Cotton film under drip irrigation method based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method for drip irrigation under mulch film in cotton, belonging to the field of agricultural water-saving irrigation technology. The method includes the following steps: constructing a multi-source heterogeneous sensing network to collect multimodal data on soil profiles, meteorological environment, and crop canopy; constructing a physical information-guided deep learning network PIDL-Net, coupling the physical priors of the crop growth mechanism model under mulch film with deep learning to predict the daily water requirement of cotton; based on automatic identification of the growth stage, using a dynamic soil moisture threshold with field capacity as the benchmark for adaptive irrigation decisions; achieving closed-loop precise execution of the drip irrigation system through independent zone control of electromagnetic valves, adaptive PID regulation of pipeline pressure, and precise water metering; and using a federated learning framework for online continuous optimization of the model. This invention achieves intelligent and precise management of the entire process of drip irrigation under mulch film in cotton, saving 20%–30% of water and improving irrigation water use efficiency by 25%–40%, balancing yield assurance and efficient water resource utilization.

Description

Technical Field

[0001] This invention relates to the field of agricultural water-saving irrigation technology, and in particular to an artificial intelligence-based method for cotton drip irrigation under mulch film. Specifically, it is a precise and intelligent control method for cotton drip irrigation under mulch film that integrates multi-source sensor data acquisition, deep neural network water demand prediction, dynamic threshold adaptive control, closed-loop execution of the drip irrigation system, and online continuous optimization of the model. Background Technology

[0002] Cotton is one of the world's most important economic crops, widely cultivated in arid and semi-arid regions of my country, such as the Xinjiang Uygur Autonomous Region. However, these regions suffer from extreme water scarcity, with agricultural irrigation accounting for over 90% of total water consumption, while the utilization rate of traditional irrigation methods is less than 40%. Drip irrigation under mulch film, a highly efficient water-saving irrigation method combining mulch film cultivation with drip irrigation systems, offers multiple advantages such as water conservation, salt suppression, temperature increase, and yield enhancement, and has been widely applied in cotton cultivation. This technology achieves highly efficient coupling of water and fertilizer through a comprehensive management system of "mulching—moisture control—uniform irrigation—fertilizer saving—salt suppression—temperature increase—weed control."

[0003] However, existing cotton drip irrigation technology still has the following prominent problems in practical applications:

[0004] (1) Data collection dimension is single and spatiotemporal resolution is low: Traditional irrigation decision-making relies on a few soil moisture sensors, lacking synchronous perception of multi-dimensional data such as deep soil salinity, crop canopy physiological state, and field micro-meteorology, resulting in a serious lack of data coverage.

[0005] (2) Irrigation decisions rely on static experience thresholds: Existing automated irrigation systems usually use fixed upper and lower limits of soil moisture as the opening and closing thresholds, without considering the differentiated response of cotton to water stress at different growth stages and the real-time changes in meteorological conditions, resulting in a mismatch between irrigation timing and water volume and actual demand.

[0006] (3) Insufficient model accuracy and lack of adaptability: Some existing irrigation decision-making methods based on machine learning either use only a single model for water demand prediction and fail to effectively integrate the synergistic effects among soil, crops and weather; or the model parameters are fixed and cannot be continuously optimized as field data accumulates.

[0007] (4) Limited control precision of drip irrigation execution system: The existing drip irrigation system lacks fine control over execution links such as pipeline pressure fluctuation and solenoid valve opening adjustment, resulting in insufficient irrigation uniformity.

[0008] In the existing technology, some intelligent irrigation solutions have been proposed. For example, patent CN120678006A discloses a farmland irrigation control method using multi-sensor fusion and edge computing nodes, but it does not specifically optimize for the special characteristics of cotton drip irrigation under plastic film, such as the soil water and heat transport patterns under plastic film mulch and the differences in water requirements of cotton at different growth stages. Patent CN120851405A uses genetic algorithms and reinforcement learning for deficit irrigation decisions, but this method mainly relies on crop growth model simulation, does not fully utilize real-time multi-source sensing data, and does not form a closed-loop linkage with the drip irrigation execution system. In addition, the covering effect of plastic film under drip irrigation significantly changes the spatiotemporal transport patterns and evaporation and transpiration processes of soil moisture. Traditional crop water requirement calculation methods based on the FAO Penman-Monteith formula are prone to significant deviations when directly applied to cotton fields under drip irrigation. Moreover, most existing artificial intelligence irrigation methods do not consider the impact of this special cultivation mode on water transport processes, leading to a decrease in the accuracy of water requirement prediction.

[0009] In summary, there is an urgent need to develop a precise and intelligent irrigation method that integrates artificial intelligence technology and is specifically designed for cotton drip irrigation under plastic film cultivation, in order to address the aforementioned shortcomings of existing technologies. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an artificial intelligence-based method for drip irrigation under mulch film in cotton. This method achieves intelligent and precise management of the entire process of drip irrigation under mulch film in cotton by collaborative sensing of multi-source heterogeneous sensor data, deep learning water demand prediction guided by physical information, adaptive dynamic threshold control during the growth period, closed-loop precise execution, and online continuous optimization of the model. Under the premise of ensuring cotton yield and quality, it maximizes water resource utilization efficiency, reduces manual management costs, and promotes the green and sustainable development of cotton planting.

