Irrigation decision-making method under visual identification of crops and multi-modal fusion of internet of things

Abstract

The application discloses an irrigation decision-making method under visual identification of crops and multi-modal fusion of Internet of Things, and belongs to the technical field of agricultural irrigation control. The method comprises the following steps: real-time access and timestamp alignment of multi-source data, access of Internet of Things environment monitoring data and visual identification result data; data preprocessing and abnormal cleaning; double-path parallel calculation of basic water demand indexes, path one is based on theoretical water demand calculated on the basis of soil-meteorological data of the Internet of Things, and path two is based on visual identification results to calculate a water stress correction coefficient and a growth correction coefficient; multi-modal fusion to generate comprehensive actual water demand; multi-level logic rule verification, including soil threshold hard constraint, rainfall prediction constraint, time period constraint, stress priority logic judgment and device pipe network constraint; output of final irrigation decision-making instruction; and closed-loop feedback self-optimization. The application takes the real physiological water shortage state of crops as the core basis, solves the misjudgment problem caused by traditional single soil humidity judgment, and realizes precise irrigation on demand.

Description

Technical Field

[0001] This invention relates to the field of intelligent irrigation control technology in agriculture, and in particular to an irrigation decision-making method based on the fusion of visual recognition and Internet of Things multimodal approaches. Background Technology

[0002] Existing agricultural irrigation control schemes have the following key technical deficiencies: Defect 1: The irrigation decision data source is singular, resulting in serious blind spots for misjudgment.

[0003] Currently, most mainstream automatic irrigation systems rely solely on soil moisture sensors for judgment, setting fixed lower and upper thresholds. Irrigation begins when soil moisture content falls below the lower threshold and stops when it exceeds the upper threshold. This system implicitly assumes a key fact: soil moisture content can completely and equivalently represent the crop's water requirement. However, in actual production, this assumption often fails, leading to the following two typical scenarios of misjudgment: Misjudgment Scenario A (Hidden Water Shortage): Soil moisture data is within the normal range (e.g., 60%–70% of field capacity), but the crop is actually in a state of water shortage and wilting due to root diseases, soil salinity stress (excessive EC inhibits root water absorption), or high transpiration rate caused by hot and low humidity weather. In this case, if only soil moisture data is used, the system will judge "no irrigation required," causing the crop to continue to be stressed and reduce yield.

[0004] Misjudgment Scenario B (Over-irrigation): Soil moisture data is below the lower threshold for irrigation, but the crop is growing vigorously, with dark green leaves and normal new shoot growth, indicating that the crop's root system is vigorous and can absorb enough water from the deeper soil layers. The low surface soil moisture content does not pose an actual stress to the crop. If irrigation is started blindly at this time, it will instead cause deep water accumulation and root rot due to lack of oxygen.

[0005] Defect 2: Irrigation decision-making lacks a feedback loop, making it impossible to continuously improve accuracy.

[0006] Traditional irrigation systems stop once irrigation is complete, without reassessing crop recovery or soil moisture changes to determine if the irrigation amount was too high or too low. The system lacks a "self-correction" mechanism, and its accuracy does not improve over time.

[0007] Defect 3: It does not take into account the spatial heterogeneity of crop growth and the differences in phenological stages.

[0008] Differences in soil fertility, drainage conditions, and crop density across different areas within the same plot lead to uneven spatial distribution of crop growth and water requirements. Traditional single-point sensor solutions cannot achieve differentiated management and can only implement uniform irrigation across the entire plot. Furthermore, the water requirements of crops differ significantly at different phenological stages (germination, vegetative growth, flowering and fruiting, and maturity), and traditional solutions cannot adaptively switch irrigation modes.

[0009] Defect 4: It does not take into account weather forecasts, and irrigation decisions lack foresight.

[0010] Traditional methods rely solely on current soil moisture data, without considering future rainfall forecasts or evaporation trends. This often results in situations where "it rains right after irrigation," leading to water waste and deep seepage. Summary of the Invention

[0011] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0012] In view of the problems existing in the prior art, the present invention is proposed.

[0013] Specifically, this invention provides the following technical solution: an irrigation decision-making method based on visual crop recognition and IoT multimodal fusion, comprising the following steps: Step S1, real-time access and timestamp alignment of multi-source data: The decision engine accesses raw data from two data sources according to a fixed polling cycle, including IoT environmental monitoring data and visual recognition result data; the IoT environmental monitoring data includes soil parameters, meteorological parameters, and equipment feedback parameters; the visual recognition result data refers to structured index data output after local inference calculation of crop canopy and leaf images captured by a camera using a lightweight model deployed on an edge computing device, rather than the original images; all data are aligned with a unified timestamp after access. Alignment ensures that environmental and visual data at the same time can participate in subsequent joint calculations; Step S2, data preprocessing and anomaly cleaning: quality verification and preprocessing are performed on the multi-source data accessed in Step S1, including removing abnormal jump values, completing missing data, filtering high-frequency noise, marking and excluding low-quality images, and normalizing data in each dimension; Step S3, dual-path parallel calculation of basic water demand indicators: two independent calculation paths are executed simultaneously. Path one is based on soil and meteorological parameters in IoT environmental data, calculating the reference crop evapotranspiration, theoretical crop evapotranspiration, and soil moisture deficit to obtain the theoretical soil-meteorological water demand; Path two is based on visual... The system identifies crop water stress level and leaf area index in the identified data, generating water stress correction coefficients and growth correction coefficients. Step S4 involves multimodal fusion to generate a comprehensive actual water requirement: the basic crop coefficient calculated in path one of step S3 is multiplied by the growth correction coefficient calculated in path two to obtain a dynamic crop coefficient. This dynamic crop coefficient is then multiplied sequentially by the reference crop evapotranspiration and the water stress correction coefficient to obtain the corrected actual crop evapotranspiration. Finally, the soil water deficit is added to the product of the corrected actual crop evapotranspiration and the planned irrigation interval days, and then the predicted effective rainfall is subtracted to obtain a comprehensive irrigation quota. This comprehensive irrigation quota simultaneously includes soil water deficit, future crop water consumption, and the predicted... Measure rainfall deduction amount; Step S5, multi-level logical rule verification: For the comprehensive irrigation quota calculated in step S4, perform multi-level constraint verification in order of priority from high to low, including soil threshold hard constraint verification, rainfall prediction constraint verification, time period constraint verification, stress priority logical judgment verification, and equipment and pipeline constraint verification. If any level of verification fails, the decision engine blocks or corrects the current irrigation decision; Step S6, output final irrigation decision instruction: After all verifications in step S5 pass, the decision engine outputs a structured irrigation decision instruction, which includes the decision conclusion, target irrigation amount, total irrigation duration, valve opening setting value, irrigation start and stop time, multi-plot rotation irrigation scheme, and accompanying alarm information;Step S7, Closed-Loop Feedback Self-Optimization: After irrigation is completed, the system automatically enters the feedback correction phase. After a preset water balance time window, soil moisture data and crop canopy images of the irrigated area are re-collected. By comparing the deviation between the actual re-measured data and the theoretically expected data, the system automatically fine-tunes the key parameters in the decision-making model and periodically updates the classification threshold of the visual recognition model using long-term accumulated historical data, forming a complete closed-loop iterative optimization.

[0014] Preferably, the visual recognition result data in step S1 includes the following structured indicators: water stress level, which is a discrete level from 0 to 4, where level 0 indicates that the crop is normal and has no stress, level 1 indicates mild water shortage, level 2 indicates moderate water shortage, level 3 indicates severe water shortage, and level 4 indicates wilting; CWSI crop water stress index, which is a continuous numerical index that quantifies the degree of water shortage of crops by analyzing the temperature difference between canopy temperature and air temperature, with a value range of 0 to 1, where a larger value indicates more severe water stress; and L-leaf surface... The canopy area index, defined as the ratio of the total area of ​​all leaves per unit land area to the land area, is calculated by performing pixel-level segmentation of the canopy image using the improved YOLO11-Seg lightweight semantic segmentation model; the NDVI growth index assesses vegetation vitality by analyzing the crop's reflectance characteristics to red and near-infrared light; canopy coverage refers to the proportion of green vegetation projection area to the total image area when viewed vertically, directly output by the segmentation model; the proportion of curled leaf area and the proportion of yellow leaf area are used to help verify the accuracy of water stress level judgment.

[0015] Preferably, the data preprocessing and anomaly cleaning in step S2 specifically includes: Sub-step S2-1, removing anomalous jump values: maintaining a sliding window for each sensor data channel, calculating the mean and standard deviation of valid data points within the window, and when the deviation of a newly accessed data point from the window mean exceeds three times the standard deviation, it is judged as an anomalous jump value and discarded; Sub-step S2-2, linear interpolation completion of missing values: when the sensor temporarily loses data due to communication interruption or power failure, a linear interpolation method is used to complete the data at the missing time point; when the missing time exceeds a preset duration or the consecutive missing data points exceed a preset proportion of the total data volume, it is determined that the current data of the sensor is unreliable, and the irrigation decision for this area is downgraded to be based solely on visual data and meteorological data. Data; Sub-step S2-3, Moving average filtering and noise reduction: The moving average filtering algorithm is used to smooth the sensor time series data, and the length of the moving window is dynamically adjusted according to the parameter change rate; Sub-step S2-4, Image quality judgment: The images captured by the camera are automatically evaluated for quality. The evaluation indicators include average brightness, contrast and occlusion ratio. Low-quality images are marked as invalid and do not participate in this irrigation decision. The decision engine uses the previous valid visual data as a substitute and marks the data timeliness; Sub-step S2-5, Data normalization: All data from different sensors are uniformly mapped to the range of 0 to 1 using the Min-Max normalization method to eliminate the influence of different units and numerical ranges on the fusion calculation.

