Intelligent seedling breeding method and system

Abstract

The application discloses an intelligent seedling breeding system, which comprises a multi-modal sensing layer, an edge computing layer, a cloud intelligent decision-making layer, a regulation and execution layer and a cluster collaborative regulation and execution module. The multi-modal sensing layer comprises an environmental sensor array, a seedling visual acquisition unit, a nutrient solution monitoring unit and a weather station unit, which are used for collecting multi-dimensional environmental data, seedling growth image data, nutrient solution parameters and outdoor weather data in a greenhouse in real time. The edge computing layer comprises a data preprocessing module, a feature extraction module, a local real-time control module and a data caching and uploading module, which are used for preprocessing, feature extraction and local real-time regulation and control of the data collected by the sensing layer. The application guarantees seedling cultivation quality from the root, realizes early detection and early prevention and control, and greatly reduces seedling loss caused by diseases and insect pests. In addition, the intelligent adaptive growth environment can accelerate seedling development, effectively shorten the seedling cultivation period and significantly improve seedling cultivation production efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural technology, specifically to an intelligent seedling propagation method and system. Background Technology

[0002] Seedling propagation is a core preliminary step in agricultural production. The growth, uniformity, and health of seedlings directly determine the subsequent growth quality, stress resistance, and final yield of crops. Standardized, precise, and intelligent seedling management is key to improving the quality and efficiency of modern agriculture. Currently, the greenhouse environmental control systems used in large-scale seedling bases in China generally rely on manual experience or fixed threshold-based on / off control modes, resulting in a low overall level of intelligence. This makes it difficult to adapt to the dynamic growth needs of seedlings throughout their entire growth cycle and fails to meet the high-standard seedling production requirements of modern facility agriculture.

[0003] Existing traditional seedling control systems suffer from core defects such as low precision in environmental control and delayed response. Traditional fixed threshold control modes can only achieve simple start-stop control and cannot dynamically adjust environmental parameters according to the seedling growth stage and changes in outdoor weather. This results in large fluctuations in key parameters such as temperature, humidity, and light within the greenhouse, leading to poor stability of the seedling growth environment. Furthermore, existing monitoring methods mostly rely on single-point, low-frequency sampling, making it difficult to capture instantaneous changes in the greenhouse microenvironment. This results in significant time delays in control decisions, making it prone to over- and under-adjustment, ultimately leading to uneven seedling growth, low seedling survival rates, and difficulty in ensuring consistent seedling quality.

[0004] Traditional seedling systems suffer from several industry pain points, including a lack of multi-factor synergistic regulation, insufficient crop perception capabilities, and weak cluster management. Environmental factors in greenhouse seedling cultivation, such as temperature, humidity, light, water, and fertilizer, exhibit strong coupling relationships. Traditional systems employ independent regulation of each factor, ignoring the interactions between parameters, resulting in low regulation efficiency and significant energy waste. More critically, existing systems only monitor environmental data and cannot perceive the actual growth status and health level of seedlings in real time, lacking the ability to predict growth trends. This passive environmental regulation fails to achieve precise, on-demand cultivation.

[0005] To solve the above-mentioned technical problems, there is a need to provide an intelligent seedling propagation method and system. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent seedling propagation method and system. Through multimodal deep learning and multi-sensor data fusion technology, it realizes intelligent perception and prediction of seedling growth status and coordinated and precise control of multiple environmental factors, significantly improving seedling quality and efficiency, and reducing energy consumption and labor costs.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent seedling breeding system, comprising: a multimodal perception layer, an edge computing layer, a cloud-based intelligent decision-making layer, a control and execution layer, and a cluster collaborative control module; The multimodal perception layer includes an environmental sensor array, a seedling visual acquisition unit, a nutrient solution monitoring unit, and a weather station unit, which are used to collect multi-dimensional environmental data, seedling growth image data, nutrient solution parameters, and outdoor meteorological data in the greenhouse in real time. The edge computing layer includes a data preprocessing module, a feature extraction module, a local real-time control module, and a data caching and uploading module, which are used to preprocess, extract features, and control the data collected by the perception layer in local real-time. The cloud-based intelligent decision-making layer includes a seedling growth prediction model, a multi-factor collaborative regulation and optimization model, a pest and disease early warning model, and a knowledge graph database, which are used to construct intelligent prediction models and optimize regulation strategies. The control execution layer includes a temperature control unit, a humidity control unit, a light control unit, a CO2 control unit, and a water and fertilizer integration control unit, which are used to execute the control strategies sent from the cloud to achieve precise regulation of the greenhouse environment; The cluster collaborative control module includes a load forecasting unit, a resource scheduling unit, and an energy efficiency optimization unit, which are used for the collaborative optimization and control of multiple greenhouses in a large-scale seedling base.