[0011] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0012] An artificial intelligence-based method for drip irrigation of cotton under mulch film includes the following steps:

[0013] Step S1: Acquisition and Fusion of Multi-Source Heterogeneous Sensing Data

[0014] A multi-source heterogeneous sensing network is constructed in cotton-growing areas. This sensing network includes the following three types of data acquisition units:

[0015] S1.1 Soil Profile Sensing Unit: Multiple soil profile monitoring points are arranged in a grid pattern within the cotton field. Each monitoring point has a multi-parameter soil sensor embedded in three layers vertically. These three layers correspond to the surface root zone (0–20 cm), the main root zone (20–40 cm), and the deep root zone (40–60 cm) of the cotton root system distribution area, respectively. The sensors at each layer collect the following soil parameters in real time: soil volumetric moisture content (accuracy ±2%), soil temperature (accuracy ±0.3℃), and soil electrical conductivity (accuracy ±5%). Data acquisition is performed every 10 minutes. Furthermore, the multi-parameter soil sensor employs a combination of a capacitive soil moisture sensor and an ion-selective electrode to achieve simultaneous monitoring of soil water and salt content.

[0016] S1.2 Meteorological Environment Sensing Unit: Automatic micro-weather stations are deployed in the core area of ​​the cotton field to collect the following meteorological parameters in real time: air temperature, relative humidity, solar radiation intensity, wind speed, wind direction, atmospheric pressure, and rainfall. Data collection frequency is once every 10 minutes.

[0017] Preferably, the micro weather station is also equipped with an evaporation dish for real-time monitoring of water evaporation in the field, providing measured data for the calibration of reference crop evapotranspiration.

[0018] S1.3 Crop Canopy Sensing Unit: A multispectral camera and a thermal infrared camera are mounted on a rotatable gimbal within the cotton field to capture images of the cotton canopy from multiple angles at preset time intervals. The preset time intervals are preferably between 10:00 AM and 2:00 PM daily. The multispectral camera acquires canopy reflectance images in five bands: blue light (450±20 nm), green light (550±20 nm), red light (650±20 nm), red edge (720±20 nm), and near-infrared (850±40 nm). The thermal infrared camera acquires spatial distribution images of the canopy surface temperature (7.5–13.5 μm). Furthermore, a drone equipped with a hyperspectral imager can be used to conduct aerial photography 1–2 times per week to acquire spectral information over a wide area of ​​the cotton field.

[0019] S1.4 Multi-source heterogeneous data fusion: Data collected by the three types of sensing units mentioned above are transmitted to edge computing nodes or cloud servers via low-power wide-area wireless communication technologies such as LoRa or NB-IoT. Preprocessing operations are performed on the raw data at the edge computing nodes, including: outlier removal and sliding window mean smoothing of soil sensor data; timestamp alignment and missing value imputation of meteorological data; radiometric calibration, geometric correction, background segmentation, and region of interest extraction of crop canopy multispectral images; and temperature calibration and noise filtering of thermal infrared images. Then, the preprocessed soil and meteorological data are constructed into structured time-series tensors, and the canopy image data is encoded into multi-dimensional image feature tensors. Through data timestamp alignment, a multi-source heterogeneous fusion dataset under a unified spatiotemporal benchmark is formed.

[0020] Step S2: Predicting water demand for cotton under drip irrigation based on physical information using a deep learning network

[0021] A Physics-Informed Deep Learning Network (PIDL-Net) is constructed to couple the prior physical knowledge of a crop growth mechanism model (preferably an improved AquaCrop model suitable for drip irrigation under mulch film) with a data-driven deep learning model to predict the daily water requirement ETc of cotton under drip irrigation under mulch film conditions.

[0022] PIDL-Net consists of the following subnetworks:

[0023] S2.1 Spatiotemporal Feature Encoding Subnetwork: This subnetwork comprises three parallel encoding branches. The first branch employs a convolutional neural network (CNN) with ResNet-50 as the backbone, extracting crop physiological state feature vectors such as canopy coverage, canopy temperature, and vegetation indices (e.g., NDVI, NDRE, CWSI) from the multispectral and thermal infrared canopy images obtained in step S1.3. The second branch employs a temporal convolutional network (TCN), containing 4–6 stacked residual blocks, each containing two dilated causal convolutional layers with dilation coefficients increasing exponentially by 2 (d=1, 2, 4, 8, 16), encoding the time-series data of soil profile parameters and meteorological parameters collected in steps S1.1 and S1.2, and extracting temporal feature vectors. The third branch employs a lightweight Transformer encoder, encoding historical irrigation events (irrigation time, irrigation volume, fertilizer application) and cotton growth stage identifiers, and extracting irrigation management feature vectors.

[0024] S2.2 Cross-modal attention fusion module: A multi-head cross-attention mechanism is used to deeply fuse the feature vectors output from the three branches mentioned above. Specifically, the soil-meteorological time-series feature vector is used as the query, and the image feature vector and irrigation management feature vector are used as the key and value, respectively. By calculating attention weights, information interaction and weighted fusion between different modal features are realized to generate a multi-modal fused feature representation.

[0025] S2.3 Mechanism Model-Guided Predictive Backbone Network: This network consists of a three-layer fully connected neural network with 256, 128, and 64 neurons in the hidden layers, respectively, and uses ReLU as the activation function. The forward computation process of the network is constrained by the crop growth mechanism model. Specifically, a locally calibrated improved AquaCrop model is used to simulate the daily evapotranspiration of drip-irrigated cotton under mulch, and the simulated ET value is used as the basis for the prediction. c-aqua With the network's output prediction value ET c-pred The difference between them is included as a physical constraint loss function term in the total loss function of network training. The total loss function is:

[0026] L total =λ1·MSE(ET c-pred ET c-obs )+ λ2·MSE(ET c-pred ET c-aqua )+ λ3·||W||2

[0027] Among them, ET c-obs The historical measured water demand (which can be obtained through the water balance method or a large lysimeter) is represented by W, which is the network weight, and λ1, λ2, and λ3 are the weight coefficients of each loss term, which are 1.0, 0.3, and 0.01, respectively.