[0016] Preferably, path one in step S3 specifically includes: sub-step S3-1-1, calculating the reference crop evapotranspiration using the FAO Penman-Monteith formula based on meteorological parameters such as air temperature, relative humidity, wind speed, and solar radiation, wherein the reference crop evapotranspiration reflects the intensity of atmospheric evaporation demand; sub-step S3-1-2, determining the basic crop coefficient based on crop type and current growth stage, and multiplying the basic crop coefficient by the reference crop evapotranspiration to obtain the theoretical crop evapotranspiration; sub-step S3-1-3, subtracting the measured volumetric water content from the field water holding capacity, and then multiplying by the planned wetting layer depth and soil bulk density to obtain the soil water deficit, wherein a value greater than zero indicates that the soil is currently in a water-deficient state, and the larger the value, the greater the water deficit.

[0017] Preferably, path two in step S3 specifically includes: sub-step S3-2-1, obtaining the water stress level and CWSI crop water stress index from the visual recognition results, and generating a water stress correction coefficient based on the water stress level: when the stress level is 0, the water stress correction coefficient is 1.0, and the original water requirement remains unchanged; when the stress level is 1, the water stress correction coefficient is between 1.1 and 1.2, and the supplementary water amount is appropriately increased; when the stress level is 2, the water stress correction coefficient is between 1.3 and 1.5, and the supplementary water amount is increased; when the stress level is 3 or above, the water stress correction coefficient is between 1.6 and 2.0, and emergency full irrigation is performed; in the same Within each level, the specific value of the water stress correction coefficient is calculated by linear interpolation based on the continuous values ​​of the CWSI crop water stress index; in sub-step S3-2-2, the L leaf area index in the visual recognition result is obtained, and this L leaf area index is compared with the standard L leaf area index of the crop under the current growth stage conditions to calculate the growth correction coefficient; when the L leaf area index is higher than the standard value, the growth correction coefficient is greater than 1, and the crop coefficient is adjusted upward to match the vigorous transpiration demand; when the L leaf area index is lower than the standard value, the growth correction coefficient is less than 1, and the crop coefficient is adjusted downward to prevent over-irrigation; the value range of the growth correction coefficient is limited to between 0.5 and 1.5.

[0018] Preferably, the stress priority logic judgment and verification in step S5 specifically includes: when there is a contradiction between the visually recognized crop water stress level and the soil moisture data, the visually recognized crop water stress level shall be the highest priority judgment basis; when the soil moisture data is within the normal range, but the visually recognized stress level is moderate or above, it is determined to be a latent physiological water shortage state, indicating that the crop may actually be in a water shortage state due to root diseases, excessive soil salinity, or excessive transpiration caused by high temperature and low humidity. At this time, a small amount of supplementary irrigation shall be performed, and the irrigation amount shall be performed according to the preset proportion of the normal calculation amount; when the soil moisture data is lower than the lower limit threshold of irrigation, but the visually recognized stress is not detected and the crop growth is dark green and the leaf area index is not lower than the preset proportion of the standard value, it is determined to be a compensatory drought resistance state of the crop, indicating that the crop roots can absorb water from the deep soil not covered by the sensor and have not been subjected to actual stress. At this time, delayed irrigation or only a small amount of maintenance water shall be selected.

[0019] Preferably, the soil threshold hard constraint verification in step S5 specifically includes: determining whether the current soil moisture content is higher than the upper limit of the field suitable moisture content for the current growth stage of the crop. If it is higher than the upper limit, irrigation is prohibited regardless of whether a slight stress level is visually detected to prevent waterlogging damage. The time period constraint verification specifically includes: determining whether the current time period is the high temperature midday period or the low temperature dew period at night. If so, the start of new irrigation is restricted. However, when severe water shortage stress is visually detected and the soil moisture content is lower than the wilting coefficient, the emergency irrigation mode is triggered, and irrigation is started immediately, overriding the time period constraint. The equipment and pipeline constraint verification specifically includes: verifying the maximum flow rate of the irrigation pump, the pressure bearing capacity of the pipeline, and the zone rotation irrigation capacity limit. When the single irrigation quota exceeds the upper limit of the equipment capacity, the current irrigation is automatically split into multiple rotation irrigations.

[0020] Preferably, step S7 specifically includes: sub-step S7-1, within the preset water balance time window after irrigation is completed, reread the soil moisture sensor data of each layer of the irrigation zone to obtain the actual water content after irrigation, and at the same time re-capture the crop canopy image and input it into the model to obtain the stress level and leaf area index after irrigation; sub-step S7-2, compare the actual soil water content after irrigation with the theoretical expected water content to calculate the deviation, and at the same time compare the visual stress level after irrigation with the stress level before irrigation to assess the degree of crop stress relief; sub-step S7-3, when a systematic deviation in irrigation volume is judged multiple times in a row, automatically adjust the mapping relationship parameters between the stress correction coefficient and the stress level and the calculation formula parameters of the growth correction coefficient corresponding to the plot; sub-step S7-4, record and store the data of the entire process of each irrigation decision in the cloud platform historical database, and when the accumulated data reaches a preset scale, use the accumulated data to iteratively update and optimize the classification threshold of the visual recognition model to form a complete closed-loop iterative mechanism.

[0021] Preferably, the system further includes a step of visually recognizing and linking crop growth with irrigation mode control: continuously monitoring the crop's growth status through a visual recognition model to identify the key phenological stages of the crop, including the germination stage, vegetative growth stage, flowering and fruiting stage, and maturity and harvest stage; when the crop is detected to be in the germination stage, which requires less water and needs precise supply to the roots, the system decides to activate the drip irrigation mode and adjusts the drip irrigation flow by controlling the opening of the electric ball valve; when the crop is detected to be in the vegetative growth stage, the system decides to adopt either drip irrigation or micro-sprinkler mode; when the crop is detected to be in the flowering and fruiting stage, which requires the greatest water and needs to moisten the canopy. To assist in cooling and humidifying, the system decides to switch to sprinkler irrigation mode and increase the opening of the electric ball valve; when the crop is detected to be in the ripening and harvesting stage, the system decides to switch back to drip irrigation mode to reduce leaf surface humidity and prevent fruit cracking and diseases; the phenological results of visual recognition and the irrigation mode decision instructions are simultaneously uploaded to the IoT cloud platform. After comprehensive verification and calculation by the cloud platform combined with soil temperature and humidity and meteorological data, the final precise control instructions are generated and issued to the execution layer; when there is a difference between the cloud calculation results and the edge decision, the cloud and edge joint arbitration mechanism is activated, and the arbitration result is pushed to the management personnel's mobile APP for final confirmation by humans.

[0022] This invention provides an irrigation decision-making method based on the fusion of visual crop recognition and Internet of Things (IoT) multimodal approaches, which has the following beneficial effects: By fusing visually recognized crop physiological stress status with IoT environmental data through multimodal feature integration, and using the crop's actual physiological water requirement as the core judgment criterion, the two key misjudgment problems of "normal soil data but crop already having implicit water shortage" and "soil being dry but crop not lacking water" are fundamentally solved.

[0023] The system employs a dual-track independent calculation approach, combining the soil-meteorological theoretical pathway and the visual physiological pathway. In the event of data discrepancies, the visual stress level is given the highest priority, allowing the system to "see" the true state of the crop, rather than simply "measuring" soil data.

[0024] Breaking through the limitations of traditional fixed-growth-period crop coefficients, this method dynamically corrects the crop coefficient K based on real-time visual observation of leaf area index (L) and canopy growth. c This allows for precise matching of evapotranspiration calculations to the actual growth status of crops, rather than rigidly relying on calendar tables.

[0025] After irrigation, soil moisture and crop images are automatically remeasured. The decision model parameters are fine-tuned in reverse by analyzing the deviation between the actual and expected results, forming a complete closed loop of "perception-computation-decision-execution-remeasurement-optimization". Irrigation accuracy continues to improve over time. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 The overall flowchart of the irrigation decision-making method provided by the present invention is shown.

[0027] Figure 2 The flowchart of the data preprocessing and anomaly cleaning method provided by the present invention is shown.

[0028] Figure 3 The flowchart is for the method of path one in step three of the present invention.

[0029] Figure 4 The flowchart is for the method of path two in step three of the present invention.

[0030] Figure 5 The overall method flowchart for step seven of this invention is shown below.

[0031] Figure 6 The flowchart of the method for visual recognition and irrigation mode linkage control provided by the present invention. Detailed Implementation

[0032] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0033] This invention provides a method for making precise on-demand irrigation decisions based on the actual physiological water stress state of crops, integrating multi-dimensional IoT environmental perception data, and using multi-modal feature fusion calculation. This method completely eliminates the blind spots caused by a single soil moisture data source and establishes a complete closed loop of "perception-computation-decision-execution-retesting-optimization".

[0034] For details, please refer to Figure 1 The irrigation decision-making method based on the fusion of visual recognition and IoT multimodal approaches includes the following steps: Step S1: Real-time access and timestamp alignment of multi-source data As the computing hub of the irrigation system, the decision engine receives raw data from two data sources according to a fixed polling cycle (e.g., once every 5 minutes) and aligns all data according to a unified timestamp to ensure that environmental data and visual data at the same time can participate in subsequent joint calculations.