[0008] Preferably, the seedling growth prediction model adopts a multimodal deep learning architecture, including an environmental time-series data input branch, a visual feature input branch, a hyperspectral feature input branch, an attention fusion module, and a prediction output layer; The environmental time-series data input branch uses a two-layer LSTM network to process environmental time-series data from the past 72 hours; the visual feature input branch uses a ResNet-50 convolutional neural network to extract seedling image features; the hyperspectral feature input branch uses a one-dimensional convolutional network to extract spectral features; the attention fusion module uses a multi-head attention mechanism to adaptively weight and fuse the features from the three branches; and the prediction output layer outputs the predicted growth rate and growth stage transition probability for the next 24, 48, and 72 hours.

[0009] Preferably, the multi-factor collaborative regulation optimization model is constructed based on a deep deterministic strategy gradient algorithm; The state space includes current environmental parameters, seedling growth status, and external weather forecast data; the action space includes temperature setpoints, humidity setpoints, light intensity and spectral ratio, CO2 concentration setpoints, nutrient solution EC value and pH value, irrigation amount and irrigation frequency; the reward function is r = w1·G - w2·E - w3·W; Where G is the seedling growth gain, E is the energy cost, W is the water resource consumption, and w1, w2, and w3 are weighting coefficients.

[0010] Preferably, the pest and disease early warning model adopts a two-stage detection architecture: the first stage uses a deep autoencoder to detect anomalies and calculate the reconstruction error of the input image. When the error exceeds a threshold, it is judged as an anomaly; the second stage uses a CNN classification network to classify the abnormal area in detail and identify the specific disease type and severity; it combines hyperspectral data analysis of physiological and biochemical changes inside the leaves to improve the sensitivity of early detection; and it classifies the risk according to the proportion of lesion area, the spread speed, and the degree of damage.

[0011] Preferably, the integrated water and fertilizer control unit includes an automatic nutrient solution mixing device and a drip irrigation system; the nutrient solution mixing device is equipped with multiple mother liquor tanks, and uses a high-precision metering pump to extract mother liquor and mix it with clean water in proportion, and monitors EC value and pH value in real time and adjusts them in a closed loop, with a mixing accuracy of ±2%; the irrigation system adopts drip irrigation, and dynamically adjusts the irrigation amount and irrigation frequency according to the seedling growth stage and water demand.

[0012] Preferably, the load forecasting unit of the cluster collaborative control module uses an LSTM network to predict the power load and water consumption of each greenhouse in the next 24 hours; the resource scheduling unit combines the time-of-use electricity pricing policy and uses a genetic algorithm to optimize the control timing of each greenhouse; and the energy efficiency optimization unit analyzes the heat exchange effect of the contiguous greenhouses and optimizes the temperature settings of adjacent greenhouses.

[0013] An intelligent seedling propagation method includes the following steps: Step 1: Multimodal data acquisition and preprocessing: Environmental parameters, seedling images, nutrient solution parameters, and meteorological data are acquired through the multimodal sensing layer and preprocessed. Step 2: Intelligent perception of seedling growth status: Extract phenotypic features from seedling images, identify seedling growth stages, and assess health status; Step 3: Seedling growth status prediction: Based on a multimodal deep learning model, predict the seedling growth rate and growth stage transition time points in the next 24-72 hours; Step 4: Multi-factor collaborative regulation decision-making: Based on deep reinforcement learning algorithms, with optimal seedling growth as the objective and minimum energy consumption as the constraint, a multi-factor collaborative regulation strategy is output. Step 5: Early warning of diseases and pests: Passed.

[0014] In summary, this invention provides an intelligent seedling propagation method and system, achieving the following technical effects: This invention achieves precise seedling environment control and crop-driven intelligent regulation, significantly improving seedling quality and survival rate. It abandons the traditional passive control mode with fixed thresholds, relying on a distributed multi-sensor array to achieve full-coverage monitoring of the greenhouse microenvironment. Combining seedling vision and hyperspectral sensing technologies, it captures seedling growth status, nutrient levels, and health status in real time, constructing an active, on-demand regulation system centered on crop growth needs. Through multimodal feature fusion and intelligent prediction models, it dynamically matches optimal environmental parameters for different seedling growth stages. Coupled with a multi-factor collaborative regulation algorithm, it completely solves the problems of large parameter fluctuations and fragmented regulation in traditional systems. Temperature control accuracy reaches ±0.5℃, and humidity control accuracy reaches ±3%RH, effectively improving seedling growth consistency and increasing seedling survival rate by 10-15%, fundamentally ensuring seedling quality.

