Farmland water-saving irrigation water system design and dynamic water distribution optimization method based on supply and demand multi-source data fusion

Abstract

The invention relates to a supply and demand multi-source data fusion farmland water-saving irrigation water system design and dynamic water distribution optimization method, which is characterized by comprising the following steps: S1, crop-soil-weather multi-source data coupling modeling, S2, water-saving water system design optimization, S3, dynamic water-saving water distribution strategy, and S4, intelligent terminal control system control. According to the scheme, accurate configuration of water resources is achieved through the artificial intelligence technology, the irrigation water consumption is reduced to the maximum extent, and meanwhile the crop growth requirement is met.

Description

Technical Field

[0001] The present invention relates to the field of smart agriculture and efficient water resource utilization technology, and specifically to an artificial intelligence-based farmland water-saving irrigation water system design and dynamic water distribution optimization method and system. It uses intelligent algorithms to achieve efficient planning and dynamic scheduling of irrigation water systems, with minimizing water resource consumption as the core goal, while taking into account crop water demand and ecological sustainability. Background Art

[0002] Traditional agricultural irrigation systems suffer from extensive design, low water resource utilization, and over-reliance on groundwater. For example, channel leakage, evaporation losses, and inappropriate irrigation timing lead to agricultural water waste rates exceeding 40%. Existing technologies often focus on single-point water distribution efficiency or cost optimization, lacking systematic design for water conservation objectives and making it difficult to dynamically adjust them based on real-time environmental data. Therefore, a multi-objective optimization method centered on water conservation is urgently needed to reduce water waste throughout the entire process, from source design to dynamic scheduling. Summary of the Invention

[0003] The purpose of this invention is to provide an irrigation water system design and dynamic water distribution optimization method with agricultural water conservation as the core, realize the precise allocation of water resources through artificial intelligence technology, minimize irrigation water consumption, and ensure crop growth needs at the same time.

[0004] (1) Full-chain water-saving optimization: Water-saving constraints are embedded in the entire process, from water system design (reducing water transmission losses), water source configuration (coordination of multiple water sources), to irrigation execution (precise water distribution on demand).

[0005] (2) Dynamic simulation of crop phenology in farmland: Based on the prediction of meteorological conditions and soil moisture, combined with the natural growth model of crops, farmland phenology is modeled to determine the ideal growth state, thereby dynamically adjusting artificial irrigation.

[0006] (3) Groundwater protection mechanism: Introduce recycled water for agricultural irrigation, increase the utilization rate of reclaimed water in the optimization algorithm, and replace groundwater extraction.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for designing a farmland water-saving irrigation system and dynamically optimizing water distribution based on the fusion of supply and demand multi-source data, the method comprising the following steps:

[0009] S1 "crop-soil-meteorology" multi-source data coupling modeling,

[0010] S2 water-saving water system design optimization,

[0011] S3 dynamic water-saving and water distribution strategy,

[0012] Controlled by S4 intelligent terminal control system.

[0013] Wherein, step S1 is specifically as follows:

[0014] Key water-saving parameters such as regional soil moisture, crop water use efficiency (WUE), root depth, and evapotranspiration coefficient are input. A "crop-soil-climate" coupled water-saving demand model is constructed to quantify the minimum ecological water requirement and the maximum allowable water shortage threshold. The implementation process is as follows:

[0015] 1.1 Multi-source data monitoring,

[0016] Multiple sensors (such as soil moisture sensors, meteorological sensors, crop monitoring sensors, etc.) are distributed in the fields to monitor soil moisture, temperature, weather, crop water requirements and other data in real time. The details are as follows: Intelligent soil moisture sensor: used to monitor the moisture content of different soil layers in the fields in real time. The sensors can be distributed in key locations in the fields and transmit data to the central control system through wireless networks.

[0017] Sensor density calculation: This sensor placement optimization method, based on an improved adaptive PSO algorithm, uses the numerical solution of the soil water transport equation (Richards' equation) as a particle search direction correction term compared to traditional methods. It also dynamically adjusts the crop root weight distribution matrix using NDVI time series data. This method involves reconstructing a multi-objective fitness function and constructing a three-dimensional evaluation index by integrating the spatial coefficient of variation (CV) of soil properties, crop root weights (RW), and meteorological influence factors (MF).

[0018] Fitness=α·CV+β·RW+γ·MF+λ·Cost

[0019] Where, Fitness is the fitness value; α, β, γ are dynamic weight coefficients; λ is the economic constraint factor; Cost is the comprehensive economic constraint term, which is used to quantify the implementation cost of the sensor deployment plan.

[0020] The second is the farmland-specific improvement strategy, including hierarchical particle coding (separately encoding the two-dimensional plane coordinates and the burial depth, and using a hybrid discrete-continuous search space), the introduction of a mutation operator (when a particle falls into a local optimum, Gaussian mutation is triggered according to the standard deviation of the crop ridge width), and an energy consumption constraint model (integrating the wireless signal attenuation equation to calculate the communication energy consumption cost). All data collected by sensors are uploaded to the data terminal control system through wireless communication technologies (such as LoRa and NB-IoT) to ensure timely data updates and remote monitoring. The transmitted real-time data can support the dynamic change monitoring of the field environment.

[0021] 1.2 Build the model architecture, as follows:

[0022] 1.2.1 Soil water transport modeling, unsaturated zone water flow model (improved Richards equation),

[0023]

[0024] θ: water content in the soil; t: time; z: depth; K(θ): unsaturated hydraulic conductivity, K sat is the saturated hydraulic conductivity, b is the gain and offset coefficient calculated by the ground calibration plate; h(θ): soil water pressure head, h e is the intake pressure head; S root (z,t): root water absorption rate, S root (z,t)=a(z)·T p (t), a(z) is the root density distribution function, T p (t) is the potential transpiration rate, the most unfavorable point determination (definition of water transport hysteresis index)

[0025]

[0026] WTLI: Temporal Water Deficit Index; θ crit : critical soil moisture content; θ(t): actual soil moisture content as a function of time t; t c : time period for observation or evaluation; In the time window [0,t c ], the cumulative deficit of soil moisture below the critical value; θ crit ·t c : Theoretical maximum deficit, 1.2.2 Crop response module

[0027] Construct a coupling algorithm between Jarvis stomatal conductance model and Feddes root water absorption function,

[0028] For the Jarvis stomatal conductance model:

[0029] g s =g smax ·f(Q)·f(Dq)·f(ψ s )·f(T)

[0030] Where, f(Q) is the photosynthetic active radiation response function; f(Dq) is the water vapor pressure difference suppression function; f(ψ s ) is the soil water potential response function; f(T) is the temperature regulation function, and for Feddes root water absorption function:

[0031] S (z) =s(ψ s )·Tp ·La (z)

[0032] Where S (z) is the water absorption of the root system at depth z per unit volume of soil; s(ψ s ) is the water stress coefficient; T p is the potential transpiration rate; L (z) is the root density distribution function,

[0033]

[0034] Where, ψ s is the soil water potential; ψ1 is the anaerobic threshold; ψ2 is the upper limit of optimal water absorption; ψ3 is the starting point of linear decline; ψ4 is the permanent wilting point,

[0035]

[0036] Where, La (z) is the root density distribution function; z s is the half-decay depth of root distribution, ranging from 20 cm to 100 cm; z is the depth; m is the distribution shape parameter, m = 1, the root density decreases slowly and linearly with the depth; m = 5, the root density decreases sharply with the depth, and the roots are concentrated in the surface layer.

[0037] 1.2.3 Weather drive module,

[0038] The Penman-Monteith dual-source model was used to distinguish the contribution of soil / vegetation evapotranspiration.

[0039] Vegetation transpiration:

[0040]

[0041] Where, ET veg is the transpiration of vegetation; Δ is the slope of the saturated water vapor pressure-temperature curve; R n,v is the canopy net radiation; ρ a is the air density; c p is the specific heat of air at constant pressure, Dq is the water vapor pressure deficit of air; is the canopy aerodynamic drag; is the total stomatal resistance of the canopy; ∈ is the psychrometric constant,

[0042] Soil evaporation:

[0043]

[0044] Where, ET soil is the soil evaporation; Δ is the slope of the saturated water vapor pressure-temperature curve; ρ a is the air density; cp is the specific heat of air at constant pressure, Dq is the air vapor pressure deficit; R n,s is the net radiation from the soil surface; is the aerodynamic resistance of the soil surface; r soil is the soil evaporation resistance; ∈ is the psychrometric constant,

[0045] 1.2.4 Establish a dynamic response function between water stress and photosynthate distribution.

[0046] Water stress index:

[0047]

[0048] Where WSI is water stress index; ET is act is the actual transpiration; ET p is the potential transpiration, corrected for total photosynthate:

[0049]

[0050] Where, P total Corrected for total photosynthetic product; P max is the maximum photosynthetic capacity (without stress); g s is the actual stomatal conductance; g smax is the maximum stomatal conductance.

[0051] In the above scheme,

[0052] By coupling the multiple processes of water transport in the unsaturated zone, stomatal response, and dual-source evapotranspiration, a dynamic feedback mechanism for the SPAC system was constructed. The Richards equation was modified to include a microbial activity correction term, enhancing the accuracy of identifying critical water points. The Jarvis-Feddes simultaneous algorithm was used to simulate root water uptake and stomatal oscillations. Furthermore, the water stress index and photosynthate partitioning function were integrated to reveal the threshold response patterns of carbon-water coupling. Compared to traditional single-process models, this model significantly reduces systematic errors and can analyze minute-scale water redistribution processes.

