A smart irrigation decision-making method for greenhouses that integrates dynamic monitoring of crop biomass.
By using multi-source image fusion and physically constrained neural networks, high-precision dynamic monitoring and rolling temporal optimization of crop biomass in greenhouses were achieved, solving the problem of the disconnect between traditional irrigation management and crop water demand patterns, and improving the accuracy of irrigation decisions and the efficiency of water resource utilization.
Patent Information
- Application Number
- CN202610991152.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-04
- Publication Date
- 2026-11-17
AI Technical Summary
In existing greenhouse cultivation, traditional irrigation management models are out of sync with the physiological water requirements of crops, making it difficult to adapt to the actual needs of crops at different growth stages and biomass accumulation states. Furthermore, multimodal phenotypic perception has not been effectively coupled, resulting in insufficient precision in irrigation decisions.
By employing multi-source image fusion technology to monitor crop biomass, and combining physical constraint neural networks and model predictive control, high-precision dynamic monitoring and rolling time-domain optimization decision-making are achieved. Through multimodal data integration and physical constraint improvement of the loss function, a rolling time-domain feedforward irrigation optimization system is constructed.
It enables high-precision dynamic monitoring of crop biomass and efficient and precise allocation of water resources, improves the accuracy of irrigation decisions and water and fertilizer utilization efficiency, adapts to the physiological needs of crops at different growth stages, and reduces ineffective water consumption.
Smart Images

Figure CN123117577A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of smart agriculture and facility horticulture technology, and relates to a smart irrigation decision-making method for greenhouses, specifically a smart irrigation decision-making method that integrates dynamic monitoring of crop biomass and multimodal data enhancement. Background Technology
[0002] With the rapid development of facility agriculture in my country, greenhouse cultivation has become an important model for ensuring a year-round supply of high-quality vegetables. Tomatoes, as one of the major facility-grown economic crops, are characterized by high economic value and complex water requirements, occupying a core position in greenhouse cultivation across the country. However, with the continuous expansion of cultivation scale and the increasing global water shortage, the traditional extensive water and fertilizer management model in greenhouse cultivation has become a key bottleneck restricting the efficient development of agriculture.
[0003] Current greenhouse crop irrigation decisions are disconnected from the actual physiological water requirements of crops. Traditional irrigation management relies heavily on empirical models of fixed times and quantities, or passive feedback based solely on environmental sensor data. It excessively depends on subjective human experience and external microclimate parameters, while traditional transpiration water consumption mechanism models generally have fixed parameters, ignoring the actual physiological water requirements of crops at different growth stages and biomass accumulation states. This makes it difficult to adapt to the complex transpiration responses caused by the dynamic evolution of the canopy under high canopy density. Furthermore, purely data-driven black-box deep learning irrigation decision models may overfit under limited greenhouse sample conditions and lack logical safeguards based on agricultural physical mechanisms when facing extreme weather, resulting in fragile prediction boundaries.
[0004] In recent years, although crop phenotypic non-destructive monitoring technology based on multi-source sensors has developed rapidly, enabling the high-precision extraction of crop structural, nutritional, and physiological activity characteristics, current map fusion and multimodal phenotyping technologies mostly remain at the monitoring and state assessment level, failing to effectively couple the real-time three-dimensional growth status of crops, canopy biomass distribution, and water transpiration mechanisms. Existing transpiration mechanism models generally have fixed parameters, and their crop coefficients are usually static empirical values, unable to adapt to the drastic changes in canopy topology and canopy spatial heterogeneity at different growth stages. This makes it difficult to form an effective closed-loop decision between multimodal phenotyping perception and refined irrigation control. Summary of the Invention
[0005] The purpose of this invention is to provide a greenhouse intelligent irrigation decision-making method that integrates dynamic monitoring of crop biomass. This method captures the macroscopic geometric and microscopic physiological characteristics of crops by integrating multi-source images including visible light, multispectral, and thermal infrared modules, achieving high-precision dynamic monitoring of individual crop biomass. Based on this, the invention deeply strengthens the algorithm model by introducing a cross-domain meteorological dataset for pre-training and initializing a lightweight time-series network. A locally modified Penman-Monteith Formula energy balance equation is used as the basis for physical information constraints, and the loss function of the time-series network is locally improved. Finally, the system combines a model predictive control strategy to output feedforward irrigation instructions based on a rolling 72-hour micro-environmental simulation sequence, achieving efficient and precise allocation of water resources.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] A smart irrigation decision-making method for greenhouses that integrates dynamic monitoring of crop biomass includes the following steps:
[0008] Step 1: Dynamic monitoring of crop biomass based on multi-source heterogeneous data fusion:
[0009] Step 1-1: Deploy a multimodal integrated imaging system that includes visible light, multispectral and thermal infrared imaging to simultaneously acquire multi-source images of crop canopy. Then, achieve pixel-level spatial registration and single-plant extraction of multi-source images through image denoising and masking.