[0035] S1.1 Access IoT environmental monitoring data IoT environmental monitoring data is divided into three main categories: (1) Soil parameters Soil sensors were deployed in layers at representative locations within each irrigation zone, collecting data at depths of 20cm (top root zone), 40cm (main root layer), and 60cm (deep reserve layer). The data collected are shown in Table 1 below: Table 1: Data Collection Table (I)

[0036] Among them, field water holding capacity θ fc and wilting coefficient θ wp These are inherent physical parameters of the soil type, which can be pre-determined through a one-time soil test and stored in the system as reference parameters for subsequent calculations.

[0037] (2) Meteorological parameters A small agricultural weather station was deployed in the center of the planting area to collect the following data, as shown in Table 2 below: Table 2: Data Collection Table (II)

[0038] (3) Equipment feedback parameters Feedback data is collected from the irrigation execution equipment to verify the execution of irrigation commands, as shown in Table 3 below: Table 3: Execution Parameter Description Table

[0039] S1.2 Access visual recognition result data In the technical solution of this invention, the vision module does not transmit the original image to the decision engine. Instead, it first performs local inference calculations on the crop canopy and leaf images captured by the camera using a lightweight model deployed on an edge computing device (such as an NVIDIA Jetson series edge board), and outputs structured index data. This approach has two advantages: first, it significantly reduces the data transmission bandwidth requirements; and second, it ensures that inference can still proceed normally in weak network or even offline environments.

[0040] The structured metrics output by visual recognition include the following, each with a clear physical meaning and numerical range. These metrics will directly participate in the quantitative calculation of irrigation decisions: (1) Water stress level (0-4) The model classifies crop water stress into five levels by analyzing the morphological characteristics of leaves and canopy, as shown in Table 4 below: Table 4: Levels of Water Stress

[0041] The model uses the MobileNetV2 lightweight convolutional neural network as the classifier backbone, performing classification inference on each canopy image on an edge device and outputting an integer level value from 0 to 4. The training data comes from a database of crop canopy images under different stress levels that have been manually annotated historically.

[0042] (2) CWSI Crop Water Stress Index The CWSI (Crop Water Stress Index) is a continuous numerical index ranging from 0 to 1, with higher values ​​indicating more severe water stress. This index quantifies the degree of water shortage in crops by analyzing the temperature difference between the canopy and the air.

[0043] Calculation principle: When crops have sufficient water, the stomata on the leaves open fully, and transpiration is strong. The heat absorbed by transpiration will cause the canopy temperature to be lower than the air temperature (similar to the human body sweating to cool down). When crops are short of water, the stomata close to reduce water loss, transpiration weakens, the canopy's heat dissipation capacity decreases, and the canopy temperature will rise, gradually approaching or even exceeding the air temperature.

[0044] The formula for calculating CWSI is:

[0045] in: T c Canopy temperature, measured by a multispectral camera or thermal infrared camera, in °C; T a Air temperature, measured by sensors at a weather station, in °C; (T) c -T a ) upper The theoretical maximum crown temperature difference when crops are severely water-deficient, in °C, is an empirical parameter that varies slightly for different crops (e.g., 35 °C for cotton and 46 °C for corn). The physical meaning of CWSI: CWSI≈0: The canopy temperature is much lower than the air temperature, resulting in strong transpiration and sufficient moisture. CWSI≈0.3~0.5: Canopy temperature is close to air temperature, moderate stress; CWSI≈0.8~1.0: Canopy temperature is higher than air temperature, indicating severe stress; (3) L leaf area index L (Leaf Area Index) is defined as the ratio of the total area of ​​all leaves per unit area of ​​land to the land area; it is a dimensionless value. For example, L = 3.0 means that every 1m² of land has a leaf area index. 2 On the land, the total leaf area is 3m² 2 .

[0046] The physical meaning of L: The smaller the L value (<1.0): fewer leaves, sparser canopy, smaller transpiration area, and lower water requirement; Medium L (1.5-3.0): Normal growth, moderate water requirement; The larger the L value (>3.5): the denser the leaves, the more closed the canopy, the larger the transpiration area, and the greater the water demand; The model improves the YOLO11-Seg lightweight semantic segmentation model to perform pixel-level segmentation of canopy images, extracts the pixel regions of green leaves, and then combines the known camera field of view and shooting height to obtain the actual land area and total leaf area corresponding to each image through geometric transformation, thereby calculating the L value.

[0047] (4) NDVI growth index NDVI (Normalized Difference Vegetation Index) is an indicator used to assess plant vigor and green biomass by analyzing the reflectance characteristics of plants to red and near-infrared light. Its calculation formula is:

[0048] in: NIR: Reflectivity in the near-infrared band (760–900 nm); Red: Reflectance in the red light band (620–700 nm); The physical meaning of NDVI: Healthy green plants strongly absorb red light for photosynthesis (which is why their leaves appear green) and strongly reflect near-infrared light (to dissipate heat), resulting in a high NDVI value (0.6–0.9). Plants subjected to water stress, pests and diseases, or aging have reduced red light absorption and near-infrared reflectance, resulting in a lower NDVI value (<0.3). The NDVI of bare soil and dead plants is usually negative or close to zero. NDVI is calculated after being captured by a multispectral camera, or it can be approximately estimated by a calibrated multispectral reconstruction algorithm from a regular high-definition camera.

[0049] (5) Canopy coverage Canopy cover refers to the proportion of the projected area of ​​green vegetation to the total area of ​​the image when viewed vertically, and is directly output by the segmentation model. Higher cover indicates a denser crop canopy, a larger transpiration area, and a greater water requirement.

[0050] (6) Percentage of leaf curling area and percentage of yellow leaf area These two indicators were identified by further subdividing the leaf regions using a segmentation model. Leaf curling is the most sensitive early morphological response of crops to water shortage stress, while yellowing is a sign of more severe stress. These indicators are used to help verify the accuracy of stress level assessment.

[0051] S1.3 Timestamp Alignment Because the acquisition times of IoT sensor data and visual recognition data may not be completely synchronized (for example, sensors collect data every 5 minutes, while cameras capture images every 15 minutes), the decision engine records the timestamp of the acquisition time for both types of data. During the fusion calculation, it selects the data with the smallest time difference for pairing. When the time difference between the two sets of data exceeds a preset threshold (for example, 30 minutes), the data quality of that fusion calculation is marked as "low confidence," and a prompt is included in the decision output.

[0052] Step S2: Data Preprocessing and Anomaly Cleaning After multi-source data is integrated into the decision engine, it must first undergo preprocessing and cleaning to remove outliers and ensure the reliability of subsequent calculations. (See also...) Figure 2 This step includes the following five sub-steps: S2.1 Remove anomalous jump values Soil moisture sensors and weather sensors may output abnormal values ​​that change abruptly and drastically when subjected to electromagnetic interference, communication packet loss, or temporary sensor malfunctions. For example, under normal circumstances, soil moisture content changes gradually (usually less than 1% per hour), but if it suddenly jumps from 30% to 5% or 80% at a certain moment, it is clearly abnormal data.

[0053] Detection method: For each sensor data channel, maintain a sliding window (e.g., the 10 most recent valid data points), and calculate the mean μ and standard deviation σ within the window. When a newly added data point deviates from the window mean μ by more than 3σ, it is judged as an abnormal jump value, marked as invalid, and discarded, and does not participate in subsequent decision calculations.

[0054] S2.2 Linear interpolation completion for missing values When the sensor experiences a brief data loss due to communication interruption, power failure, or other reasons (the loss time does not exceed 2 hours), a linear interpolation method is used to complete the data.

[0055] The formula for linear interpolation:

[0056] in: θ t1 and θ t2 : These represent the values ​​of the preceding and following valid data points in the missing time period, respectively; t1 and t2: timestamps of the previous and next valid data points, respectively; t: Missing time points that need to be filled in When the missing time exceeds 2 hours or the number of consecutive missing data points exceeds 20% of the total data, the current data of the sensor is determined to be unreliable, and the irrigation decision for that area is downgraded to the "visual + meteorological data only" mode.

[0057] S2.3 Moving average filtering for noise reduction Sensor data always contains some high-frequency noise (e.g., instantaneous fluctuations in wind speed cause fluctuations in evapotranspiration calculations), requiring data smoothing. This solution uses a moving average filtering algorithm:

[0058] x n The original data value at time n; : The filtered data value; N: Sliding window length (in this scheme, N=5, that is, the average of the most recent 5 data points); The choice of N value needs to be dynamically adjusted according to the rate of parameter change: for parameters that change rapidly (such as wind speed and instantaneous rainfall), N should be a small value (3-5) to preserve the trend of change; for parameters that change slowly (such as soil moisture content), N can be a slightly larger value (7-10) to obtain a better smoothing effect.

[0059] S2.4 Image Quality Judgment Images captured by the camera under conditions such as backlighting, heavy fog, lens contamination, and nighttime without supplemental lighting are of poor quality, resulting in low confidence levels for the model's inference results on these images. Therefore, it is necessary to automatically assess image quality before inference. Evaluation metric 1 - Average brightness: Calculate the average grayscale value of all pixels in the image. If it is below the threshold (too dark) or above the threshold (overexposed), it is judged as a low-quality image. Evaluation metric 2 – Contrast: Calculate the standard deviation of the image pixel grayscale. If it is lower than the threshold (image is blurry, heavy fog), it is judged as a low-quality image. Evaluation Criterion 3 – Occlusion Ratio: Determine whether there are foreign objects (such as insects, water droplets, or mud) obstructing the core area of ​​the lens through simple color block detection; Low-quality images are marked as "invalid" and will not be included in the irrigation decision this time. The decision engine will use the previous valid visual data as a substitute (but will mark the timeliness of the data).

