A precision irrigation decision-making method and system for an arid region farm under water resource constraints

Abstract

This invention discloses a method and system for precision irrigation decision-making in arid farms under water resource constraints, belonging to the field of smart agriculture technology. The method includes: collecting multi-source data, calculating leaf area index using Swin-ConvNeXt, obtaining soil moisture content by combining UAV imagery and ground sensor data and performing spatial correction using the inverse distance weighting method, calibrating the MAIZSIM model through Morris screening and particle swarm optimization, generating long and short agricultural condition sequences using an ensemble smoother, outputting irrigation instructions through NSGA-III, and assessing decision risk using virtual simulation and uncertainty quantification. This invention integrates agricultural condition data collected from UAVs and ground sensors across the entire region, accurately calculates crop parameters and calibrates the model using deep learning and data assimilation, and outputs periodic and plot-specific irrigation instructions by combining water resource constraints and multi-objective optimization. It assesses decision risk through virtual simulation and uncertainty quantification, realizing the transformation of arid farm irrigation from experience-based to quantitative and intelligent methods, ensuring crop yield and water use efficiency.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method and system for precision irrigation decision-making in arid farms under water resource constraints. Background Technology

[0002] my country's arid-region arable land is widely distributed, and these areas generally suffer from low natural precipitation and a scarcity of total water resources. The imbalance between water supply and demand has long constrained the large-scale planting and production of grain crops. Agricultural water use accounts for a high proportion of total water consumption in these regions. As modern agriculture develops towards large-scale planting, how to balance crop yield and irrigation water demand under limited water resources has gradually become a problem that needs to be addressed in farmland management. Corn, as one of the main crops grown in arid regions, has a continuous requirement for soil moisture during its growth process. Irrigation management directly affects crop growth, final yield, and water resource utilization efficiency. In the process of modern agricultural development, farmland irrigation is gradually moving towards refined management, and various field sensing devices, crop growth models, and intelligent optimization algorithms are beginning to be applied in the field of farmland management. Utilizing remote sensing imagery and ground sensors for crop condition monitoring, combined with crop mechanism models to simulate growth states, and formulating irrigation plans based on monitoring and prediction results, has become an important development direction for irrigation management in arid regions.

[0003] Traditional irrigation management models for arid farmland still have many significant shortcomings. Large-scale farm irrigation generally suffers from limited monitoring coverage, insufficient accuracy in predicting crop growth status, delayed irrigation decisions, and inadequate real-time adjustment capabilities. Conventional field monitoring relies heavily on manual sampling at fixed points and the deployment of single sensors, resulting in limited monitoring coverage and difficulty in achieving full coverage and round-the-clock collection of crop information across the entire farmland. The acquired crop growth and soil moisture data exhibit spatial discontinuities, limiting data completeness and timeliness. Traditional crop models mostly use general parameter systems without adapting to regional field environments and planting conditions, leading to significant discrepancies between simulation results and actual farmland conditions, failing to accurately reflect regional crop growth patterns. Past irrigation scheduling relied heavily on the experience of management personnel, simply referencing soil moisture status to formulate irrigation plans without comprehensively considering regional water resources, the load on water conveyance facilities, and the water needs of crops at different growth stages, easily resulting in water waste or insufficient water supply. Meanwhile, most plans lack the analysis process for different hydrological conditions, observation errors, changes in water inflow, and changes in management measures, making it difficult to judge the possible changes in yield, irrigation water use, and economic benefits after the implementation of the irrigation plan, and it is also difficult to assess the applicability of the irrigation plan under different conditions. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method and system for precision irrigation decision-making in arid farms under water resource constraints. It acquires farmland observation data through unmanned aerial vehicles (UAVs) and ground sensors, utilizes deep learning methods to retrieve crop leaf area index and soil moisture content information, and combines a localized MAIZSIM model with an iterative ensemble smoother to assimilate and predict crop conditions, generating long-term and short-term crop condition sequences. Under constraints such as water availability, irrigation capacity, and crop growth stage, a multi-objective optimization algorithm is employed to solve irrigation schemes. Scenario simulation and uncertainty analysis are used to evaluate changes in yield, irrigation water consumption, and economic benefits under different schemes, providing a basis for decision-making in irrigation management for arid farms.

