Plant environment intelligent management method based on multi-sensor and image recognition and its application for popular science

Abstract

The application discloses a plant environment intelligent management method based on multi-sensor and image recognition and application thereof to science popularization, comprising the following steps: collecting plant images through a camera, recognizing growth stages by using a convolutional neural network, extracting stage-specific physiological characteristics such as leaf area, and storing sensitive factors in each stage; establishing a dynamic mapping relationship of physiological characteristics, correcting sensor data deviation, fusing data according to sensitive factor weights and dynamically adjusting sensor calibration weights; constructing a response model based on fused data and growth stages, calculating the matching degree of environmental parameters and optimal threshold values, and generating a step-by-step regulation scheme according to the priority of sensitive factor weights when the matching degree is lower than the threshold value; realizing precise, dynamic and adaptive management of the plant growth environment, ensuring that the environment regulation strategy is highly matched with the actual physiological needs of each growth stage of the plant, thereby improving the overall precision and efficiency of plant environment management, and providing precise and adaptive environmental protection for high-quality growth of the plant.

Description

Technical Field

[0001] This invention relates to an intelligent plant environment management method, and more particularly to an intelligent plant environment management method based on multi-sensor and image recognition, and its application in popular science. Background Technology

[0002] Plant growth is closely related to environmental parameters (such as light, temperature, humidity, soil nutrients, and CO2 concentration), and different growth stages (seed germination, seedling stage, flowering stage, etc.) have significantly specific environmental requirements. For example, seed germination depends on suitable soil moisture and temperature, while the flowering and fruiting stages are more sensitive to light duration and nutrient content. Therefore, plant environmental management requires both real-time monitoring of changes in plant physiological characteristics and dynamic adjustment of environmental parameters according to growth stages to meet their stage-specific needs, ultimately improving yield and quality.

[0003] Currently, plant environmental management methods mainly fall into three categories: First, those driven by human experience, where growers observe plant growth with the naked eye, judge environmental needs based on experience, and manually adjust them, such as watering and supplemental lighting. This approach is highly subjective and difficult to quantify stage-specific needs. Second, static control using single or multiple sensors, which collects data such as temperature, humidity, and light intensity through sensors and triggers control based on fixed thresholds, but does not incorporate dynamic calibration based on plant physiological characteristics. Third, basic model-assisted methods, which introduce simple models to associate environmental parameters with growth status. However, these models are mostly trained on fixed datasets and are not optimized for different growth stages.

[0004] However, existing methods have significant limitations and present a prominent contradiction with the need for precise plant environmental management: First, they lack stage specificity, failing to distinguish the physiological characteristics of different stages such as seed germination and flowering, and adopting a uniform control strategy, resulting in a mismatch between environmental parameters and the actual needs of plants; Second, sensor data bias is not corrected, ignoring the impact of plant physiological activities on the local microenvironment, resulting in low accuracy of raw data, and they are not dynamically calibrated in conjunction with real-time physiological characteristics; Third, the models have poor adaptability, as the environmental response models are not dynamically updated with the plant growth cycle, making it difficult to adapt to the growth characteristics of different varieties and seasons, leading to a decline in long-term management accuracy, resulting in resource waste or lagging regulation. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a plant environment intelligent management method based on multiple sensors and image recognition.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a plant environment intelligent management method based on multi-sensor and image recognition, comprising the following steps:

[0007] S1. Acquire plant images using a camera, identify growth stages based on a pre-trained convolutional neural network, extract stage-specific physiological features, and store the sensitive factors for each stage, where the total weight of the sensitive factors is... In the formula, For the first The weights of each sensitivity factor, This represents the total number of sensitive factors.

[0008] S2. Establish a dynamic mapping relationship of physiological characteristics, calculate the influence value of transpiration using the Penman-Monteith model, correct sensor data deviations in real time, and perform weighted fusion based on the weights of sensitivity factors. Dynamically adjust the sensor calibration weights; where, For the first The fusion weights of each environmental parameter, For the corrected sensor data, Adjust the calibration weights of each sensor to match the total number of environmental parameters; This is the result after weighted fusion;

[0009] S3. Based on the fused data and growth stage, construct a phased response model, calculate the matching degree between environmental parameters and the optimal threshold, and if the matching degree is lower than the threshold, generate a step-by-step control scheme according to the weight priority of sensitive factors.

[0010] In a preferred embodiment of the present invention, step S1, which involves identifying growth stages and extracting stage-specific physiological features using a convolutional neural network, includes: cropping, normalizing, and denoising plant images captured by a camera; inputting the processed images into a ResNet, AlexNet, or VGG model pre-trained through transfer learning, extracting deep features of the images through the model's convolutional layers, and classifying and identifying growth stages using fully connected layers; simultaneously, extracting pixel information of leaf and stem regions from the feature maps output by the convolutional layers using image segmentation algorithms, obtaining quantized values ​​of leaf area, plant height, and stem diameter through pixel counting and size calibration, and obtaining RGB or HSV color parameters of the leaves through color space conversion; when storing the sensitive factors for each growth stage, a database table is established with the growth stage as the index, containing the sensitive factor type, initial weight value, and threshold range corresponding to that stage, with the initial weight value determined through statistical analysis of historical growth data, and the threshold range set based on conclusions from plant physiological research.

[0011] In a preferred embodiment of the present invention, the growth stages include seed germination period, seedling period, vegetative growth period, flowering period, and fruiting period;

[0012] Among them, the stage-specific physiological characteristics of seed germination include radicle length, coleoptile height and germination rate, and the corresponding sensitive factors are soil moisture, soil temperature and oxygen content.

[0013] The stage-specific physiological characteristics of seedlings include cotyledon expansion, number of true leaves and stem base diameter, and the corresponding sensitive factors are light intensity, soil nitrogen content and air humidity.

[0014] The stage-specific physiological characteristics of the vegetative growth period include total leaf area, plant height and number of branches, and the corresponding sensitive factors are CO2 concentration, soil phosphorus and potassium content and diurnal temperature range.

[0015] The stage-specific physiological characteristics of the flowering period include the number of flower buds, flower diameter and flowering rate, and the corresponding sensitive factors are light duration, air temperature and relative humidity.

[0016] The stage-specific physiological characteristics during the fruiting period include the number of fruits, the volume of a single fruit, and the coloring degree of the fruit. The corresponding sensitive factors are soil moisture, calcium content, and light intensity.