[0060] S2.5 Data Normalization Data from different sensors have different dimensions and numerical ranges, such as soil moisture content between 0 and 50%, wind speed between 0 and 20 m / s, stress level between 0 and 4, and NDVI between -11. If data is directly spliced ​​together for calculation without normalization, the larger dimensions will dominate the calculation results and mask the contribution of the smaller dimensions.

[0061] This scheme uses the Min-Max normalization method to uniformly map all indicators to the [0, 1] interval:

[0062] Where x min and x max These are the theoretical minimum (or historical minimum) and theoretical maximum (or historical maximum) values ​​for the indicator, which need to be pre-set based on crop type and climate conditions. To facilitate the restoration of certain dimensions in subsequent calculations, the normalized data includes its original dimensional information and normalization parameters.

[0063] Step S3: Parallel calculation of basic water demand indicators via dual paths The preprocessed multi-source data is fed into two parallel computing paths to calculate basic water requirement indicators from the "environment-soil" dimension and the "crop physiology" dimension, respectively.

[0064] Path 1: Soil-meteorological theoretical water demand calculation based on IoT environmental data Calculate the amount of water needed by crops to maintain normal transpiration based on meteorological conditions, and then combine this with the current soil moisture level to determine the water shortage. (See also...) Figure 3 : Sub-step S3-1-1: Calculate the reference crop evapotranspiration ET0 Reference Evapotranspiration (ET0) refers to the evapotranspiration rate of a hypothetical reference crop (assuming a uniform green grassland 12 cm high, with a surface resistance of 70 s / m, and sufficient moisture without stress) under given meteorological conditions, measured in mm / day. It reflects the "atmospheric evaporation demand intensity," that is, the atmosphere's "ability" to extract water from vegetation and soil. The hotter and drier the weather, the stronger the wind, and the more intense the sunlight, the larger the ET0.

[0065] This scheme uses the FAO Penman-Monteith formula to calculate ET0. The weighted average of the sum of the "energy term" and the "aerodynamic term" is as follows:

[0066] The physical meaning and acquisition method of each parameter in the formula are explained below: The first term of the molecule: energy term (radiation-driven term). 0.408: Unit conversion factor, converting radiation flux from MJ / m² / day to equivalent evaporation depth in mm / day; Δ: The slope of the saturated vapor pressure-temperature curve, in kPa / ℃. Its physical meaning is: how much more water vapor the air can hold for every 1℃ increase in temperature. Δ increases with increasing temperature—the higher the temperature, the stronger the air's water storage capacity and the greater its evaporation potential. The calculation formula is:

[0067] Where T is the daily average temperature (°C), provided by the meteorological station. The exponential formula in parentheses is the saturated vapor pressure formula.

[0068] R n Net radiation on the crop canopy surface, in MJ / m² / day. Obtained from total solar radiation measured by a solar radiation sensor after reflectivity correction: R n = (1-α)×R s -R nl Where α is the crop canopy reflectance (approximately 0.23), and R... s To measure the total solar radiation, R nl This is the net longwave radiation (which can be estimated based on temperature and humidity).

[0069] G: Soil heat flux, in MJ / m² / day. For diurnal calculations, G≈0 (because the heat absorbed by the soil during the day is roughly balanced with the heat released at night).

[0070] The second term in the molecule: aerodynamics (wind-driven term) γ: Wet and dry constant, in kPa / ℃.

[0071] The ratio of air's "drying capacity" to its "heat-carrying capacity". γ≈0.066×P / 100, where P is atmospheric pressure (kPa), which can be measured by a barometric pressure sensor at a weather station or estimated based on altitude.

[0072] 900: Empirical constant, which is a comprehensive coefficient that combines physical parameters such as altitude and gas constant, and is applicable under standard conditions.

[0073] T: Average daily temperature, in °C.

[0074] u2: Wind speed at 2m height, unit: m / s. If the wind speed sensor at the weather station is not installed at a height of 2m, height correction is required.

[0075] Where u z The measured wind speed at the sensor installation height z.

[0076] e s : Saturated vapor pressure, unit: kPa. That is, the maximum amount of water vapor that air can hold at the current temperature.

[0077] e a Actual vapor pressure, unit kPa. Calculated from relative humidity (RH): e a = e s ×RH / 100; (e s -e a ): Saturation difference, unit: kPa. The larger the difference, the drier the air, and the stronger the "pull" of evaporation. For example, in the desert... s -e a It's very large (the air is very dry), while in the humid rainforest... s -e a Very small (the air is almost saturated).

[0078] The physical meaning of the denominator: The denominator Δ + γ × (1 + 0.34 × u²) is an "impedance" term. The larger the wind speed u², the larger the denominator, indicating a stronger ability of the airflow to carry away water vapor. However, the energy term and the aerodynamic term have different sensitivities to the denominator—under high temperature and high radiation conditions (large Δ), the energy term dominates; under strong winds and dry conditions (large γ term in the denominator), the aerodynamic term dominates. This weighting mechanism allows the formula to adapt to various climatic conditions.

[0079] Summary of the physical meaning of ET0: Under weather conditions where ET0 = 5 mm / day, a well-watered grassland will consume 5 mm of water per day (equivalent to 5 liters of water per m² area). This is the basis for subsequent calculations of crop water requirements.

[0080] Sub-step S3-1-2: Calculate the theoretical crop evapotranspiration ET c The actual water requirement of a crop is not equal to that of a reference crop because different crops consume water at different rates, and the water requirement also varies at different growth stages of the same crop. Therefore, the crop coefficient Kc0 (basic crop coefficient) is introduced for conversion: ET c =Kc0×ET0; in: ET c Theoretical crop evapotranspiration, in mm / day; Kc0: Basic crop coefficient, dimensionless, with a value between 0.4 and 1.2; The value of Kc0 depends on the crop type and growth stage. For example: During the seedling / germination stage of crops: fewer leaves, smaller transpiration area, Kc0≈0.4~0.6; During the vegetative growth stage of crops: leaves expand rapidly, Kc0≈0.7~0.9; During the flowering and fruiting stage of crops: leaf area is largest, transpiration is strongest, Kc0≈1.0~1.2; During the crop's maturity and harvest period: leaves begin to age and fall off, Kc0≈0.6~0.8; Kc0 is stored in the system's crop variety parameter library. After the user selects a crop variety, the system automatically matches the Kc0 value for the corresponding growth stage.

[0081] Sub-step S3-1-3: Calculate the soil moisture deficit D Soil moisture deficit (D) reflects how much water is currently lacking in the soil that is available for crop use. The calculation formula is: D=(θ fc -θ real )×H×ρ; in: D: Soil moisture deficit, in mm (i.e., the depth of water layer that needs to be replenished). θ fc Field water holding capacity, measured in m³ / m³. It is a soil physical property parameter, pre-measured and stored in the system parameters. θ real : The measured volumetric moisture content, in m³ / m³, is collected in real time by a soil moisture sensor; H: Planned wetting layer depth, in mm. It refers to the depth of the main root layer of the crop, and this value varies for different crops and different growth stages. For example: during the seedling stage, the root system is shallow, H≈200-300mm (approximately 20-30cm); during the fruiting stage, the root system is deep, H≈400-600mm; ρ: Soil bulk density, unit: g / cm³. It reflects the compaction of the soil. For sandy soil, ρ≈1.3~1.5, and for clay soil, ρ≈1.1~1.3. Assume the field water holding capacity θ of a certain plot of land fc = 0.35 (35%), measured moisture content θ real = 0.25 (25%), planned wetting layer depth H = 400mm, soil bulk density ρ = 1.4. Then: D=(0.35-0.25)×400×1.4 = 0.10×400×1.4=56mm; This means that the soil is currently lacking the moisture equivalent to a water layer depth of 56 mm, and that much water needs to be added to restore the soil to field capacity levels.

[0082] Path 2: Visual-based calculation of crop stress correction coefficient This path outputs two correction coefficients: K s (Water stress correction factor) and KL (growth vigor correction factor). See also Figure 4 : Sub-step S3-2-1: Generate the water stress correction coefficient K based on the stress level. s K s Physical meaning of K: s > 1 indicates that the crop is already showing signs of water stress and needs more water than theoretically calculated to compensate for the physiological deficit already caused; K s = 1 indicates no stress, and irrigation can be carried out according to the theoretical calculation.

[0083] The visual output provides discrete stress levels (0-4), but irrigation decisions require continuous correction coefficients. Therefore, a mapping function from stress levels to coefficients is established, as shown in Table 5 below: Table 5: Mapping Function from Grade to Coefficient

[0084] Within the same level, K s The specific values ​​within the range are linearly interpolated based on continuous CWSI values ​​to achieve a smoother correction effect. K s =1.0+(level_index)×0.25+(CWSI-CWSI min_level ) / (CWSI) max_level -CWSI min_level ) × 0.2; In the formula, level_index is the stress level value (0~4), CWSI min_level and CWSI max_level This represents the CWSI range boundary corresponding to this level.

[0085] Sub-step S3-2-2: Generate L correction coefficient KL based on L growth potential. This reflects the impact of the "ratio of actual leaf area to standard leaf area" on water requirement. The more leaves (larger L), the larger the transpiration area of ​​the crop, and the more water it requires, so the crop coefficient needs to be adjusted upwards; the fewer leaves (smaller L), the smaller the transpiration area, and the less water it requires, so the crop coefficient needs to be adjusted downwards.