[0053] Among them, the S2 water-saving water system design optimization is as follows:

[0054] 2.1 Construction of system framework terrain data: Extract terrain slope and confluence network based on digital elevation model (DEM) to drive gradient optimization of gravity water delivery path;

[0055] Soil data: Calculate channel leakage rate and regional infiltration loss using the spatial distribution of soil permeability coefficients. Crop data: Divide differentiated irrigation scheduling units based on the spatial heterogeneity of crop types and water requirements during their growth periods.

[0056] Meteorological data: Combined with rainfall contour interpolation results, quantify the regional rainwater resource utilization potential and guide the spatial layout of rainwater collection facilities;

[0057] 2.2 Multi-objective optimization, minimizing water loss, maximizing irrigation uniformity, and maximizing recycled water utilization are selected as the goals, and a multi-objective water-saving irrigation optimization model is established, as follows:

[0058] (1) Water loss objective function (reducing leakage and evaporation through channel anti-seepage design and pipeline transformation):

[0059]

[0060] Where N c is the total number of channel segments; L i is the length of the i-th channel; q i is the leakage per unit length of the i-th channel; is the channel anti-seepage efficiency coefficient; N p is the total number of pipeline segments; P j is the surface area of ​​the jth section of open-air pipeline; E j is the evaporation rate per unit area; is the pipe coverage correction factor,

[0061] q i =K s,i W i ·H i

[0062] Where q i is the leakage per unit length of the i-th channel; K s,i is the saturated permeability coefficient of the soil below the i-th channel section; W i is the channel bottom width; H i Design water depth for the channel,

[0063] P j =π·D j ·L j

[0064] Where, P j is the surface area of ​​the jth section of open-air pipeline; D j is the pipe diameter; L j is the pipe length.

[0065]

[0066] Where, is the wind speed correction coefficient; Δe j is the air saturation vapor pressure difference; P atm is atmospheric pressure,

[0067] (2) Irrigation uniformity objective function (using drip irrigation and sprinkler irrigation zoning design to reduce local excessive irrigation):

[0068]

[0069] constraint:

[0070]

[0071] Where n is the total number of irrigation units; d k is the water supply of the kth irrigation unit; μ d is the average water supply per irrigation unit, ET c A is the water requirement of crops; total is the total irrigated area,

[0072] (3) Recycled water utilization objective function (integrating rainwater collection systems with reclaimed water reuse facilities to replace groundwater extraction):

[0073]

[0074] Where η r For rainwater collection efficiency, D total is the total irrigation water requirement, D total =ET c ·A total t,V r is the volume of the rainwater collection tank; U w is the daily treatment capacity of reclaimed water;

[0075] Ic is the irrigation cycle; A c is the catchment area; Dr is the design rainfall; Rc is the runoff coefficient.

[0076] In the above scheme,

[0077] For the first time, GIS spatial heterogeneity analysis and multi-objective optimization algorithms are coupled. Through DEM-driven water system topology optimization and collaborative modeling of soil-crop-rainfall multi-source data, precise Pareto frontier searches are achieved for water transmission losses (joint quantification of Darcy infiltration and Penman evaporation), irrigation uniformity (Christiansen coefficient constraints) and recycled water utilization (integration of rainwater and greywater systems). This solves the problem of traditional single-objective optimization that ignores spatial explicit constraints and resource coordination, and significantly improves water-saving efficiency.

[0078] Among them, the S3 dynamic water-saving and water distribution strategy is as follows:

[0079] 3.1 Soil Moisture Prediction (Lag-PhysFusion):

[0080] (1) Historical time series data:

[0081]

[0082] Input features: X t =[H t ,L t ,W t ,I t ,V t ] T

[0083] Where, X t is the input feature vector at time t; S t is the measured soil moisture value at time t; T is the total length of the time series, that is, the maximum time step of the historical data; H t Soil moisture calculated by the improved Richards equation (stratified data, dimension Da); L t is the hysteresis index (scalar, reflecting the delayed effect of water migration); W t is meteorological data (temperature, precipitation, evaporation, etc.); I t is the irrigation volume; V t is the vegetation index; S t is the measured soil moisture value, the output value (S t ∈R).

[0084] (2) Physical feature extraction:

[0085] Extracting physical dynamics by improving the discrete solution of Richards equation:

[0086]

[0087] Discrete form (finite difference method):

[0088]

[0089] Where SW is soil water potential; t is time; K(SW) is unsaturated hydraulic conductivity; zz is vertical spatial coordinate; YH is source-sink term; SW t (d) is the moisture of the dth layer of soil at time t; Δt is the time step; Δzz 2 is the spatial step length; is the hydraulic conductivity of the ddth layer at time t; is the source and sink term of the dth layer at time t, and the physical model output and the hysteresis index are fused into enhanced features:

[0090]

[0091] Where, is the vector after physical feature encoding; ReLU is the activation function; [H t ;L t ] is the concatenation of input features; W p is the weight matrix; b p is the bias vector,

[0092] (3) Lag-perceived time coding (the lag index L t Dynamically embedded time position coding):

[0093]

[0094] Where EB t is the standard Transformer position encoding; W L is a learnable parameter used to adjust the degree of time distortion of the lag effect; L t is the lag index,

[0095] (4) Spatiotemporal Attention Mechanism:

[0096]

[0097] Where C X is the query matrix; J is the key matrix; Z is the value matrix; M L is the lagged similarity matrix, τ is the temperature parameter that controls the sensitivity of the hysteresis difference, LZ i , LZ j is the hysteresis index; d k is the key vector dimension,

[0098] (5) Multi-scale temporal convolution:

[0099]

[0100] Where, is the output feature of the k-th convolution at time t; is the sth weight parameter of the kth layer convolution; is the time step ts×d (k) The input feature vector of d (k) is the expansion factor of the kth layer; S is the number of time steps covered by the convolution kernel,

[0101] (6) Dynamic fusion prediction:

[0102] Combined with physical characteristics P t With time series feature C t :

[0103]

[0104] Fusion weight ξ t is dynamically generated by the hysteresis index:

[0105] ξ t =σ(w ξ ZH t +b ξ )

[0106] Where, is the predicted output; t is the dynamic fusion weight; f phy (P t ) is the physical model branch; f seq (C t ) is the sequence model branch; ZH t is the hysteresis index; σ is the Sigmoid function, which controls the contribution ratio of the physical model and the data-driven model; w ξ is the weight parameter; b ξ is the bias parameter,

[0107] 3.2 Annual rainfall forecast (MPF-Net model):

[0108] (1) Input data:

[0109] Target variable: Y t ∈R y (historical annual rainfall series, y is the year);

[0110] Climate factor: X t ∈R (y×m) (m climate indices, such as ENSO, PDO, AMO, etc.);

[0111] Geographical features: G∈R (p) (p static geographic parameters, such as latitude, longitude, elevation, etc.).

[0112] (2) Wavelet multi-scale decomposition: t Perform Mallat wavelet decomposition:

[0113]

[0114] Where Y t is the target variable; is the j-th scale detail component (capturing interannual variability); A J (t) is the approximate component (representing the long-term trend); J is the number of decomposition levels,

[0115] (3) Physical Constraint Feature Engineering:

[0116] Energy constraints of the climate system (introducing the mass-energy conservation equation as a regularization term):

[0117]

[0118] Where, is the climate system energy; v is the atmospheric vertical velocity field; HD is the water vapor flux divergence; λ1 and λ2 are learnable constraint coefficients.

[0119] Geographical features embedded:

[0120] Γ=ReLU(W g G+b g )⊙Sigmoid(W t Ω+b t )

[0121] Where W g G, W t T is the learnable weight matrix; Ω is the temporal encoding vector,

[0122] (4) Multimodal spatiotemporal attention mechanism:

[0123] Climate factors cross attention: Where Q = W q [A J (t); Γ]; K = W k [X t ;G];V=W v [X t ; G],

[0124] Where, Attn Q is the output attention feature, reflecting the spatiotemporal modulation effect of climate factors on rainfall variability; O is the dynamic query matrix; is the wavelet detail component, representing the interannual variability signal; d k is the scaling factor; VZ is the value matrix; A J (t) is the wavelet approximation component; Γ is the geographic modulation matrix; W q is the learnable weight matrix; G is the geographic feature vector; W k , W v is the learnable weight matrix,

[0125] Time-frequency dual-domain gating:

[0126] g t =σ(W g [D j (t); Attn Q ])

[0127]

[0128] Where g t is the gate vector; σ is the Sigmoid activation function; W gis the learnable gating weight matrix; For the corrected detail components, data-driven features are integrated with physical constraints.