[0028] S2.4 Model Training and Validation: Multi-source data and corresponding measured water requirement data from cotton fields for three or more consecutive growing seasons were collected and divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The Adam optimizer was used for training with an initial learning rate of 0.001, employing a cosine annealing learning rate decay strategy, a batch size of 32, and 200–300 training epochs. After training, the daily water requirement prediction sequence ET for cotton over the next 1–7 days was output. c-pred (t+1)~ET c-pred (t+7).

[0029] Step S3: Adaptive irrigation decision based on dynamic thresholds for reproductive period identification

[0030] Based on the water demand sequence predicted in step S2, and combined with the current growth stage of the cotton, a dynamic irrigation decision scheme is generated.

[0031] S3.1 Automatic Identification of Cotton Growth Stages: Using the canopy multispectral image obtained in step S1.3, two time-series indices, Normalized Difference Vegetation Index (NDVI) and Canopy Cover (CC), are calculated to construct a growth stage discrimination model. Specifically, when NDVI ≤ 0.35 and CC ≤ 15%, it is determined to be the seedling stage; when 0.35 < NDVI ≤ 0.65 and 15% < CC ≤ 45%, it is determined to be the budding stage; when 0.65 < NDVI ≤ 0.85 and 45% < CC ≤ 80%, it is determined to be the flowering and boll-forming stage; when NDVI > 0.85 and CC > 80%, it is determined to be the peak boll-forming stage; when NDVI begins to decrease and CC tends to stabilize, it is determined to be the boll-opening stage. The automatic identification results of the growth stages are used to determine the irrigation threshold in subsequent steps.

[0032] S3.2 Calculation of Dynamic Soil Moisture Thresholds for Subsurface Drip Irrigation: Considering the differentiated response characteristics of cotton to water stress at different growth stages, stage-specific dynamic soil moisture upper and lower limit thresholds based on field capacity (FC) are set. The preferred soil moisture control thresholds for each growth stage in this invention are as follows:

[0033] Reproductive stage Lower limit of irrigation (%FC) Irrigation limit (%FC) Seedling stage 55～65 75～85 budding period 60～70 80～90 Flowering season 70～80 85～100 Peak season 65～75 80～95 boll opening period 50～60 65～75

[0034] Furthermore, the above thresholds are not fixed but are dynamically adjusted based on the real-time meteorological conditions collected in step S1.2. When no effective rainfall is forecast for the next 3 days and the maximum temperature exceeds 35℃, the lower limit threshold for irrigation at each growth stage is automatically increased by 5-8 percentage points; when rainfall is forecast to exceed 5 mm in the next 24 hours, the lower limit threshold for irrigation is automatically decreased by 5-10 percentage points. Simultaneously, when the soil conductivity monitored in step S1.1 exceeds the crop salt tolerance threshold (e.g., 2.5 dS / m), the salt leaching irrigation mode is triggered, increasing the irrigation volume by 15%-25% from the normal decision volume to leach salt below the root zone.

[0035] S3.3 Calculation of Single Irrigation Quota: An irrigation event is triggered when the real-time soil volumetric moisture content in the main root zone (20–40 cm) of the cotton field is lower than the irrigation lower limit threshold corresponding to the current growth stage. The single irrigation quota I (m³ / acre) is calculated using the following formula:

[0036] I = H (θ 上限 -θ 实测 ) · S · γ · η · 10

[0037] In the formula: θ 上限 θ represents the upper limit of soil volumetric moisture content (m³ / m³) for irrigation during the current growth stage. 实测The measured soil volumetric moisture content (m³ / m³) at the time of triggering irrigation is given, H is the planned wetting layer depth in the main root zone (taken as 0.4 m), S is the unit area (mu), γ is the soil bulk density (g / cm³), η is the drip irrigation water utilization coefficient (taken as 0.90~0.95 under drip irrigation under film), and 10 is the unit conversion factor.

[0038] S3.4 Irrigation Timing Optimization: After the irrigation event is triggered, the specific execution time of irrigation is optimized by combining the water demand for the next 3 days predicted in step S2 and the weather forecast information. If moderate or above rainfall (rainfall ≥ 10 mm) is forecast within the next 48 hours, irrigation will be postponed until after the rainfall ends to avoid deep seepage and water waste caused by immediate rainfall after irrigation. If the current soil moisture content is close to the lower limit of irrigation and the future forecast is for continuous high temperature and drought, irrigation will be carried out 1-2 days in advance to prevent crop water stress.

[0039] Step S4: Closed-loop precision execution and adaptive adjustment of the drip irrigation system

[0040] The irrigation decision scheme generated in step S3 is transformed into precise control instructions for the drip irrigation system and executed in a closed loop.

[0041] S4.1 Independent Control of Solenoid Valves by Zone: The cotton field drip irrigation network is divided into several independent control zones according to a preset irrigation zoning scheme. Each zone covers an area of ​​10–50 mu (approximately 6.7–3.3 hectares). Each zone is equipped with an intelligent electric regulating valve with valve position feedback function and a flow meter. The irrigation controller generates differentiated irrigation instructions for each zone based on the real-time soil moisture content of each zone and the decision results from step S3, including valve opening degree (0–100%) and irrigation duration accurate to the minute. Preferably, the dripper spacing is 30 cm, the dripper design flow rate is 1.8–2.6 L / h, and the capillary spacing is 60–90 cm, i.e., one pipe in two rows or one pipe in four rows. More preferably, for machine-harvested cotton fields, a 1-film, 2-pipe, 6-row arrangement is adopted, with a dripper flow rate of 2.6 L / h.