[0005] The above objectives can be achieved through the following approach:

[0006] A precision irrigation decision-making method for arid farms under water resource constraints, specifically including:

[0007] S1. Multi-source images and soil moisture content data of cornfields are collected using UAVs and ground sensors. The leaf area index is calculated from the images using the Swin-ConvNeXt deep learning model. Soil moisture content information is obtained by combining UAV images and ground sensor data. Spatial correction is performed using the inverse distance weighting method to form basic agricultural observation data.

[0008] S2. Using field test data and obtained basic agricultural observation data as input, the parameters of the MAIZSIM model are screened through Morris global sensitivity analysis, and the parameters of the MAIZSIM model are calibrated through particle swarm optimization algorithm to obtain the localized MAIZSIM model.

[0009] S3. Unify the spatiotemporal scale of the localized MAIZSIM model and basic agricultural observation data, use an iterative ensemble smoother to fuse the basic agricultural observation data to calibrate the model, and connect to the multi-source meteorological data-driven model to generate a 90-day long-term agricultural sequence and a 15-day short-term agricultural sequence.

[0010] S4. Using the generated long-term and short-term agricultural conditions sequences as inputs, and constrained by water availability, irrigation infrastructure capacity and crop growth stage, the NSGA-Ⅲ multi-objective optimization algorithm is used to solve for the field irrigation execution instructions.

[0011] S5. Based on the obtained localized MAIZSIM model, construct a virtual farmland simulation scenario, set up combined simulation scenarios and carry out batch simulations, and use uncertainty quantification algorithm to statistically analyze the relative yield loss, irrigation volume increase and economic benefit loss between yield and irrigation water use efficiency, and complete the irrigation decision risk assessment.

[0012] Optionally, in step S1, the drone is divided into multispectral drone and thermal infrared drone. The two types of drones carry out routine flight operations throughout the entire growth cycle of corn. The flight range covers all independent irrigated fields in the farm. Multiple fixed time periods are set every day to complete image acquisition. The types of images acquired include RGB images, multispectral images and thermal infrared images.

[0013] Ground soil moisture sensors are arranged in a grid pattern inside each irrigated field, and meteorological sensors are simultaneously deployed in the field area. The meteorological sensors continuously collect field environmental data, and the soil moisture sensors continuously collect soil moisture content data at different soil layers.

[0014] The drone image acquisition work was carried out simultaneously with the ground sensor acquisition work.

[0015] Optionally, in step S1, the Swin-ConvNeXt deep learning model is composed of an input adaptation layer, a dual-branch feature extraction layer, a feature fusion layer, a global average pooling layer, and a single-value regression layer connected in sequence. The model internally sets up a dual-branch feature extraction structure with the Swin Transformer branch and the ConvNeXt branch in parallel. The image samples input to the model are first processed uniformly and then enter the input adaptation layer to carry out the integration and transformation of heterogeneous image channels. The two independent branches extract features from the image data respectively. All the extracted feature data are sent to the feature fusion layer for integration. The integrated feature vector is passed to the global average pooling layer to achieve spatial dimension compression. The processed feature data is finally input to the single-value regression layer to complete the corresponding numerical output.

[0016] Optionally, in step S2, the field trial data comes from the full-cycle collection content of the multi-group zonal trial. The trial area is divided into trial areas with different irrigation treatment combinations, trial areas with different planting density combinations, and exclusive trial areas distinguished by sowing date. During the sowing stage, soil samples are collected in layers and indoor tests are completed to form soil-related data. Throughout the entire growth cycle of maize, crop phenological information, plant morphology information, root distribution information, and aboveground and underground biomass information are collected and recorded sequentially according to different growth stages. During the crop harvest stage, grain yield-related information is collected uniformly. Irrigation operation information of each zone is recorded synchronously throughout the entire trial. All collected content is uniformly classified and organized according to the trial zone and collection time sequence to form a complete field trial dataset.

[0017] Optionally, in step S2, the MAIZSIM is a maize-specific mechanism model. It adopts a two-dimensional spatial architecture and performs simulation calculations with a fixed time step. The model is equipped with soil hydrothermal coupling calculation units, solute transport units, crop growth and development units, and root growth calculation units. Soil-related calculations rely on spatial discrete grids to carry out global numerical analysis. The root growth part adopts a calculation architecture corresponding to biomass allocation. The model synchronously integrates plant morphology units, growth process units, and dry matter accumulation units. Data communication and synchronous calculations are realized between various functional units to form an integrated model calculation system.