[0129] (5) Adaptive learning of mutation points:

[0130] Construct a change point probability model for Bayesian change point detection:

[0131]

[0132] Where c t ∈{0,1} represents the change point indicator variable; h t is hidden state; W c is the learnable weight matrix; h t-1 ;h t is the hidden state vector of the adjacent time step; σ is the Sigmoid function,

[0133] Dynamic parameter adjustment:

[0134]

[0135] In the formula, MO is the key parameter of the model, which realizes the mutation adaptation of the parameter space; MO t is the key parameter of the model at the current moment; MO0 is the initial parameter value; ΔMO is the parameter adjustment step; c i is the cumulative sum of the historical change point indicator variables,

[0136] (6) Physical Information Neural Network Architecture:

[0137]

[0138] Where, is the prediction error; is the KL divergence regularization term for change point detection; Loss of physical restraint; is the adaptive weight parameter,

[0139] (7) Uncertainty Quantification (Monte Carlo Physics Dropout):

[0140] Keep dropout turned on during prediction and perform N samplings:

[0141]

[0142] Where, is the model prediction output of the i-th Monte Carlo sampling; CN is the number of sampling times; is the network parameter after randomly discarding some neurons during the i-th sampling; σ 2 is the prediction variance,

[0143] 3.3 Short-term rainfall prediction (MDSTF model):

[0144] (1) Select key indicators and observe them: air humidity, temperature, wind speed, air pressure, etc.; (2) Decompose the time series data of each indicator:

[0145] 3.4 Simulation of ideal crop growth state,

[0146] Based on predictions of future rainfall, soil moisture and other data, in order to predict the ideal growth state of crops in fields under given environmental conditions, an innovative framework integrating crop growth mechanism model, time series deep learning and dynamic feedback optimization is proposed, named Hybrid Crop Growth Digital Twin (HCG-DT). In the above solution,

[0147] This method innovatively combines climate prediction optimization of deep reinforcement learning with the real-time dynamics of crop growth, realizing closed-loop irrigation decision-making at multiple time scales. Compared with traditional static water distribution or single time scale methods, it significantly improves the global utilization efficiency of water resources and the sustainability of groundwater while ensuring crop yields. This method first proposed the Lag-PhysFusion model, MPF-Net model, MDSTF model, HCG-DT model and deep reinforcement learning DRL model to predict soil moisture, weather, and ideal growth status of crops, with significant results. Among them, the S4 intelligent terminal control system is controlled: drones regularly fly to monitor the fields, take real-time images of crops, and upload real-time data to the intelligent control terminal for unified decision-making. According to real-time data, prediction results and calculation results of the irrigation model, the terminal controller controls the intelligent sprinkler equipment to ensure the accuracy of irrigation operations.

[0148] The algorithms in this control terminal include judging the crop growth conditions in each area through image analysis; judging the moisture content of the soil in each area through soil moisture movement; making an overall judgment on multiple data based on future rainfall conditions, and controlling smart valves in various locations to carry out irrigation.

[0149] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for designing a farmland water-saving irrigation system and for dynamic water distribution optimization based on the fusion of multi-source supply and demand data is implemented.

[0150] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the farmland water-saving irrigation water system design and dynamic water distribution optimization method based on the fusion of supply and demand multi-source data.

[0151] Compared to existing technologies, the present invention offers the following advantages: The dynamic water distribution irrigation system proposed in this application achieves theoretical innovation and technological breakthroughs in precision irrigation through the deep integration of multi-physics field coupling modeling and multi-scale intelligent decision-making. The system uses a modified Richards equation combined with the Lag-PhysFusion model to describe non-equilibrium water transport. The system corrects the unsaturated hydraulic conductivity parameterization bias using a hysteresis kernel function, significantly reducing soil moisture prediction errors compared to traditional methods. Based on this, a combined Jarvis-Feddes model is constructed, coupling the dynamic response of stomatal conductance with root water uptake heterogeneity. A dynamic response function for photosynthate allocation is introduced to quantify the effects of water stress, significantly improving transpiration efficiency. The Penman-Monteith dual-source model is used to separate soil and vegetation evapotranspiration contributions. Combined with canopy temperature and leaf area index features analyzed from drone multispectral imagery (CNN classification accuracy of 92.4%), a crop growth digital twin (HCG-DT model) is established, achieving accurate characterization of evapotranspiration with an error of ≤0.3 mm / day. To meet the needs of multi-timescale regulation, the MPF-Net model is used to predict annual rainfall types and drive deep reinforcement learning (DRL) to generate long-term water-saving strategies. Through multi-objective optimization, water transmission losses are reduced by more than 20% and the utilization rate of recycled water reaches more than 80%. At the same time, an MDSTF model is established to achieve short-term high-precision rainfall forecasts (TS score above 0.8). Combined with model predictive control (MPC), a rolling optimization mechanism is constructed. Based on real-time soil moisture data and the most unfavorable point identification algorithm, the irrigation decision response time is compressed. The closed-loop control system achieves real-time field-level regulation (delay <1s) through edge computing. Three years of field verification have shown that the system has significantly improved water-saving efficiency, yield-increasing effects, water use efficiency, and irrigation uniformity coefficient compared to traditional irrigation. Through the multi-dimensional collaboration of mechanism models and data-driven, this solution overcomes the problem of delayed irrigation decisions caused by spatiotemporal heterogeneity, providing a verifiable engineering paradigm for smart agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0152] Figure 1 Schematic diagram of the operating mechanism of the intelligent control terminal. DETAILED DESCRIPTION

[0153] In order to deepen the understanding of the present invention, this embodiment is described in detail below with reference to the accompanying drawings.

[0154] Example 1: See Figure 1 A method for designing a water system for farmland water-saving irrigation and optimizing dynamic water distribution based on the fusion of multi-source data on supply and demand is provided. The method comprises the following steps:

[0155] S1 “Crop-Soil-Meteorology” Multi-Source Data Coupling Modeling

[0156] Key water-saving parameters such as input regional soil moisture, crop water use efficiency (WUE), root depth, and evapotranspiration coefficient are used; a "crop-soil-meteorology" coupled water-saving demand model is constructed to quantify the minimum ecological water requirement and the maximum allowable water shortage threshold.

[0157] The details are as follows:

[0158] 1.1 Multi-source data monitoring

[0159] Multiple sensors (such as soil moisture sensors, weather sensors, and crop monitoring sensors) are deployed throughout the field to monitor soil moisture, temperature, weather conditions, crop water requirements, and other data in real time. The following details are provided: Intelligent Soil Moisture Sensors: These sensors monitor the moisture content of different soil layers in the field in real time. These sensors can be placed at key locations throughout the field and transmit data to a central control system via a wireless network.

[0160] Meteorological data collection device: including temperature and humidity sensors, wind speed and direction meters, air pressure sensors, etc., which collect high-temporal and spatial resolution satellite precipitation data, three-dimensional distribution of solar radiation, canopy microclimate correction coefficient and other meteorological conditions in real time to provide accurate environmental data.

[0161] Crop growth monitoring sensor: Through the plant growth monitor, real-time data such as crop growth and leaf moisture are collected to assist in determining the water requirement of crops.

[0162] Soil moisture sensor: monitors changes in soil moisture, especially changes in moisture in deep soil layers, to further improve the accuracy of irrigation decisions.

[0163] Sensor density calculation: This sensor placement optimization method, based on an improved adaptive PSO algorithm, uses the numerical solution of the soil water transport equation (Richards' equation) as a particle search direction correction term compared to traditional methods. It also dynamically adjusts the crop root weight distribution matrix using NDVI time series data. This method involves reconstructing a multi-objective fitness function and constructing a three-dimensional evaluation index by integrating the spatial coefficient of variation (CV) of soil properties, crop root weights (RW), and meteorological influence factors (MF).

[0164] Fitness=α·CV+β·RW+γ·MF+λ·Cost

[0165] Where Fitness is the fitness value; α, β, and γ are dynamic weight coefficients; λ is the economic constraint factor; and Cost is the comprehensive economic constraint term, which is used to quantify the implementation cost of the sensor deployment plan.

[0166] The second is the farmland-specific improvement strategy, including hierarchical particle coding (separately encoding the two-dimensional plane coordinates and the burial depth, and adopting a hybrid discrete-continuous search space), the introduction of mutation operators (when the particles fall into the local optimum, the Gaussian mutation is triggered according to the standard deviation of the crop ridge width) and the energy consumption constraint model (integrating the wireless signal attenuation equation to calculate the communication energy consumption cost).

[0167] All data collected by sensors are uploaded to the data terminal control system through wireless communication technologies (such as LoRa, NB-IoT, etc.) to ensure timely data updates and remote monitoring. The transmitted real-time data can support the dynamic change monitoring of the field environment.

[0168] 1.2 Model Architecture

[0169] 1.2.1 Soil Moisture Movement Modeling

[0170] Unsaturated zone flow model (improved Richards equation)

[0171]

[0172] θ: water content in the soil; t: time; z: depth; K(θ): unsaturated hydraulic conductivity, K sat is the saturated hydraulic conductivity, b is the gain and offset coefficient calculated by the ground calibration plate; h(θ): soil water pressure head, h e is the intake pressure head; S root (z,t): root water absorption rate, S root (z,t)=a(z)·T p (t), a(z) is the root density distribution function, T p (t) is the potential transpiration rate. Determination of the most unfavorable point (definition of water migration hysteresis index)

[0173]

[0174] WTLI: Temporal Water Deficit Index; θ crit : critical soil moisture content; θ(t): actual soil moisture content as a function of time t; t c : time period for observation or evaluation; In the time window [0,t c ], the cumulative deficit of soil moisture below the critical value; θ crit ·t c : Theoretical maximum deficit.

[0175] 1.2.2 Crop Response Module

[0176] Construct a coupling algorithm between Jarvis stomatal conductance model and Feddes root water absorption function.

[0177] For the Jarvis stomatal conductance model:

[0178] g s =g smax ·f(Q)·f(Dq)·f(ψ s )·f(T)

[0179] Where, f(Q) is the photosynthetic active radiation response function; f(Dq) is the water vapor pressure difference suppression function; f(ψ s ) is the soil water potential response function; f(T) is the temperature regulation function.

[0180] For Feddes root water absorption function:

[0181] S (z) =s(ψ s )·T p ·La (z)

[0182] Where S (z) is the water absorption of the root system at depth z per unit volume of soil; s(ψ s ) is the water stress coefficient; T p is the potential transpiration rate; L (z) is the root density distribution function.

[0183]

[0184] Where, ψ s is the soil water potential; ψ1 is the anaerobic threshold; ψ2 is the upper limit of optimal water absorption; ψ3 is the starting point of linear decline; and ψ4 is the permanent wilting point.