[0010] Steps 1-2: Extract geometric structural features from visible light images, nutrient density features from multispectral images, and temperature and vitality features from thermal infrared images to construct a multidimensional phenotypic feature set. Specific steps are as follows: Extract the geometric and textural features of crop height and canopy coverage from visible light images; simultaneously, extract wide and narrow band vegetation indices from multispectral images; map a clean canopy mask onto the thermal infrared image to extract canopy temperature and normalized relative canopy temperature; perform feature filtering on the extracted features, including significance testing and the Borutas algorithm; the selected features are used to construct a single-plant biomass model.
[0011] Steps 1-3: Use feature selection algorithm to perform dimensionality reduction and optimization of multidimensional phenotypic feature set, and input the optimized feature subset into Gaussian process regression (GPR) network for training to establish a multi-source fusion absolute biomass estimation model and dynamically output single plant biomass;
[0012] Step 2: Evapotranspiration prediction based on physical constraint neural network mechanism reinforcement:
[0013] Step 2-1: In the data preparation stage, receive multi-dimensional environmental and phenotypic data as input features, including temperature, relative humidity, water vapor pressure deficit, photosynthetically active radiation, and real-time crop biomass obtained through multi-source monitoring. Construct a time-series sliding window with a fixed sequence length to provide a standardized data input format for subsequent model training.
[0014] Step 2-2: Before formal training, the Transformer network performs a pre-training task using a publicly available greenhouse dataset to learn the coupling relationship between meteorological variables and soil moisture.
[0015] Steps 2-3: In the local algorithm adaptation and training phase, the classic FAO-Penman-Monteith formula is localized for a greenhouse scenario, adjusting aerodynamic drag parameters to adapt to weak wind conditions. Subsequently, the modified mechanism equation is transformed into a hard physical constraint generator. For each batch of input samples, the prior model simultaneously derives the theoretical transpiration rate that conforms to the laws of thermodynamics. During network backpropagation, the total loss function is reconstructed as a weighted combination of data-driven loss and physical constraint loss.
[0016] Step 3: Rolling time-domain irrigation optimization control based on model predictive control:
[0017] After obtaining a high-precision dynamic evapotranspiration baseline, in order to achieve the mathematical transformation from evapotranspiration prediction to single irrigation volume, model predictive control is used to perform rolling time-domain feedforward irrigation optimization. The decision layer combines the microenvironment prediction network to perform rolling optimization on the rolling prediction sequence of temperature, relative humidity, and photosynthetically active radiation for the next 72 hours. The specific steps are as follows:
[0018] Step 3-1: Construct the state-space update equation:
[0019] Hours as the distance from the walk Taking into account the water-holding characteristics of the cultivation medium and the dynamic trends of weather evolution, a physical state evolution model is established between root zone water content and transpiration, deep seepage, and irrigation volume:
[0020]
[0021] In the formula, For the first The actual water content of the root zone at hours; Hourly evapotranspiration predicted by the fusion model; This represents the hourly deep seepage rate; The control variable is the amount of water used per irrigation.