[0018] Optionally, in step S3, at the spatial level, the two-dimensional grid data output by the localized MAIZSIM model is aggregated according to the independent irrigated fields within the farm, and the aggregated data and the basic agricultural observation data use the same spatial division unit; at the temporal level, the simulation step size of the localized MAIZSIM model is matched one by one with the collection time of the basic agricultural observation data.

[0019] The localized MAIZSIM model starts its preheating operation on a fixed date each year. During the preheating phase, measured soil moisture content data and groundwater level data of the entire farm are recorded. During the model retrospective operation phase, hourly meteorological data collected from field stations are accessed. During the model prediction operation phase, long-term and short-term meteorological data released by meteorological forecasting agencies are accessed. These two types of meteorological data correspond to the 90-day agricultural sequence generation stage and the 15-day agricultural sequence generation stage, respectively.

[0020] Optionally, in step S3, the iterative set smoother generates multiple sets of initial parameter samples based on the distribution characteristics of the model parameters. The iterative calculation process adopts the Gauss-Newton calculation method. The calculation process integrates all time-series basic agricultural observation data. The model parameter set is continuously updated during the iteration process. When the numerical change generated by multiple adjacent iterations reaches the preset judgment standard, the iteration process ends and the model calibration is completed.

[0021] Optionally, in step S4, the constraints include the farm's total annual irrigation water volume, time-series data of cross-regional water diversion, maximum and minimum storage capacity of the reservoir, daily water delivery volume of the irrigation system, range of single-time irrigation water volume for a single field, and lower limit threshold of soil moisture content corresponding to different growth stages of maize. The calculation is divided into long-cycle planning calculation and short-cycle real-time calculation. The long-cycle planning uses 15 days as a time unit, while the short-cycle real-time calculation uses a natural day as a time unit. The long-cycle planning generates phased water allocation data and reservoir water storage data, while the short-cycle real-time calculation generates irrigation water volume data and irrigation time period data for each field. After integration, a complete field irrigation execution instruction is formed.

[0022] Optionally, in step S5, the combined simulation scenario includes a hydrological type scenario, an observation error scenario, an external water diversion scenario, and a field management scenario. The hydrological type is divided into three categories: abundant water, normal water, and low water. The observation error is divided into three levels. The external water diversion is configured with multiple sets of random time-series data. The field management is matched with different plant densities and different sowing times. The simulation process uses the localized MAIZSIM model as the simulation carrier and simultaneously sets the traditional irrigation mode as the comparison group. Multiple sets of parallel simulations are carried out based on the independent irrigated fields inside the farm as the basic computing unit. After the simulation is completed, the yield data and irrigation water volume data of each group of simulations are extracted in batches to complete the corresponding probability distribution statistics.

[0023] Based on the same inventive concept, the present invention also provides a precision irrigation decision-making system for arid farms under water resource constraints. The system includes a multi-source data acquisition module, a data processing and storage module, a model localization module, an agricultural condition sequence generation module, an irrigation decision solving module, a simulation risk assessment module, and an instruction issuance module.

[0024] The multi-source data acquisition module is used to carry a multispectral drone, a thermal infrared drone, a soil moisture sensor and a meteorological sensor to simultaneously acquire multi-source images of cornfields, stratified soil moisture content, field meteorology and full-cycle data of field experiments, and upload the acquired data to the data processing and storage module in real time.

[0025] The data processing and storage module is equipped with a built-in Swin-ConvNeXt deep learning model and an inverse distance weight interpolation algorithm. It is used to calculate the leaf area index based on UAV images, generate layered soil moisture content based on UAV images and ground sensor data, complete data sorting, classification and storage, and construct and store basic agricultural observation datasets and field test datasets.

[0026] The model localization module interfaces with the data processing and storage module, and integrates the Morris global sensitivity analysis algorithm and the particle swarm optimization algorithm to screen and calibrate the MAIZSIM model parameters and output the localized MAIZSIM model.