[0017] In a preferred embodiment of the present invention, the process of establishing the dynamic mapping relationship includes: based on the growth stage identified in step S1, training a phased association model using historical data. The association model takes the environmental parameters surrounding the plant as input and the physiological characteristics of the plant corresponding to the growth stage as output. The model parameters are updated in real time using a sliding window algorithm so that the mapping relationship changes dynamically with plant growth. Then, based on the leaf area and stomatal conductance in the physiological characteristics of the current stage, the influence value of plant transpiration on the local microenvironment is calculated in combination with the Penman-Monteith transpiration model. The influence value is superimposed on the original sensor data to eliminate bias.

[0018] In a preferred embodiment of the present invention, when data fusion is performed based on plant sensitivity factors at different growth stages, a weighted average algorithm is used, and the fusion weight of each environmental parameter is consistent with the weight of the corresponding sensitivity factor at the current stage. When adjusting the sensor calibration weight, the deviation value between the fused data and the theoretical environmental parameters predicted by the dynamic mapping relationship is calculated. When the deviation value of a parameter corresponding to a certain sensor exceeds the threshold range, the calibration weight is reduced according to the deviation ratio, and the reduced weight value is distributed to the relevant sensors according to the weight ratio of other sensitivity factors, thus forming a dynamic weight adjustment mechanism.

[0019] In a preferred embodiment of the present invention, the process of constructing the response model includes:

[0020] S301. Determine the input and output variables. The input variables are the calibrated and fused environmental data in step S2 and the growth stage labels identified in step S1. The output variables are the predicted values ​​of plant physiological characteristics under the corresponding growth stage.

[0021] S302. A phased modeling strategy is adopted, and sub-models are constructed for different stages of the plant. Each sub-model adopts an improved LSTM neural network structure and strengthens the influence weight of the sensitive factors of the current stage on physiological characteristics by adding an attention mechanism.

[0022] S303. Using time-series samples from historical monitoring data as the training set, and employing root mean square error as the loss function, the network parameters are iteratively updated using the Adam optimizer until the prediction error on the validation set is less than 5%.

[0023] S304. Integrate the sub-models of each stage, automatically call the corresponding sub-model through the growth stage label, and combine the sliding time window to incorporate newly collected environmental and physiological data in real time, and dynamically update the model parameters to adapt to the nonlinear changes in the plant growth process.

[0024] In a preferred embodiment of the present invention, the calculation process of the matching degree includes:

[0025] S311. Based on the conclusions of plant physiology research and historical best growth data, determine the threshold range of each environmental parameter required for the optimal physiological state at the corresponding growth stage, and store the threshold range as a standard reference value.

[0026] S312. Output the predicted values ​​of plant physiological characteristics under the current environmental parameters through the response model, and calculate the deviation rate between the current environmental parameters and the standard reference value. If the current value is within the threshold range, the absolute value of the deviation rate is taken. If it exceeds the range, a penalty coefficient is added on the absolute value.

[0027] S313. Combining the weights of the sensitive factors in the current growth stage in step S1, the deviation rates of each environmental parameter are weighted and summed to obtain the matching sub-items at the environmental parameter level; the similarity between the physiological characteristic values ​​predicted by the response model and the standard values ​​of the optimal physiological characteristics at this stage is calculated to obtain the matching sub-items at the physiological characteristic level; the matching sub-items at the two levels are weighted and fused proportionally to obtain the overall matching degree.

[0028] In a preferred embodiment of the present invention, the process of generating the environmental control scheme includes:

[0029] S321. Sort the environmental parameters to be regulated according to the weight priority of the sensitive factors in step S1, and prioritize the regulation of the parameter corresponding to the sensitive factor with the highest weight.

[0030] S322. Based on the difference between the current environmental parameters and the optimal threshold and the physiological characteristic change trend predicted by the response model, calculate the adjustment amount of each parameter.

[0031] S323. Combine the maximum adjustment capability of the auxiliary device to limit the adjustment amount, generate an executable step-by-step control command, and gradually approach the optimal threshold in stages.

[0032] In a preferred embodiment of the present invention, the method further includes step S4, which involves classifying and storing the original image data, sensor data, control schemes, and physiological characteristic change curves according to growth stages, and using time series analysis algorithms to mine the correlation between environmental parameter fluctuations and physiological characteristic mutations within the same growth stage; adjusting the initial weights of sensitive factors for each growth stage based on newly added data each month, and updating the training sample set of the response model at the same time.

[0033] In a preferred embodiment of the present invention, an application of a plant environment intelligent management method based on multi-sensor and image recognition for popular science is also provided. Based on the aforementioned plant environment intelligent management method, the method includes: acquiring and processing plant images via a camera to generate real-time dynamic animations of growth stages and a health dashboard; synchronously displaying environmental parameters and labeling sensitive factors of growth stages through a visual interface using multi-sensor data; designing interactive experiments, manually adjusting virtual environmental parameters, and providing real-time feedback on changes in plant physiological characteristics and sensor data correction processes; constructing a growth simulator using a predictive model, and designing empirical log activities for teenagers, transforming technical principles into visual experiences, interactive operations, and empirical analysis, intuitively conveying the scientific understanding of the dynamic adaptation of plant growth to the environment.

[0034] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0035] (1) By combining convolutional neural networks and image segmentation technology, the plant growth stages are accurately identified and stage-specific physiological characteristics are extracted, solving the problem of "one-size-fits-all" management in traditional management. Sensitive factors are stored through growth stage indexing to ensure that environmental regulation matches the real-time needs of plants.

[0036] (2) Based on the Penman-Monteith transpiration model and stage physiological characteristics, the environmental data deviation caused by plant physiological activities of the sensor is corrected in real time; the data of multiple sensors are dynamically fused by the weight of sensitive factors, and the sensor calibration weight is adjusted according to the deviation ratio, which solves the problems of unreliable single sensor data and the disconnect between environmental parameters and the actual microenvironment of plants.

[0037] (3) An improved LSTM neural network and attention mechanism are used to construct a response model in stages and strengthen the influence weight of sensitive factors in the current stage. The difference between the current environment and the optimal state is quantified by matching degree calculation. A step-by-step control scheme is generated according to the priority of sensitive factors to avoid over-control and improve resource utilization efficiency.