[0042] S4.2 Adaptive PID Control for Pipeline Pressure: Pressure sensors are installed at key nodes in the main and branch pipelines of the drip irrigation system to monitor pipeline pressure in real time. An incremental PID control algorithm is used to perform closed-loop regulation of the variable frequency pump speed. The input to the PID controller is the deviation between the set pressure value and the measured pressure value, and the output is the frequency adjustment of the frequency converter. When a fluctuation in pipeline pressure is detected due to the opening or closing of a certain zone valve, the PID controller completes pressure compensation regulation within 5 to 10 seconds, stabilizing the pipeline pressure within ±5% of the set working pressure, thereby ensuring the stability and uniformity of the dripper flow rate.

[0043] S4.3 Precise Measurement and Anomaly Diagnosis of Drip Irrigation Water Volume: During irrigation, the irrigation volume is accumulated in real time through the flow meters of each zone. When the accumulated irrigation volume reaches 95% to 100% of the target irrigation quota calculated in step S3.3, the control system automatically closes the solenoid valve of the corresponding zone. At the same time, the operating status parameters of the drip irrigation system are continuously monitored during irrigation. When the following abnormalities occur, an alarm is triggered and corresponding measures are taken: (a) If the flow rate of a certain zone is continuously lower than 70% of the set value and the pipeline pressure is normal, it is judged to be dripper blockage, and flushing and maintenance are prompted; (b) If the flow rate of a certain zone is continuously higher than 130% of the set value, it is judged to be pipeline damage and leakage, the valve of that zone is automatically closed and an alarm message is pushed; (c) If the frequency of the variable frequency pump exceeds 95% of the rated frequency and still cannot reach the set pressure, it is judged to be head blockage or excessive filter pressure difference, and backflushing operation is prompted.

[0044] Step S5: Online continuous optimization of the model based on the federated learning framework

[0045] To improve the model's generalization ability and long-term adaptability, a federated learning framework is used to continuously optimize the PIDL-Net model online.

[0046] S5.1 Local Model Incremental Update: After each irrigation cycle (e.g., 7 days), the PIDL-Net model is incrementally trained on edge computing nodes using newly accumulated measured data from that cycle, including soil moisture changes, meteorological data, canopy images, and actual ETC values ​​calculated using the water balance method. Incremental training only updates some parameters of the model (mainly the parameters of the fully connected layers and attention modules), while the backbone network parameters are fine-tuned to adapt to changes in the field environment while maintaining model stability.

[0047] S5.2 Federated Aggregation Update: After completing the incremental update of the local model, the edge computing nodes of each cotton field encrypt and upload the model parameters to the cloud server. The cloud server uses the FedAvg algorithm to weighted aggregate the local model parameters from multiple cotton fields to generate a globally optimized model.

[0048] W global (t+1) = ∑ k=1 k ·(n k / N)·W k (t+1)

[0049] Where K is the number of cotton fields participating in federated learning, and n k Let W be the number of local training samples in the k-th cotton field, N be the total number of samples, and W be the number of samples in the k-th cotton field. k (t+1) These are the local update model parameters for the k-th cotton field.

[0050] S5.3 Global Model Distribution and Version Management: The cloud server distributes the aggregated global model parameters to the edge computing nodes of each cotton field, replacing the original local model. The cloud server maintains the model's version history. When a significant decrease in the prediction accuracy of a new version model is detected, it can quickly roll back to the previous stable version, ensuring the reliability of irrigation decisions.

[0051] Step S6: Visual monitoring and human-machine collaborative intervention

[0052] Build a WebGIS-based visualization monitoring and management platform to realize visualization monitoring and human-machine collaborative intervention throughout the entire irrigation process.

[0053] S6.1 Real-time Data Visualization: The platform displays the geographical location and zoning of each cotton field in map form, and shows in real time the soil moisture content, temperature, electrical conductivity of each monitoring point, as well as various parameters of the weather station and crop canopy growth indicators. The data refresh rate is 1 minute.

[0054] S6.2 Irrigation Plan Push and Confirmation: After each irrigation decision plan is generated, the platform pushes irrigation suggestions to the administrator via mobile app or SMS. The suggestions include: irrigation zones, suggested irrigation start time, irrigation quota, and estimated irrigation duration. Administrators can choose "Automatic Execution" (default mode) or "Execution after Manual Review." In manual review mode, administrators can modify irrigation parameters and issue execution instructions after confirmation.

[0055] S6.3 Historical Data Statistics and Analysis: The platform automatically records detailed data such as the time, zone, irrigation volume, and irrigation duration of each irrigation event, generating irrigation water use statistics reports and water-saving effect comparison charts. Furthermore, based on the cumulative irrigation volume, effective rainfall, and final yield data during the cotton growing season, the platform automatically calculates irrigation water use efficiency (IWUE) and water use efficiency (WUE), providing a reference for irrigation management in the next season.

[0056] Beneficial effects

[0057] Compared with the prior art, the present invention has the following significant advantages:

[0058] (1) High accuracy in water demand prediction: By constructing a physical information-guided deep learning network (PIDL-Net), the prior physical knowledge of the crop growth mechanism model under drip irrigation under film is combined with a data-driven deep neural network, making full use of the effective information in multi-source heterogeneous data, which significantly improves the prediction accuracy of water demand for cotton under drip irrigation. Compared with pure data-driven deep learning methods, the root mean square error (RMSE) of prediction is reduced by about 18% to 25%.

[0059] (2) Strong adaptability of irrigation decision: The present invention adopts a dynamic threshold irrigation decision method based on automatic identification of growth stage, which overcomes the defect that the traditional fixed threshold irrigation method cannot adapt to the dynamic changes in cotton growth, and makes the irrigation timing and water volume highly matched with the actual water demand of cotton.