[0015] This invention possesses predictive control and early pest and disease warning capabilities, effectively shortening the seedling cycle and reducing production energy consumption and losses. Combining outdoor weather forecasting and seedling growth status prediction, this invention anticipates environmental changes and seedling growth needs in advance, employing a proactive, predictive control strategy to avoid energy losses caused by frequent and large-scale parameter adjustments. Compared to traditional systems, it can reduce overall seedling energy consumption by 20-30%. Simultaneously, relying on deep learning image recognition and hyperspectral analysis technology, it can detect anomalies in the early stages of pest and disease outbreaks that are invisible to the naked eye, providing early warnings of over 48 hours, enabling early detection and control, and significantly reducing seedling losses caused by pests and diseases. Furthermore, the intelligently adaptable growth environment accelerates seedling development, effectively shortening the seedling cycle and significantly improving seedling production efficiency.

[0016] This invention is adapted to large-scale seedling cultivation scenarios, achieving cluster collaborative optimization and system adaptive iterative upgrades. Addressing the needs of parallel production in multiple greenhouses at large seedling bases, this invention adds a cluster collaborative control module. Through load forecasting, time-sharing resource scheduling, and greenhouse heat exchange efficiency optimization, it achieves peak shaving and valley filling across multiple greenhouses, and shared utilization of water, fertilizer, and energy resources, effectively improving the overall operational efficiency of large-scale seedling cultivation by 15-20% and reducing large-scale production costs. Simultaneously, the intelligent decision-making model of this invention supports online learning and dynamic parameter updates. It can continuously optimize control strategies based on accumulated seedling data, possessing self-learning and self-iterative characteristics. Over long-term use, its control accuracy and adaptability continuously improve, adapting to the seedling propagation needs of different varieties and environments, making it highly valuable for industrial application. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall architecture of the intelligent seedling propagation system of the present invention; Figure 2 This is a schematic diagram of the multimodal sensing layer of the present invention; Figure 3 This is a flowchart illustrating the cloud-based intelligent decision-making layer of the present invention. Figure 4 This is a schematic diagram of the structure of the seedling growth prediction model of the present invention; Figure 5 This is a schematic diagram illustrating the working principle of the multi-factor synergistic regulation optimization model of the present invention; Figure 6 This is a flowchart illustrating the intelligent seedling propagation method of the present invention. Detailed Implementation

[0018] To make the inventive objectives, complete technical solutions, and outstanding beneficial effects compared to the prior art clearer and more detailed, and to facilitate the complete replication and implementation of this solution by those skilled in the art, the device structure, implementation steps, and supporting technical details disclosed in the embodiments of this application will be comprehensively, clearly, and completely described below with reference to the accompanying drawings corresponding to multiple embodiments of this application.

[0019] Example 1: As Figures 1 to 6 As shown in this embodiment, an intelligent seedling propagation system is disclosed. The overall architecture sequentially includes a multimodal perception layer, an edge computing layer, a cloud-based intelligent decision-making layer, a control and execution layer, and a cluster collaborative control module. Each layer cooperates to complete the intelligent perception, intelligent decision-making, and precise control of seedling propagation. The multimodal perception layer mainly consists of an environmental sensor array, a seedling visual acquisition unit, a nutrient solution monitoring unit, and a weather station unit, used to comprehensively collect environmental, seedling, nutrient solution, and outdoor meteorological data during the seedling cultivation process. The environmental sensor array is uniformly deployed in a grid pattern within the greenhouse, with one sensor node set for every 100 square meters. Each node integrates temperature, humidity, light intensity, and carbon dioxide concentration detection elements. An additional soil temperature and humidity sensor is provided for the dedicated area of ​​the seedling bed. All sensor nodes use LoRa wireless communication to transmit data, uniformly set to a 2Hz sampling frequency, enabling comprehensive and real-time coverage of the environmental conditions in different areas of the greenhouse.