[0038] (4) By storing data throughout the entire cycle and analyzing time series, we can explore the correlation between environmental fluctuations and physiological mutations. We can adjust the weight of sensitive factors and update the model training set every month, so that the response model can adapt to the growth characteristics of different varieties and seasons. This solves the problems of static and poor adaptability of traditional models and ensures long-term management accuracy. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a preferred embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0043] It should be noted that image recognition technology acquires plant images through cameras. After preprocessing such as cropping, normalization, and noise reduction, it uses pre-trained convolutional neural networks like ResNet and AlexNet, trained through transfer learning, to accurately classify growth stages (recognizing the entire cycle from seed germination to fruiting). Combined with image segmentation algorithms, it extracts quantitative physiological characteristics such as leaf area, plant height, and stem diameter, as well as detailed information like RGB / HSV color parameters. This process not only achieves a "visual interpretation" of plant growth status but also provides a "plant-centric" basis for subsequent management by extracting stage-specific physiological characteristics. For example, identifying the cotyledon unfolding degree and the number of true leaves in the seedling stage clarifies the high sensitivity of this stage to light intensity and soil nitrogen content, providing direction for sensor data processing.

[0044] Multi-sensor data processing constructs a "precise sensing network" for environmental information. To address potential biases in raw sensor data, the method uses a dynamic mapping relationship (combined with the Penman-Monteith transpiration model to calculate the impact of plant transpiration on the microenvironment) to correct these biases in real time, eliminating local environmental disturbances caused by plant physiological activities. Then, based on growth stage sensitivity factors identified through image recognition, a weighted average algorithm is used to fuse multi-source environmental data (e.g., focusing on CO2 concentration and soil phosphorus and potassium content data during the vegetative growth stage). A dynamic weight adjustment mechanism is used to proportionally reduce the weight of a sensor when its data deviation exceeds the limit and allocate it to other reliable sensors, ensuring the adaptability and reliability of the data fusion.

[0045] The synergy between the two forms a closed loop of "physiological characteristics - environmental data - regulation strategy": image recognition provides a stage-specific benchmark for sensor data processing, while the data processed by multiple sensors provides high-quality input for response model construction. Ultimately, this enables the environmental regulation scheme to not only accurately match the physiological needs of the plant at its current growth stage, but also dynamically adapt to environmental changes, achieving an upgrade from "experience-driven" to "data and image dual-drive".

[0046] like Figure 1 The method for intelligent plant environment management based on multi-sensor and image recognition includes the following steps:

[0047] S1. Acquire plant images via camera, perform image cropping, size normalization, and noise reduction preprocessing; input the processed images into a ResNet, AlexNet, or VGG model pre-trained through transfer learning, extract deep features using convolutional layers, and classify and identify growth stages (including seed germination, seedling, vegetative growth, flowering, and fruiting stages) using fully connected layers; simultaneously, combine image segmentation algorithms to extract leaf and stem region pixel information from the convolutional layer feature maps, and quantify the leaf area as... (in The pixel-to-actual area conversion coefficient is obtained through size calibration experiments. (This refers to the number of pixels in the leaf area); the RGB parameters of the leaf are obtained through color space conversion, and the formula for calculating the brightness feature value is: (in , , These are the pixel values ​​for the red, green, and blue channels, ranging from 0 to 255. The leaf area, plant height, and stem diameter are obtained by converting pixel counts and size calibration. The RGB or HSV color parameters of the leaves are obtained by color space conversion, which are used as the stage-specific physiological characteristics of each growth stage.

[0048] Among them, the stage-specific physiological characteristics of seed germination include radicle length, coleoptile height, and germination rate. In the formula, The number of germinating seeds, The total number of seeds tested;

[0049] The corresponding sensitivity factor weights satisfy In the formula, Soil moisture weighting, Soil temperature weighting, Weighted by oxygen content;

[0050] The stage-specific physiological characteristics of seedlings include cotyledon spread, number of true leaves, and stem base diameter, with corresponding sensitivity factor weights satisfying [the specified criteria]. In the formula, Weighted by light intensity. As a weight for soil nitrogen content, Weighted by air humidity;

[0051] Stage-specific physiological characteristics during the vegetative growth period include total leaf area, plant height, and number of branches, with corresponding sensitive factor weights satisfying [the following criteria]. In the formula, Weighted by CO2 concentration, Assigning weights to soil phosphorus and potassium content. Weighted by diurnal temperature range;

[0052] Stage-specific physiological characteristics during the flowering period include the number of flower buds, flower diameter, and flowering rate. In the formula, The number of open flowers, The total number of flower buds; the corresponding sensitivity factor weights satisfy... In the formula, Weighted by duration of illumination, Weighted by air temperature, Weighted by relative humidity;

[0053] The stage-specific physiological characteristics of the fruiting period include fruit quantity, single fruit volume, and fruit coloring intensity, with corresponding sensitivity factor weights satisfying the following conditions: In the formula, Soil moisture weight, Weighted by calcium content. Light intensity is used as the weight; a database table is created with growth stage as the index to store the sensitivity factors corresponding to each stage.

[0054] The specific sensitive factors are as follows: during seed germination, soil moisture, soil temperature, and oxygen content; during seedling stage, light intensity, soil nitrogen content, and air humidity; during vegetative growth stage, CO2 concentration, soil phosphorus and potassium content, and diurnal temperature range; during flowering stage, light duration, air temperature, and relative humidity; and during fruiting stage, soil moisture, calcium content, and light intensity. The table includes the type of sensitive factor, the initial weight values ​​determined through statistical analysis of historical growth data, and the threshold ranges set based on plant physiological research conclusions.

[0055] S2. Establish a dynamic mapping relationship of stage-specific physiological characteristics: Based on the growth stages identified in S1, train a stage-specific association model using historical data. In the formula, For physiological feature vectors, For environmental parameter vectors, Labels for growth stages. For the first The parameter set of the stage model; the sliding window algorithm is used to update the model parameters in real time, and the parameter update formula is as follows: In the formula, This is a smoothing coefficient, with a value ranging from 0.6 to 0.8. These are the parameters from the previous time step; (New parameters estimated within the current window) allow the mapping relationship to change dynamically with plant growth; based on leaf area and stomatal conductance in the current physiological characteristics, the impact of transpiration on the local microenvironment is calculated using the Penman-Monteith transpiration model:

[0056] In the formula, The slope of the saturated water vapor pressure curve is expressed in kPa / ℃. Net radiation from crop canopy, in MJ / (m²・d); Soil heat flux, in MJ / (m²・d); The constant of the wet / dry surface is expressed in kPa / ℃. The average daily temperature is expressed in degrees Celsius (°C). Wind speed at a height of 2m, in m / s; The pressure is the saturated vapor pressure, in kPa. This is the actual water vapor pressure, in kPa. This influencing value is superimposed on the original sensor data to eliminate bias; the corrected data is... In the formula, The stage-specific correction coefficient (valued from 0.1 to 0.3 depending on the growth stage) is added to the original sensor data to eliminate bias.