[0027] The agricultural sequence generation module is connected to the model localization module and the external meteorological data source respectively. It is used to unify the spatiotemporal scale of the localized MAIZSIM model and the observation data, call the iterative ensemble smoother to complete the model calibration, and input long-term and short-term meteorological data to generate a 90-day long-term agricultural sequence and a 15-day short-term agricultural sequence.

[0028] The irrigation decision-solving module, with constraints such as water availability, irrigation infrastructure capacity, and crop growth stage, integrates the NSGA-Ⅲ multi-objective optimization algorithm, receives the crop condition sequence output by the crop condition sequence generation module, solves and outputs field irrigation execution instructions;

[0029] The simulation risk assessment module is connected to the model localization module. It builds a virtual farmland simulation scene based on the localized MAIZSIM model, configures multiple combined simulation scenarios such as hydrology, observation error, external water diversion, and field management, and conducts batch simulations in parallel. It calls the uncertainty quantification algorithm to complete the statistics and assessment of risk indicators related to yield, irrigation volume, and economic benefits.

[0030] The instruction issuing module is connected to the irrigation decision solving module and the field irrigation execution equipment, respectively. It is used to receive irrigation execution instructions and issue the instructions to the irrigation facilities in each field of the farm to complete the automated irrigation operation scheduling.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] I. This invention achieves comprehensive agricultural data collection by integrating multi-source UAV imagery and ground sensors. It utilizes deep learning models and spatial correction technology to accurately calculate leaf area index and stratified soil moisture content, constructing a standardized and structured observation dataset. This overcomes the limitations of traditional field monitoring data, which suffers from single-dimensionality and insufficient spatial coverage. Combined with field experimental data, it completes parameter selection and localization calibration of crop mechanism models. Data assimilation technology unifies spatiotemporal scales and calibrates the model. By connecting multi-source meteorological data, it generates agricultural data sequences for different periods, making the model simulation results more closely match the actual growth environment of arid farmland. The entire data collection and model optimization chain significantly improves the accuracy of agricultural data perception and the reliability of simulation, providing real, continuous, and comprehensive data support for subsequent irrigation decisions. It is suitable for operational scenarios in arid areas with scattered plots and uneven water and fertilizer conditions.

[0033] Second, this invention sets multiple constraints by combining the total regional water resources, the carrying capacity of engineering facilities, and crop growth patterns. It uses a multi-objective optimization algorithm to output irrigation execution instructions in stages and plots, taking into account both total water use control and precise water allocation needs in the field. This effectively alleviates the problems of water shortage and unreasonable irrigation allocation in arid areas. At the same time, it builds virtual simulation scenarios based on localized crop models, sets up multiple types of combined scenarios to carry out batch simulations, and uses uncertainty quantification methods to complete the risk assessment of irrigation decisions from multiple dimensions such as yield, water use efficiency, and economic benefits. This not only realizes the transformation of irrigation schemes from experience-based to quantitative and intelligent, but also enables the prediction of potential risks of decision-making schemes under different environmental and management conditions. It helps farm managers flexibly adjust irrigation strategies and ensure stable crop output and comprehensive benefits under limited water resources.

[0034] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims, and drawings. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating a precision irrigation decision-making method for arid farms under water resource constraints, according to an embodiment of the present invention.

[0037] Figure 2 This is a heatmap of water stress time windows and irrigation decision results for various plots of farmland according to an embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of the structure of a precision irrigation decision-making system for arid farms under water resource constraints, according to an embodiment of the present invention.

[0039] Figure 4 This is a flowchart of the drone image preprocessing and dataset construction process.

[0040] Figure 5 This is the architecture diagram of the Swin-ConvNeXt model.

[0041] Figure 6 This is a schematic diagram of the calculation of soil moisture content correction coefficient based on the inverse distance weighting method.

[0042] Figure 7 This is a schematic diagram of the assimilation simulation of multi-source remote sensing from UAVs and the MAIZSIM crop model.

[0043] Figure 8 This is a farm-scale, long-term and short-term two-level linkage irrigation decision-making framework diagram.