[0060] (3) Significant water-saving and yield-increasing effects: Through the accurate prediction of PIDL-Net and the adaptive decision of dynamic threshold, as well as the closed-loop precise execution of the drip irrigation system, this invention can achieve water saving of 20% to 30% while maintaining or even increasing cotton yield, and improve irrigation water use efficiency by 25% to 40%.

[0061] (4) Strong continuous optimization capability of the system: By introducing a federated learning framework, the model is continuously optimized online. Under the premise of protecting the data privacy of each cotton field, new field data is continuously absorbed to improve the model performance, making irrigation decisions more and more accurate.

[0062] (5) Good adaptability of drip irrigation under film: This invention fully considers the special influence of film covering on soil moisture transport and evaporation and transpiration under drip irrigation under film conditions. It has made special designs for drip irrigation under film in terms of data acquisition scheme, mechanism model selection, model calibration method and irrigation threshold setting, so as to ensure the applicability and reliability of the method in practical applications.

[0063] (6) Human-machine collaboration is friendly: the entire irrigation process is transparently monitored and humanized through a visual management platform, which not only ensures the high degree of automation of the system, but also retains the necessary intervention authority for managers, thus taking into account both intelligence and controllability. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the overall process of the cotton drip irrigation method based on artificial intelligence according to the present invention.

[0065] Figure 2 This is a schematic diagram of the structure of the multi-source heterogeneous sensing network in this invention.

[0066] Figure 3 This is a schematic diagram of the architecture of the Physical Information Guided Deep Learning Network (PIDL-Net) in this invention.

[0067] Figure 4 This is a flowchart of the dynamic threshold irrigation decision-making process based on reproductive stage identification in this invention.

[0068] Figure 5 This is a schematic diagram of the closed-loop execution control of the drip irrigation system in this invention.

[0069] Figure 6 This is a schematic diagram of the architecture of the federated learning framework in this invention.

[0070] Figure 7 This is a schematic diagram of the interface of the visualization monitoring and management platform of the present invention. Detailed Implementation

[0071] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. It should be noted that these embodiments are only used to explain the present invention and are not intended to limit the present invention.

[0072] The core working mechanism of this invention's artificial intelligence-based cotton drip irrigation method lies in: using real-time acquisition and fusion of multi-source heterogeneous sensing data as a foundation, employing a deep learning-based water demand prediction model guided by physical information as the decision-making core, and combining automatic identification of cotton growth stages with a dynamic threshold adaptive adjustment strategy. Through the closed-loop precise execution of the drip irrigation system and continuous model optimization under a federated learning framework, a complete intelligent closed-loop system is formed, encompassing perception, cognition, decision-making, execution, and evolution. Its specific working process is as follows:

[0073] (1) Multi-source data perception and fusion mechanism

[0074] This invention achieves a three-dimensional perception of the soil-meteorology-crop system by deploying soil profile sensing units, meteorological environment sensing units, and crop canopy sensing units within cotton fields. The soil profile sensing units simultaneously collect soil moisture, temperature, and electrical conductivity at three key depths (0–20 cm, 20–40 cm, and 40–60 cm) along the cotton root system distribution to capture the water transport patterns and salt accumulation status in the vertical profile under drip irrigation conditions. The meteorological environment sensing units record micrometeorological parameters such as solar radiation, temperature, humidity, wind speed, and rainfall in real time, providing data support for calculating reference crop evapotranspiration. The crop canopy sensing units utilize multispectral and thermal infrared imaging technology to obtain the spatial distribution of canopy reflectance and temperature, thereby retrieving key physiological indicators such as NDVI, canopy coverage, and crop water stress index (CWSI). The aforementioned multi-source data is transmitted to edge computing nodes via LoRa or NB-IoT wireless communication technologies. Under a unified timestamp benchmark, data cleaning, interpolation, and spatial registration are completed to form a spatiotemporally aligned multi-source heterogeneous fusion dataset, providing a high-fidelity data foundation for subsequent water demand forecasting and irrigation decisions.

[0075] (2) The principle of water demand prediction guided by physical information

[0076] Traditional crop water requirement prediction methods either rely on purely empirical formulas (such as the FAO Penman-Monteith formula multiplied by the crop coefficient), which fail to adequately consider the inhibitory effect of mulch film on evapotranspiration under drip irrigation conditions; or they rely on purely data-driven deep learning models, lacking constraints from crop growth mechanisms, which are prone to overfitting or predictions that do not conform to physical laws when data volume is limited. The Physical Information Guided Deep Learning Network (PIDL-Net) constructed in this invention organically couples these two approaches. Its working principle is as follows: First, it uses a CNN branch to extract crop physiological state features from canopy images, a TCN branch to capture the temporal evolution of water consumption from soil-meteorological time-series data, and a Transformer branch to encode structured information such as irrigation management and growth stages. Then, it achieves deep fusion of cross-modal features through a multi-head cross-attention mechanism, enabling the model to fully explore the synergistic effects among soil, crop, and meteorology. Finally, during forward propagation, the water requirement simulated by a locally calibrated improved AquaCrop mechanism model is used as physical prior knowledge, constraining the optimization direction of network parameters in the form of an additional loss function term. This mechanism-data dual-driven learning paradigm enables the model to learn complex nonlinear mapping relationships from historical data while also following the basic physical laws of crop water transport, thereby significantly improving prediction accuracy and generalization ability under limited sample conditions.

