected layers to output results such as pest and disease risks, water requirements, and nutrient requirements.

[0027] Transfer learning implementation: Select the ResNet50 model pre-trained on public plant datasets (such as ImageNet, PlantVillage, etc.) and vineyard data from different regions. For the grape varieties (such as Cabernet Sauvignon) in a specific vineyard and the local climate conditions (such as an average annual temperature of 15°C and an annual precipitation of 600 mm), part of the data from this vineyard (15%-20% of the total data) is used as the fine-tuning dataset. During the fine-tuning process, first freeze the first 100 convolutional layers of the pre-trained model, and only update the parameters of the subsequent convolutional layers and fully connected layers. During training, gradually reduce the learning rate from the initial 0.001 to 0.0001. After 10 rounds of training, unfreeze all convolutional layers and continue training for 10 rounds to enable the model to adapt to the actual situation of the specific vineyard and improve the generalization ability.

[0028] Model training implementation: Preprocess the collected historical environmental data and image data, including data cleaning, normalization, etc. For image data, normalize the pixel values from 0-255 to 0-1; for environmental data, use the Z-score normalization method to standardize the data to a distribution with a mean of 0 and a standard deviation of 1. Then divide it into a training set, a validation set, and a test set according to the ratio of 7:2:1 or 6:2:2. Use the training set to train the model, set the number of training rounds to 100, and the batch size to 32. Use the Adam or AdamW optimizer, and the learning rate adopts a dynamic decay strategy, such as halving every 10 epochs or triggering decay based on the validation loss. The loss function uses the weighted sum of mean squared error and cross entropy, and the weight ratio is 0.5:0.5. After each round of training, use the validation set to evaluate the model and monitor indicators such as the loss value and accuracy of the model. If the loss value on the validation set does not decrease for 5 consecutive rounds, stop training and save the model parameters.

[0029] III. Implementation of the intelligent control system Linkage implementation of the automatic actuator: Connect the supplementary light, drip irrigation system, and UAV spraying equipment to the output results of the deep learning model. Set the control logic in the intelligent control system. When the model predicts insufficient light (such as the sunshine duration is less than 6 hours), automatically send an on command to the supplementary light and adjust the brightness of the supplementary light according to the predicted light intensity requirement (such as 1000-5000 lux). For the drip irrigation system, according to the water demand predicted by the model (such as the predicted water demand for the next 24 hours is 5 cubic meters per hectare), automatically control parameters such as the opening time (such as 10-60 minutes) and flow rate (such as 1-5 liters per minute) of the drip irrigation equipment. When the model detects a high risk of pests and diseases (such as the probability is greater than 0.7), automatically trigger the UAV spraying equipment, and set parameters such as the flight path of the UAV and the spraying dose (such as 100-500 milliliters per mu) according to the type of pests and diseases (such as aphids, powdery mildew, etc.) and the severity, to achieve precise spraying.

[0030] Implementation of the hierarchical warning mechanism: Set at least three levels of alarm thresholds in the intelligent control system. For example, different levels of thresholds are set for indicators such as temperature and humidity, pest and disease risks, and water requirements. For temperature, the mild anomaly threshold is set to be higher than 30°C or lower than 10°C, the moderate anomaly threshold is set to be higher than 35°C or lower than 5°C, and the severe anomaly threshold is set to be higher than 40°C or lower than 0°C. For the pest and disease risk index, the mild anomaly threshold is set to 0.3 - 0.5, the moderate anomaly threshold is set to 0.5 - 0.7, and the severe anomaly threshold is set to be above 0.7. When the monitored data or the model prediction results exceed the corresponding thresholds, the system automatically triggers alarms at different levels. The first-level alarm (such as a yellow warning) pushes notifications to farmers through a mobile application. The second-level alarm (such as an orange warning) sends text messages in addition to the APP push. The third-level alarm (such as a red warning), on the basis of the first two notification methods, when the high-temperature warning reaches the highest level (the daily maximum temperature exceeds 38°C), the system automatically triggers the spray cooling system and starts the spraying operation (spraying time range: 10 - 30 minutes, spraying pressure range: 0.2 - 0.5 MPa) to ensure the normal growth of grape plants.

[0031] IV. Collaborative Implementation Example of Edge Computing and Cloud Platform (I) Edge Computing Implementation Device Deployment: Deploy 10 ESP32 edge computing nodes at the vineyard site. Each node is responsible for connecting 50 surrounding sensors and image acquisition devices. The ESP32 nodes connect to the data acquisition module of the sensors through the SPI interface and connect to the image acquisition devices through the USB interface.