[0044] Figure 9 This is a schematic diagram of the process for generating virtual observation information from unmanned aerial vehicles (UAVs) and the framework for assessing the impact of its uncertainties. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Reference Figure 1One embodiment of the present invention proposes a precision irrigation decision-making method for arid farms under water resource constraints. It adopts a full-process technical system including air-ground collaborative multi-source crop condition monitoring, localized calibration of crop mechanism model parameters, multi-source data assimilation for crop condition prediction, multi-constraint multi-objective optimization decision-making, and virtual simulation risk assessment. It can achieve precise, dynamic, and intelligent decision-making for farm-scale cornfield irrigation under arid conditions with limited water resources, significantly improving irrigation water use efficiency and crop yield stability.

[0047] The method described in this embodiment specifically includes:

[0048] S1. Multi-source images and soil moisture content data of cornfields are collected using UAVs and ground sensors. The leaf area index is calculated from the images using the Swin-ConvNeXt deep learning model. Soil moisture content information is obtained by combining UAV images and ground sensor data. Spatial correction is performed using the inverse distance weighting method to form basic agricultural observation data.

[0049] S2. Using field test data and obtained basic agricultural observation data as input, the parameters of the MAIZSIM model are screened through Morris global sensitivity analysis, and the parameters of the MAIZSIM model are calibrated through particle swarm optimization algorithm to obtain the localized MAIZSIM model.

[0050] S3. Unify the spatiotemporal scale of the localized MAIZSIM model and basic agricultural observation data, use an iterative ensemble smoother to fuse the basic agricultural observation data to calibrate the model, and connect to the multi-source meteorological data-driven model to generate a 90-day long-term agricultural sequence and a 15-day short-term agricultural sequence.

[0051] S4. Using the generated long-term and short-term agricultural conditions sequences as inputs, and constrained by water availability, irrigation infrastructure capacity and crop growth stage, the NSGA-Ⅲ multi-objective optimization algorithm is used to solve for the field irrigation execution instructions.

[0052] S5. Based on the obtained localized MAIZSIM model, construct a virtual farmland simulation scenario, set up combined simulation scenarios and carry out batch simulations, and use uncertainty quantification algorithm to statistically analyze the relative yield loss, irrigation volume increase and economic benefit loss between yield and irrigation water use efficiency, and complete the irrigation decision risk assessment.

[0053] Specifically, the overall logic of this invention lies in constructing a closed-loop irrigation decision-making system that encompasses precise perception, scientific prediction, optimized decision-making, and risk management. This system first acquires high-precision leaf area index and stratified soil moisture content data covering the entire farm area through multi-source observation methods involving both air and ground, addressing the shortcomings of traditional single-point monitoring (insufficient representativeness) and poor timeliness of satellite remote sensing. Then, relying on the maize-specific mechanism model MAIZSIM, and combining field trial data, it completes localized calibration of model parameters, ensuring the model accurately represents maize growth and soil moisture transport processes in the study area. Based on this, an iterative ensemble smoother integrates real-time observation data into the model, achieving dynamic calibration of agricultural conditions and accurate prediction of future trends. Furthermore, with multiple constraints including total water resources, engineering capacity, and crop requirements, a multi-objective optimization algorithm generates irrigation execution instructions for different time periods and fields. Finally, through virtual simulation and uncertainty quantification analysis, it assesses the potential risks of decision-making schemes under different scenarios, providing a comprehensive scientific basis for irrigation decisions.

[0054] Optionally, in step S1, the drones are divided into multispectral drones and thermal infrared drones. Both types of drones carry out routine flight operations throughout the entire growth cycle of corn, with the flight range covering all independent irrigated fields in the farm. Multiple fixed time periods are set up each day to complete image acquisition. The types of images acquired include RGB images, multispectral images, and thermal infrared images. Ground soil moisture sensors are deployed in a grid layout inside each irrigated field, and meteorological sensors are deployed simultaneously in the field area. The meteorological sensors and soil moisture sensors continuously collect field environmental data and soil moisture content data at different soil layers. The drone image acquisition work and the ground sensor data acquisition work are carried out simultaneously.