[0022] Step 3-2: Set periodic operation constraints:
[0023] Based on the agronomical practice of irrigation every three days in actual greenhouse production, and to avoid the hardware constraint of frequent pump start-stop, the control cycle of physical execution is set to 72 hours, and the control variables are... Apply strict periodic pulse input constraints:
[0024]
[0025] In the formula, To maximize physical output irrigation, the optimizer only has decision-making power at the beginning of each 72-hour cycle (3 days), and the system relies on the physical water retention capacity of the cultivation medium for dynamic buffering at other times.
[0026] Step 3-3: Construct an optimization cost function covering multiple time domain steps:
[0027] Actively read the real physical state fed back by sensors at each decision-making moment. As the initial boundary conditions, the solution covers the next 72 hours (i.e., the prediction step size). The rolling optimization cost function for (=72) is:
[0028]
[0029] In the formula, This is a projected hourly water content trajectory for the future, based on the current state. The optimal moisture content reference threshold is used to ensure that tomatoes are not subjected to water stress. , , These are the weighted penalty coefficients for water tracking error, irrigation water consumption, and deep seepage, respectively.
[0030] Steps 3-4: Continuously output optimal control commands:
[0031] At the start of each decisionable control cycle (i.e. (At time), combined with the predicted water consumption trajectory, the above cost function is minimized through optimization algorithms such as sequential quadratic programming. Solve for the current optimal single irrigation amount When time progresses to the next decision moment (i.e. At any given moment, force a reread of the true physical value. As the new decision-making benchmark, the prediction window slides backward as a whole, and finally dynamically outputs feedforward irrigation instructions that take into account both the physiological safety of crop moisture content and the minimization of water consumption, thus completing the integrated and precise closed-valve regulation of water and fertilizer.
[0032] Compared with the prior art, the present invention has the following advantages:
[0033] (1) By integrating dynamic monitoring of crop biomass, this invention overcomes the signal saturation bottleneck of traditional single-phenotype perception. Traditional machine vision is prone to structural information extraction distortion in the middle and late stages of crop growth due to severe canopy occlusion. This invention employs a multimodal integrated imaging system that deeply integrates macroscopic geometric volume of visible light, biochemical nutrient density of multiple spectra, and transpiration metabolic activity characteristics of thermal infrared light. This multi-dimensional feature complementarity effectively compensates for the physical occlusion defects of single modalities, achieving high-precision, non-destructive dynamic monitoring of single-plant biomass throughout the entire growth period under small sample conditions, providing reliable physiological feedback on crop ontology for precision irrigation.
[0034] (2) The introduction of a physically constrained neural network enhances the mechanism and improves the generalization ability and boundary robustness of evapotranspiration water loss prediction. Traditional data-driven deep learning irrigation decision models are prone to overfitting under limited greenhouse samples. This invention innovatively introduces a multi-moisture gradient meteorological dataset to pre-train the time-series network. Based on this, the locally modified Penman-Monteith energy balance formula is transformed into a hard physical constraint and embedded into the total loss function of the deep neural network. This mechanism forces the model's parameter updates and final prediction output to be constrained by the laws of agricultural thermodynamics, suppressing prediction distortion under abnormal climate or sensor noise interference.
[0035] (3) A feedforward collaborative control system in the rolling time domain was constructed, realizing cost reduction and efficiency improvement in greenhouse water management. Existing greenhouse irrigation mostly relies on passive environmental threshold feedback or static expert experience, with the response lagging behind the actual peak water demand of crops. This invention utilizes a multi-model collaborative system, including microclimate time series extrapolation and water consumption prediction, to transform the microclimate evolution trajectory of the next 72 hours and the dynamic water consumption benchmark of crops into feedforward constraints, and solves the optimal cost function in the multi-step time domain through model predictive control. This strategy effectively reduces ineffective deep seepage water consumption, more accurately ensures the physiological needs of crops during key phenological periods, and improves the overall utilization efficiency of integrated water and fertilizer management. Attached Figure Description
[0036] Figure 1 A roadmap for automatic monitoring of greenhouse tomato biomass based on multi-source heterogeneous image feature fusion;
[0037] Figure 2 Architecture and system deployment diagram of evaporation prediction algorithm that integrates physical constraints and Transformer;
[0038] Figure 3 This is a comparison and verification figure showing the prediction accuracy of a dual-driven model that couples biomass characteristics with both mechanistic and data-driven approaches.