[0077] (3) Dynamic threshold adaptive decision-making principle

[0078] Cotton exhibits significant differences in its sensitivity to water stress and water requirement intensity at different growth stages. During the seedling stage, vegetative growth is dominant, and moderate drought promotes root development, allowing for a lower water threshold. The budding and boll-forming stages are critical for yield formation and are extremely sensitive to water stress, requiring a higher soil moisture level. During the boll-opening stage, moderate water control is necessary to promote boll opening and lint release. This invention utilizes canopy spectral information to automatically identify the current growth stage and accordingly set staged upper and lower soil moisture thresholds based on field capacity. These thresholds are not static but dynamically adjusted according to real-time weather forecasts: when a heatwave is forecast, the lower irrigation threshold is automatically raised to prevent water stress; when effective rainfall is forecast, the threshold is automatically lowered to avoid water waste caused by immediate rainfall after irrigation. Furthermore, when soil conductivity exceeds the crop's salt tolerance threshold, the system automatically triggers a salt-leaching irrigation mode, increasing irrigation by 15%–25% on top of the normal amount, leaching salt from the root zone to deeper layers and maintaining a suitable salinity environment in the root zone. This three-level adaptive decision-making mechanism, which combines "basic threshold for reproductive stage + dynamic meteorological adjustment + salt stress correction," ensures a precise match between irrigation timing and water volume and the actual needs of cotton.

[0079] (4) Closed-loop precision execution principle of drip irrigation system

[0080] After the irrigation decision plan is generated, it needs to be precisely executed by the drip irrigation system to translate into actual water-saving and yield-increasing effects. This invention employs a triple mechanism of independent zone control, adaptive pressure adjustment, and precise water metering to achieve closed-loop precise execution. First, the cotton field drip irrigation network is divided into several independent control zones. Each zone generates differentiated irrigation instructions based on its measured soil moisture and the decision results, achieving refined management tailored to each zone. Second, during irrigation, pipeline pressure sensors monitor network pressure changes in real time. An incremental PID controller adjusts the speed of the variable frequency pump based on the deviation between the set pressure and the actual pressure, completing pressure compensation within 5-10 seconds to ensure the pipeline pressure remains stable within ±5% of the set working pressure, thus guaranteeing uniform and stable dripper flow. Finally, the flow meter accumulates the irrigation volume in real time. When the target irrigation quota is reached (95%-100%), the solenoid valve automatically shuts off, completing the irrigation task. Meanwhile, the system continuously monitors operating parameters such as flow rate and pressure, and automatically diagnoses and alarms abnormalities such as dripper blockage, pipeline leakage, and excessive filter pressure difference, ensuring the long-term stable and reliable operation of the drip irrigation system.

[0081] (5) The principle of online continuous optimization of models under the federated learning framework

[0082] The cotton field environment is a dynamic and complex system, with soil physicochemical properties, crop variety characteristics, and irrigation regimes shifting over time and space. If the PIDL-Net model is fixed after training, its prediction accuracy will gradually decrease with environmental changes. This invention introduces a federated learning framework to achieve continuous online optimization. Its working principle is as follows: After each irrigation cycle, each cotton field edge computing node incrementally trains its local model using newly accumulated measured data from that cycle, updating only some network layer parameters to maintain model stability while adapting to local environmental changes. Subsequently, each node uploads the encrypted model parameters to the cloud server, where a federated averaging algorithm is used for weighted aggregation to generate a globally optimized model. Finally, the cloud distributes the aggregated global model to each edge node, completing the model update. This mechanism, while protecting the data privacy of each cotton field, fully utilizes distributed data resources to continuously improve model performance, making irrigation decisions increasingly accurate. The cloud server also maintains the model's version history; when a new version of the model experiences a decrease in accuracy, it can quickly roll back to the previous stable version, ensuring the reliability of irrigation decisions.

[0083] This invention constructs a complete closed-loop intelligent management system for cotton drip irrigation under mulch film through the organic synergy of five links: multi-source sensing, intelligent prediction, dynamic decision-making, precise execution, and continuous optimization. It realizes full-chain intelligence from data to decision-making to execution, and provides a systematic solution for the efficient use of water resources in cotton planting in arid and semi-arid regions.

[0084] Example 1

[0085] This example was conducted in a cotton field in Xinjiang, covering an area of ​​200 mu. The soil type was sandy loam, the field water holding capacity (FC) was 0.28 m³ / m³, the soil bulk density was 1.42 g / cm³, and the cotton variety was "Xinluzhong 66".

[0086] I. Deployment of Multi-Source Sensing Network

[0087] Deploy the sensing network according to the requirements of step S1:

[0088] Soil profile sensing unit: Eight soil monitoring points are set up in a 50 m × 50 m grid. At each point, a capacitive soil three-parameter sensor is buried at a depth of 20 cm, 40 cm and 60 cm. Data is uploaded through a LoRa wireless module, and the sampling frequency is once every 10 minutes.

[0089] Meteorological environment sensing unit: An automatic weather station is set up in the center of the cotton field, equipped with air temperature and humidity sensors, solar radiation sensors, wind speed and direction sensors, rain gauges and evaporation pans, and the data collection frequency is once every 10 minutes.

[0090] Crop canopy sensing unit: Two high-pole rotatable gimbals are erected at the boundary of the cotton field, each equipped with a multispectral camera and a thermal infrared camera, which automatically capture canopy images at 12:00 noon every day.

[0091] II. PIDL-Net Model Training

[0092] Multi-source data and corresponding measured water requirement data for the first three growing seasons (2023–2025) of the cotton field were collected. The measured water requirement was obtained using both the water balance method and data from a large lysimeter (2.5 m in diameter, 2.0 m in depth). The improved AquaCrop model was calibrated using local experimental data, achieving a coefficient of determination (R²) of 0.856. The dataset was partitioned into 7:1.5:1.5 subsets, and model training was performed on an NVIDIA A100 GPU server. During training, the validation set loss converged after 120 epochs. The final root mean square error (RMSE) of the model prediction on the test set was 0.38 mm / d, and the coefficient of determination (R²) was 0.91.