[0185]

[0186] Where, La (z) is the root density distribution function; z s is the half-decay depth of root distribution, ranging from 20 cm to 100 cm; z is the depth; m is the distribution shape parameter, when m = 1, the root density decreases slowly and linearly with depth; when m = 5, the root density decreases sharply with depth, and the roots are concentrated in the surface layer.

[0187] 1.2.3 Weather Drive Module

[0188] The Penman-Monteith two-source model was applied to distinguish the contributions of soil and vegetation evapotranspiration.

[0189] Vegetation transpiration:

[0190]

[0191] Where, ET veg is the transpiration of vegetation; Δ is the slope of the saturated water vapor pressure-temperature curve; R n,v is the canopy net radiation; ρ a is the air density; c p is the specific heat of air at constant pressure, Dq is the water vapor pressure deficit of air; is the canopy aerodynamic drag; is the total stomatal resistance of the canopy; ∈ is the psychrometric constant.

[0192] Soil evaporation:

[0193]

[0194] Where, ET soil is the soil evaporation; Δ is the slope of the saturated water vapor pressure-temperature curve; ρ a is the air density; c p is the specific heat of air at constant pressure, Dq is the air vapor pressure deficit; R n,s is the net radiation from the soil surface; is the aerodynamic resistance of the soil surface; r soil is the soil evaporation resistance; ∈ is the psychrometric constant.

[0195] 1.2.4 Establishing a dynamic response function for water stress and photosynthate distribution

[0196] Water stress index:

[0197]

[0198] Where WSI is water stress index; ET is act is the actual transpiration; ET p is the potential transpiration.

[0199] Total photosynthate correction:

[0200]

[0201] Where, P total Corrected for total photosynthetic product; P max is the maximum photosynthetic capacity (without stress); g s is the actual stomatal conductance; g smax is the maximum stomatal conductance.

[0202] Improvements and advantages:

[0203] By coupling the multiple processes of water transport in the unsaturated zone, stomatal response, and dual-source evapotranspiration, a dynamic feedback mechanism for the SPAC system was constructed. The Richards equation was modified to include a microbial activity correction term, enhancing the accuracy of identifying critical water points. The Jarvis-Feddes simultaneous algorithm was used to simulate root water uptake and stomatal oscillations. Furthermore, the water stress index and photosynthate partitioning function were integrated to reveal the threshold response patterns of carbon-water coupling. Compared to traditional single-process models, this model significantly reduces systematic errors and can analyze minute-scale water redistribution processes.

[0204] S2 water-saving water system design optimization

[0205] 2.1 System Framework Terrain data: Extract terrain slope and confluence network based on digital elevation model (DEM) to drive gradient optimization of gravity water delivery path;

[0206] Soil data: Calculate channel leakage rate and regional infiltration loss using the spatial distribution of soil permeability coefficients. Crop data: Divide differentiated irrigation scheduling units based on the spatial heterogeneity of crop types and water requirements during their growth periods.

[0207] Meteorological data: Combined with rainfall contour interpolation results, quantify the regional rainwater resource utilization potential and guide the spatial layout of rainwater collection facilities.

[0208] 2.2 Multi-objective optimization

[0209] Taking minimizing water loss, maximizing irrigation uniformity, and maximizing recycled water utilization as the goals, a multi-objective water-saving irrigation optimization model is established, as follows:

[0210] (1) Water loss objective function (reducing leakage and evaporation through channel anti-seepage design and pipeline transformation):

[0211]

[0212] Where N c is the total number of channel segments; L i is the length of the i-th channel; q i is the leakage per unit length of the i-th channel; is the channel anti-seepage efficiency coefficient; N p is the total number of pipeline segments; P j is the surface area of ​​the jth section of open-air pipeline; E j is the evaporation rate per unit area; is the pipe coverage correction factor.

[0213] q i =K s,i W i ·H i

[0214] Where qi is the leakage per unit length of the i-th channel; K s,i is the saturated permeability coefficient of the soil below the i-th channel section; W i is the channel bottom width; H i Design water depth for the channel.

[0215] P j =π·D j ·L j

[0216] Where, P j is the surface area of ​​the jth section of open-air pipeline; D j is the pipe diameter; L j is the pipe length.

[0217]

[0218] Where, is the wind speed correction coefficient; Δe j is the air saturation vapor pressure difference; P atm is atmospheric pressure.

[0219] (2) Irrigation uniformity objective function (using drip irrigation and sprinkler irrigation zoning design to reduce local excessive irrigation):

[0220]

[0221] constraint:

[0222]

[0223] Where n is the total number of irrigation units; d k is the water supply of the kth irrigation unit; μ d is the average water supply per irrigation unit, ET c A is the water requirement of crops; total is the total irrigated area.

[0224] (3) Recycled water utilization objective function (integrating rainwater collection systems with reclaimed water reuse facilities to replace groundwater extraction):

[0225]

[0226] Where η r For rainwater collection efficiency, D total is the total irrigation water requirement, D total =ET c ·A total t,V r is the volume of the rainwater collection tank; U w is the daily treatment capacity of reclaimed water;

[0227] Ic is the irrigation cycle; A c is the catchment area; Dr is the design rainfall; Rc is the runoff coefficient.

[0228] Improvements and advantages:

[0229] For the first time, GIS spatial heterogeneity analysis and multi-objective optimization algorithms are coupled. Through DEM-driven water system topology optimization and collaborative modeling of soil-crop-rainfall multi-source data, precise Pareto frontier searches are achieved for water transmission losses (joint quantification of Darcy infiltration and Penman evaporation), irrigation uniformity (Christiansen coefficient constraints) and recycled water utilization (integration of rainwater and greywater systems). This solves the problem of traditional single-objective optimization that ignores spatial explicit constraints and resource coordination, and significantly improves water-saving efficiency.

[0230] S3 dynamic water-saving and water distribution strategy

[0231] Long-term rainfall forecasts are conducted to determine the drought, average, or bumper harvest conditions in the planned year and adjust the planning scheme. Simulate the ideal growth state of crops under given conditions based on soil moisture and weather forecasts. Based on real-time soil moisture and short-term rainfall forecasts, irrigation plans are adjusted in advance, and irrigation duration and water volume are dynamically calculated to achieve "drip irrigation on demand." The reward function for optimizing water distribution strategies includes:

[0232] Minimize water consumption per unit of output;

[0233] Groundwater extraction is below sustainable thresholds;

[0234] Maximize the utilization rate of rainwater resources.

[0235] 3.1 Soil Moisture Prediction (Lag-PhysFusion):

[0236] (1) Historical time series data:

[0237]

[0238] Input features: X t =[H t ,L t ,W t ,I t ,V t ] T

[0239] Where, X t is the input feature vector at time t; S t is the measured soil moisture value at time t; T is the total length of the time series, that is, the maximum time step of the historical data; H tSoil moisture calculated by the improved Richards equation (stratified data, dimension Da); L t is the hysteresis index (scalar, reflecting the delayed effect of water migration); W t is meteorological data (temperature, precipitation, evaporation, etc.); I t is the irrigation volume; V t is the vegetation index; S t is the measured soil moisture value, the output value (S t ∈R).

[0240] (2) Physical feature extraction:

[0241] Extracting physical dynamics by improving the discrete solution of Richards equation:

[0242]

[0243] Discrete form (finite difference method):

[0244]

[0245] Where SW is soil water potential; t is time; K(SW) is unsaturated hydraulic conductivity; zz is vertical spatial coordinate; YH is source-sink term; SW t (d) is the moisture of the dth layer of soil at time t; Δt is the time step; Δzz 2 is the spatial step length; is the hydraulic conductivity of the ddth layer at time t; is the source and sink term of the dth layer at time t. The physical model output is fused with the hysteresis index to form an enhanced feature:

[0246]

[0247] Where, is the vector after physical feature encoding; ReLU is the activation function; [H t ;L t ] is the concatenation of input features; W p is the weight matrix; b p is the bias vector.

[0248] (3) Lag-perceived time coding (the lag index L t Dynamically embedded time position coding):

[0249]

[0250] Where EB t is the standard Transformer position encoding; W L is a learnable parameter used to adjust the degree of time distortion of the lag effect; Lt is the lag index.

[0251] (4) Spatiotemporal Attention Mechanism:

[0252]

[0253] Where C X is the query matrix; J is the key matrix; Z is the value matrix; M L is the lagged similarity matrix, τ is the temperature parameter that controls the sensitivity of the hysteresis difference, LZ i , LZ j is the hysteresis index; d k is the key vector dimension.

[0254] (5) Multi-scale temporal convolution:

[0255]

[0256] Where, is the output feature of the k-th convolution at time t; is the sth weight parameter of the kth layer convolution; is the time step ts×d (k) The input feature vector of d (k) is the expansion factor of the kth layer; S is the number of time steps covered by the convolution kernel.

[0257] (6) Dynamic fusion prediction:

[0258] Combined with physical characteristics P t With time series feature C t :

[0259]

[0260] Fusion weight ξ t Dynamically generated by the lagging index:

[0261] ξ t =σ(w ξ ZH t +b ξ )

[0262] Where, is the predicted output; t is the dynamic fusion weight; f phy (P t ) is the physical model branch; f seq (C t ) is the sequence model branch; ZH tis the hysteresis index; σ is the Sigmoid function, which controls the contribution ratio of the physical model and the data-driven model; w ξ is the weight parameter; b ξ is the bias parameter.

[0263] 3.2 Annual rainfall forecast (MPF-Net model):

[0264] (1) Input data:

[0265] Target variable: Y t ∈R y (historical annual rainfall series, y is the year);

[0266] Climate factor: X t ∈R (y×m) (m climate indices, such as ENSO, PDO, AMO, etc.);

[0267] Geographical features: G∈R (p) (p static geographic parameters, such as latitude, longitude, elevation, etc.).