[0032] Data Processing: The ESP32 nodes receive the raw data collected by the sensors in real time. First, the data is filtered using the Kalman filter algorithm to remove noise interference. Then, outlier detection is performed, and the 3σ criterion is used to identify and mark the data that exceeds the normal range. For image data, simple image compression processing is performed using the JPEG compression algorithm, and the image quality factor is set to 80% to reduce the data transmission volume. The processed data is transmitted to the cloud platform through the Wi-Fi module, and at the same time, the processing results (such as whether the data is abnormal, the preliminary extraction results of image features) are fed back to the intelligent control system for timely adjustment of the control strategy.

[0033] (II) Cloud Platform Implementation Platform Construction: Use an ECS server to build the cloud platform. The server configuration is an 8-core CPU, 16GB of memory, and a 500GB hard drive. Install the Hadoop big data processing framework and the Spark distributed computing framework for storing and analyzing the data uploaded by the edge computing nodes.

[0034] Data analysis and services: After the cloud platform receives the data uploaded by the edge computing nodes, it first stores the data in the Hadoop Distributed File System (HDFS). Then, it uses the Spark framework to deeply mine the historical data, adopts the clustering analysis algorithm to analyze the yield change rules in different seasons and years, and adopts the association rule mining algorithm to analyze the relationship between the occurrence of pests and diseases and environmental factors. According to the analysis results, planting suggestions are provided to farmers. For example, in June-July every year (the grape fruit swelling period), according to the soil nutrient content and historical yield data, it is recommended that farmers top-dress 50 kg of high-potassium compound fertilizer per mu; according to the occurrence rules of pests and diseases, in mid-May every year (before the high incidence of pests and diseases), farmers are advised in advance to take pest and disease prevention measures, such as installing insect-proof nets and hanging traps. At the same time, farmers can log in to the cloud platform through the web end or mobile application to query information such as the historical data of the vineyard, real-time monitoring data, growth status analysis results, and the implementation status of intelligent control measures.

[0035] Multi-vineyard collaborative management: The cloud platform integrates the data of multiple vineyards, generates regionalized planting suggestions through clustering analysis, helps farmers learn from the successful experiences of surrounding vineyards, and optimize the planting management strategy. The edge computing nodes are interconnected through the Mesh network to achieve data sharing and redundant backup, improve the reliability and security of the data, and ensure the normal operation of the system even when some nodes fail.

[0036] User interaction module: Develop a mobile terminal application to display the vineyard monitoring data, model prediction results, and the execution status of control measures in real time, so that farmers can understand the situation of the vineyard anytime and anywhere. Farmers can manually adjust the alarm threshold or intervene in the control device parameters through the interface, and the system records the operation logs and synchronizes them to the cloud platform for subsequent query and analysis, realizing good interaction between farmers and the system.

[0037] The specific implementation cases of this embodiment are as follows: Case 1: A medium-sized vineyard in Bordeaux, France Deployment of multi-source data acquisition system: In this 30-hectare vineyard, according to the standard of one monitoring point per 400 square meters, temperature and humidity sensors (DHT22), light sensors (BH1750), and CO2 sensors (MH-Z19B) are evenly installed. These sensors are connected to the data acquisition module via the RS485 bus, and then the data is transmitted to the edge computing node in real time through the LoRa wireless transmission module. Every month, using the ASD FieldSpec4 portable ground object spectrometer, 60 sampling points are randomly selected in the garden to collect soil samples at a depth of 0-20 cm, analyze the reflectance spectral data of the soil in the 350-2500 nm band, and use the partial least squares method to invert the soil nutrient content and distribution.

[0038] Install high-definition cameras at heights of 1.5 meters and 2 meters respectively beside the grape plants, and automatically photograph the above-ground parts of the grape plants every 15 minutes with a resolution of 1920×1080 pixels. Use a drone to fly along a preset route at a height of 50 meters and a speed of 5 m / s every week to collect overall images of the vineyard. Install an optimized fog cultivation device near the roots of each grape plant, with an OV2710 high-definition camera built-in, and collect root images every 24 hours. By adjusting the spray interval (15 - 30 seconds) and adapting the nutrient solution formula (nitrogen-phosphorus-potassium ratio of 15:10:10 during the germination period and 10:8:15 during the fruit expansion period), create a suitable environment for root growth and monitor the growth amount.