[0039] Figure 4 Verification chart for multi-step rolling prediction results of core microenvironmental parameters (temperature, humidity, light radiation) in greenhouses;
[0040] Figure 5 A comparison chart showing the implementation effects of intelligent optimization closed-loop decision-making and multi-benchmark strategy irrigation. Detailed Implementation
[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.
[0042] This invention provides a smart irrigation decision-making method for greenhouses that integrates dynamic monitoring of crop biomass, the method comprising the following steps:
[0043] Step 1: Dynamic monitoring of crop biomass based on multi-source heterogeneous data fusion:
[0044] This invention constructs a multi-source image fusion biomass estimation system for small sample conditions. The system deploys an integrated visual terminal that combines visible light, multispectral, and thermal infrared technologies to simultaneously acquire multi-source images of crop canopies.
[0045] In the multi-source image acquisition and preprocessing stage, a multimodal integrated imaging system is used to simultaneously acquire high-resolution visible light, multispectral, and thermal infrared images of the crop canopy. Subsequently, the system removes soil and background noise by using Laplacian operator-based image halo removal and relative depth map masking, achieving pixel-level spatial registration and pure single-plant extraction of multi-source images.
[0046] In the multidimensional phenotypic feature extraction stage, the system extracts "volume density-vibration" multidimensional phenotypic features from the registered images. Specifically, it extracts geometric and textural features such as crop height and canopy coverage from visible light images to characterize the three-dimensional "volume" of the plant. The formula for calculating canopy coverage is:
[0047]
[0048] In the formula, Canopy coverage; Indicates the image in coordinates The binarized mask pixel values at the location (1 for crop areas, 0 for background areas); and These represent the effective pixel width and height of the region of interest, respectively.
[0049] Simultaneously, broad and narrow band vegetation indices were extracted from multispectral images to analyze canopy chlorophyll content and nutrient "density"; a purified canopy mask was mapped onto thermal infrared images to extract the average canopy temperature and normalized relative canopy temperature (NRVT) to sensitively characterize crop transpiration metabolism "activity". Normalized relative canopy temperature (NRVT) The formula for calculating ) is:
[0050]
[0051] In the formula, The average temperature (°C) within the target canopy region; The highest temperature (°C) within the target canopy area; The lowest temperature (°C) within the target canopy area.
[0052] In the model building and local application phases, to address the collinearity problem in the extracted high-dimensional phenotypic feature sets, the Boruta feature selection algorithm is introduced for encapsulated dimensionality reduction, eliminating redundant noise and locking in the core feature subset. Faced with the constraints of limited labeled samples and the dynamic changes in the contribution of each modality feature at different growth stages in greenhouse scenarios, this invention inputs the optimized feature subset into a Gaussian process regression (GPR) network for nonlinear mapping training, establishing a multi-source fusion absolute biomass estimation model. In actual operation, the system receives real-time multi-source feature inputs, robustly predicts and outputs single-plant biomass, thus providing reliable data feedback for subsequent irrigation decisions based on ontological physiological water demand patterns. The overall roadmap is as follows: Figure 1 As shown.
[0053] Step 2: Evapotranspiration Irrigation Prediction Based on Physically Constrained Neural Network Mechanism Enhancement:
[0054] This invention proposes a framework for predicting transpiration irrigation based on physically constrained neural networks. The core of this framework lies in the deep integration of deep learning's temporal feature extraction capabilities with plant physiological mechanisms. Its complete system application and algorithm adaptation process is as follows: Figure 2 As shown. Combined with Figure 2 The specific structure of this prediction framework includes three core modules: a sliding window input and temporal coding layer, a pre-trained weighted multi-head self-attention mechanism layer, and a physical prior joint constraint decoding and parameter update layer. The specific application process is as follows:
[0055] During the data preparation phase, the system receives multi-dimensional environmental and phenotypic data as input features, including temperature (T), relative humidity (RH), vapor pressure deficit (VPD), photosynthetically active radiation (PAR), and real-time crop biomass obtained through multi-source monitoring. A time-series sliding window of fixed sequence length (L=12) is constructed to provide a standardized data input format for subsequent model training. Addressing the challenge of scarce labeled samples in greenhouse scenarios, the Transformer network performs pre-training using publicly available greenhouse datasets before formal training. This pre-training process learns the coupling relationship between meteorological variables and soil moisture. Through this pre-training process, the model's multi-head attention mechanism and feedforward network (FFN) parameters receive reasonable initial weights, effectively breaking down the barriers of inconsistent basic meteorological features across different data sources and laying a solid foundation for feature extraction priors for local adaptation.