[0268] (2) Wavelet multi-scale decomposition: t Perform Mallat wavelet decomposition:

[0269]

[0270] Where Y t is the target variable; is the j-th scale detail component (capturing interannual variability); A J (t) is the approximate component (representing the long-term trend); J is the number of decomposition levels.

[0271] (3) Physical Constraint Feature Engineering:

[0272] Energy constraints of the climate system (introducing the mass-energy conservation equation as a regularization term):

[0273]

[0274] Where, is the climate system energy; v is the atmospheric vertical velocity field; HD is the water vapor flux divergence; λ1 and λ2 are learnable constraint coefficients.

[0275] Geographical features embedded:

[0276] Γ=ReLU(W g G+b g )⊙Sigmoid(W t Ω+b t )

[0277] Where W gG, W t T is the learnable weight matrix; Ω is the temporal encoding vector.

[0278] (4) Multimodal spatiotemporal attention mechanism:

[0279] Climate factors cross attention: Where Q = W q [A J (t); Γ]; K = W k [X t ;G];V=W v [X t ; G].

[0280] Where, Attn Q is the output attention feature, reflecting the spatiotemporal modulation effect of climate factors on rainfall variability; O is the dynamic query matrix; is the wavelet detail component, representing the interannual variability signal; d k is the scaling factor; VZ is the value matrix; A J (t) is the wavelet approximation component; Γ is the geographic modulation matrix; W q is the learnable weight matrix; G is the geographic feature vector; W k , W v is the learnable weight matrix.

[0281] Time-frequency dual-domain gating:

[0282] g t =σ(W g [D j (t); Attn Q ])

[0283]

[0284] Where g t is the gate vector; σ is the Sigmoid activation function; W g is the learnable gating weight matrix; For the corrected detail component, data-driven features are fused with physical constraints.

[0285] (5) Adaptive learning of mutation points:

[0286] Construct a change point probability model for Bayesian change point detection:

[0287]

[0288] Where c t ∈{0,1} represents the change point indicator variable; h t is hidden state; W c is the learnable weight matrix; ht-1 ;h t is the hidden state vector of the adjacent time step; σ is the Sigmoid function.

[0289] Dynamic parameter adjustment:

[0290]

[0291] In the formula, MO is the key parameter of the model, which realizes the mutation adaptation of the parameter space; MO t is the key parameter of the model at the current moment; MO0 is the initial parameter value; ΔMO is the parameter adjustment step; c i is the cumulative sum of the historical change point indicator variable.

[0292] (6) Physical Information Neural Network Architecture:

[0293]

[0294] Where, is the prediction error; is the KL divergence regularization term for change point detection; Loss of physical restraint; is the adaptive weight parameter.

[0295] (7) Uncertainty Quantification (Monte Carlo Physics Dropout):

[0296] Keep dropout turned on during prediction and perform N samplings:

[0297]

[0298] Where, is the model prediction output of the i-th Monte Carlo sampling; CN is the number of sampling times; is the network parameter after randomly discarding some neurons during the i-th sampling; σ 2 is the prediction variance.

[0299] 3.3 Short-term rainfall prediction (MDSTF model):

[0300] (1) Select key indicators and observe them:

[0301] Air humidity, temperature, wind speed, air pressure, etc.

[0302] (2) Decompose the time series data of each indicator:

[0303] For one of the indicator time series x i (t):

[0304] Generate a noisy signal:

[0305] x (i) (t) = x(t) + β0w (i) (t)(i=1,2,…,I)

[0306] w (i) (t): Gaussian white noise; β0: initial noise coefficient; x(t) is the original signal.

[0307] Decompose and average:

[0308]

[0309] Where, EMD1 is the empirical mode decomposition of the signal; is the first-order IMF obtained by the i-th decomposition of the noisy signal; IMF1(t) is the average of the first-order IMFs of all I-th decomposition results.

[0310] Update residuals:

[0311] r1(t)=x(t)-IMF1(t)

[0312] Where r1(t) is the residual signal after removing the first-order IMF.

[0313] k-th order decomposition:

[0314]

[0315] Where r k-1 (t) is the residual signal of the k-1th stage; E k (w (i) (t)) is the Gaussian white noise w (i) (t) kth order IMF obtained after EMD decomposition; β k-1 is the noise coefficient of the k-1th stage.

[0316]

[0317] Where, is the k-th order IMF obtained by decomposing the i-th noise residual; IMF k (t) is the average value of the k-th order IMF of all I-time decomposition results.

[0318] r k (t) = r k-1 (t)-IMF k (t)

[0319] Where r k (t) is the new residual signal after removing the k-th order IMF.

[0320] Final decomposition results:

[0321]

[0322] Where, is the linear combination of all IMFs; R(t) is the residual.

[0323] (3) Short-term forecast:

[0324] Decompose data and reconstruct (assuming the original rainfall series is P(t)):

[0325]

[0326] Where C i (t) is the i-th intrinsic mode function; R(t) is the residual, reflecting the long-term trend or low-frequency component; P(t) is the original rainfall time series; is the total number of IMFs obtained by EMD decomposition.

[0327] Frequency division prediction:

[0328] High-frequency IMF (bidirectional LSTM+TCN model):

[0329]

[0330] Where, is the time window length; is the forecast value of the ith IMF at time t+1; is the i-th IMF in the time window Historical data within is the window length; BiLSTM is a bidirectional long short-term memory network, which is used to capture the bidirectional temporal dependencies; TCN is a temporal convolutional network, which extracts local mutation features through dilated convolution.

[0331] Low-frequency IMF (adaptively weighted ARIMA model):

[0332]

[0333] Where, is the ARIMA coefficient; is the predicted value of the jth low-frequency IMF at time t+1; is the autoregressive (AR) order; is the moving average (MA) order; is the nth moving average coefficient; ∈(t) is the white noise term; is the predicted value of the residual term at time t+1; is the autoregressive coefficient; is the time trend coefficient; is the intercept term; t is the time.

[0334] Dynamic attention fusion mechanism:

[0335]

[0336] Where h i (t) is the implicit state of the ith IMF prediction model; H(t) is the global context; W i is the trainable weight matrix; is the attention weight of the i-th IMF, indicating its importance to the final prediction; W i is the trainable weight matrix; [h i (t); H(t)] is a vector concatenation operation, which transforms the implicit state h of IMF into i (t) is merged with the global context H(t); H(t) is the global context vector; is the final rainfall forecast value at time t+1; is the forecast value of the ith IMF; is the predicted value of the residual term.

[0337] Residual correction and uncertainty quantification:

[0338]

[0339] Where, is the adaptive correction coefficient; Δ(t+1) is the correction term used to adjust the predicted value MLP is a multi-layer perceptron; is the absolute value of the recent forecast error; σ(t) is the standard deviation of the recent forecast error, reflecting uncertainty; P final (t+1) is the final predicted value after error correction.

[0340] (4) Substitute into the rainfall prediction model;

[0341]

[0342] P is the predicted rainfall, mm; ∈ is the model error term; t is the time; ψ is the water vapor flux divergence, kg·m -2 ·s -1 , q is the specific humidity, is the wind speed vector, g is the acceleration of gravity, p s , p t is the surface pressure and the integrated top pressure, is the pressure gradient; Temperature gradient, K·km -1 , is the vertical temperature lapse rate, v T is the horizontal temperature advection, Cp RH is the specific heat of dry air at constant pressure; Γ : relative humidity correction term; RH: relative humidity, %; Γ: Clapeyron correction coefficient, L v is the latent heat of water vapor, R v is the water vapor gas constant, R d is the dry air gas constant, is the ambient temperature; Pressure adjustment factor; P a is the normalized sea level pressure; ω: vertical velocity, CAPE: Convective Effective Potential Energy, LFC is the free convection altitude, EL is the equilibrium altitude, It is caused by the deficiency of qi and fever. is the ambient temperature; RI: Richardson number, is the potential temperature, u and v are the horizontal wind speed components; Cloud water content ratio, C w is the actual cloud water content, C w0 is the critical cloud water content.

[0343] 3.4 Simulation of ideal crop growth state

[0344] Based on predictions of future rainfall, soil moisture and other data, in order to predict the ideal growth state of crops in fields under given environmental conditions, an innovative framework named Hybrid Crop Growth Digital Twin (HCG-DT) was proposed that integrates crop growth mechanism model, time series deep learning and dynamic feedback optimization.

[0345] (1) Input data standardization

[0346]

[0347] Where, X norm is the standardized data; μx is the mean of the original data; σx is the standard deviation of the original data; X is the original environmental data.

[0348] (2) Mechanism model construction (light-temperature potential baseline)

[0349]

[0350] Where, P g is the total photosynthetic rate; is the light energy conversion efficiency; is the temperature correction coefficient; is photosynthetically active radiation; is the half-saturation constant of the light response curve; LAI is the leaf area index.

[0351] (3) Deep learning correction (environmental stress modeling)

[0352] δ(t)=f LSTM (soil moisture (t), rainfall (t), T (t), H (t))

[0353] Corrected growth index = mechanism model output (t) × (1-δ(t))

[0354] Where δ(t) is the stress correction coefficient of the time step (0-1, indicating the growth attenuation ratio); f LSTM is the nonlinear function modeled by the LSTM network; soil moisture (t) is the soil moisture content; rainfall (t) is the predicted rainfall; T (t) is the temperature; H (t) is the relative humidity.

[0355] (4) Dynamic feedback optimization (field strategy generation)

[0356]

[0357] Where S w (t) is the water stress index; S T (t) is the temperature stress index; is the weight coefficient; To optimize the time range.