[0057] Data fusion was performed based on plant sensitivity factors at different growth stages, using the following formula: In the formula, For the first The weights of each sensitive factor; For the first Correction data from each sensor; The number of sensors is specified. A weighted average algorithm is used, with the fusion weight of each environmental parameter consistent with the weight of the corresponding sensitivity factor at the current stage. Sensor calibration weights are adjusted by calculating the deviation between the fused data and the theoretical environmental parameters predicted by the dynamic mapping relationship. In the formula, These are the theoretical environmental parameters predicted by the model; when ( When the preset deviation threshold is met, the calibration weight is reduced proportionally to the deviation, and the new weight is... In the formula, The weighting adjustment factor ranges from 0.2 to 0.5; simultaneously, the reduced weight values ​​are distributed to the relevant sensors according to the weight ratio of other sensitivity factors, using the following distribution formula: In the formula, For the first New weights are assigned to relevant sensors to form a dynamic weight adjustment mechanism.

[0058] S3. The input variables are the calibrated and fused environmental data from S2 and the growth stage labels identified in S1. The output variables are the predicted values ​​of plant physiological characteristics at the corresponding growth stage. A staged modeling strategy is adopted, constructing sub-models for different growth stages. Each sub-model uses an improved LSTM neural network structure, and an attention mechanism is added to strengthen the influence weight of sensitive factors on physiological characteristics at the current stage. The attention score is calculated using the following formula: In the formula, For the first The attention scores for each sensitive factor are calculated through a fully connected layer. For the first The attention weights of each sensitive factor satisfy the following conditions: .

[0059] Using time-series samples from historical monitoring data as the training set, and employing root mean square error as the loss function, the loss function is: In the formula, The number of samples; These are the physiological characteristic values ​​predicted by the model; The actual physiological characteristics are used; the network parameters are iteratively updated through the Adam optimizer until the prediction error on the validation set is less than 5%; the sub-models of each stage are integrated, and the corresponding sub-models are automatically called through the growth stage labels, and the newly collected environmental and physiological data are incorporated in real time by combining the sliding time window, and the model parameters are dynamically updated to adapt to the nonlinear changes in plant growth.

[0060] Based on the findings of plant physiology research and historical best growth data, the threshold ranges of various environmental parameters required for the optimal physiological state at the corresponding growth stage were determined. The predicted values ​​of plant physiological characteristics under the current environmental parameters are output through the response model. The deviation rate between the current environmental parameters and the standard reference value is calculated (if the current value is within the threshold range, the absolute value is taken; if it exceeds the range, a penalty coefficient is added to the absolute value). Combining the sensitivity factor weights of the current growth stage in S1, the deviation rates of each environmental parameter are weighted and summed to obtain the environmental parameter level matching sub-item. The similarity between the physiological characteristic value predicted by the response model and the standard value of the optimal physiological characteristic at this stage is calculated to obtain the physiological characteristic level matching sub-item. The two sub-items are weighted and fused proportionally to obtain the overall matching degree.

[0061] If the matching degree is lower than the preset threshold, the environmental parameters to be regulated are sorted according to the weight priority of the sensitive factors in S1, and the parameter with the highest weight is regulated first; the adjustment amount of each parameter is calculated based on the difference between the current environmental parameter and the optimal threshold and the physiological characteristic change trend predicted by the response model; the adjustment amount is limited by the maximum adjustment capacity of the auxiliary device to generate a step-by-step control command, which gradually approaches the optimal threshold in stages.

[0062] S4. Store raw image data, sensor data, control schemes, and physiological characteristic change curves according to growth stages; use time series analysis algorithms to mine the correlation between environmental parameter fluctuations and physiological characteristic mutations within the same growth stage; adjust the initial weights of sensitive factors for each growth stage monthly based on newly added data, using the following formula: In the formula, The updated weights; The current weight; The optimal weights are estimated based on the newly added data; The correction coefficient is set to 0.3; the training sample set of the response model is updated simultaneously to ensure that the model is adaptable to the growth characteristics of plants in different seasons and varieties.

[0063] By recognizing plant transpiration characteristics (such as the opening and closing of leaf stomata and the degree of leaf curling) through images, and establishing a dynamic calibration model with data such as temperature, humidity, light, and CO2 concentration collected by multiple sensors, errors caused by sensor drift due to environmental fluctuations (such as sudden changes in temperature and humidity) can be corrected in real time, ensuring the accuracy of input data.

[0064] Growth stage-specific fusion systems require data fusion based on plant sensitivity factors at different growth stages (e.g., cotyledon stage is sensitive to water, true leaf stage is sensitive to light). If the input data has been calibrated as described above, "incorrect sensitivity factor identification" caused by sensor errors can be avoided, allowing for more accurate allocation of stage-specific weights (e.g., after calibration, light data from the true leaf stage can more accurately determine whether supplemental lighting is needed).

[0065] The growth stage-specific fusion system determines the current growth stage of a plant through image recognition and outputs the core physiological characteristics of that stage (such as low transpiration rate during the cotyledon stage and significantly increased transpiration rate during the true leaf stage).

[0066] The dynamic calibration model can dynamically adjust calibration parameters based on the specific physiological characteristics of these stages: for example, during the cotyledon stage, transpiration is weak and the image recognition shows that the stomata of the leaves are mainly closed. At this time, the calibration weight of the temperature and humidity sensors can be reduced (because the feedback of transpiration to the environment is weak); during the true leaf stage, transpiration is strong and the stomata of the leaves are wide open. The image features are strongly correlated with temperature, humidity and light. At this time, the calibration weight of the corresponding sensors can be increased, so that the calibration is more in line with the physiological laws of the current stage and avoids insufficient accuracy caused by uniform calibration parameters to be adapted to all stages.

[0067] Furthermore, the growth stages are divided into the following five stages according to the physiological development process of plants, and the definitions and parameter settings for each stage are as follows:

[0068] (1) Seed germination period

[0069] Specific physiological characteristics: radicle breaking through seed coat length ≥2mm (pixel-level measurement of seed longitudinal section using image segmentation algorithm), coleoptile elongation rate ≥0.5mm / day (morphological analysis based on time series images).

[0070] Sensitive factors:

[0071] ① Soil moisture (60-70% field capacity, based on the 3rd edition of the "Crop Water Requirements Atlas")

[0072] ② Soil temperature (18~25℃, based on data from the 2023 germination experiment of temperature-sensitive crops conducted by the Chinese Academy of Agricultural Sciences)

[0073] ③Oxygen content (≥15% by volume, refer to FAO soil aeration standards).