[0039] Deep learning model construction and training: Use Python and PyTorch as tools to build a spatio-temporal fusion deep learning network integrating LSTM and CNN. Set 3 hidden layers in LSTM, with 128 neurons in each layer; for CNN, use the ResNet50 backbone network, followed by an SPP module. Select the ResNet50 model pre-trained on the ImageNet and PlantVillage datasets, and fine-tune the parameters of the last few convolutional layers and fully connected layers of the model with 20% of the local data for the main variety Cabernet Sauvignon in the garden and the local maritime climate.

[0040] After normalizing and standardizing the collected historical environmental data and image data, divide them into a training set, a validation set, and a test set according to a ratio of 7:2:1. Use the Adam optimizer with an initial learning rate of 0.001, which is halved every 10 epochs, and the loss function is the weighted sum of mean squared error and cross-entropy (weights 0.5:0.5). After 100 rounds of training, the model can accurately predict the pest and disease risks, water demand, and nitrogen-phosphorus-potassium nutrient demand within the next 72 hours.

[0041] Operation of the intelligent control system: Connect the LED supplementary light, drip irrigation system, and drone spraying equipment to the deep learning model. When the model predicts insufficient light (sunshine duration less than 6 hours), the supplementary light automatically turns on and is adjusted to a brightness of 1000 - 5000 lux. According to the predicted water demand, the drip irrigation system automatically controls the opening time (10 - 60 minutes) and flow rate (1 - 5 liters / minute). If a pest and disease risk is detected (such as the risk probability of downy mildew being greater than 0.7), the drone spraying equipment quickly starts, and sets the flight path and adjusts the spraying dose (100 - 500 ml / mu) according to the type and severity of the pest and disease for precise control.

[0042] Set three levels of alarm thresholds to monitor indicators such as temperature, humidity, and pest and disease risks. When data is abnormal, the first-level alarm (yellow warning) will notify farmers through the APP push notification; the second-level alarm (orange warning) will add SMS notifications; the third-level alarm (red warning) will automatically trigger the spray cooling system in case of high temperature (maximum daily temperature exceeding 38°C), spraying for 10-30 minutes, and the pressure is maintained at 0.2-0.5 MPa to ensure the normal growth of grape plants.

[0043] Edge computing and cloud platform work together: 10 ESP32 edge computing nodes are deployed in the vineyard, and each node is responsible for processing data from 40-50 sensors and image acquisition devices in the surrounding area. The nodes receive data in real time, de-noise it through Kalman filtering, detect abnormal values through the 3σ criterion, compress the image data (JPEG quality factor 80%), and transmit it to the cloud platform via Wi-Fi. At the same time, the processing results are fed back to the intelligent control system.

[0044] The cloud platform uses ECS servers (8-core CPU, 16GB memory, 500GB hard disk), equipped with Hadoop and Spark frameworks. After receiving edge node data, it is stored in HDFS and deeply analyzed using Spark. For example, cluster analysis found that grapes are prone to disease during the high temperature period from July to August each year, and farmers are advised to take precautions in advance; data for many years is stored for farmers to query and compare at any time to assist in planting decisions.

[0045] Case 2: A large vineyard in Xinjiang, China Deployment of multi-source data acquisition system: The vineyard covers an area of 100 hectares. Temperature and humidity sensors (AM2302), light sensors (TSL2591) and carbon dioxide sensors (SCD30) are installed at a density of one monitoring point per 600 square meters. The data is aggregated to the edge computing node through ZigBee wireless transmission technology. The PSR+3500 hyperspectral spectrometer is used every two weeks to collect soil spectral data at 80 sampling points in the vineyard, and the principal component regression algorithm is used to analyze soil nutrients.

[0046] High-definition cameras are installed at 1m, 2m, and 3m in the vineyard to capture images of the aboveground parts every 12 minutes. A drone is used to fly at an altitude of 60m and a speed of 6m / s every 3 days to collect overall images. An independently developed optimized mist cultivation device is installed at the roots of the grapes, with a built-in OV5647 camera to collect root system images every day. Root growth is monitored by adjusting the spray interval (20-40 seconds) and the formula of the nutrient solution (nitrogen, phosphorus, and potassium ratios of 12:10:8 during the flowering period and 8:6:15 during the color change period).