[0358] (5) Visualization of coercion risk

[0359]

[0360] Where R d The risk probability of is the indicator function (1 when the condition is met, otherwise 0); τ w , τ T is the water / temperature stress threshold; is the sample size or the number of forecast days.

[0361] (6) Growth status prediction output

[0362] Y(t)=Y 机理 (t)·(1-δ(t))+η(t)

[0363] Where Y(t) is the predicted growth index; Y 机理 (t) is the predicted value of the mechanism model; η(t) is the Gaussian noise term. 3.5 Long-term water-saving scheduling (deep reinforcement learning DRL model)

[0364] (1) State space

[0365] s t =[QX t ,W res ,W gnd ,G his ,S t ,R f ]

[0366] Where, QX t is the climate prediction vector (precipitation, temperature, evaporation, etc.); W res is the current water storage capacity of the reservoir; W gnd is the recoverable reserves of groundwater; G his is the historical groundwater extraction volume; S t is the crop growth stage; R f The capacity of rainwater collection facilities.

[0367] (2) Action Space

[0368]

[0369] Where, Reservoir water allocation ratio; Groundwater allocation ratio; Rainwater utilization ratio.

[0370] (3) Reward Function

[0371] J L =v1R WUE +v2R GND +v3R RAIN

[0372] Where v1, v2, and v3 are weight coefficients used to balance the priorities of different objectives.

[0373] Minimize water consumption per unit of output:

[0374]

[0375] Where, YC is crop yield; R RAIN is the amount of rainwater collected during the forecast period; the negative sign is to penalize high water consumption.

[0376] Groundwater sustainability constraints:

[0377]

[0378] Where G threshold is the threshold for sustainable groundwater extraction; R is the penalty coefficient for over-exploitation.

[0379] Maximize the utilization of rainwater resources:

[0380]

[0381] Where R total is the total available rainwater resources.

[0382] (4) State transition equation

[0383]

[0384] Where R inflow is the reservoir inflow; L evap is the evaporation loss of the reservoir; R infil is the natural groundwater recharge; η is the recharge efficiency coefficient.

[0385] 3.6 Short-term precision irrigation (model predictive control MPC)

[0386] Rolling optimization objective function:

[0387]

[0388] wxya min ≤XW k ≤XW field

[0389] 0≤I k ≤I max

[0390] Where NY is the prediction time domain; QJ, RJ are weight matrices, balancing moisturizing and water saving; XW field is the field water holding capacity (the upper limit of crop water requirement); XW min is the wilting coefficient (lower limit of water requirement); XW target is the target soil moisture content; I max The maximum water supply rate for the drip irrigation system.

[0391] Improvements and advantages:

[0392] This method innovatively combines climate forecast optimization using deep reinforcement learning with the real-time dynamics of crop growth, enabling closed-loop irrigation decision-making at multiple timescales. Compared to traditional static water allocation or single-timescale approaches, this significantly improves global water resource utilization efficiency and groundwater sustainability while ensuring crop yields. This method primarily proposes the Lag-PhysFusion model, the MPF-Net model, the MDSTF model, the HCG-DT model, and the deep reinforcement learning (DRL) model to predict soil moisture, weather, and ideal crop growth conditions, achieving significant results.

[0393] S4 Intelligent Terminal Control System: Drones regularly monitor fields, capturing real-time images of crops and uploading this data to the intelligent control terminal for unified decision-making. Based on real-time data, forecasts, and irrigation model calculations, the terminal controller controls the intelligent sprinkler equipment, ensuring precise irrigation operations.

[0394] The control terminal's algorithms analyze images to determine crop growth in each area, analyze soil moisture movement to determine soil moisture content in each area, and combine this with future rainfall data to make an overall assessment and control smart valves in various locations for irrigation.

[0395] like Figure 1 shown.

[0396] 4.1 Determining whether crops are in ideal growth state:

[0397] 4.1.1 Image preprocessing:

[0398] Radiometric correction. Use the Empirical Linear Method (ELM) to eliminate the effects of illumination variations:

[0399]

[0400] Where DN is the original pixel value; as and bs are the gain and offset coefficients calculated by the ground calibration plate.

[0401] 4.1.2 Image stitching:

[0402] SIFT feature matching and fade-in and fade-out fusion algorithms were used to generate orthophotos (error < 0.5 pixels).

[0403] 4.1.3 Image Analysis

[0404] (1) Crop segmentation:

[0405] Use the improved Mask R-CNN model (the backbone network is ResNet-101-FPN):

[0406] Input: RGB image (resolution 1024×1024)

[0407] Output: Pixel-level mask of individual crops (accuracy >92%)

[0408] (2) Growth stage identification:

[0409] ResNet-50 classification model based on time series images:

[0410] Input: 3 consecutive crop canopy images (time step is 3 days)

[0411] Output: Probability distribution of growth period (e.g. 85% for emergence period, 12% for jointing period)

[0412] (3) Key parameter extraction:

[0413] Leaf Area Index (LAI): Inverted by canopy porosity (Beer-Lambert law)

[0414] Canopy temperature: radiometric calibration of thermal infrared images (accuracy ±0.5°C)

[0415] Biomass estimation: regression model of NDVI and ground truth data (R 2 >0.8)

[0416] 4.2 Irrigation demand judgment and decision-making:

[0417] Through the images scanned and processed by drones, it is possible to determine whether the actual growth trend of crops meets the ideal growth trend, determine whether crops in different areas need irrigation, and optimize the irrigation timing.

[0418] 4.3 Intelligent control and precision irrigation:

[0419] Based on the decision results, the terminal controller controls the intelligent sprinkler system for precise irrigation, ensuring that the irrigation amount meets the actual needs of the crops and avoiding over-irrigation or under-irrigation.

[0420] 4.4 Data storage and remote monitoring:

[0421] All data and system status will be stored on the cloud platform, which is convenient for users to conduct long-term monitoring and analysis and supports remote control.

[0422] 4.5 Closed-loop feedback on water-saving effects

[0423] Deploy IoT devices (flow meters, smart water meters, etc.) to monitor water consumption and water-saving indicators in real time; use digital twin technology to simulate the water-saving potential of different irrigation strategies, conduct dynamic optimization, and ensure continuous improvement of water-saving targets.

[0424] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present invention.

Claims

1. A method for designing a water system for farmland water-saving irrigation and optimizing dynamic water distribution based on the fusion of multi-source data on supply and demand, characterized by: The method comprises the following steps: S1 "crop-soil-meteorology" multi-source data coupling modeling, S2 water-saving water system design optimization, S3 dynamic water-saving and water distribution strategy, Controlled by S4 intelligent terminal control system.

2. The method for designing a water system for farmland water-saving irrigation and optimizing dynamic water distribution based on multi-source data fusion of supply and demand according to claim 1 is characterized in that: Step S1 is specifically as follows: 1.1 Multi-source data monitoring, Multiple sensors are distributed throughout the fields to monitor soil moisture, temperature, weather conditions, and crop water requirements in real time, as shown below: Sensor density calculation: A sensor placement optimization method based on an improved adaptive PSO algorithm uses the numerical solution of the soil water transport equation (Richards equation) as a particle search direction correction term. The crop root weight distribution matrix is ​​dynamically adjusted using NDVI time series data. This includes multi-objective fitness function reconstruction and the construction of a three-dimensional evaluation index by integrating the spatial coefficient of variation (CV) of soil properties, crop root distribution weight (RW), and meteorological influence factor (MF). Fitness=α·CV+β·RW+γ·MF+λ·Cost Where, Fitness is the fitness value; α, β, and γ are dynamic weight coefficients; λ is the economic constraint factor; Cost is the comprehensive economic constraint term, which is used to quantify the implementation cost of the sensor deployment plan; CV is the spatial variation coefficient of soil properties; RW is the weight of crop root distribution; MF is the meteorological influence factor, The second is the farmland-specific improvement strategy, including hierarchical particle coding (separately encoding the two-dimensional plane coordinates and the burial depth, using a hybrid discrete-continuous search space), the introduction of mutation operators (when particles fall into a local optimum, Gaussian mutation is triggered according to the standard deviation of the crop ridge width), and an energy consumption constraint model (integrating the wireless signal attenuation equation to calculate the communication energy consumption cost). All data collected by sensors are uploaded to the data terminal control system through wireless communication technology to ensure timely data updates and remote monitoring. The transmitted real-time data can support the dynamic change monitoring of the field environment. 1.2 Build the model architecture, as follows: 1.2.1 Soil moisture transport modeling, Unsaturated zone flow model, θ: water content in the soil; t: time; z: depth; K(θ): unsaturated hydraulic conductivity, K sat is the saturated hydraulic conductivity, b is the gain and offset coefficient calculated by the ground calibration plate; h(θ): soil water pressure head, h e is the intake pressure head; S root (z,t): root water absorption rate, S root (z,t)=a(z)·T p (t), a(z) is the root density distribution function, T p (t) is the potential transpiration rate, determined at the most unfavorable point, WTLI: Temporal Water Deficit Index; θ crit : critical soil moisture content; θ(t): actual soil moisture content as a function of time t; t c : time period for observation or evaluation; In the time window [0,t c ], the cumulative deficit of soil moisture below the critical value; θ crit ·t c : Theoretical maximum deficit, 1.2.2 Crop Response Module, Construct a coupling algorithm between Jarvis stomatal conductance model and Feddes root water absorption function, For the Jarvis stomatal conductance model: g s =g smax ·f(Q)·f(dq)·f(ψ s )·f(T) Where, f(Q) is the photosynthetic active radiation response function; f(Dq) is the water vapor pressure difference suppression function; f(ψ s ) is the soil water potential response function; f(T) is the temperature regulation function, For Feddes root water absorption function: S (z) =s(ψ s )·T p ·La (z) Where S (z) is the water absorption of the root system at depth z per unit volume of soil; s(ψ s ) is the water stress coefficient; T p is the potential transpiration rate; L (z) is the root density distribution function, Where, ψ s is the soil water potential; ψ1 is the anaerobic threshold; ψ2 is the upper limit of optimal water absorption; ψ3 is the starting point of linear decline; ψ4 is the permanent wilting point, Where, La (z) is the root density distribution function; z s is the half-decay depth of root distribution, ranging from 20 cm to 100 cm; z is the depth; m is the distribution shape parameter, m = 1, the root density decreases slowly and linearly with the depth; m = 5, the root density decreases sharply with the depth, and the roots are concentrated in the surface layer. 1.2.3 Weather drive module, The Penman-Monteith dual-source model was used to distinguish the contribution of soil / vegetation evapotranspiration. Vegetation transpiration: Where, ET veg is the transpiration of vegetation; Δ is the slope of the saturated water vapor pressure-temperature curve; R n,v is the canopy net radiation; ρ a is the air density; c p is the specific heat of air at constant pressure, Dq is the water vapor pressure deficit of air; is the canopy aerodynamic drag; is the total stomatal resistance of the canopy; ∈ is the psychrometric constant, Soil evaporation: Where, ET soil is the soil evaporation; Δ is the slope of the saturated water vapor pressure-temperature curve; ρ a is the air density; c p is the specific heat of air at constant pressure, Dq is the air vapor pressure deficit; R n,s is the net radiation from the soil surface; is the aerodynamic resistance of the soil surface; r soil is the soil evaporation resistance; ∈ is the psychrometric constant, 1.2.4 Establish a dynamic response function between water stress and photosynthate distribution. Water stress index: Where WSI is water stress index; ET is act is the actual transpiration; ET p is the potential transpiration, Total photosynthate correction: Where, P total Corrected for total photosynthetic product; P max is the maximum photosynthetic capacity (without stress); g s is the actual stomatal conductance; g smax is the maximum stomatal conductance.