[0086] The formula for calculating KL is: K L =Lreal / L standard ; in: L real The leaf area index currently measured by the visual model; L standard The standard leaf area index (the value that should be reached under normal growth conditions) of this crop under the current growth stage is stored in the crop variety parameter database; Example: Standard L for a certain crop during the fruiting period standard = 3.0, if vision detects the current L real = 3.6 (vigorous growth), then K L = 3.6 / 3.0=1.2, crop coefficient increased by 20%; if L real = 2.0 (weak growth), then K L = 2.0 / 3.0≈0.67, the crop coefficient is reduced by 33% to avoid root rot caused by excessive irrigation.

[0087] K L Set protective boundaries: to prevent extreme situations (such as L) real Overcorrection due to abnormally small or large measurement errors (causing K to be overcorrected) L Set upper and lower limits: 0.5≤K L ≤1.5.

[0088] Step S4: Multimodal fusion to generate comprehensive actual water demand After the previous step completed the independent calculation of the two paths, this step integrates the theoretical calculation results of the soil-meteorological path with the physiological correction coefficient of the visual path to generate the final comprehensive irrigation quota.

[0089] S4.1 Calculate the dynamic crop coefficient K c In traditional irrigation schemes, the crop coefficient is fixed (determined solely by looking up a table based on the growth period). This invention uses real-time visual observation of crop growth (L) to dynamically and adaptively correct the basic crop coefficient Kc0, generating a dynamic crop coefficient K. c : K c =Kc0×K L ; in: Kc0 (basic crop coefficient): represents "the standard water requirement characteristics of the crop at the current growth stage", which is an empirical value under conditions of full irrigation and standard growth. K L (L correction coefficient): Represents the "deviation between the current actual growth and the standard growth", which is the result of real-time visual observation; K c(Dynamic crop coefficient): The product represents the "water requirement coefficient of the crop under the current condition" after correction for the current actual growth status; For example: For a crop at the fruiting stage, the standard Kc0 = 1.0, and the visual measurement L... real =L standard (If growth is normal), then K L =1.0, K c = 1.0 × 1.0 = 1.0, executed according to standard coefficient.

[0090] However, if visual measurement shows L real > L standard (Growing exceptionally vigorously), K L > 1, K c > 1. The crop coefficient has been increased to match the greater transpiration demand.

[0091] S4.2 Calculate the corrected crop evapotranspiration ETc adj Traditional ETC only considers meteorological factors (reflected by ETO) and crop factors (reflected by K). c This method reflects the crop's condition (e.g., crop performance), but it overlooks whether the crop is already under water stress. In step S4.1, after obtaining the dynamic crop coefficient Kc, this invention further multiplies it by the stress correction coefficient K. s The corrected actual crop evapotranspiration was obtained: ETc adj =K c ×ET0×K s ; The physical meaning of the formula: K c ×ET0: Represents "how much water a crop at its current growth stage needs each day to maintain normal physiological activities under current weather conditions"; Multiply by K s : represents "if the crop is already short of water (K) s > 1), then additional water needs to be added to make up for the deficit that has already occurred; In fact, ETC adj It includes not only the water needed to maintain current transpiration, but also a portion of "compensatory" water—water that makes up for the water deficit already caused by previous water stress in the crop. This is why, when stress is identified, K... s > 1 will increase irrigation volume.

[0092] S4.3 Calculate the comprehensive single irrigation quota W Combining the soil moisture deficit D from path S3-1 and the corrected evapotranspiration calculated in step S4.2, and then subtracting the predicted future rainfall, the final water requirement for this irrigation is obtained: W=max{0,(D+ET c-adj ×TP forecast ×η)}; The physical meaning of each parameter is as follows: W: Single irrigation quota, in mm. If W≤0, it means that natural rainfall can cover the demand, and irrigation is not required. The water depth in mm can be easily converted into water demand per unit area: 1mm / m² = 1L / m² = 0.001m³ / m²; D: Soil moisture deficit (mm), from S3-1-3. It represents "how much water is currently needed in the soil to restore it to field capacity". ETc adj ×T: The amount of water (mm) that the crop will consume in the next T days. ETc adj It is the daily water consumption rate (mm / day), and T is the planned irrigation interval in days (i.e., how long after this irrigation is planned to wait before the next irrigation).

[0093] P forecast ×η: Future predicted effective rainfall (mm). P forecast η is the total rainfall (mm) given in the weather forecast, and η is the effective utilization coefficient (0, 1), because some rainfall will be lost through surface runoff or deep infiltration after exceeding the field capacity. Usually, η is taken as 0.7-0.85, depending on the ground slope and soil infiltration performance.

[0094] Calculation example: Assume the following data: Soil moisture deficit D = 56mm (the soil is currently short of water by 56mm). Corrected crop evapotranspiration ETc adj = 4.5mm / day (crops consume 4.5mm of water per day); The planned irrigation interval T = 5 days (the next irrigation is planned 5 days later); Predicted rainfall P forecast = 12mm, effective utilization coefficient η = 0.75; calculate: The soil is currently short of water + the water consumption in the next 5 days = 56 + 4.5 × 5 = 56 + 22.5 = 78.5 mm; Effective rainfall deduction: 12 × 0.75 = 9 mm; Irrigation quota: 78.5 - 9 = 69.5mm; Because 69.5 > 0, irrigation is needed, and a water layer depth of 69.5mm needs to be added; If P forecastVery large: For example, if the predicted rainfall is 50mm, 50 × 0.75 = 37.5mm; Irrigation quota: 78.5 - 37.5 = 41mm; This indicates that the rainfall has already covered part of the demand, and only an additional 41mm is needed; If P forecast Very large: Forecast rainfall 120mm, 120 × 0.75 = 90mm; Irrigation quota: 78.5 - 90 = -11.5 < 0; If we set max{0, negative} = 0, then no irrigation is needed; we can simply wait for rain.

[0095] Step S5: Multi-level logical rule verification (safety gate against false positives) After calculating W, the decision engine does not immediately issue irrigation commands. Instead, it uses multi-layered logical rules to verify the data and prevent unreasonable irrigation in special circumstances. Verifications are performed sequentially from highest to lowest priority. If any level of verification fails, the irrigation decision will be blocked or corrected.

[0096] S5.1 Soil Threshold Hard Constraint (Highest Priority) Rule description: If the current soil moisture content is already higher than the upper limit of the field suitable moisture content for the current growth stage of the crop, irrigation is prohibited regardless of whether a slight stress level is visually detected.

[0097] Conditional logic: IF θ real >θ upper_limit ⇒BlockIrrigation; in: θ real Current measured volumetric moisture content; θ upper_limit The upper limit of suitable moisture content during this growth period is usually taken as 85% to 90% of field capacity. For example, if the field capacity is 35%, then the upper limit is 35% × 0.9 = 31.5%. When the soil moisture content has exceeded the suitable upper limit, it means that the water in the soil pores is close to saturation. Continuing to irrigate will lead to a serious decrease in soil aeration, root hypoxia, and will inhibit crop growth. Even if you can visually identify slight wilting of the leaves (which may be caused by other factors such as disease or salt stress), you should not irrigate, because further irrigation will aggravate the damage caused by hypoxia and root rot.

[0098] Special case explanation: If θ real >θ upper_limit However, if the visual system detects severe stress (level ≥ 3), it will trigger an "abnormal alarm" instead of irrigation—the system will determine that it may be a sensor malfunction or poor soil drainage, and remind manual intervention to check.

[0099] S5.2 Rainfall Forecasting Constraints Rule description: If there is effective rainfall within the next 24 to 48 hours according to the weather forecast, the irrigation amount will be reduced proportionally to the rainfall; if the predicted rainfall has fully met the needs, the irrigation will be cancelled.

[0100] Judgment logic: If the predicted effective rainfall P forecast If ×η ≥ the irrigation quota W calculated in this case, then irrigation is cancelled. If P forecast If ×η < W, then the corrected irrigation quota is: W' = W - P forecast ×η This rule aims to prevent waste caused by the combined effects of irrigation and natural rainfall, leading to deep seepage. It gives the system a "proactive" approach, rather than mechanically relying on current data.

[0101] S5.3 Time Constraints Rule description: Irrigation should be restricted during the following two periods to reduce ineffective water loss and the risk of disease occurrence: (1) High temperature noon period (11:00-15:00) If the current time is between 11:00 and 15:00 and the temperature is above 30°C, new irrigation should not be started; irrigation that is already in progress will not be interrupted, but no additional irrigation will be added.

[0102] During the midday heat, solar radiation is strongest, and the transpiration pull of the air is greatest. Irrigating at this time results in a large amount of water evaporating and dissipating into the air before it can be absorbed by the crops, leading to extremely low water utilization (sometimes even below 30%). Simultaneously, irrigating with cold water at high temperatures can cause "cold shock" damage to crop roots, causing stomata to close and inhibiting photosynthesis. Therefore, this period is designated as an "uneconomical period" for irrigation, and new irrigation should not be initiated.

[0103] (2) Nighttime low temperature condensation period (22:00~6:00) If the current time is between 22:00 and 6:00, and the relative humidity of the air is higher than 85% (close to saturation and prone to condensation), then irrigation should be restricted.

[0104] Under low temperature and high humidity conditions at night, dew easily forms on the surface of crop leaves. Irrigation will increase the duration of canopy humidity, significantly increasing the risk of fungal diseases (such as downy mildew and gray mold). At the same time, transpiration is extremely weak at night, and irrigation water cannot be effectively absorbed and utilized by crops, with most of the water seeping deep into the soil and being wasted. Therefore, it is also forbidden to start new irrigation during this period.