[0074] (2) Seedling stage

[0075] Specific physiological characteristics: cotyledon fully unfolded angle ≥120° (using Hough transform to detect leaf edge contour), number of true leaves ≥2 (based on YOLOv8 leaf detection model counting), stem base diameter ≥3mm (calculated by fitting lidar point cloud data).

[0076] Sensitive factors:

[0077] ① Light intensity (10,000~15,000 Lux, recommended value from the "Manual of Light Management for Facility Agriculture")

[0078] ② Soil nitrogen content (available nitrogen 80~120mg / kg, according to NY / T1121.24~2018 testing standard)

[0079] ③Air humidity (65~75%RH, based on the optimization results of the plant stomatal conductance model).

[0080] (3) Vegetative growth period

[0081] Specific physiological characteristics: total leaf area ≥ 500 cm² (calculated using U~Net++ semantic segmentation model), daily plant height increase ≥ 5 cm (using background difference method combined with scale calibration), number of branches ≥ 3 (based on graph theory stem connection point detection).

[0082] Sensitive factors:

[0083] ①CO2 concentration (800~1200ppm, based on data from the IPCC study on the benefits of greenhouse crops).

[0084] ② Soil phosphorus and potassium content (available phosphorus 20~40mg / kg, available potassium 150~250mg / kg, according to the secondary index of the "Soil Nutrient Classification Standard").

[0085] ③ Diurnal temperature range (8~12℃, referring to the thermodynamic model of photosynthetic product accumulation).

[0086] (4) Flowering period

[0087] Specific physiological characteristics: flower bud color development rate ≥80% (statistically obtained through HSV color space threshold segmentation), flower diameter ≥5cm (measured using ellipse fitting algorithm), and flowering rate ≥30% (based on time series flower detection model).

[0088] Sensitive factors:

[0089] ① Photoperiod duration (12~14h / d, based on data from a study on photoperiod regulation of flowering gene expression).

[0090] ②Air temperature (20~28℃, set according to pollen viability temperature response curve).

[0091] ③Relative humidity (50~60%RH, refer to the correlation test between corolla unfolding rate and humidity).

[0092] (5) Result period

[0093] Specific physiological characteristics: Fruit coloring Lab* value ≥ (50, 30, 20) (calibrated by color matching algorithm), single fruit volume ≥ 200 cm³ (calculated using three-dimensional point cloud reconstruction technology), and abscisic acid content in the fruit stalk ≥ 0.5 μg / g (referenced from the fruit ripening biochemical index database).

[0094] Sensitive factors:

[0095] ① Soil moisture (70~80%FC, optimized based on root water absorption kinetics model).

[0096] ② Calcium content (exchangeable calcium 800~1200mg / kg, refer to the nutritional standards for fruit cracking prevention).

[0097] ③ Light intensity (8000~10000 Lux, based on data from studies on the transport efficiency of photosynthetic products).

[0098] Based on the above, this paper proposes an application of a multi-sensor and image recognition-based intelligent plant environment management method in the field of science popularization. Through technological visualization and interactive experiences, the public can intuitively understand the scientific principles of plant growth and environmental regulation. Plant images collected by cameras can be processed to generate real-time dynamic displays of growth stages. For example, on the screen of a science popularization exhibition hall, the length of the radicle during seed germination and the number of true leaves during the seedling stage will be presented in animation as they change over time. The RGB color parameters of the leaves can also be converted into a "plant health dashboard," allowing visitors to intuitively see the correlation between leaf color depth and light and nutrients. Environmental data collected by multiple sensors can be displayed synchronously through a visual interface, showing real-time values ​​of parameters such as soil moisture and CO2 concentration, and annotating sensitive factors at the current growth stage. For example, during the vegetative growth stage, the interface will highlight parameters such as "CO2 concentration" and "soil phosphorus and potassium content," accompanied by text explaining that "at this time, plant leaves grow rapidly, and these elements are like 'nutritional meals,'" allowing the public to understand the "differences in needs" of plants at different stages.

[0099] Furthermore, the dynamic calibration and data fusion process of this technology can be designed as an interactive experiment: visitors can manually adjust the light intensity or soil moisture in the virtual environment, and the system will provide real-time feedback on changes in plant physiological characteristics (such as slower or faster leaf area growth) through image recognition, while demonstrating how the sensor data is corrected. For example, when simulating "enhanced transpiration," the interface will display the difference between the original and corrected data from the temperature and humidity sensors, explaining that "plant activity affects the surrounding microenvironment, just as sweating makes the surroundings humid." The predictive function of the response model can be transformed into a "plant growth simulator." By inputting different combinations of environmental parameters, the system will predict the future plant height, flowering rate, etc., based on phased sub-models, allowing the public to understand through experimentation that "precise regulation can make plants grow better."

[0100] For science popularization among teenagers, the data storage and analysis mechanisms in technology can be combined to design a "Plant Growth Log" activity: students collect images and environmental data of campus plants using simple equipment. After the system categorizes and organizes the data according to growth stages, it generates physiological characteristic change curves and environmental parameter fluctuation graphs, guiding students to discover patterns such as "insufficient light duration during the flowering period leads to a decrease in flowering rate." This transforms abstract plant physiology knowledge into observable and analyzable empirical data. This approach, which breaks down complex technology into visual experiences, interactive experiments, and empirical investigations, not only allows the public to understand the scientific logic of intelligent management but also conveys the core understanding that "plant growth requires dynamic adaptation to the environment," achieving a dual popularization of technological principles and scientific knowledge.

[0101] Furthermore, the intelligent plant environment management method based on multi-sensor and image recognition also includes: utilizing multimodal sensors, including RGB cameras, infrared cameras, temperature and humidity sensors, soil moisture sensors, light intensity sensors, and CO2 concentration sensors, to collect comprehensive data on garden plants. Preprocessing operations are performed on data from different sources, including but not limited to data cleaning, noise filtering, standardization and normalization, and time alignment. Convolutional neural networks are used to extract morphological features of garden plants from RGB images, such as leaf shape, quantity, color changes, and the thickness and curvature of branches. Time series analysis methods are used to extract environmental change features from environmental data such as temperature, humidity, light intensity, and CO2 concentration. Weighted, stacked, or cascaded strategies are employed to fuse features from different modalities into a unified feature representation, thereby training an AI model to accurately assess the growth status of garden plants. Furthermore, the data fusion strategy and weights can be dynamically adjusted based on real-time collected data and the AI ​​model's evaluation results to achieve accurate monitoring and long-term tracking of garden plant growth status.