[0056] During the local algorithm adaptation and training phase, this invention addresses the black-box limitations of purely data-driven models by providing them with physical common sense. For example... Figure 2 As shown in the "Loss Function Constraint (Physics-Informed Loss)" module below, the system first performs a localized correction to the classic FAO-Penman-Monteith formula for a greenhouse scenario (reconstructing the conversion relationship between dynamic drag and stomatal drag in the standard formula for the typical weak wind and limited airflow environment of a greenhouse). The corrected localized theoretical transpiration baseline (denoted as...) The calculation formula is as follows:
[0057]
[0058] In the formula, This is the revised theoretical transpiration baseline. Net radiation at the canopy surface; Soil heat flux; Air temperature; It is the saturated vapor pressure; This is the actual water vapor pressure, where This is a vapor pressure deficit (VPD). The slope of the saturated water vapor pressure-temperature curve; 1 is the dry and wet constant; constants 628 and 1.24 are the aerodynamic drag and pore drag conversion coefficients after correction for weak wind environment in greenhouse.
[0059] Subsequently, the modified mechanistic equation was transformed into a hard physical constraint generator. For each batch of input samples, the prior model simultaneously derived the theoretical evaporation rate (denoted as ) that conforms to the laws of thermodynamics. During network backpropagation, the total loss function is reconstructed as a weighted combination of data-driven loss and physical constraint loss:
[0060]
[0061] In the formula, This represents the total loss function value optimized during the backpropagation process of the neural network. Data-driven loss refers to the predicted evapotranspiration value output by the model. ) and the measured true value of transpiration ( The mean square error between ) These are the physical constraint weighting coefficients, used to balance the regularization strength between data-driven fitting and physical logic penalties; The physical constraint loss refers to the predicted evaporation value output by the model. ) and the theoretical reference quantity calculated from the above-mentioned modified thermodynamic mechanism formula ( The mean square error penalty term between ().
[0062] This method, which introduces physical constraint loss as a regularization term, reconstructs the total loss function, subjecting the network to strong physiological priors during parameter updates. This forces the predicted output to logically conform to the physiological water consumption patterns of plants. This mechanism effectively endows the model with agricultural physics common sense, thereby suppressing prediction anomalies under extreme weather conditions or sensor noise interference.
[0063] To verify the effectiveness of this framework, a rigorous comparative ablation experiment was designed. For example... Figure 3 As shown, in a purely meteorological driven model that does not incorporate biomass characteristics, the predictive determination coefficient (R²) is limited by the lack of perception of the crop's intrinsic state. 2 The coefficient of determination (R²) is only 0.9161. In contrast, the physical constraint model, which integrates meteorological and biomass-driven factors, has a much higher coefficient of determination (R²). 2 The accuracy was significantly improved to 0.9751, and the root mean square error was greatly reduced. This fully demonstrates that the triple mechanism of "pre-training weighting, physical regularization constraint, and multimodal feature fusion" constructed in this invention not only significantly improves the prediction accuracy of evapotranspiration water consumption, but also effectively enhances the boundary robustness of the model under extreme weather conditions, providing solid and reliable data support and physical constraint guarantee for subsequent precise irrigation decisions.