[0055] Specifically, the multispectral drone is equipped with an RGB camera and a 5-channel multispectral sensor, which can simultaneously acquire image data in the blue, green, red, red-edge, and near-infrared bands for crop leaf area index retrieval. The thermal infrared drone is equipped with a 640×512 resolution thermal infrared imager to acquire canopy temperature data to assist in soil moisture content retrieval. The flight frequency is dynamically adjusted according to the maize growth stage: once every 3 days during the seedling and maturity stages, and once a day during the faster growth stages from jointing to grain filling. The daily flight times are set at 9:00, 11:00, 13:00, 15:00, and 17:00 to cover different lighting conditions and build a multi-scene dataset. Ground soil moisture sensors are deployed in a 50m×50m grid, with one set of sensors buried at the center of each grid to monitor soil moisture content at depths of 10cm, 20cm, and 30cm, with a sampling interval of 5 minutes. Field weather stations simultaneously collect meteorological data such as temperature, relative humidity, solar radiation, wind speed, and precipitation, with a sampling interval of 5 minutes. All data acquisition devices are synchronized via GPS to ensure strict alignment between image data and ground sensor data in the time dimension. After collection, all data is transmitted in real time to the farm's data center via a wireless communication network, where it is categorized and stored according to field number and collection time, forming a raw observation database.

[0056] For example, the farm has a total area of ​​120 hectares, divided into 24 independent irrigated plots, each approximately 5 hectares in size. Soil moisture sensors are deployed in a 50m × 50m grid, with a total of 480 sensor sets. Each sensor set monitors soil moisture content at depths of 10cm, 20cm, and 30cm. During the corn jointing stage, drone flights are conducted daily at 9:00, 11:00, 13:00, 15:00, and 17:00, covering all 24 plots each time. The flight altitude is 30m, the flight speed is 5m / s, the forward overlap is 80%, and the lateral overlap is 70%. The RGB image resolution for each flight is 2.5cm, the multispectral image resolution is 10cm, and the thermal infrared image resolution is 20cm. Ground sensors and a weather station collect data synchronously, and all data is transmitted in real-time to the farm's data center via a 4G network. The daily data storage volume is approximately 15GB. Figure 4 As shown.

[0057] Optionally, in step S1, the Swin-ConvNeXt deep learning model is composed of an input adaptation layer, a dual-branch feature extraction layer, a feature fusion layer, a global average pooling layer, and a single-value regression layer connected in sequence. The model internally sets up a dual-branch feature extraction structure with the Swin Transformer branch and the ConvNeXt branch in parallel. The image samples input to the model are first processed uniformly and then enter the input adaptation layer to carry out the integration and transformation of heterogeneous image channels. The two independent branches extract features from the image data respectively. All the extracted feature data are sent to the feature fusion layer for integration. The integrated feature vector is passed to the global average pooling layer to achieve spatial dimension compression. The processed feature data is finally input to the single-value regression layer to complete the corresponding numerical output.

[0058] Specifically, the input adaptation layer uses 3×3 convolutional kernels to uniformly map the heterogeneous input channels of RGB + multispectral + vegetation index to a 64-dimensional feature channel. Cross-channel weighted fusion is used to initially mine complementary information between different spectral and index channels. In the dual-branch feature extraction layer, the Swing Transformer branch consists of four stacked Swing Transformer Blocks, each containing a windowed multi-head self-attention module and a feedforward neural network, responsible for extracting global spatial correlation features of the image; the ConvNeXt branch consists of four stacked ConvNeXt Blocks, each containing depthwise separable convolutions, layer normalization, and a GELU activation function, responsible for extracting local spectral detail features of the image. The feature fusion layer uses a concatenation fusion method, concatenating the feature vectors output from the two branches into a 128-dimensional fused feature vector. The global average pooling layer compresses the 128-dimensional spatial feature vector into a 128-dimensional global feature vector, directly mapping the global features of the entire image and matching the population attributes of the leaf area index. The single-valued regression layer consists of two fully connected layers. The first fully connected layer maps the 128-dimensional feature vector to a 64-dimensional vector, and the second fully connected layer maps the 64-dimensional feature vector to a single continuous value, namely the average leaf area index of the field corresponding to the image. During model training, ImageNet pre-trained weights are used to initialize the dual-branch backbone. The HuberLoss loss function is selected, and the AdamW optimizer is used with a learning rate of 5e-5. A weight decay and cosine annealing learning rate scheduling strategy is employed to suppress model overfitting. Figure 5 As shown.