[0093] III. Irrigation Implementation and Effectiveness Evaluation

[0094] The trained PIDL-Net model was deployed on an edge computing node (NVIDIA Jetson Orin) and communicated with the drip irrigation controller via the Modbus TCP protocol. The drip irrigation system used a 1-film, 2-pipe, 6-row arrangement, with a dripper flow rate of 2.6 L / h and a dripper spacing of 30 cm. During the 2026 growing season, irrigation was triggered 12 times throughout the growing season, with a total irrigation volume of 245 m³ / mu, saving 23.4% of water compared to traditional irrigation (320 m³ / mu). The final seed cotton yield was 485 kg / mu, a slight increase (+2.8%) compared to the yield of the traditional irrigation method (472 kg / mu). The irrigation water use efficiency improved from 1.48 kg / m³ of the traditional method to 1.98 kg / m³, an improvement of 33.8%.

[0095] IV. Continuous Optimization of Federated Learning

[0096] During the growing season, the system incrementally updates the local model every 7 days using newly accumulated measured data, and performs federated aggregation with edge nodes of three surrounding cotton fields at a fixed time each month. After a full growing season of online optimization, the model's predicted RMSE further decreased to 0.31 mm / d in the second growing season, with improvements in both water-saving effect and yield stability.

[0097] Example 2

[0098] This embodiment further verifies the effectiveness of the dynamic threshold adjustment strategy based on Embodiment 1. A control area (using fixed threshold irrigation: the lower limit of irrigation throughout the entire growth period was uniformly set at 65% FC) was set up in the same cotton field. The control area covered 30 mu (approximately 2 hectares), and the other conditions were the same as in Embodiment 1. The results showed that in the main area using the dynamic threshold irrigation strategy of this invention, the average soil moisture content during the budding and boll-forming stages was maintained at approximately 78% FC and 85% FC, respectively, which highly matched the needs of cotton at these stages. In contrast, due to the limitation of the fixed threshold, the control area experienced mild water stress during the budding stage (average moisture content dropped to 61% FC), leading to an increased boll shedding rate. Ultimately, the seed cotton yield in the main area (485 kg / mu) was significantly higher than that in the control area (452 ​​kg / mu), an increase of 7.3%, while the irrigation amount in the main area (245 m³ / mu) was slightly lower than that in the control area (258 m³ / mu). This result fully demonstrates the superiority of the dynamic threshold irrigation strategy of this invention.

[0099] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications, equivalent substitutions, and improvements to the technical solutions of the present invention without departing from the principles and spirit of the invention, and all such modifications, equivalent substitutions, and improvements should be covered within the protection scope of the present invention.

Claims

1. A method for subsurface drip irrigation of cotton based on artificial intelligence, characterized in that: Includes the following steps: Step S1: Construct a multi-source heterogeneous sensing network, including a soil profile sensing unit, a meteorological environment sensing unit, and a crop canopy sensing unit. Collect soil parameters, micro-meteorological parameters of cotton fields, and multispectral and thermal infrared images of crop canopy at different depths in the cotton root zone. Perform timestamp alignment and preprocessing on the collected multi-source data to form a multi-source heterogeneous fusion dataset. Step S2: Construct a physical information-guided deep learning network PIDL-Net, which couples the physical prior knowledge of the cotton growth mechanism model under drip irrigation with the deep learning model. Using the multi-source heterogeneous fusion dataset obtained in Step S1 as input, predict the daily water requirement sequence of cotton under drip irrigation. Step S3: Based on the water demand sequence predicted in Step S2, and combined with the current growth stage of cotton automatically identified based on canopy spectral information, determine the upper and lower limits of dynamic soil moisture based on field capacity. When the measured soil moisture content is lower than the lower limit of irrigation for the current growth stage, trigger an irrigation event and calculate the single irrigation quota. Step S4: Transform the irrigation decision generated in step S3 into control commands for the drip irrigation system. Achieve closed-loop precision execution through independent control of solenoid valve zones, adaptive PID regulation of pipeline pressure, and precise metering of drip irrigation water. Step S5: Using a federated learning framework, the PIDL-Net model is locally incrementally updated on edge computing nodes using newly accumulated measured data from each irrigation cycle. The model parameters of each cotton field are uploaded to the cloud for federated aggregation, and a globally optimized model is generated and distributed to each edge node to achieve online continuous optimization of the model.

2. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 1, characterized in that, In step S1: The soil profile sensing unit has multiple monitoring points arranged in a grid pattern in the cotton field. Each monitoring point has three layers of soil multi-parameter sensors buried in the vertical direction. The three layers correspond to the surface root zone (0-20 cm), the main root zone (20-40 cm), and the deep root zone (40-60 cm), respectively, and collect soil volumetric water content, soil temperature, and soil electrical conductivity in real time. The crop canopy sensing unit uses a multispectral camera and a thermal infrared camera to acquire canopy reflectance images and canopy surface temperature distribution images in five bands: blue light, green light, red light, red edge, and near infrared. Data is transmitted to edge computing nodes or cloud servers via low-power wide-area wireless communication technology, and preprocessing operations, including outlier removal, sliding window smoothing, missing value interpolation, and image radiometric geometric correction, are performed on the edge computing nodes.

3. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 1, characterized in that, The physical information-guided deep learning network PIDL-Net mentioned in step S2 includes: The spatiotemporal feature coding subnetwork is composed of parallel CNN, TCN and Transformer branches, which respectively extract crop physiological state feature vectors from canopy images, extract temporal feature vectors from soil profile parameters and meteorological parameter time series, and extract irrigation management feature vectors from historical irrigation events and growth stage identifiers. The cross-modal attention fusion module uses a multi-head cross-attention mechanism to deeply fuse the feature vectors output by the three branches; the mechanistic model-guided prediction backbone network is composed of a multi-layer fully connected neural network, and its training loss function introduces a physical constraint term to account for the difference between the simulated value of the mechanistic model and the predicted value of the network.

4. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 3, characterized in that, The physical constraint term is based on a locally calibrated improved AquaCrop model, which is specifically calibrated for crop evapotranspiration under drip irrigation conditions. The mean square error between the model's output mechanistic simulated water demand and the network-predicted water demand is used as the physical constraint loss term. The total loss function is: L total =λ1·MSE(ET c-pred , AND c-obs )+ λ2·MSE(ET c-pred , AND c-aqua )+ λ3·||W||2 Among them ET c-pred ET c-pred To measure actual water demand, ET c-aqua To simulate water demand using a mechanistic model, λ1, λ2, and λ3 are the network weights, and λ1, λ2, and λ3 are the weight coefficients.

5. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 1, characterized in that, The automatic identification method for cotton growth stages in step S3 is as follows: using the canopy multispectral image obtained in step S1, two time-series indicators, Normalized Difference Vegetation Index (NDVI) and Canopy Coverage (CC), are calculated to construct a growth stage discrimination model based on decision rules; the stage-specific dynamic soil moisture thresholds are based on Field Holding Capacity (FC), and the lower and upper limits of irrigation for each growth stage are as follows: seedling stage 55%–65% FC and 75%–85% FC, bud stage 60%–70% FC and 80%–90% FC, boll-forming stage 70%–80% FC and 85%–100% FC, full boll stage 65%–75% FC and 80%–95% FC, and boll-opening stage 50%–60% FC and 65%–75% FC.

6. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 5, characterized in that, Step S3 also includes dynamically fine-tuning the soil moisture threshold based on real-time meteorological conditions: when there is no effective rainfall forecast for the next 3 days and the maximum temperature exceeds 35℃, the lower limit threshold for irrigation at each growth stage is automatically increased by 5 to 8 percentage points; when rainfall is forecast to exceed 5 mm in the next 24 hours, the lower limit threshold for irrigation is automatically decreased by 5 to 10 percentage points; when the monitored soil conductivity exceeds the crop salt tolerance threshold, the salt leaching irrigation mode is triggered, increasing the irrigation amount by 15% to 25% based on the normal decision amount.

7. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 1, characterized in that, In step S3, the single irrigation quota I is calculated using the following formula: I = H (θ 上限 -θ 实测 ) · S · c · h · 10 In the formula: θ 上限 θ represents the upper limit of soil volumetric moisture content for irrigation during the current growth stage. 实测 To determine the measured soil volumetric water content at the time of irrigation, H is the planned wetting layer depth of the main root zone, taken as 0.4 m, S is the unit area, γ is the soil bulk density, η is the water utilization coefficient of drip irrigation under film, taken as 0.90 to 0.95, and 10 is the unit conversion factor. Furthermore, the timing of irrigation is optimized by combining the predicted future water demand and weather forecast from step S2: if the predicted rainfall is ≥10 mm in the next 48 hours, irrigation is postponed until after the rainfall ends; if the current soil moisture content is close to the lower limit of irrigation and the forecast is for continuous high temperature and drought, irrigation is carried out 1 to 2 days in advance.

8. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 1, characterized in that, In step S4: The solenoid valve zone independent control divides the cotton field drip irrigation network into several independent control zones. Each zone is equipped with an intelligent electric regulating valve and a flow meter. Differentiated irrigation instructions are generated based on the real-time soil moisture content and irrigation decisions of each zone. The drip irrigation tape adopts an arrangement scheme with a dripper spacing of 30 cm, a dripper design flow rate of 1.8 to 2.6 L / h, and a capillary tube spacing of 60 to 90 cm. The adaptive PID control of pipeline pressure is achieved by monitoring pipeline pressure in real time and using an incremental PID control algorithm to adjust the speed of the variable frequency water pump in a closed loop, so that the pipeline pressure is stabilized within ±5% of the set working pressure of 0.08 to 0.15 MPa; at the same time, the operating status parameters are continuously monitored, and abnormal conditions such as dripper blockage, pipeline leakage and filter blockage are automatically diagnosed and alarmed.

9. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 1, characterized in that, The specific process of the federated learning framework described in step S5 is as follows: After each irrigation cycle ends, the newly accumulated measured data during that cycle is used to incrementally train the PIDL-Net model on the edge computing node, updating only some network layer parameters. Each cotton field edge computing node uploads encrypted model parameters to the cloud server, and the cloud uses the FedAvg algorithm to perform weighted aggregation to generate a global optimization model; The cloud server maintains the model version history and automatically rolls back to the previous stable version when the prediction accuracy of the new version model drops significantly.

10. The method for subsurface drip irrigation of cotton based on artificial intelligence according to claim 1, characterized in that, It also includes step S6: building a WebGIS-based visualization monitoring and management platform to display real-time data on soil, weather and crop canopy in each cotton field, push irrigation decision suggestions to managers and support two modes: automatic execution or execution after manual review, automatically record detailed data of each irrigation event and generate irrigation water statistical reports and water-saving effect comparison and analysis charts, and automatically calculate irrigation water use efficiency and water use efficiency based on the cumulative irrigation water volume, effective rainfall and final yield data during the growing season.

Images