[0020] The seedling visual acquisition unit consists of a 20-megapixel high-definition industrial camera and a visible-near-infrared hyperspectral imaging device. The entire unit is mounted two meters above the seedbed via a sliding track, allowing it to move at a uniform speed along the length of the seedbed. The camera's movement speed is set to 0.5 m / s, achieving an image resolution of 0.5 mm / pixel. The hyperspectral imaging device has a spectral detection range of 400-1000 nm and a spectral resolution of 5 nm. The device can adaptively adjust the acquisition frequency according to the seedling growth stage: acquiring image data every two hours during germination, every four hours during the cotyledon stage, and every six hours during the true leaf and seedling stages, achieving visualized monitoring of the entire seedling growth cycle. The nutrient solution monitoring unit is centrally deployed in the nutrient solution storage tank and irrigation main pipeline, integrating four types of sensors: EC value, pH value, dissolved oxygen, and nutrient solution temperature. It completes data sampling once per minute, monitoring the nutrient solution water-fertilizer ratio and water quality status in real time. The meteorological station unit is located in an open area around the greenhouse without any obstructions. It is equipped with outdoor temperature and humidity, light, wind speed and direction, and rainfall sensors. Outdoor meteorological data is collected every five minutes to provide external environmental data for predictive control of the greenhouse.

[0021] The edge computing layer is deployed on a local ARM-based edge gateway device in the greenhouse. This device is equipped with 8GB of RAM and 128GB of storage, runs a Linux operating system, and includes a data preprocessing module, a feature extraction module, a local real-time control module, and a data caching and uploading module. Its main functions include local data processing and emergency control. The data preprocessing module standardizes multi-source, heterogeneous collected data, using the 3σ criterion to remove abnormal sensor data, employing a Kalman filter algorithm for denoising and smoothing time-series data, and using linear interpolation to synchronize data from different sampling frequencies and fill in short-term missing data. For long-term missing data, it combines historical data from the same period and trend extrapolation to complete the data, ensuring the accuracy and consistency of the input data. The feature extraction module uses a lightweight deep learning model, employing a U-Net semantic segmentation network to accurately distinguish seedlings from the background area, extracting the outline of individual seedlings, and then calculating morphological phenotypic features such as plant height, leaf area, and number of leaves. Simultaneously, leaf color index and yellowing degree are extracted through image color space conversion and histogram analysis, and the relative chlorophyll content and nitrogen nutrition level of seedlings are estimated by combining hyperspectral vegetation index, comprehensively extracting seedling growth characteristic information. The local real-time control module is responsible for emergency control of the greenhouse. It presets the safe environmental threshold for seedling growth. When parameters such as temperature and humidity in the greenhouse exceed the safe range, it can quickly trigger the start and stop control of the equipment within one second, without waiting for cloud instructions, ensuring the safety of seedling growth from the source. The data caching and upload module can cache the pre-processed effective data locally and upload it to the cloud every five minutes. It is also equipped with a breakpoint resume mechanism, which automatically re-uploads the unuploaded data after the network is restored, effectively ensuring the integrity and continuity of data transmission.

[0022] The cloud-based intelligent decision-making layer, deployed on a cloud server cluster, comprises a seedling growth prediction model, a multi-factor collaborative regulation and optimization model, a pest and disease early warning model, and a knowledge graph database, serving as the core carrier for the system's intelligent decision-making. The seedling growth prediction model employs a multimodal deep learning architecture, building three input branches: environmental time series, visual features, and hyperspectral features. It processes nearly 72 hours of environmental time series data and extracts time series variation features using two LSTM networks with 128 hidden units each. A ResNet-50 convolutional neural network is used to extract visual phenotypic features from seedling images. A one-dimensional convolutional network is used to analyze hyperspectral curves and extract spectral features. A multi-head attention mechanism is then used to adaptively weight and fuse the three types of features. Finally, a fully connected output layer yields the seedling growth rate and growth stage transition probability for the next 24, 48, and 72 hours. The model is trained using a transfer learning approach combining pre-training on public datasets with fine-tuning on local seedling data, and an early stopping strategy is employed to prevent overfitting.

[0023] The multi-factor collaborative regulation and optimization model is built based on a deep deterministic strategy gradient algorithm. It uses real-time greenhouse environmental parameters, seedling growth status, and outdoor weather forecast data as the state space, and continuous regulation variables such as temperature and humidity setpoints, light parameters, carbon dioxide concentration, nutrient solution parameters, and irrigation parameters as the action space. A reward function is constructed with seedling growth gain as a positive reward and energy and water consumption as negative constraints. An Actor network outputs the optimal regulation action, and a Critic network evaluates the action value. The regulation strategy is continuously optimized through simulation pre-training and online fine-tuning. The pest and disease early warning model adopts a two-stage detection architecture. First, a deep autoencoder models normal seedling images, using image reconstruction errors to perform initial screening for pest and disease anomalies. Then, a CNN classification network further classifies abnormal areas, accurately identifying the types and severity of diseases such as downy mildew, powdery mildew, and gray mold. Simultaneously, hyperspectral data is used to analyze physiological characteristics such as chlorophyll degradation and water changes within leaves, improving detection sensitivity in the early stages of disease. Risk classification is completed based on the proportion of lesions and the spread rate, and corresponding prevention and control recommendations are pushed. The knowledge graph database builds a dedicated knowledge system for the seedling industry, covering entities and relationships such as crop varieties, growth stages, suitable environmental parameters, types of pests and diseases, and prevention and control measures. It can be dynamically updated and improved based on actual seedling data, providing professional knowledge support for intelligent decision-making.