[0102] Based on data collected from various garden plant species and growth stages, as well as data analyzed by multimodal sensors and AI models, intelligent decision-making rules for the growth environment of garden plants are constructed. A control model is established based on the physical characteristics and environmental data of the garden, formulating specific regulation strategies for factors such as light, temperature, humidity, and soil nutrients in the environment of the garden plants. For example, when it is detected that garden plants in a certain area are in flowering stage and the light intensity is below a suitable threshold, the system automatically controls the retraction of shading facilities or the activation of supplemental lighting; if the soil moisture is below a set range, the system automatically starts the irrigation system to provide appropriate water. During the execution of regulation, the latest environmental data is acquired in real time through the sensor network and compared with the desired environmental state for the plant at that growth stage. Once a deviation exists between the actual state and the desired state, the system can automatically adjust the control strategy based on an adaptive algorithm, executing new regulatory actions to ensure that the garden plants are always in the most suitable growth environment, promoting their healthy growth and achieving the desired landscape effect.

[0103] In another embodiment, the rose, a common plant in gardens, is selected as the application object. Combined with a plant environment intelligent management technology solution based on multi-sensor and image recognition, precise management of the entire growth cycle is carried out.

[0104] First, high-definition industrial cameras and multi-sensor arrays are deployed in the rose planting area. The cameras are installed 1.5 meters above the planting area and collect images of the roses every 2 hours. At the same time, sensors for soil moisture, soil temperature, light intensity, air temperature and humidity, CO2 concentration and soil nutrients (nitrogen, phosphorus, potassium and calcium) are activated and a set of environmental data is collected every hour to ensure that the data covers the entire growth period of the roses.

[0105] The rose images captured by the camera were preprocessed: irrelevant garden background areas were cropped, the image size was normalized to 224×224 pixels to fit the model input, and Gaussian filtering was used to eliminate noise interference such as leaf reflection. The processed images were then input into a ResNet-50 model pre-trained through transfer learning. The model's convolutional layers extracted deep features such as leaf texture and bud morphology, and the fully connected layers completed the classification of growth stages. Based on the growth characteristics of roses, the original "seed germination period" was adjusted to the "rooting period of cuttings" which is more in line with the garden seedling cultivation scenario. The remaining stages, including seedling stage, vegetative growth stage, flowering stage, and fruiting stage (rose fruit development stage), were retained. Simultaneously, the U-Net++ image segmentation algorithm was used to segment the root zone of cuttings, the leaf zone of seedlings, and the stem and flower bud zone of mature plants from the feature map output by the convolutional layer. Pixel counting combined with actual size calibration (pre-measuring leaves of known size with a ruler and recording the corresponding pixel count, calculating the pixel-to-actual-size conversion coefficient k = 0.02cm² / pixel) was used to convert the root length (≥3cm for normal rooting), callus area (≥1cm² for healthy state) of cuttings during the rooting stage, the number of true leaves (≥3) and stem base diameter (≥4mm) during the seedling stage, the total leaf area (≥800cm²) and plant height (≥30cm) during the vegetative growth stage, the number of flower buds (≥5) and flower diameter (≥6cm) during the flowering stage, and the number of rose fruits and single fruit volume (≥10cm³) during the fruiting stage. Finally, the brightness value V of the leaves was extracted through RGB color space conversion. max(R,G,B) / 255 (V≥0.6 is the leaf health threshold) is used as an auxiliary physiological indicator.A database table was then created using growth stages as the index to store the sensitive factors for each stage: For the rooting stage of cuttings, the sensitive factors were soil moisture (65%~75% field capacity), soil temperature (20~25℃), and oxygen content (≥18% by volume), with initial weights of 0.4, 0.35, and 0.25, respectively; for the seedling stage, the sensitive factors were light intensity (12000~16000 Lux), soil nitrogen content (available nitrogen 90~130 mg / kg), and air humidity (60%~70% RH), with weights of 0.4, 0.3, and 0.3; and for the vegetative growth stage, the sensitive factors were CO2 concentration (900~1300 ppm) and soil phosphorus and potassium content (available phosphorus 25~45 mg / kg). The weights for the following parameters were: available potassium (160-260 mg / kg), diurnal temperature range (10-14℃), and light intensity (9000-11000 Lux), with weights of 0.35, 0.4, and 0.25, respectively; for flowering period, light duration (13-15 h / d), air temperature (22-26℃), and relative humidity (55%-65%RH), with weights of 0.45, 0.3, and 0.25, respectively; and for fruiting period, soil moisture (75%-85% field capacity), calcium content (exchangeable calcium 900-1300 mg / kg), and light intensity (9000-11000 Lux), with weights of 0.4, 0.3, and 0.3, respectively. All threshold ranges were set based on the conclusions of rose cultivation physiology research and the historical optimal growth data of garden roses.

[0106] Next, a dynamic mapping relationship between the physiological characteristics of roses and environmental parameters is established: based on the identified growth stages, a stage-specific correlation model is trained using environmental-physiological data of garden roses from the past 3 years. The model input consists of environmental parameter vectors such as soil temperature and humidity and light, and the output consists of physiological characteristic vectors such as root length and leaf area for the corresponding stage. The model parameters are updated in real time using a sliding window algorithm (window size set to 7 days, smoothing coefficient α = 0.7) to ensure that the mapping relationship is dynamically adjusted with the growth of roses. Meanwhile, based on the total leaf area at the current stage (e.g., 800 cm² during the vegetative growth period) and stomatal conductance (calculated by leaf RGB lightness values, with lightness V = 0.7 corresponding to stomatal conductance of 0.5 mol / (m²·s)), the Penman-Monteith model is used to calculate the impact of rose transpiration on the local microenvironment. For example, the transpiration rate ET0 = 4.2 mm / d is calculated during the summer vegetative growth period. This value is multiplied by the stage correction coefficient β = 0.25 (the transpiration intensity of rose leaves is moderate, so β ​​is taken as the median value) to obtain a correction amount of 1.05 mm / d. This correction is then added to the original data from the soil moisture sensor to eliminate the bias of low local soil moisture caused by transpiration. The corrected data e'_j = original data + 1.05 mm / d. Subsequently, data fusion is performed based on the weights of sensitive factors at each stage, using a weighted average algorithm. For example, when fusing CO2 concentration and soil phosphorus and potassium content data during the vegetative growth stage, the fusion value e_fusion = 0.35 × CO2 correction data + 0.4 × (phosphorus data + potassium data) / 2 + 0.25 × diurnal temperature difference correction data. If the data collected by a soil phosphorus sensor deviates from the theoretical phosphorus content predicted by the model by δ = 12 mg / kg, exceeding the preset threshold δ_threshold = 8 mg / kg, its calibration weight is reduced proportionally to the deviation. The new weight w'_i = original weight 0.2 × (1 - 0.3 × 12 / 8) = 0.11. The reduced weight of 0.09 is then allocated to the CO2 sensor (0.09 × 0.35 / (0.35 + 0.25) = 0.0525) and the diurnal temperature difference sensor (0.09 × 0.25 / 0.6 = 0.0375) according to the weight ratio of other sensitive factors, forming a dynamic weight adjustment mechanism.