[0064] Step 3: Rolling time-domain irrigation optimization control based on model predictive control:
[0065] Obtaining a high-precision dynamic transpiration baseline is a prerequisite for achieving accurate feedforward control. This invention, based on the principle of water balance, constructs a high-precision transpiration acquisition system comprising a weighing module for substrate weight change, a weighing module for crop weight change, and a weighing module for inflow and outflow of liquid. This system continuously monitors changes in irrigation volume, outflow volume, substrate weight, and crop weight at a high-frequency rate down to the minute level, thereby accurately calculating hourly transpiration as the dynamic transpiration baseline. The formula for calculating transpiration is as follows:
[0066]
[0067] In the formula: The transpiration rate of the crop (g); Irrigation volume (g); The matrix weight (g); Discharge volume (g); The weight of the crop is in grams (g).
[0068] In practical applications, the benchmark data must meet strict specific requirements: First, it must be smoothed by a sliding window to remove weight spikes caused by agricultural operations such as leaf removal and harvesting; second, it must be combined with the operating status of the solenoid valve to strictly separate the weight disturbances caused by irrigation liquid inflow and deep drainage; finally, the true value sequence must be downsampled at the timestamp with the corresponding multidimensional meteorological and phenotypic features to ensure a high degree of consistency of the monitoring signal in time and space.
[0069] After obtaining a high-precision dynamic evapotranspiration baseline, this invention utilizes a microenvironment prediction model in conjunction with predictive control to perform feedforward irrigation optimization in the rolling time domain, such as... Figure 4As shown, the decision-making layer combines a microenvironment prediction network to construct an optimized cost function covering multiple time domains by generating rolling prediction sequences for temperature, relative humidity (RH), and photosynthetically active radiation (PAR) over the next 72 hours. By solving for the optimal control sequence of the system, the model output 72-hour predicted trajectory achieves high-precision fitting with the measured ground truth across all core microenvironment dimensions. Specifically: in temperature prediction, the model accurately captures the periodic fluctuations of the greenhouse temperature between approximately 12°C and 30°C during the day and night, with the slope changes of temperature rise and fall highly synchronized with the measured ground truth; in relative humidity prediction, the model accurately extrapolates the dynamic evolution from high humidity (approximately 80%–90%) at night to low humidity (approximately 35%–50%) during the day; and in photosynthetically active radiation (PAR) prediction, the model highly reproduces the peak profile of daytime light radiation of approximately 500–620 W / m². Although there were slight deviations between the predicted and measured values at local troughs in the latter half of the process, the overall maximum deviation was effectively constrained within a reasonable range, and the core trend was fully captured, providing reliable data support for subsequent feedforward irrigation decisions in the rolling time domain. Ultimately, the system dynamically outputs feedforward irrigation instructions that balance crop physiological safety moisture content with minimizing water consumption, achieving integrated and precise closed-valve control of water and fertilizer.
[0070] Example:
[0071] This invention was systematically validated in a solar greenhouse at the Babai Facility Agriculture Base in Inner Mongolia, targeting the entire spring crop growth period of tomato varieties. By coupling biomass sensing results obtained from multi-source heterogeneous data with a predictive control algorithm, the system underwent long-term practical operation testing. The actual performance of two fully automated data-driven decision groups (ML and TS) and two traditional control groups (PM and EXP) is as follows: Figure 5 As shown in the figure, this figure clearly presents the evolution trend of irrigation amount for each strategy group during the entire growth period of tomatoes from 35 to 125 days after transplanting, and superimposes the dynamic changes of crop biomass and microenvironment vapor pressure deficit (VPD).
[0072] The evolution of the actual implementation curves reveals distinct characteristics in the response of traditional decision-making strategies to greenhouse environment and crop status. The Expert Experience Group (EXP) exhibits a stepped, fixed irrigation pattern, failing to dynamically respond to the continuous growth rhythm of crops and high-frequency microclimate changes. The Traditional Mechanism Model Group (PM), lacking feedback constraints from the crop's own growth status (such as biomass accumulation), sees its irrigation volume highly dependent on and following short-term jumps in environmental parameters (such as VPD), resulting in frequent and dramatic fluctuations in the implementation curves.