[0024] The control and execution layer is the carrier for implementing system commands. It includes temperature control, humidity control, light control, carbon dioxide control, and fertigation control units, and can accurately execute various control commands issued from the cloud. The temperature control unit is equipped with heating equipment and evaporative cooling systems, employing a composite algorithm of PID and fuzzy control. Under large temperature difference conditions, it quickly responds to temperature adjustment through fuzzy control, and switches to PID high-precision voltage stabilization when approaching the target temperature, achieving a temperature control accuracy of ±0.5℃. The humidity control unit is equipped with high-pressure spray humidification and dehumidification equipment, which can dynamically adjust the spray volume and dehumidification power according to the humidity requirements of different growth stages of seedlings, with humidity control accuracy within ±3%RH.

[0025] The light control unit consists of full-spectrum LED supplemental lighting and an electric shading net. The LED supplemental lighting can adjust the red-to-blue light ratio from 1:1 to 8:1 and the light intensity from 0 to 1000 μmol·m⁻²·s⁻¹. The shading net can automatically open and close according to the natural light intensity, dynamically matching the optimal light environment based on the seedlings' light requirements. The carbon dioxide control unit relies on a liquid carbon dioxide storage tank and an adjustable solenoid valve to precisely control the carbon dioxide concentration in conjunction with the greenhouse ventilation strategy, maintaining a stable control range of 400 to 1500 ppm. During ventilation, the unit automatically suspends gas supply to avoid resource waste. The integrated water and fertilizer control unit includes multiple sets of mother liquor mixing tanks, high-precision metering pumps, and an intelligent drip irrigation system. It can automatically mix nitrogen, phosphorus, potassium, and trace element nutrient solutions according to the seedlings' growth needs, perform real-time closed-loop calibration of EC and pH values, achieving a mixing accuracy of ±2%, and dynamically adjust irrigation time and volume to achieve precise water and fertilizer supply.

[0026] The cluster-based collaborative control module is adapted to large-scale multi-greenhouse seedling cultivation scenarios and includes a load forecasting unit, a resource scheduling unit, and an energy efficiency optimization unit. The load forecasting unit, based on an LSTM network, combines historical control data from each greenhouse with future weather forecasts to accurately predict the electricity and water consumption load of each greenhouse over the next 24 hours. The resource scheduling unit, relying on a genetic algorithm and combined with time-of-use pricing policies, optimizes the control timing of each greenhouse, scheduling high-energy-consuming control operations during off-peak hours to achieve peak shaving and valley filling, and reduce operating costs. The energy efficiency optimization unit, considering the characteristics of contiguous greenhouse layouts, analyzes the heat exchange patterns between greenhouses, optimizes the temperature setpoints of adjacent greenhouses, utilizes natural temperature differences to reduce overall temperature control energy consumption, and simultaneously achieves unified scheduling and shared utilization of water and fertilizer resources, significantly improving resource utilization and overall operational energy efficiency in large-scale seedling cultivation.

[0027] like Figure 1 As shown, this embodiment provides an intelligent seedling breeding system, including a multimodal perception layer, an edge computing layer, a cloud-based intelligent decision-making layer, a control and execution layer, and a cluster collaborative control module.

[0028] like Figure 2As shown, the multimodal perception layer includes an environmental sensor array, a seedling visual acquisition unit, a nutrient solution monitoring unit, and a weather station unit.

[0029] The environmental sensor array is deployed in a grid pattern within the greenhouse, with one sensor node per 100 square meters. Each node integrates a temperature sensor, humidity sensor, light intensity sensor, and CO2 concentration sensor; additional soil temperature and humidity sensors are deployed in the seedling bed area. All sensor nodes use LoRa wireless communication with a sampling frequency of 2Hz to ensure real-time data transmission and spatial coverage.

[0030] The seedling visual acquisition unit includes a 20-megapixel high-definition industrial camera and a visible-near-infrared hyperspectral imaging device. Deployed 2 meters above the seedbed via a sliding rail mechanism, it can move along the length of the seedbed to achieve full-coverage image acquisition of the entire seedbed. The camera's moving speed is 0.5 m / s, and the image resolution is 0.5 mm / pixel. The hyperspectral imaging device has a spectral range of 400-1000 nm and a spectral resolution of 5 nm. The acquisition frequency is adaptively adjusted according to the seedling growth stage: once every 2 hours during germination, once every 4 hours during the cotyledon stage, and once every 6 hours during the true leaf and seedling stages.