[0105] Exceptions to time-limited constraints: When severe water shortage stress (level ≥ 3) is visually detected and the soil moisture content is below the wilting coefficient θwp When the "emergency irrigation mode" is triggered, it overrides time constraints and immediately starts irrigation to save crop lives. This indicates that emergency risk avoidance takes precedence over economic considerations and disease control.

[0106] S5.4 Coercion Priority Logical Judgment When there is a discrepancy between the visually recognized crop water stress level and the soil moisture data, the visual stress level shall be the highest priority criterion.

[0107] Judgment logic can be divided into two typical contradictory scenarios: Contradictory Scenario A: Moderate soil moisture (θ) real (within the normal range), but crop stress was identified. If (θ) lower_limit ≤θ real ≤θ upper_limit If the stress level is ≥2, then it is determined as "latent physiological dehydration"; Implementation: Small-volume supplemental irrigation; The irrigation amount is adjusted to: W' = W×0.5 (50% of the normal calculated amount); Explanation: Soil moisture data is normal, but crop leaves are observed to be wilting and curling. This indicates that the crop may be suffering from root disease, excessive soil salinity (high EC) hindering root water absorption, or excessive transpiration due to high temperature and low humidity, resulting in a situation where the crop has soil but no water (the soil contains water, but the roots cannot absorb it). In this case, judging based on soil data would lead to the incorrect conclusion that "irrigation is unnecessary." The system identifies this as "latent water shortage" and administers small-volume irrigation—the volume should not be too large (too much water will worsen root hypoxia), but water must be provided to alleviate the immediate problem of leaf wilting.

[0108] Contradictory Scenario B: Low soil moisture (θ) real Below the lower threshold, but no stress was identified (level ≤ 1) and the crop growth is dark green. If (θ) real <θ lower_limit If the result is (stress level ≤ 1) AND (L ≥ L_standard × 0.8), then it is determined as "crop compensatory drought resistance"; Execution: Delayed irrigation; The irrigation amount is adjusted to: W' = W×0 (temporarily suspend irrigation) or W' = W×0.3 (small amount of maintenance water replenishment); Explanation: Soil data indicates water shortage, but the crop leaves are fully extended, dark green, and growing vigorously, showing no signs of stress. This suggests that the crop has a well-developed root system, capable of absorbing water from deeper soil layers (those not covered by sensors), or that the crop variety has a strong drought tolerance, meaning the surface soil is dry but not causing actual stress to the crop. In this case, indiscriminate irrigation could lead to deep waterlogging and root rot. The system identifies this as "compensatory drought tolerance" and suggests delaying irrigation or providing only a small amount of maintenance water.

[0109] S5.5 Equipment and Piping Constraints Rule description: Verify the upper limit of the irrigation equipment's capacity. If the single irrigation quota exceeds the equipment's capacity, the irrigation task will be automatically split.

[0110] The judgment logic includes the following three dimensions: (1) Maximum flow rate constraint of a single pump The rated flow rate of a single water pump is Q. pump (m³ / h), if the planned irrigation water volume is V (m³), the minimum irrigation duration is: T min =V / Q pump ; If T min Exceeding the system's maximum allowed single continuous run time T max_pump (For example, if the water pump should not run continuously for more than 4 hours), then irrigation needs to be divided into multiple irrigation cycles, with each irrigation lasting no more than T hours. max_pump .

[0111] (2) Pipeline pressure bearing capacity constraints The system determines whether the pipeline network is within the normal pressure range based on the current pressure value fed back by the pipeline pressure sensor. If the pressure is abnormally low (possibly due to pipeline damage or blockage), irrigation is stopped and a pipeline fault alarm is triggered; if the pressure is close to the upper limit, the valve opening is appropriately reduced to decrease the flow rate and prevent the pipeline from bursting.

[0112] (3) Regional rotation irrigation capacity constraints When the system has multiple irrigation zones, the total irrigation volume after each zone makes independent decisions may exceed the instantaneous water supply capacity of the water source. In this case, the zones should be prioritized according to their stress level: Rotation irrigation sorting rules: 1. Zones with higher coercion levels → irrigate first (coercion levels > 2 take precedence); 2. When stress levels are the same, irrigate the soil with the larger soil moisture deficit (D) first; 3. If all of the above are the same → irrigate in rotation according to the zone number order; Based on the sorting results and the pump station's water supply capacity, the system automatically generates start-up and shutdown schedules and valve switching plans for each zone, ensuring that at most 1 to 2 zones are irrigated simultaneously, without exceeding the total water supply capacity limit.

[0113] Step S6: Output the final irrigation decision command After the above layers of verification, the decision engine generates a structured irrigation decision instruction, which is then sent to the controller in the execution control layer via the communication network. The instruction contains the following complete information: (See details) Figure 5 : S6.1 Decision Conclusion The decision conclusion is a ternary outcome, taking only one of the following three cases, as shown in Table 6 below: Table 6: Decision Conclusion Table

[0114] S6.2 Irrigation Execution Parameters For the "Irrigation" conclusion, the following detailed execution parameters are output: (1) Target irrigation amount Output in mm: W (mm) Convert to volume: V area = W×A / 1000 (m³), where A is the area of ​​the irrigated zone (m²) (2) Total irrigation duration

[0115] in: TT: Total irrigation time, in hours; W: Irrigation quota, unit mm; A: Irrigation zone area, in m²; Q design Design flow rate, in m³ / h; ε: Irrigation uniformity coefficient (0.8~0.95), used to compensate for the flow deviation of the emitters; (3) Valve opening setting For electric ball valves, there is a non-linear relationship between valve opening and flow rate (approximately an S-shaped curve). The system queries a valve characteristic curve table pre-stored in the controller, based on the target flow rate Q. target Retrieve the corresponding valve opening percentage.

[0116] For example, the characteristic curve of a certain electric ball valve shows that when the opening degree is 30%, the flow rate is approximately 20% of the maximum flow rate; when the opening degree is 60%, the flow rate is approximately 60% of the maximum flow rate; and when the opening degree is 90%, the flow rate is approximately 95% of the maximum flow rate. The system obtains the opening degree value by looking up a table based on the target flow rate percentage.

[0117] (4) Irrigation start and stop times Based on time constraints (avoiding midday and nighttime), irrigation should be scheduled during the two optimal irrigation windows: early morning (6:00–8:00) or late afternoon (16:00–19:00).

[0118] Specific calculations: If the current time is within the allowed irrigation window AND irrigation duration TT ≤ remaining window time, then irrigation will start immediately; otherwise, it will be postponed until the start of the next irrigation window.

[0119] Emergency irrigation mode is not subject to this restriction.

[0120] S6.3 Multi-plot rotational irrigation scheme When the system manages multiple irrigation zones, the decision engine generates independent irrigation plans for each zone and coordinates the irrigation rotation sequence. The output format is shown in Table 7 below: Table 7: Output Format Table

[0121] S6.4 Includes alarm information The decision command is accompanied by the following alarm information, which is pushed to the system software platform and mobile APP in real time, as shown in Table 8 below: Table 8: Alarm Information Push Table

[0122] Step S7: Closed-loop feedback self-optimization (continuous iteration of the decision engine) Irrigation is not a one-time open-loop control, but a continuous closed-loop iterative process. After each irrigation cycle, the system automatically enters a feedback correction phase, making the decisions increasingly accurate. (See also...) Figure 6 : S7.1 Post-irrigation data retesting After irrigation is completed, the system does not immediately close the current decision cycle. Instead, it waits for a preset "moisture balance time window" (usually 2-4 hours after irrigation to allow water to fully diffuse and evenly distribute in the soil) before automatically performing the following retesting operation: Soil moisture retest: Read the soil moisture sensor data for each layer of this zone again to obtain the θ after irrigation. real_post value Visual retest: Take another image of the crop canopy and input it into the model to re-obtain stress level and L and other indicators. S7.2 Comparison and Analysis of Actual Results The system compares the retested data with the expected data to determine the effectiveness of the irrigation. (1) Soil moisture response assessment Theoretical expectation: The soil moisture content after irrigation should reach: θ expected =θ real_pre +W / H×ρ; Where θ real_pre This represents the measured moisture content before irrigation.

[0123] Actual value reached: θ real_post ; Calculate the deviation: Δθ = θ real_post -θ expected ; If Δθ ≈ 0 (deviation within ±2%), it indicates that the irrigation amount is calculated accurately and the water infiltration and distribution are normal. If Δθ >> 0 (actual moisture content is significantly higher than expected): this indicates that the irrigation amount is too large, or the actual field capacity is lower than the set value, resulting in deep seepage and waste. If Δθ << 0 (actual moisture content is significantly lower than expected): this indicates insufficient irrigation, or problems such as surface runoff loss or pipeline leakage. (2) Assessment of crop stress mitigation Compare the changes in visually recognized stress levels before and after irrigation: The stress level has recovered from level 2 or above to level 0-1: This indicates that irrigation has effectively alleviated the stress. No significant change in stress level: This indicates that irrigation is still insufficient, or the cause of stress is not a water problem (such as disease). The stress level actually increased: This may be due to irrigation causing root hypoxia and exacerbating the stress. S7.3 Automatic parameter fine-tuning When the system identifies systematic biases through multiple evaluations, it automatically adjusts the key parameters in the decision-making model: (1) Adaptive correction of Ks mapping relationship If the system detects a stress level of 2, press K. s =1.4 After irrigation, the crop still showed signs of stress (insufficient relief), indicating that K s If the value is too small, the system will automatically adjust the K value corresponding to coercion level 2. s The lower limit has been raised from 1.3 to 1.35, and the upper limit has been raised from 1.5 to 1.55.