3. The method for designing a water system for farmland water-saving irrigation and optimizing dynamic water distribution based on multi-source data fusion of supply and demand according to claim 1 is characterized in that: S2 water-saving water system design optimization, as follows: 2.1 Build a system framework, including terrain data, soil data, crop data, and meteorological data; 2.2 Multi-objective optimization, Taking minimizing water loss, maximizing irrigation uniformity, and maximizing recycled water utilization as the goals, a multi-objective water-saving irrigation optimization model is established, as follows: (1) Water loss objective function (reducing leakage and evaporation through channel anti-seepage design and pipeline transformation): Where N c is the total number of channel segments; L i is the length of the i-th channel; q i is the leakage per unit length of the i-th channel; is the channel anti-seepage efficiency coefficient; N p is the total number of pipeline segments; is the surface area of ​​the jth section of open-air pipeline; E j is the evaporation rate per unit area; is the pipe coverage correction factor, q i =K s,i ·W i ·H i Where q i is the leakage per unit length of the i-th channel; K s,i is the saturated permeability coefficient of the soil below the i-th channel section; W i is the channel bottom width; H i Design water depth for the channel, Where, is the surface area of ​​the jth section of open-air pipeline; D j is the pipe diameter; L j is the pipe length, Where, is the wind speed correction coefficient; Δe j is the air saturation vapor pressure difference; is atmospheric pressure, (2) Irrigation uniformity objective function: constraint: Where n is the total number of irrigation units; d k is the water supply of the kth irrigation unit; μ d is the average water supply per irrigation unit, ET c A is the water requirement of crops; total is the total irrigated area, (3) Recycled water utilization objective function: Where η r For rainwater collection efficiency, D total is the total irrigation water requirement, D total =ET c ·A total t,V r is the volume of the rainwater collection tank; U w is the daily treatment capacity of reclaimed water; Ic is the irrigation cycle; A c is the catchment area; Dr is the design rainfall; Rc is the runoff coefficient.