[0031] The nutrient solution monitoring unit is deployed in the nutrient solution storage tank and irrigation pipeline, and includes an EC value sensor, a pH value sensor, a dissolved oxygen sensor and a nutrient solution temperature sensor. The sampling frequency is 1 time / minute, and it monitors various parameters of the nutrient solution in real time.

[0032] Example 2: This example discloses a seedling cultivation method based on the above-mentioned intelligent seedling propagation system. The entire method relies on a multimodal perception, edge computing, cloud-based intelligent decision-making, and cluster collaborative control system to achieve intelligent, precise, and predictive cultivation of seedlings throughout their entire growth cycle. The overall process is coherent and closed-loop, covering seven core links: data acquisition and preprocessing, growth status perception, growth trend prediction, collaborative control decision-making, pest and disease early warning, execution feedback optimization, and multi-greenhouse cluster control. In the data acquisition and preprocessing stage, the system continuously collects environmental data such as temperature, humidity, light, carbon dioxide, and soil moisture in the greenhouse through a distributed sensor array at a frequency of 2Hz. It also adaptively acquires RGB and hyperspectral images of seedlings according to their growth stages through a sliding-rail vision device. Nutrient solution monitoring equipment collects EC value, pH value, dissolved oxygen, and temperature parameters of the nutrient solution every minute. At the same time, it collects outdoor meteorological data every five minutes through an outdoor weather station, completing the comprehensive acquisition of multi-dimensional and multi-modal data. Subsequently, the edge computing layer preprocessing module uniformly completes data cleaning, noise reduction, time synchronization, standardization, and missing value imputation, eliminating interfering and abnormal data and ensuring that all input data is accurate and consistent, laying a data foundation for subsequent intelligent analysis and decision-making.

[0033] In the intelligent perception of seedling growth status, the edge computing layer feature extraction module accurately segments the target region of the seedling image using the U-Net semantic segmentation algorithm, extracting core phenotypic features such as plant height, leaf area, number of leaves, and leaf color index. Combined with a pre-trained ResNet classification model, it accurately determines the growth stage of the seedling, such as germination, cotyledon, true leaf, and seedling stage. At the same time, it analyzes the chlorophyll content and nitrogen nutrition level of the seedling based on the hyperspectral vegetation index to determine the water stress status of the seedling. It calculates a seedling health index from 0 to 100 points by combining environmental parameters and seedling phenotypic data, and counts the coefficient of variation of plant height and stem diameter of the whole bed of seedlings. It accurately grasps the uniformity of seedling growth and health status, realizing a comprehensive upgrade from environmental monitoring to crop status monitoring. Based on this, seedling growth status prediction is carried out. The system inputs nearly 72 hours of environmental time-series data, seedling visual image data, and hyperspectral data into the corresponding network branches, extracts multi-dimensional feature vectors, and adaptively fuses various effective features through a multi-head attention mechanism. Relying on the trained multimodal prediction model, it accurately predicts the seedling growth rate and growth stage transition time in the next 24 hours, 48 ​​hours, and 72 hours, and simultaneously outputs the optimal environmental parameter adaptation scheme for the next growth stage, realizing the transformation from passive regulation to predictive active regulation.

[0034] In the intelligent control decision-making and execution phase, the cloud-based multi-factor collaborative control optimization model takes optimal seedling growth as its core objective and minimizes energy and water consumption as constraints. Combining real-time seedling status, greenhouse environment, and outdoor meteorological data, it uses a deep deterministic strategy gradient algorithm to solve for the optimal control scheme, outputting collaborative control parameters for multiple factors such as temperature, humidity, light, carbon dioxide, and water and fertilizer supply. Based on the varying response speeds of different equipment such as heating, ventilation, and irrigation, it formulates time-segmented collaborative execution plans to ensure that all environmental factors are synchronously and accurately adapted to the seedling growth needs. Throughout the seedling cultivation process, the pest and disease early warning model continuously reconstructs and detects anomalies in seedling images, initially screening for abnormal plant conditions, conducting detailed classification and identification of suspected disease areas, and combining hyperspectral data analysis of internal physiological pathological characteristics of leaves to achieve unobtrusive early warning of pest and disease outbreaks. Based on the extent and severity of disease spread, it classifies risks into mild, moderate, and severe, promptly pushing corresponding prevention and control measures to user terminals to reduce pest and disease losses at the source.