[0124] Correction algorithm (moving average update): K s_new (level) = (1-α) × K s_old (level) + α × K s_optimal (level); in: α is the learning rate (ranging from 0.1 to 0.3), which controls the adjustment speed. K s_optimal (level) represents the optimal value for this level, deduced from the observed results. (2) Adaptive correction of KL correction coefficient If the system detects: when L real / L standard Press K when larger L After adjusting the crop coefficient, the actual evapotranspiration (calculated by reverse calculation of soil moisture consumption before and after irrigation) was lower than expected, indicating that the actual transpiration was not as high as expected. L If it's too large, the system will automatically fine-tune K. L The scaling factor in the calculation formula.

[0125] (3) Adaptive correction of soil parameters If the system detects that the change in soil moisture content after multiple irrigations consistently deviates from the theoretical expectation, and the deviation is in the same direction (always too high or always too low), it indicates that the pre-stored soil parameters (such as field capacity θ) are inaccurate. fc The soil bulk density (ρ) may not match the actual value. After five consecutive instances of deviation in the same direction, the system automatically issues parameter calibration suggestions and corrects the stored parameter values ​​according to the back-calculation results.

[0126] S7.4 Long-term iteration of model threshold The system records and stores all data from each irrigation decision process—including environmental data before the decision, visual data, decision instructions, execution status, and post-irrigation retest data—into the historical database of the cloud platform.

[0127] Once the accumulated data reaches a certain scale (typically one growing season's worth of data is recommended), the system uses this data to perform incremental training or threshold optimization of the visual recognition model. Threshold calibration for stress classification models: Compare the stress level output by the model with the actual recovery effect after irrigation, and adjust the decision boundary of the classifier. For example, if the model frequently misclassifies "mild stress" as "moderate stress" (leading to K... s If the threshold is too high (e.g., excessive irrigation), the classification threshold will be fine-tuned to make the model's judgments more closely reflect actual results. Calibration of the L estimation model: Compare the deviation between the visually estimated L value and the ground-based measured value (manual sampling measurement), and adjust the model parameters. This forms a complete closed loop of "perception-computation-decision-execution-retesting-optimization". The irrigation accuracy of the system continues to improve over time, adapting to the actual conditions of different plots, different years and different climates.

[0128] To help understand the entire process, a complete running example is provided below: Setting: An apple orchard, divided into two irrigation zones, A and B, each covering an area of ​​1 acre (approximately 667 m²). The variety is Red Fuji, currently in its fruiting stage.

[0129] At 6:00 AM, the decision engine begins a new round of decision-making: Step S1 - Data Access, as shown in Table 9 below: Table 9: Data Access Table

[0130] Step S3 - Dual-path calculation: Path 1 calculation (taking partition A as an example): Calculate ET0: According to the Penman-Monteith method, substitute T... a =26℃, RH=60%, u2=2.5, R n =12, so ET0≈4.8mm / day; If Kc0 = 1.0 in the result period, then ET c = 1.0 × 4.8 = 4.8 mm / day; Soil deficit D (taking the 40cm layer): (0.35-0.30)×400×1.4 = 28mm; Path 2 calculation (taking partition A as an example): Stress level = 2 (moderate water shortage), corresponding to K s = 1.4 (take the midpoint of the interval); L correction: K L = 3.2 / 3.5 = 0.91 (L is slightly lower than the standard, so it should be adjusted downwards appropriately); Step S4 - Unified Computation (Partition A): K c = 1.0 × 0.91 = 0.91; ETc adj = 0.91×4.8×1.4 = 6.12mm / day (higher than the theoretical ETC, because the crop is already stressed and requires additional compensation). The planned interval is T=3 days, the predicted rainfall is P=5mm, and η=0.8; W = max{0, (28 + 6.12×3 - 5×0.8)} = max{0, (28+18.36-4)} = 42.36mm; Step S5 - Logical Verification: Partition A: Soil threshold hard constraint: θ real =30% (40cm layer), upper limit = 35% × 0.9 = 31.5%, 30% < 31.5%, passed; Rainfall forecast: The effective rainfall is predicted to be 4mm, W'=42.36-4=38.36mm>0, irrigation is still required; Time constraint: Currently at 6:00, within the allowed window, pass; Stress priority: No contradiction (drier soil + stress detected, consistent); Equipment constraints: Checked according to pump capacity, no problems found; Conclusion: Zone A requires irrigation, with an irrigation amount of 38.36 mm, which translates to a water volume of approximately 38.36 × 667 / 1000 = 25.6 m³. Partition B: Soil deficit D: (0.35-0.35)×400×1.4=0mm (the 40cm layer is completely at field capacity); ETc adj = K c ×ET0×K s K L =3.8 / 3.5=1.09, K c =1.0 × 1.09 = 1.09, K s =1.0 (without coercion), ETc adj =1.09×4.8×1.0=5.23mm / day; W = max{0, (0+5.23×3-4)} = max{0, (15.69-4)} = 11.69mm; Soil threshold: θ real =35%, maximum 31.5%, 35% > 31.5% → Irrigation prohibited! Verification failed; irrigation is prohibited for partition B this time. Conclusion: Irrigate zone A, do not irrigate zone B.

[0131] Step S6 - Output command (partition A): Irrigation volume: 25.6 m³; Irrigation duration: Assuming a design flow rate Q = 10 m³ / h and a uniformity coefficient of 0.9, TT = 25.6 / (10 × 0.9) ≈ 2.84 h ≈ 2 hours and 50 minutes; Valve opening degree: Determined by referring to a table based on the characteristic curve; Startup time: Start immediately (6:05), end time approximately 8:55; Phenological stage: Fruiting stage → Default switch to sprinkler irrigation mode; Step S7 - Closed-loop feedback (3 hours after irrigation): The moisture content of the 40cm layer in zone A was retested and increased from 30% to 34.2% (theoretical expectation 34.3%). The deviation Δθ = 34.2% - 34.3% = -0.1%, which is within the allowable range, indicating precise irrigation. The stress level was retested and decreased from level 2 to level 0, indicating that the stress was effectively relieved. No parameter correction is required.

[0132] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An irrigation decision-making method based on visual crop recognition and multimodal fusion of the Internet of Things, characterized in that, Includes the following steps: Step S1, Real-time access and timestamp alignment of multi-source data: The decision engine accesses raw data from two data sources, including IoT environmental monitoring data and visual recognition result data, according to a fixed polling cycle; The IoT environmental monitoring data includes soil parameters, meteorological parameters, and device feedback parameters; the visual recognition result data refers to the structured index data output by performing local inference calculations on crop canopy and leaf images captured by cameras using a lightweight model deployed on edge computing devices, rather than the original images; all data are aligned according to a unified timestamp after access to ensure that environmental data and visual data at the same time can participate in subsequent joint calculations. Step S2, Data Preprocessing and Anomaly Cleaning: Perform quality verification and preprocessing on the multi-source data accessed in Step S1, including removing abnormal jump values, completing missing data, filtering high-frequency noise, marking and excluding low-quality images, and normalizing data in each dimension. Step S3, Parallel Calculation of Basic Water Demand Indicators via Dual Paths: Two independent calculation paths are executed simultaneously. Path 1 calculates the theoretical soil-meteorological water demand by calculating the reference crop evapotranspiration, theoretical crop evapotranspiration, and soil moisture deficit based on soil and meteorological parameters from IoT environmental data. Path 2 generates water stress correction coefficients and growth correction coefficients based on crop water stress level and leaf area index from visual recognition results data. Step S4, Multimodal fusion to generate comprehensive actual water demand: Multiply the basic crop coefficient calculated by path one in step S3 with the growth correction coefficient calculated by path two to obtain the dynamic crop coefficient. Then multiply the dynamic crop coefficient with the reference crop evapotranspiration and the water stress correction coefficient in turn to obtain the corrected actual crop evapotranspiration. Finally, add the soil water deficit to the product of the corrected actual crop evapotranspiration and the planned irrigation interval days, and then subtract the predicted effective rainfall to obtain the comprehensive irrigation quota. This comprehensive irrigation quota includes the soil water deficit, the crop water consumption in the future period, and the predicted rainfall deduction. Step S5, Multi-level logical rule verification: For the comprehensive irrigation quota calculated in step S4, multi-level constraint verification is performed in order of priority from high to low, including soil threshold hard constraint verification, rainfall prediction constraint verification, time period constraint verification, stress priority logical judgment verification and equipment network constraint verification. If any level of verification fails, the decision engine will block or correct the current irrigation decision. Step S6, output the final irrigation decision instruction: After all the verifications in step S5 are passed, the decision engine outputs a structured irrigation decision instruction, which includes the decision conclusion, target irrigation amount, total irrigation duration, valve opening setting value, irrigation start and stop time, multi-plot rotation irrigation scheme and accompanying alarm information. Step S7, Closed-loop feedback self-optimization: After irrigation is completed, the system automatically enters the feedback correction stage. After the preset water balance time window, the soil moisture data and crop canopy images of the irrigation zone are collected again. By comparing the deviation between the actual retest data and the theoretical expected data, the key parameters in the decision model are automatically fine-tuned, and the classification threshold of the visual recognition model is iteratively updated periodically using the historical data accumulated over a long period of time, forming a complete closed-loop iterative optimization.