[0073] In contrast, both the data-driven groups (ML and TS) incorporating algorithmic models exhibited higher fit to the actual water requirements of crops, but their regulatory logic and response characteristics differed significantly. The machine learning decision-making group (ML) focused on coupling the macroscopic physiological rhythms of steady crop biomass accumulation, resulting in a smooth and stable irrigation curve. During periods of high humidity and low transpiration (when VPD declines), it could restrainedly reduce irrigation, demonstrating robust regulatory characteristics based on crop ontological state feedback. The pure time series prediction group (TS), relying on rolling projections of the microenvironment (temperature, light, and humidity) for the next 72 hours, could keenly track high-frequency fluctuations in environmental factors. This manifested in the irrigation curve as timely following of short-term jumps in parameters such as VPD. Especially in the mid-to-late stages of crop growth (65 days after transplanting) facing complex meteorological disturbances, the TS group maintained high environmental response sensitivity, fully demonstrating the efficient adaptability of the forward-looking rolling optimization strategy to dynamic changes in the microenvironment.
[0074] Experimental results objectively demonstrate that the closed-loop decision-making model, which couples continuous crop status perception with dynamic algorithms, can more sensitively respond to microenvironmental changes while ensuring the crop's water needs in the mid-to-late stages. This control strategy, which integrates multimodal data feedback, effectively avoids the drawbacks of traditional mechanical watering and blindly following environmental fluctuations, significantly reduces ineffective water consumption, and improves the overall utilization efficiency of greenhouse water management, showing promising prospects for engineering implementation.
Claims
1. A greenhouse intelligent irrigation decision-making method integrating dynamic monitoring of crop biomass, characterized in that... The method includes the following steps: Step 1: Dynamic monitoring of crop biomass based on multi-source heterogeneous data fusion: Step 1-1: Deploy a multimodal integrated imaging system that includes visible light, multispectral and thermal infrared imaging to simultaneously acquire multi-source images of crop canopy. Then, achieve pixel-level spatial registration and single-plant extraction of multi-source images through image denoising and masking. Steps 1-2: Extract the geometric structure features from the visible light image, the nutrient density features from the multispectral image, and the temperature and vitality features from the thermal infrared image to construct a multidimensional phenotypic feature set; Steps 1-3: Use feature selection algorithm to perform dimensionality reduction and optimization of multidimensional phenotypic feature set, and input the optimized feature subset into Gaussian process regression network for training to establish a multi-source fusion absolute biomass estimation model and dynamically output single plant biomass. Step 2: Evapotranspiration prediction based on physical constraint neural network mechanism reinforcement: Step 2-1: In the data preparation stage, receive multi-dimensional environmental and phenotypic data as input features, including temperature, relative humidity, water vapor pressure deficit, photosynthetically active radiation, and real-time crop biomass obtained through multi-source monitoring. Construct a time-series sliding window with a fixed sequence length to provide a standardized data input format for subsequent model training. Step 2-2: Before formal training, the Transformer network performs a pre-training task using a publicly available greenhouse dataset to learn the coupling relationship between meteorological variables and soil moisture. Steps 2-3: In the local algorithm adaptation and training phase, the classic FAO-Penman-Monteith formula is localized for a greenhouse scenario, adjusting aerodynamic drag parameters to adapt to weak wind conditions. Subsequently, the modified mechanism equation is transformed into a hard physical constraint generator. For each batch of input samples, the prior model simultaneously derives the theoretical transpiration rate that conforms to the laws of thermodynamics. During network backpropagation, the total loss function is reconstructed as a weighted combination of data-driven loss and physical constraint loss. Step 3: Rolling time-domain irrigation optimization control based on model predictive control: After obtaining a high-precision dynamic evapotranspiration baseline, in order to realize the mathematical transformation of evapotranspiration prediction to single irrigation volume, the model predictive control is used to perform rolling time-domain feedforward irrigation optimization. The decision layer combines the microenvironment prediction network to perform rolling optimization on the rolling prediction sequence of temperature, relative humidity and photosynthetically active radiation in the next 72 hours.
2. The greenhouse intelligent irrigation decision-making method integrating dynamic monitoring of crop biomass as described in claim 1, characterized in that... The specific steps of steps 1-2 are as follows: extract the geometric and textural features of crop height and canopy coverage from the visible light image; simultaneously, extract the wide and narrow band vegetation indices from the multispectral image; map the pure canopy mask onto the thermal infrared image to extract the canopy temperature and normalized relative canopy temperature; perform feature screening on the extracted features, including significance testing and the Borutas algorithm, and use the screened features for constructing a single-plant biomass model.