[0035] Based on the control strategies issued from the cloud, the system drives each execution unit to complete precise control operations. Through a composite control algorithm, it achieves high-precision temperature and humidity regulation, dynamically matches light intensity and spectral ratio, stabilizes carbon dioxide concentration, and intelligently proportions and supplies nutrient solution. Simultaneously, the local real-time control module at the edge monitors the greenhouse's safety status throughout the process. In case of sudden changes in environmental parameters or other emergencies, it can immediately trigger emergency control measures to ensure a stable growth environment for seedlings. All control execution results and environmental change data are transmitted back to the cloud model in real time, driving the model to learn online, iteratively update network parameters, and continuously optimize the control strategy to achieve closed-loop optimization of the control effect. For large-scale seedling bases, the entire method also includes a multi-greenhouse cluster collaborative control process. The system predicts the energy and water consumption needs of each greenhouse through a load prediction unit, optimizes the control timing of each greenhouse in conjunction with time-of-use electricity pricing policies, avoids peak energy consumption periods, and utilizes the heat exchange effect of contiguous greenhouses to optimize the overall temperature control strategy. It also unifies the scheduling and sharing of water, fertilizer, and energy resources, significantly improving the overall production efficiency and resource utilization efficiency of large-scale seedling cultivation, making it suitable for large-scale, standardized, and intelligent seedling propagation and production scenarios.

[0036] Example 3: This example is a field verification experiment. Tomato, the main variety of greenhouse vegetables, is used as the seedling object. A standard glass greenhouse of a large-scale seedling base is selected for the control experiment. The experiment comprehensively verifies the control accuracy, seedling effect, energy-saving benefits and pest and disease control capabilities of the intelligent seedling propagation system of the present invention in actual seedling scenarios, and proves the practicality and inventiveness of the technical solution of the present invention.

[0037] Experimental location: Standardized glass greenhouse No. 3 at the large-scale intelligent seedling nursery in the Modern Agricultural High-tech Industrial Demonstration Zone, Shouguang City, Shandong Province. The greenhouse has an overall size of 500m². 2 It is an integrated intelligent glass greenhouse, equipped with conventional greenhouse environmental control equipment and water and fertilizer irrigation equipment. The ventilation and lighting conditions are uniform, meeting the standards for large-scale seedling breeding experiments.

[0038] The trial period from sowing to seedling emergence was 30 days.

[0039] Experimental groups: An experimental group (using the intelligent seedling propagation system of this invention) and a control group (using the traditional threshold control system) were set up, with an area of ​​250 square meters for each group and other conditions being the same. This experiment strictly adhered to the principle of a single variable. Except for the control system, the external conditions of the two greenhouses were completely identical, including seedling substrate, sowing density, basic water and fertilizer application, ventilation duration, and manual operation of seedling management. Throughout the experiment, greenhouse environmental parameters, seedling growth indicators, equipment energy consumption, and pest and disease occurrence were continuously monitored. After the experiment, all data were compiled to form a quantitative comparison result. Experimental Conclusion: Compared to traditional fixed-threshold greenhouse environmental control systems, the AI-based intelligent seedling propagation system of this invention possesses significant technological advancements and practicality. This invention, through multimodal data perception, deep learning prediction, and reinforcement learning collaborative control technologies, completely solves industry pain points such as poor precision in traditional seedling environment control, fragmented factor regulation, lack of predictive capabilities, high energy consumption, uneven seedling growth, and lagging pest and disease control. In actual large-scale seedling propagation scenarios, it can effectively improve the precision of greenhouse temperature and humidity control, significantly enhance seedling growth consistency and seedling quality, shorten the seedling cycle, reduce production energy consumption and pest and disease losses, and adapt to the needs of large-scale, standardized, and intelligent seedling propagation production, possessing extremely high industrialization and promotion value. Furthermore, the experimental data fully verify the completeness, inventiveness, and sufficiency of disclosure of the technical solution of this invention, meeting the requirements for invention patent authorization.

[0040] The embodiments described in this invention are for illustrative purposes only and do not constitute a limitation on the scope of the claims. Other substantially equivalent substitutions that can be conceived by those skilled in the art are all within the scope of protection of this invention.