4. The method for designing a water system for farmland water-saving irrigation and optimizing dynamic water distribution based on multi-source data fusion of supply and demand according to claim 1 is characterized in that: S3 dynamic water-saving and water distribution strategy is as follows: 3.1 Soil Moisture Prediction (Lag-PhysFusion): (1) Historical time series data: Input features: X t =[H t ,L t ,W t ,I t ,V t ] T Where, X t is the input feature vector at time t; S t is the measured soil moisture value at time t; T is the total length of the time series, that is, the maximum time step of the historical data; H t Soil moisture calculated by improving Richards equation; L t is the hysteresis index; W t is meteorological data; I t is the irrigation volume; V t is the vegetation index; S t is the measured soil moisture value, the output value (S t ∈R), (2) Physical feature extraction: Extracting physical dynamics by improving the discrete solution of Richards equation: Discrete form: Where SW is soil water potential; t is time; K(SW) is unsaturated hydraulic conductivity; zz is the vertical spatial coordinate; YH is the source-sink term; is the moisture of the dth layer of soil at time t; Δt is the time step; Δzz 2 is the spatial step length; is the hydraulic conductivity of the ddth layer at time t; is the source and sink term of the dth layer at time t, and the physical model output and the hysteresis index are fused into enhanced features: Where, is the vector after physical feature encoding; ReLU is the activation function; [H t ;L t ] is the concatenation of input features; W p is the weight matrix; b p is the bias vector, (3) Delayed Perceptual Time Coding: Where EB t is the standard Transformer position encoding; W L is a learnable parameter used to adjust the degree of time distortion of the lag effect; L t is the lag index, (4) Spatiotemporal Attention Mechanism: Where C X is the query matrix; J is the key matrix; Z is the value matrix; M L is the lagged similarity matrix, τ is the temperature parameter that controls the sensitivity of the hysteresis difference, LZ i , LZ j is the hysteresis index; d k is the key vector dimension, (5) Multi-scale temporal convolution: Where, is the output feature of the k-th convolution at time t; is the sth weight parameter of the kth layer convolution; is the time step ts×d (k) The input feature vector of d (l) is the expansion factor of the kth layer; S is the number of time steps covered by the convolution kernel, (6) Dynamic fusion prediction: Combined with physical characteristics P t With time series feature C t : Fusion weight ξ t Dynamically generated by the lagging index: x t =σ(w ξ ZH t +b ξ ) Where, is the predicted output; t is the dynamic fusion weight; f phy (P t ) is the physical model branch; f seq (C t ) is the sequence model branch; ZH t is the hysteresis index; σ is the Sigmoid function, which controls the contribution ratio of the physical model and the data-driven model; w ξ is the weight parameter; b ξ is the bias parameter, 3.2 Annual rainfall forecast: (1) Input data: Target variable: Y t ∈R y (historical annual rainfall series, y is the year); Climate factor: X t ∈R (y×m) , m climate indices, Geographical features: G∈R (p) , p static geographic parameters, (2) Wavelet multi-scale decomposition: t Perform Mallat wavelet decomposition: Where Y t is the target variable; is the j-th scale detail component; A J (t) is the approximate component; J is the number of decomposition layers, (3) Physical Constraint Feature Engineering: Climate system energy constraints: Where, is the climate system energy; v is the atmospheric vertical velocity field; HD is the water vapor flux divergence; λ1, λ2 are learnable constraint coefficients, Geographical features embedded: Γ=ReLU(W g G+b g )⊙Sigmoid(W t Ω+b t ) Where W g G, W t T is the learnable weight matrix; Ω is the temporal encoding vector, (4) Multimodal spatiotemporal attention mechanism: Climate factors cross attention: Where Q = W q [A J (t); Γ]; K = W k [X t ; G]; V = W v [X t ; G], Where, Attn Q is the output attention feature, reflecting the spatiotemporal modulation effect of climate factors on rainfall variability; O is the dynamic query matrix; is the wavelet detail component, representing the interannual variability signal; d k is the scaling factor; VZ is the value matrix; A J (t) is the wavelet approximation component; Γ is the geographic modulation matrix; W q is the learnable weight matrix; G is the geographic feature vector; W k , W v is the learnable weight matrix, Time-frequency dual-domain gating: g t =σ i (W g [D j (t);Attn Q ]) Where g t is the gate vector; σ i is the Sigmoid activation function; W g is the learnable gating weight matrix; For the corrected detail components, data-driven features are integrated with physical constraints. (5) Adaptive learning of mutation points: Construct a change point probability model for Bayesian change point detection: Where c t ∈{0,1} represents the change point indicator variable; h t is hidden state; W c is the learnable weight matrix; h t-1 ;h t is the hidden state vector of the adjacent time step; σ i is the Sigmoid function, Dynamic parameter adjustment: In the formula, MO is the key parameter of the model, which realizes the mutation adaptation of the parameter space; MO t is the key parameter of the model at the current moment; MO0 is the initial parameter value; ΔMO is the parameter adjustment step; c i is the cumulative sum of the historical change point indicator variables, (6) Physical Information Neural Network Architecture: Where, is the prediction error; is the KL divergence regularization term for change point detection; Loss of physical restraint; is the adaptive weight parameter, (7) Uncertainty quantification: Keep dropout turned on during prediction and perform N samplings: Where, is the model prediction output of the i-th Monte Carlo sampling; CN is the number of samplings; n i is the network parameter after randomly discarding some neurons during the i-th sampling; σ 2 is the prediction variance, 3.3 Short-term rainfall forecast: (1) Select key indicators and observe them: air humidity, temperature, wind speed, and air pressure; (2) Decompose the time series data of each indicator: For one of the indicator time series x i (t): Generate a noisy signal: x (i) (t)=x(t)+β0w (i) (t)(i=1,2,…,I) w (i) (t): Gaussian white noise; β0: initial noise coefficient; x(t) is the original signal, Decompose and average: Where, EMD1 is the empirical mode decomposition of the signal; is the first-order IMF obtained by the i-th decomposition of the noisy signal; IMF1(t) is the average of the first-order IMFs of all I-th decomposition results. Update residuals: r1(t)=x(t)-IMF1(t) Where r1(t) is the residual signal after removing the first-order IMF, k-th order decomposition: Where r k-1 (t) is the residual signal of the k-1th stage; E k (w (i) (t)) is the Gaussian white noise w (i) (t) kth order IMF obtained after EMD decomposition; β k-1 is the noise coefficient of the k-1th stage, Where, is the k-th order IMF obtained by decomposing the i-th noise residual; IMF k (t) is the average value of the k-th order IMF of all I-time decomposition results, r k (t)=r k-1 (t)-IMF k (t) Where r k (t) is the new residual signal after removing the k-th order IMF, Final decomposition results: Where, is the linear combination of all IMFs; R(t) is the residual, (3) Short-term forecast: Decompose data and reconstruct (assuming the original rainfall series is P(t)): Where C i (t) is the i-th intrinsic mode function; R(t) is the residual, reflecting the long-term trend or low-frequency component; P(t) is the original rainfall time series; is the total number of IMFs obtained by EMD decomposition, Frequency division prediction: High-frequency IMF (bidirectional LSTM+TCN model): Where, is the time window length; is the forecast value of the ith IMF at time t+1; is the i-th IMF in the time window Historical data within is the window length; BiLSTM is a bidirectional long short-term memory network, which is used to capture the bidirectional temporal dependencies; TCN is a temporal convolutional network, which extracts local mutation features through dilated convolution. Low-frequency IMF (adaptively weighted ARIMA model): Where, is the ARIMA coefficient; is the predicted value of the jth low-frequency IMF at time t+1; is the autoregressive (AR) order; is the moving average (MA) order; is the nth moving average coefficient; ∈(t) is the white noise term; is the predicted value of the residual term at time t+1; is the autoregressive coefficient; is the time trend coefficient; is the intercept term; t is the time, Dynamic attention fusion mechanism: Where h i (t) is the implicit state of the ith IMF prediction model; H(t) is the global context; W i is the trainable weight matrix; is the attention weight of the i-th IMF; W i is the trainable weight matrix; [h i (t); H(t)] is a vector concatenation operation, which transforms the implicit state h of IMF into i (t) is merged with the global context H(t); H(t) is the global context vector; is the final rainfall forecast value at time t+1; is the forecast value of the ith IMF; is the predicted value of the residual term, Residual correction and uncertainty quantification: Where, is the adaptive correction coefficient; Δ(t+1) is a correction term used to adjust the predicted value MLP is a multi-layer perceptron; is the absolute value of the recent forecast error; σ(t) is the standard deviation of the recent forecast error, reflecting uncertainty; P final (t+1) is the final predicted value after error correction, (4) Substitute into the rainfall prediction model; P is the predicted rainfall, mm; ∈ is the model error term; ψ is the water vapor flux divergence, kg·m -2 ·s -1 , q is the specific humidity, is the wind speed vector, g is the acceleration of gravity, p s , p t is the surface pressure and the integrated top pressure, is the pressure gradient; Temperature gradient, is the vertical temperature lapse rate, v T is the horizontal temperature advection, C p is the specific heat capacity of dry air at constant pressure; RH Γ : relative humidity correction term; RH: relative humidity, %; Γ: Clapeyron correction coefficient, L v is the latent heat of water vapor, R v is the water vapor gas constant, R d is the dry air gas constant, is the ambient temperature; Pressure adjustment factor; P a is the normalized sea level pressure; ω : vertical speed, CAPE: Convective Effective Potential Energy, LFC is the free convection altitude, EL is the equilibrium altitude, It is caused by the deficiency of qi and fever. It is caused by the false temperature of the environment; RI: Richardson number, is the potential temperature, u and v are the horizontal wind speed components; Cloud water content ratio, C w is the actual cloud water content, C w0 is the critical cloud water content, 3.4 Simulation of ideal crop growth state, (1) Input data standardization Where, X norm is the standardized data; μx is the mean of the original data; σx is the standard deviation of the original data; X is the original environmental data, (2) Construction of mechanism model Where, P g is the total photosynthetic rate; is the light energy conversion efficiency; is the temperature correction coefficient; is photosynthetically active radiation; is the half-saturation constant of the light response curve; is the leaf light absorption coefficient; LAI is the leaf area index, (3) Deep learning correction δ(t)=f LSTM (soil moisture (t), rainfall (t), T (t), H (t)) Corrected growth index = mechanism model output (t) × (1-δ(t)) Where δ(t) is the stress correction coefficient of the time step (0-1, indicating the growth attenuation ratio); f LSTM is the nonlinear function modeled by the LSTM network; soil moisture (t) is the soil moisture content; rainfall (t) is the predicted rainfall; T (t) is the temperature; H (t) is the relative humidity of the air, (4) Dynamic feedback optimization Where S w (t) is the water stress index; S T (t) is the temperature stress index; is the weight coefficient; To optimize the time range, (5) Visualization of coercion risk Where R d The risk probability of is the indicator function (1 when the condition is met, otherwise 0); τ w , τ T is the water / temperature stress threshold; is the sample size or forecast days, (6) Growth status prediction output Y(t)=Y 机理 (t)·(1-δ(t))+η(t) Where Y(t) is the predicted growth index; Y 机理 (t) is the predicted value of the mechanism model; η(t) is the Gaussian noise term, 3.5 Long-term water-saving scheduling, (1) State space s t =[QX t ,W res ,W gnd ,G his ,S t ,R f ] Where, QX t is the climate prediction vector (precipitation, temperature, evaporation, etc.); W res is the current water storage capacity of the reservoir; W gnd is the recoverable reserves of groundwater; G his is the historical groundwater extraction volume; S t is the crop growth stage; R f is the capacity of rainwater collection facilities, (2) Action Space Where, Reservoir water allocation ratio; Groundwater allocation ratio; Rainwater utilization ratio, (3) Reward Function I L =v1R WUE +v2R GND +v3R RAIN Where ν1, ν2, and ν3 are weight coefficients used to balance the priorities of different objectives. Minimize water consumption per unit of output: Where, YC is crop yield; R RAIN is the amount of rainwater collected during the forecast period; the negative sign is to penalize high water consumption, Groundwater sustainability constraints: Where G threshold threshold for sustainable groundwater extraction; is the over-exploitation penalty coefficient, Maximize the utilization of rainwater resources: Where R total is the total available rainwater resources, (4) State transition equation Where R inflow is the reservoir inflow; L evap is the evaporation loss of the reservoir; R infil is the natural recharge of groundwater; η is the recharge efficiency coefficient, 3.6 Short-term precision irrigation, Rolling optimization objective function: Where NY is the prediction time domain; QJ, RJ are weight matrices, balancing moisturizing and water saving; XW field is the field water holding capacity; XW min is the wilting coefficient; XW target is the target soil moisture content; I max The maximum water supply rate for the drip irrigation system.

5. The method for designing a water system for farmland water-saving irrigation and optimizing dynamic water distribution based on multi-source data fusion of supply and demand according to claim 1 is characterized in that: The S4 intelligent terminal control system is controlled as follows: 4.1 Determining whether crops are in ideal growth state: 4.1.1 Image preprocessing: Radiometric correction, using the Empirical Linear Method (ELM) to eliminate the effects of illumination variations: Where DN is the original pixel value; as, bs are the gain and offset coefficients calculated by the ground calibration plate, 4.1.2 Image stitching: SIFT feature matching and fade-in and fade-out fusion algorithms are used to generate orthophotos (error < 0.5 pixels). 4.1.3 Image Analysis: (1) Crop segmentation: Using the improved Mask R-CNN model: Input: RGB image Output: Pixel-level mask of individual crops (accuracy >92%) (2) Growth stage identification: ResNet-50 classification model based on time series images: Input: 3 consecutive crop canopy images Output: Probability distribution of reproductive period (3) Key parameter extraction: Leaf Area Index (LAI): Inversion via Canopy Porosity Canopy temperature: radiometric calibration of thermal infrared imagery Biomass estimation: regression model of NDVI and ground truth data (R 2 >0.8) 4.2 Irrigation demand judgment and decision-making, Through the images scanned and processed by drones, it is possible to determine whether the actual growth trend of crops meets the ideal growth trend, determine whether crops in different areas need irrigation, and optimize the irrigation timing. 4.3 Intelligent control and precise irrigation, Based on the decision results, the terminal controller controls the intelligent sprinkler system for precise irrigation, ensuring that the irrigation volume meets the actual needs of the crops and avoiding over-irrigation or under-irrigation. 4.4 Data storage and remote monitoring, All data and system status will be stored on the cloud platform, which is convenient for users to conduct long-term monitoring and analysis, and supports remote control. 4.5 Closed-loop feedback on water-saving effects Deploy IoT devices to monitor water consumption and water-saving indicators in real time; use digital twin technology to simulate the water-saving potential of different irrigation strategies, conduct dynamic optimization, and ensure continuous improvement in water-saving targets.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for designing a farmland water-saving irrigation system and dynamically optimizing water distribution by integrating supply and demand multi-source data as described in any one of claims 1 to 5 above is implemented.

7. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the farmland water-saving irrigation water system design and dynamic water distribution optimization method based on the fusion of supply and demand multi-source data as described in any one of claims 1 to 5 is implemented.

Images