3. The greenhouse intelligent irrigation decision-making method integrating dynamic monitoring of crop biomass as described in claim 2, characterized in that... The formula for calculating the canopy coverage is: In the formula, Canopy coverage; Indicates the image in coordinates The binary mask pixel value at the location is 1 for the crop area and 0 for the background area; and These represent the effective pixel width and height of the region of interest, respectively.
4. The greenhouse intelligent irrigation decision-making method integrating dynamic monitoring of crop biomass as described in claim 3, characterized in that... The normalized canopy relative temperature The calculation formula is: In the formula, The average temperature within the target canopy region; The highest temperature within the target canopy region; The lowest temperature within the target canopy region.
5. The greenhouse intelligent irrigation decision-making method integrating dynamic monitoring of crop biomass according to claim 1, characterized in that... The The calculation formula is as follows: In the formula, This is the revised theoretical transpiration baseline. Net radiation at the canopy surface; Soil heat flux; Air temperature; It is the saturated vapor pressure; This is the actual water vapor pressure, where This is due to a water vapor pressure deficit; The slope of the saturated water vapor pressure-temperature curve; 1 is the dry and wet constant; constants 628 and 1.24 are the aerodynamic drag and pore drag conversion coefficients after correction for weak wind environment in greenhouse.
6. The greenhouse intelligent irrigation decision-making method integrating dynamic monitoring of crop biomass according to claim 1, characterized in that... The expression for the total loss function is: In the formula, This represents the total loss function value optimized during the backpropagation process of the neural network. Data-driven loss; These are the physical constraint weighting coefficients; For physical constraint loss, The transpiration prediction value output by the model. This represents the true value of measured transpiration.
7. The greenhouse intelligent irrigation decision-making method integrating dynamic monitoring of crop biomass according to claim 1, characterized in that... The specific steps of step 3 are as follows: Step 3-1: Construct the state-space update equation: Hours as the distance from the walk Taking into account the water-holding characteristics of the cultivation medium and the dynamic trends of weather evolution, a physical state evolution model is established between root zone water content and transpiration, deep seepage, and irrigation volume: In the formula, For the first The actual water content of the root zone at hours; Hourly evapotranspiration predicted by the fusion model; This represents the hourly deep seepage rate; The control variable is the amount of water irrigated per irrigation. Step 3-2: Set periodic operation constraints: Based on the agronomical practice of irrigation every three days in actual greenhouse production, and to avoid the hardware constraint of frequent pump start-stop, the control cycle of physical execution is set to 72 hours, and the control variables are... Apply strict periodic pulse input constraints: In the formula, To maximize physical output irrigation, the optimizer only has decision-making power at the beginning of each 72-hour cycle, and the system relies on the physical water retention capacity of the cultivation medium for dynamic buffering at other times. Step 3-3: Construct an optimization cost function covering multiple time domain steps: Actively read the real physical state fed back by sensors at each decision-making moment. As the initial boundary condition, the solution covers the next 72 hours, i.e., the prediction step size. Rolling optimization cost function with a value of 72: In the formula, This is a projected hourly water content trajectory for the future, based on the current state. The optimal moisture content reference threshold is used to ensure that tomatoes are not subjected to water stress. , , These are the weighted penalty coefficients for water tracking error, irrigation water consumption, and deep seepage, respectively. Steps 3-4: Continuously output optimal control commands: At the start of each decisionable control cycle, i.e. At the same time, combining the predicted water consumption trajectory, the above cost function is minimized through optimization algorithms such as sequential quadratic programming. Solve for the current optimal single irrigation amount When time progresses to the next decision moment, that is At time 1, force a reread of the true physical value. As the new decision-making benchmark, the prediction window slides backward as a whole, and finally dynamically outputs feedforward irrigation instructions that take into account both the physiological safety of crop moisture content and the minimization of water consumption, thus completing the integrated and precise closed-valve regulation of water and fertilizer.