Claims

1. An intelligent seedling propagation system, characterized in that, include: Multimodal perception layer, edge computing layer, cloud-based intelligent decision-making layer, regulation and execution layer, and cluster collaborative regulation module; The multimodal perception layer includes an environmental sensor array, a seedling visual acquisition unit, a nutrient solution monitoring unit, and a weather station unit, which are used to collect multi-dimensional environmental data, seedling growth image data, nutrient solution parameters, and outdoor meteorological data in the greenhouse in real time. The edge computing layer includes a data preprocessing module, a feature extraction module, a local real-time control module, and a data caching and uploading module, which are used to preprocess, extract features, and control the data collected by the perception layer in local real-time. The cloud-based intelligent decision-making layer includes a seedling growth prediction model, a multi-factor collaborative regulation and optimization model, a pest and disease early warning model, and a knowledge graph database, which are used to construct intelligent prediction models and optimize regulation strategies. The control execution layer includes a temperature control unit, a humidity control unit, a light control unit, a CO2 control unit, and a water and fertilizer integration control unit, which are used to execute the control strategies sent from the cloud to achieve precise regulation of the greenhouse environment; The cluster collaborative control module includes a load prediction unit, a resource scheduling unit, and an energy efficiency optimization unit, which are used for the collaborative optimization and control of multiple greenhouses in a large-scale seedling base.

2. The intelligent seedling propagation system according to claim 1, characterized in that, The seedling growth prediction model adopts a multimodal deep learning architecture, including an environmental time-series data input branch, a visual feature input branch, a hyperspectral feature input branch, an attention fusion module, and a prediction output layer; The environmental time-series data input branch uses a two-layer LSTM network to process environmental time-series data from the past 72 hours; the visual feature input branch uses a ResNet-50 convolutional neural network to extract seedling image features; the hyperspectral feature input branch uses a one-dimensional convolutional network to extract spectral features; the attention fusion module uses a multi-head attention mechanism to adaptively weight and fuse the features from the three branches; and the prediction output layer outputs the predicted growth rate and growth stage transition probability for the next 24, 48, and 72 hours.

3. The intelligent seedling propagation system according to claim 1, characterized in that, The multi-factor collaborative regulation optimization model is constructed based on a deep deterministic strategy gradient algorithm; The state space includes current environmental parameters, seedling growth status, and external weather forecast data; the action space includes temperature setpoints, humidity setpoints, light intensity and spectral ratio, CO2 concentration setpoints, nutrient solution EC value and pH value, irrigation amount and irrigation frequency; the reward function is r = w1·G - w2·E - w3·W; Where G is the seedling growth gain, E is the energy cost, W is the water resource consumption, and w1, w2, and w3 are weighting coefficients.

4. The intelligent seedling propagation system according to claim 1, characterized in that, The pest and disease early warning model adopts a two-stage detection architecture: the first stage uses a deep autoencoder to detect anomalies and calculates the reconstruction error of the input image. When the error exceeds the threshold, it is judged as an anomaly. The second stage uses a CNN classification network to further classify abnormal areas, identify specific disease types and severity; combines hyperspectral data analysis to analyze physiological and biochemical changes inside leaves, improving early detection sensitivity; and classifies risks based on the proportion of lesion area, spread rate, and degree of damage.

5. The intelligent seedling propagation system according to claim 1, characterized in that, The integrated water and fertilizer control unit includes an automatic nutrient solution mixing device and a drip irrigation system. The nutrient solution mixing device is equipped with multiple mother liquor tanks. It uses a high-precision metering pump to extract mother liquor and mix it with clean water in proportion. It monitors EC and pH values ​​in real time and adjusts them in a closed loop. The mixing accuracy can reach ±2%. The irrigation system adopts drip irrigation and dynamically adjusts the irrigation amount and frequency according to the seedling growth stage and water requirements.

6. The intelligent seedling propagation system according to claim 1, characterized in that, The load forecasting unit of the cluster collaborative control module uses an LSTM network to predict the power load and water consumption of each greenhouse in the next 24 hours; the resource scheduling unit combines the time-of-use electricity pricing policy and uses a genetic algorithm to optimize the control timing of each greenhouse; the energy efficiency optimization unit analyzes the heat exchange effect of the contiguous greenhouses and optimizes the temperature settings of adjacent greenhouses.

7. An intelligent seedling propagation method, characterized in that, Includes the following steps: Step 1: Multimodal data acquisition and preprocessing: Environmental parameters, seedling images, nutrient solution parameters, and meteorological data are acquired through the multimodal sensing layer and preprocessed. Step 2: Intelligent perception of seedling growth status: Extract phenotypic features from seedling images, identify seedling growth stages, and assess health status; Step 3: Seedling growth status prediction: Based on a multimodal deep learning model, predict the seedling growth rate and growth stage transition time points in the next 24-72 hours; Step 4: Multi-factor collaborative regulation decision-making: Based on deep reinforcement learning algorithms, with optimal seedling growth as the objective and minimum energy consumption as the constraint, a multi-factor collaborative regulation strategy is output. Step 5: Early warning of diseases and pests: Passed.

Images