[0107] A phased response model for roses was constructed based on fused data and growth stages. The input variables were fused environmental data and growth stage labels (such as the "flowering period" label), and the output variables were predicted values ​​of physiological characteristics such as the number of flower buds and the flowering rate. An improved LSTM sub-model was constructed for each rose stage. An attention mechanism was added to the flowering period sub-model, and the attention score s_i = 0.8 (higher than other factors) for light duration was calculated through a fully connected layer. The attention weight a_i = exp(0.8) / (exp(0.8) + exp(0.3) + exp(0.2)) = 0.62, strengthening the influence of light duration on the flowering rate. Using time-series monitoring data of garden roses from the past three years (a total of 1200 samples) as the training set, the root mean square error was used as the loss function. The network parameters were iteratively updated using the Adam optimizer. After 150 training rounds, the prediction error on the validation set decreased to 4.2% (below 5%), meeting the accuracy requirements. The sub-models of each stage are integrated, and the corresponding sub-model is automatically called through the growth stage label. For example, after the rose enters the flowering period, the flowering period sub-model is automatically started. The latest collected light and temperature data are incorporated in combination with the sliding time window (window size 5 days) to dynamically update the model parameters to adapt to nonlinear changes such as the sudden reduction of light during the rainy season.

[0108] When calculating the matching degree between environmental parameters and optimal thresholds, the optimal reference values ​​for each stage are first determined, such as the optimal light duration of 14h / d, air temperature of 24℃, and relative humidity of 60%RH during the flowering period. The response model outputs a predicted flowering rate of 65% for the rose under the current environment. Simultaneously, the deviation rate between the current light duration of 12h / d and the optimal value is calculated to be 14.3% (within the threshold, the absolute value is taken), and the deviation rate for air temperature of 28℃ is 16.7% (exceeding the threshold, a penalty coefficient of 1.2 is added, making the deviation rate 20.04%). Combining the weights of the flowering period sensitivity factors, the matching sub-item at the environmental parameter level = 0.45×14.3% + 0.3×20.04% + 0.25×5% = 12.46%. The similarity between the predicted flowering rate of 65% and the optimal flowering rate of 85% is calculated to be 76.47%, meaning the matching sub-item at the physiological characteristic level = 76.47%. The two sub-items are then weighted and merged in a 4:6 ratio, resulting in an overall matching degree = 0.4 × 12.46% + 0.6 × 76.47% = 49.06%, which is lower than the preset threshold of 60%, requiring the generation of a control plan. Prioritizing sensitive factors by weight, the light duration is controlled first (weight 0.45 is the highest). Based on the 2-hour difference between the current 12 hours of light and the optimal 14 hours, and considering the trend predicted by the response model that "for every 1 hour of light increase, the flowering rate increases by 8%", the adjustment amount is calculated to be 2 hours. Considering that the maximum adjustment capacity of the garden supplemental lighting equipment is 1.5 hours / day (to avoid scorching leaves with strong light), after limiting the adjustment amount, a step-by-step control instruction is generated: On the first day, the supplemental lighting duration is increased from 0 to 1 hour, bringing the total light duration to 13 hours; on the third day, it is increased by another 0.5 hours, bringing the total light duration to 13.5 hours; on the fifth day, it is increased to 1.5 hours, bringing the total light duration to 14 hours, gradually approaching the optimal threshold.

[0109] Finally, the original images of roses, sensor data, control schemes, and physiological characteristic change curves are stored according to growth stages. Time series analysis algorithms are used to mine patterns. For example, it was found that when the CO2 concentration is below 900ppm during the vegetative growth stage, the daily leaf growth rate decreases from 0.8cm to 0.3cm. The initial weights of sensitive factors are adjusted monthly based on 30 new sets of data. For example, the weight of calcium element during the fruiting period is adjusted from 0.3 to 0.32 (because calcium deficiency is found to lead to an increase in the cracking rate of rose fruits). At the same time, the training sample set of the response model is updated to ensure that the model adapts to the temperature and light changes of the garden in all four seasons. In terms of popular science applications, a visual screen is set up in the garden visitor center to play dynamic animations of rose growth stages in real time (such as the process of flower buds gradually opening during the flowering period) and a health dashboard (the RGB brightness values ​​of the leaves correspond to the three-color labels of "healthy-normal-warning"). It also displays data from multiple sensors and labels the sensitive factors at the current stage. An interactive experimental device is designed so that after visitors manually adjust the virtual lighting duration, the screen provides real-time feedback on changes in the rose flowering rate and the sensor data correction process. For example, when the lighting is dimmed, it displays "Insufficient light, the sensor detected that the leaf stomata are closed, and the air humidity has increased by 0.5% after correction". A "Rose Growth Log" activity is designed for teenagers, providing simple image acquisition equipment and soil moisture detection paper to guide students to record the correlation between changes in rose leaf area and environmental parameters, intuitively conveying the scientific knowledge of plant growth and dynamic adaptation to the environment.

[0110] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for intelligent plant environment management based on multi-sensor and image recognition, characterized in that, Includes the following steps: S1. Acquire plant images using a camera, identify growth stages based on a pre-trained convolutional neural network, extract stage-specific physiological features, and store the sensitive factors for each stage, where the total weight of the sensitive factors is... ; S2. Establish a dynamic mapping relationship of physiological characteristics, calculate the influence value of transpiration using the Penman-Monteith model, correct sensor data deviations in real time, and perform weighted fusion based on the weights of sensitivity factors. Dynamically adjust sensor calibration weights; S3. Based on the fused data and growth stage, construct a phased response model, calculate the matching degree between environmental parameters and the optimal threshold, and if the matching degree is lower than the threshold, generate a step-by-step control scheme according to the weight priority of sensitive factors.

2. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: In step S1, the process of identifying growth stages and extracting stage-specific physiological features through a convolutional neural network includes: cropping, normalizing, and denoising plant images captured by a camera; inputting the processed images into a ResNet, AlexNet, or VGG model pre-trained through transfer learning, extracting deep features of the images through the model's convolutional layers, and classifying and identifying growth stages using fully connected layers; simultaneously, extracting pixel information of leaf and stem regions from the feature maps output by the convolutional layers using image segmentation algorithms, obtaining quantized values ​​of leaf area, plant height, and stem diameter through pixel counting and size calibration, and obtaining RGB or HSV color parameters of the leaves through color space conversion; when storing the sensitive factors for each growth stage, a database table is established with the growth stage as the index, containing the sensitive factor type, initial weight value, and threshold range corresponding to that stage. The initial weight value is determined through statistical analysis of historical growth data, and the threshold range is set based on the conclusions of plant physiological research.

3. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: The growth stages include seed germination, seedling stage, vegetative growth stage, flowering stage, and fruiting stage; Among them, the stage-specific physiological characteristics of seed germination include radicle length, coleoptile height and germination rate, and the corresponding sensitive factors are soil moisture, soil temperature and oxygen content. The stage-specific physiological characteristics of seedlings include cotyledon expansion, number of true leaves and stem base diameter, and the corresponding sensitive factors are light intensity, soil nitrogen content and air humidity. The stage-specific physiological characteristics of the vegetative growth period include total leaf area, plant height and number of branches, and the corresponding sensitive factors are CO2 concentration, soil phosphorus and potassium content and diurnal temperature range. The stage-specific physiological characteristics of the flowering period include the number of flower buds, flower diameter and flowering rate, and the corresponding sensitive factors are light duration, air temperature and relative humidity. The stage-specific physiological characteristics during the fruiting period include the number of fruits, the volume of a single fruit, and the coloring degree of the fruit. The corresponding sensitive factors are soil moisture, calcium content, and light intensity.

4. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: The process of establishing the dynamic mapping relationship includes: based on the growth stage identified in step S1, training a phased association model using historical data. The association model takes the environmental parameters around the plant as input and the physiological characteristics of the plant corresponding to the growth stage as output. The model parameters are updated in real time using a sliding window algorithm so that the mapping relationship changes dynamically with plant growth. Then, based on the leaf area and stomatal conductance in the physiological characteristics of the current stage, the Penman-Monteith transpiration model is used to calculate the impact value of plant transpiration on the local microenvironment. The impact value is then superimposed on the original sensor data to eliminate bias.

5. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: When fusing data based on plant sensitivity factors at different growth stages, a weighted average algorithm is used, with the fusion weight of each environmental parameter consistent with the weight of the corresponding sensitivity factor at the current stage. When adjusting sensor calibration weights, the deviation between the fused data and the theoretical environmental parameters predicted by the dynamic mapping relationship is calculated. When the deviation of a parameter corresponding to a certain sensor exceeds the threshold range, the calibration weight is reduced proportionally to the deviation. At the same time, the reduced weight is distributed to the relevant sensors according to the weight ratio of other sensitivity factors, forming a dynamic weight adjustment mechanism.

6. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: The process of constructing the response model includes: S301. Determine the input and output variables. The input variables are the calibrated and fused environmental data in step S2 and the growth stage labels identified in step S1. The output variables are the predicted values ​​of plant physiological characteristics under the corresponding growth stage. S302. A phased modeling strategy is adopted, and sub-models are constructed for different stages of the plant. Each sub-model adopts an improved LSTM neural network structure and strengthens the influence weight of the sensitive factors of the current stage on physiological characteristics by adding an attention mechanism. S303. Using time-series samples from historical monitoring data as the training set, and employing root mean square error as the loss function, the network parameters are iteratively updated using the Adam optimizer until the prediction error on the validation set is less than 5%. S304. Integrate the sub-models of each stage, automatically call the corresponding sub-model through the growth stage label, and combine the sliding time window to incorporate newly collected environmental and physiological data in real time, and dynamically update the model parameters to adapt to the nonlinear changes in the plant growth process.

7. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: The matching degree calculation process includes: S311. Based on the conclusions of plant physiology research and historical best growth data, determine the threshold range of each environmental parameter required for the optimal physiological state at the corresponding growth stage, and store the threshold range as a standard reference value. S312. Output the predicted values ​​of plant physiological characteristics under the current environmental parameters through the response model, and calculate the deviation rate between the current environmental parameters and the standard reference value. If the current value is within the threshold range, the absolute value of the deviation rate is taken. If it exceeds the range, a penalty coefficient is added on the absolute value. S313. Combining the weights of the sensitive factors in the current growth stage in step S1, the deviation rates of each environmental parameter are weighted and summed to obtain the matching sub-items at the environmental parameter level; the similarity between the physiological characteristic values ​​predicted by the response model and the standard values ​​of the optimal physiological characteristics at this stage is calculated to obtain the matching sub-items at the physiological characteristic level; the matching sub-items at the two levels are weighted and fused proportionally to obtain the overall matching degree.

8. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: The process of generating the environmental control scheme includes: S321. Sort the environmental parameters to be regulated according to the weight priority of the sensitive factors in step S1, and prioritize the regulation of the parameter corresponding to the sensitive factor with the highest weight. S322. Based on the difference between the current environmental parameters and the optimal threshold and the physiological characteristic change trend predicted by the response model, calculate the adjustment amount of each parameter. S323. Combine the maximum adjustment capability of the auxiliary device to limit the adjustment amount, generate an executable step-by-step control command, and gradually approach the optimal threshold in stages.

9. The intelligent plant environment management method based on multi-sensor and image recognition according to claim 1, characterized in that: It also includes step S4, which involves classifying and storing the original image data, sensor data, control schemes, and physiological characteristic change curves according to growth stages, and using time series analysis algorithms to mine the correlation between environmental parameter fluctuations and physiological characteristic mutations within the same growth stage; and adjusting the initial weights of sensitive factors for each growth stage monthly based on newly added data, while updating the training sample set of the response model.

10. The application of a plant environment intelligent management method based on multi-sensor and image recognition for popular science, based on the plant environment intelligent management method according to any one of claims 1 to 9, is characterized in that, include: Plant images are captured and processed by a camera to generate real-time dynamic animations of growth stages and a health dashboard; environmental parameters are displayed synchronously through a visual interface, and sensitive factors of growth stages are labeled. Design an interactive experiment to manually adjust virtual environment parameters and then provide real-time feedback on changes in plant physiological characteristics and sensor data correction process. By combining predictive models to build a growth simulator and designing empirical log activities for teenagers, the technical principles are transformed into visual experiences, interactive operations and empirical analysis, intuitively conveying the scientific understanding of plant growth and dynamic adaptation to the environment.

Images