[0047] Deep learning model construction and training: Using Python and the TensorFlow framework, a spatio-temporal fusion deep learning network integrating bidirectional LSTM and 3D convolutional neural network (with 3DResNeXt as the backbone network) is constructed. The bidirectional LSTM has 2 hidden layers, with 256 neurons in each layer. Select the 3DResNeXt model pre-trained on the PlantVillage, PlantClef datasets and data from multiple domestic vineyards. For the main variety Thompson Seedless in the orchard and the local temperate continental climate, fine-tune the parameters of the last two 3D convolutional blocks and the fully connected layer of the model with 15% local data.

[0048] After preprocessing the data, divide the dataset according to 6:2:2. Use the AdamW optimizer with an initial learning rate of 0.0005. When the validation set loss does not decrease for 5 consecutive epochs, the learning rate is reduced to 0.2 times the original. The loss function is selected as the focal loss function. After multiple rounds of training, the model can accurately predict the indicators related to the growth status of the vineyard in the next 72 hours.

[0049] Operation of the intelligent control system: Connect the LED plant growth lights, sprinkler irrigation system and drone spraying equipment to the deep learning model. When the model predicts insufficient light (sunshine duration less than 7 hours), the plant growth lights are automatically turned on and adjusted to a brightness of 2000 - 8000 lux. The sprinkler irrigation system automatically controls the opening time (30 - 120 minutes) and sprinkler irrigation pressure (0.3 - 0.8 MPa) according to the predicted water demand. If the pest and disease risk is relatively high (such as the risk probability of spider mites is greater than 0.8), the drone spraying equipment is started, and the flight path and spraying dose (200 - 800 ml / mu) are set according to the pest and disease situation.

[0050] Set four-level alarm thresholds for monitoring indicators such as temperature, humidity, and pest and disease risks. The first-level alarm (blue warning) is pushed through the APP; the second-level alarm (yellow warning) adds SMS notification; the third-level alarm (orange warning) automatically adjusts the equipment parameters; the fourth-level alarm (red warning) in case of extreme high temperature (daily maximum temperature exceeds 40°C), automatically unfolds the sunshade net (sunshade rate 30% - 70%) to protect the grape plants. 4. Edge computing and cloud platform collaborative operation: Deploy 20 edge computing nodes using Raspberry Pi 4B to be responsible for collecting and processing data from surrounding devices. After filtering and detecting anomalies in the data, upload it to the cloud platform through the 4G network, and at the same time feedback the processing results to the intelligent control system. The cloud platform is built on Tencent Cloud servers, and uses big data analysis tools to mine data value. For example, analyzing data over the years finds that strong winds in spring are likely to cause pests and diseases, and it is recommended that farmers install windbreaks in advance and strengthen pest and disease monitoring.

[0051] In the description of this specification, the descriptions referring to the terms "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in a suitable manner in any one or more embodiments or examples.

[0052] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and utilize the present invention well. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A real-time monitoring and intelligent regulation method for the growth status of vineyards based on deep learning, characterized in that, It includes the following steps: S1. Construct a multi-source data acquisition system, including: An environmental data acquisition module, which is used to collect the micro-environment data of the vineyard by deploying temperature and humidity sensors, light sensors and CO2 sensors, and obtain the soil nutrient content and distribution through soil hyperspectral analysis technology; An image data acquisition module, which is used to obtain the above-ground part image of the grape plant, the overall image of the vineyard and the root image through a multi-angle high-definition camera, a drone and a root image acquisition device respectively; A root system data acquisition module, which is used to regulate the root growth environment by optimizing the fog cultivation device and monitor the root growth amount in real time based on an image processing algorithm; S2. Build and train a spatio-temporal fusion deep learning network, including: Construct a deep learning model that combines a long short-term memory network (LSTM) and a convolutional neural network (CNN), and input the historical environmental data, real-time image data and root growth amount data collected by the multi-source data acquisition system; Based on the pre-trained model, fine-tune the deep learning model through transfer learning technology to adapt to different grape varieties and climate conditions; Train the deep learning model so that it can predict the pest and disease risks, water requirements and nutrient requirements in the vineyard within the next 72 hours; S3. According to the output result of step S2, judge whether regulation is needed. If so, trigger the intelligent regulation system to perform the following operations: Link the supplementary light, drip irrigation system and drone spraying equipment to automatically adjust the environmental parameters according to the predicted insufficient light, water requirements and pest and disease risks; Set multi-level alarm thresholds. When the monitored data or prediction results exceed the thresholds, notify the farmers through the mobile terminal and trigger emergency measures.