2. The irrigation decision-making method based on visual recognition and IoT multimodal fusion according to claim 1, characterized in that, The visual recognition result data in step S1 includes the following structured indicators: Water stress level, which is a discrete level from 0 to 4, where level 0 indicates that the crop is normal and has no stress, level 1 indicates mild water shortage, level 2 indicates moderate water shortage, level 3 indicates severe water shortage, and level 4 indicates wilting. The CWSI Crop Water Stress Index is a continuous numerical index that quantifies the degree of water shortage in crops by analyzing the temperature difference between the canopy and the air. The value ranges from 0 to 1, with a higher value indicating more severe water stress. The L-leaf area index is defined as the ratio of the total area of ​​all leaves per unit land area to the land area. It is calculated by performing pixel-level segmentation on the canopy image using the improved YOLO11-Seg lightweight semantic segmentation model. The NDVI growth index assesses vegetation vitality by analyzing the crop's reflectance characteristics to red and near-infrared light. Canopy coverage refers to the proportion of the projected area of ​​green vegetation to the total area of ​​the image when viewed from a vertical direction, and is directly output by the segmentation model. The percentage of leaf curling area and the percentage of yellow leaf area are used to help verify the accuracy of the water stress level assessment.

3. The irrigation decision-making method based on visual recognition and IoT multimodal fusion according to claim 2, characterized in that, Step S2, data preprocessing and anomaly cleaning, specifically includes: Sub-step S2-1, Remove abnormal jump values: Maintain a sliding window for each sensor data channel, calculate the mean and standard deviation of the valid data points in the window, and when the deviation of a newly connected data point from the mean of the window exceeds three times the standard deviation, it is judged as an abnormal jump value and discarded. Sub-step S2-2, missing value linear interpolation completion: When the sensor temporarily loses data due to communication interruption or power failure, the linear interpolation method is used to complete the data at the missing time point; when the missing time exceeds the preset duration or the consecutive missing data points exceed the preset proportion of the total data, the current data of the sensor is determined to be unreliable, and the irrigation decision of this area is downgraded to be based solely on visual data and meteorological data. Sub-step S2-3, moving average filtering noise reduction: The moving average filtering algorithm is used to smooth the sensor time series data, and the length of the moving window is dynamically adjusted according to the parameter change rate; Sub-step S2-4, Image quality judgment: Automatic quality assessment of images captured by the camera. Assessment indicators include average brightness, contrast and occlusion ratio. Low-quality images are marked as invalid and do not participate in this irrigation decision. The decision engine uses the previous valid visual data as a substitute and marks the data timeliness. Sub-step S2-5, data normalization: All data from different sensors are uniformly mapped to the range of 0 to 1 using the Min-Max normalization method to eliminate the influence of different units and numerical ranges on the fusion calculation.

4. The irrigation decision-making method based on visual recognition and IoT multimodal fusion according to claim 3, characterized in that, Step S3, path one specifically includes: Sub-step S3-1-1: Based on meteorological parameters such as air temperature, relative humidity, wind speed, and solar radiation, calculate the reference crop evapotranspiration using the FAO Penman-Monteith formula. The reference crop evapotranspiration reflects the intensity of atmospheric evaporation demand. Sub-step S3-1-2: Determine the basic crop coefficient based on the crop type and current growth stage, and multiply the basic crop coefficient by the reference crop evapotranspiration to obtain the theoretical crop evapotranspiration. Sub-step S3-1-3: Subtract the measured volumetric water content from the field water holding capacity, and then multiply by the planned wetting layer depth and soil bulk density to obtain the soil water deficit. A value greater than zero indicates that the soil is currently in a water-deficient state, and the larger the value, the greater the water deficit.

5. The irrigation decision-making method based on visual recognition and IoT multimodal fusion according to claim 4, characterized in that, Step S3, path two specifically includes: Sub-step S3-2-1: Obtain the water stress level and CWSI crop water stress index from the visual recognition results, and generate a water stress correction coefficient based on the water stress level: when the stress level is 0, the water stress correction coefficient is 1.0, and the original water requirement remains unchanged; when the stress level is 1, the water stress correction coefficient is between 1.1 and 1.2, and the supplementary water amount is appropriately increased; when the stress level is 2, the water stress correction coefficient is between 1.3 and 1.5, and the supplementary water amount is increased; when the stress level is 3 or above, the water stress correction coefficient is between 1.6 and 2.0, and emergency full irrigation is implemented; within the same level, the specific value of the water stress correction coefficient is calculated by linear interpolation based on the continuous values ​​of the CWSI crop water stress index; Sub-step S3-2-2: Obtain the L-leaf area index from the visual recognition result, compare the L-leaf area index with the standard L-leaf area index of the crop under the current growth stage conditions, and calculate the growth correction coefficient; when the L-leaf area index is higher than the standard value, the growth correction coefficient is greater than 1, and the crop coefficient is adjusted upward to match the vigorous transpiration demand; when the L-leaf area index is lower than the standard value, the growth correction coefficient is less than 1, and the crop coefficient is adjusted downward to prevent over-irrigation; the value range of the growth correction coefficient is limited to between 0.5 and 1.

5.

6. The irrigation decision-making method based on visual recognition and IoT multimodal fusion according to claim 5, characterized in that, The coercion priority logic judgment and verification in step S5 specifically includes: When there is a discrepancy between the visually identified crop water stress level and the soil moisture data, the visually identified crop water stress level shall be the highest priority criterion. When soil moisture data is within the normal range, but a moderate or higher level of stress is visually detected, it is judged as a latent physiological water shortage state. This indicates that the crop may actually be in a water shortage state due to root diseases, excessive soil salinity, or excessive transpiration caused by high temperature and low humidity. At this time, small-volume supplementary irrigation should be carried out, and the irrigation volume should be carried out according to the preset ratio of the normal calculation volume. When soil moisture data is below the lower limit threshold for irrigation, but no visual stress is detected and the crop is lush and green with a leaf area index not lower than the preset proportion of the standard value, it is determined to be a compensatory drought-resistant state of the crop. This indicates that the crop roots can absorb water from the deep soil not covered by the sensor and have not yet been subjected to actual stress. At this time, it is advisable to delay irrigation or only do a small amount of maintenance water.

7. The irrigation decision-making method based on visual recognition and IoT multimodal fusion according to claim 6, characterized in that, The soil threshold hard constraint verification in step S5 specifically includes: determining whether the current soil moisture content is higher than the upper limit of the field suitable moisture content for the current growth stage of the crop. If it is higher than the upper limit, irrigation is prohibited regardless of whether the slight stress level is visually identified, in order to prevent waterlogging damage. The time period constraint verification specifically includes: determining whether the current time period is a high-temperature noon period or a low-temperature dew period at night. If so, the initiation of new irrigation is restricted. However, when severe water shortage stress is visually detected and the soil moisture content is lower than the wilting coefficient, the emergency irrigation mode is triggered, which overrides the time period constraint and immediately starts irrigation. The equipment pipeline constraint verification specifically includes: verifying the maximum flow rate of the irrigation pump, the pressure bearing capacity of the pipeline, and the capacity limit of the zonal rotation irrigation. When the single irrigation quota exceeds the upper limit of the equipment capacity, the current irrigation will be automatically split into multiple rotation irrigations.

8. The irrigation decision-making method based on visual recognition and IoT multimodal fusion according to claim 7, characterized in that, Step S7 specifically includes: Sub-step S7-1: Within the preset water balance time window after the irrigation is completed, reread the soil moisture sensor data of each layer of the irrigation zone to obtain the actual water content after irrigation, and at the same time re-capture the crop canopy image and input it into the model to obtain the stress level and leaf area index after irrigation. Sub-step S7-2 compares the actual soil moisture content after irrigation with the theoretical expected moisture content to calculate the deviation, and compares the visual stress level after irrigation with the stress level before irrigation to assess the degree of crop stress relief. Sub-step S7-3: When a systematic deviation in irrigation volume is determined to exist multiple times in a row, the mapping relationship parameter between the stress correction coefficient and the stress level and the calculation formula parameter of the growth correction coefficient for the corresponding plot are automatically adjusted. Sub-step S7-4 records and stores the entire process data of each irrigation decision in the cloud platform's historical database. When the accumulated data reaches a preset scale, the classification threshold of the visual recognition model is iteratively updated and optimized using the accumulated data, forming a complete closed-loop iterative mechanism.

9. The irrigation decision-making method based on visual crop recognition and IoT multimodal fusion according to claim 8, characterized in that, It also includes the step of linking visual recognition with irrigation mode control: The growth status of crops is continuously monitored through a visual recognition model to identify the key phenological stages of the crops, including the budding stage, the vegetative growth stage, the flowering and fruiting stage, and the maturity and harvest stage. When the system detects that the crop is in the germination stage, the water requirement is small and needs to be precisely supplied to the roots. The system decides to start the drip irrigation mode and adjusts the drip irrigation flow by controlling the opening of the electric ball valve. When the system detects that the crop is in the vegetative growth stage, it decides whether to use drip irrigation or micro-sprinkler irrigation. When the system detects that the crop is in the flowering and fruiting stage, which requires the most water and needs to moisten the canopy to assist in cooling and humidification, the system decides to switch to sprinkler irrigation mode and increase the opening of the electric ball valve. When the system detects that the crop is at maturity and ready for harvest, it decides to switch back to drip irrigation mode to reduce leaf humidity and prevent fruit cracking and disease. The results of crop phenology and irrigation mode decision instructions obtained from visual recognition are simultaneously uploaded to the Internet of Things cloud platform. After comprehensive verification and calculation by the cloud platform in combination with soil temperature and humidity and meteorological data, the final precise control instructions are generated and issued to the execution layer. When there is a discrepancy between the cloud computing results and the edge decision, a joint arbitration mechanism between the cloud and the edge is initiated, and the arbitration result is pushed to the administrator's mobile APP for final confirmation by humans.

Images