2. The real-time monitoring and intelligent regulation method for the growth state of a vineyard based on deep learning according to claim 1, characterized in that, In the step S1: The soil hyperspectral analysis technology uses a portable ground object spectrometer to collect soil reflection spectral data, and inverses the soil nutrient content through partial least squares method or principal component regression algorithm; The root image acquisition module includes: Optimize the fog cultivation device, and provide a controllable growth environment for the roots through spray interval adjustment and dynamic adaptation of the nutrient solution formula; An built-in high-definition camera periodically collects root images, and uses an improved U-Net or MaskR-CNN algorithm for image segmentation to calculate root length, surface area and volume indicators.

3. The real-time monitoring and intelligent regulation method for the growth status of vineyards based on deep learning according to claim 1, characterized in that In the step S2: The spatio-temporal fusion deep learning network includes ResNet50 as the CNN backbone network, combined with a spatial pyramid pooling (SPP) module to adapt to input images of different sizes; The transfer learning technology is specifically: freeze some convolutional layers of the pre-trained model, fine-tune the fully connected layer based on specific vineyard data, and gradually unfreeze the convolutional layers to improve the model generalization ability.

4. The real-time monitoring and intelligent regulation method for the growth status of vineyards based on deep learning according to claim 3, characterized in that, In the transfer learning technology, select a model pre-trained on a public plant dataset and vineyard data in different regions. When fine-tuning for a specific vineyard, the fine-tuning dataset accounts for 15%-20% of the total data, and fine-tune the parameters of the last few convolutional layers and fully connected layers of the pre-trained model.

5. The real-time monitoring and intelligent regulation method for the growth state of a vineyard based on deep learning according to claim 3, characterized in that, The training process in the step S2 includes: Data preprocessing uses normalization and Z-score standardization, and the ratio of the training set, validation set and test set is 7:2:1 or 6:2:2; The loss function is the weighted sum of mean squared error and cross entropy, with a weight ratio of 0.5:0.

5. The optimizer uses Adam or AdamW, and the learning rate dynamic decay strategy is to halve it every 10 epochs or trigger decay based on the validation loss.

6. The real-time monitoring and intelligent control method for the growth state of a vineyard based on deep learning according to claim 1, characterized in that In the step S3: The multi-level alarm thresholds include at least three levels, corresponding to different degrees of abnormality respectively. The triggering methods include mobile application push, SMS notification and automatic activation of emergency devices; The emergency measures include a spray cooling system, unfolding of sunshade nets or increasing the frequency of drone spraying, and the parameter adjustment range is dynamically adjusted according to the prediction results.

7. The real-time monitoring and intelligent regulation method for the growth status of vineyards based on deep learning according to claim 1, characterized in that, It also includes the collaboration between edge computing and the cloud platform: Edge computing nodes are deployed on-site in the vineyard to receive and process the data collected by sensors in real time, and conduct preliminary screening, cleaning and analysis to reduce the amount and frequency of data transmitted to the cloud; The cloud platform receives the data uploaded by the edge nodes for in-depth big data analysis, mines the long-term trends of the vineyard growth status, provides planting suggestions for farmers and stores historical data.

8. The real-time monitoring and intelligent regulation method for the growth status of vineyards based on deep learning according to claim 7, characterized in that, This method supports the collaborative management of multiple vineyards. Specifically: The cloud platform integrates the data of multiple vineyards and generates regionalized planting suggestions through cluster analysis; The edge computing nodes are interconnected through a Mesh network to achieve data sharing and redundant backup.

9. The real-time monitoring and intelligent regulation method for the growth status of vineyards based on deep learning according to claim 1, characterized in that It also includes a user interaction module. Specifically: The mobile terminal displays the vineyard monitoring data, model prediction results and the execution status of the control measures in real time; Farmers can manually adjust the alarm thresholds or intervene in the control device parameters through the interface, and the system records the operation logs and synchronizes them to the cloud platform.

10. The real-time monitoring and intelligent regulation method for the growth status of vineyards based on deep learning according to claim 1, characterized in that, The drone spraying equipment in the intelligent control system includes: A dynamic path planning algorithm to generate the optimal flight path according to the vineyard terrain and the distribution of pests and diseases; the spraying dose and frequency are adjusted based on the pest and disease risk level predicted by the model, with the dose range of 100 - 800 ml / mu and the spraying height of 1 - 3 meters.

Images