[0059] For example, when constructing the leaf area index (LAI) inversion model, UAV imagery and ground-based measured data were collected throughout the entire growth period of maize, resulting in a total of 3600 valid samples. These samples were divided into training, validation, and test sets in a 7:2:1 ratio. During training, the batch size was set to 32, and the number of iterations was set to 100. After training, the model's coefficient of determination R on the test set was calculated.2 The mean square error (RMSE) was 0.92, the root mean square error (RMSE) was 0.31, and the mean absolute error (MAE) was 0.24. For a 5m×5m quadrat, with the corresponding 256×256 pixel image block as input, the model output leaf area index was 4.2, with a relative error of 2.4% compared to the ground-measured value of 4.1.

[0060] Optionally, in step S1, the calculation formulas used to calculate the leaf area index and correct the soil moisture content are as follows: The leaf area calculation formula is:

[0061]

[0062] Among them, the leaf area of ​​a single plant , This represents the number of corn plants within the quadrat. This refers to the number of leaves on a single corn plant. For the first sample plot Total leaf area of ​​corn plants The area of ​​the sample plot. For corn Maximum blade width, For corn Blade length;

[0063] The formula for calculating soil moisture content deviation is:

[0064]

[0065] In the formula, For the first The deviation coefficient of each monitoring quadrat. Soil moisture content collected by ground sensors. Soil moisture content obtained from drone inversion;

[0066] The formula for calculating the normalized weights of neighboring sensors is:

[0067]

[0068] In the formula, For the sample plots to be calibrated With the Normalized weights for each sensor sample plot, Let the distance between the two sample squares be denoted as . This is the weight decay coefficient. neighbor

[0069] Number of sensor quadrats within the domain. To achieve a balance with the silent element;

[0070] The corrected formula for calculating soil moisture content is:

[0071]

[0072] In the formula, To obtain soil moisture content from unmanned aerial vehicle (UAV) data for the sample plots to be calibrated, This is the corrected soil moisture content.

[0073] Specifically, during ground-based measurements of leaf area index (LAI), three representative maize plants were randomly selected within each 5m × 5m quadrat. The maximum width and length of each leaf were measured using a measuring tape, and the leaf area per plant was calculated. The average LAI was then calculated and combined with the plant density within the quadrat to obtain the average LAI for that quadrat. For soil moisture content correction, the deviation coefficient of each sensor quadrat was first calculated to reflect the relative deviation between the UAV-retrieved value and the ground-based measured value. Then, taking the quadrat to be corrected as the center, all sensor quadrats within a 100m radius were selected as neighborhood samples. The normalized weight of each neighborhood sample was calculated based on spatial distance; the closer the distance, the greater the weight, and the weight attenuation coefficient k was set to 2. Finally, the deviation coefficients and corresponding weights of the neighborhood samples were used to perform weighted correction on the soil moisture content retrieved by the UAV, resulting in the corrected soil moisture content. This method effectively integrates high-precision single-point data from ground sensors with high spatial resolution data from the UAV, improving the overall accuracy and spatial continuity of soil moisture content retrieval. Figure 6 As shown.

[0074] For example, the soil moisture content at a depth of 10 cm retrieved by UAV for sample plot p to be corrected is 18.2%. There are three sensor sample plots in its neighborhood, at distances of 30 m, 60 m, and 90 m, with corresponding bias coefficients of 0.05, -0.03, and 0.02, respectively. According to the formula for calculating the normalized weights of neighborhood sensors, the weights of the three sample plots are calculated to be 0.73, 0.18, and 0.09, respectively. Substituting these values ​​into the formula for calculating the corrected soil moisture content, the corrected soil moisture content is obtained as follows: Compared with the ground-measured value of 18.9% for the same sample plot, the relative error was 0.58%, which is a significant improvement compared with the relative error of 3.7% before correction.

[0075] Optionally, in step S2, the field trial data comes from the full-cycle collection content of the multi-group zonal trial. The trial area is divided into trial areas with different irrigation treatment combinations, trial areas with different planting density combinations, and exclusive trial areas distinguished by sowing date. During the sowing stage, soil samples are collected in layers and indoor tests are completed to form soil-related data. Throughout the entire growth cycle of maize, crop phenological information, plant morphology information, root distribution information, and aboveground and underground biomass information are collected and recorded sequentially according to different growth stages. During the crop harvest stage, grain yield-related information is collected uniformly. Irrigation operation information of each zone is recorded synchronously throughout the entire trial. All collected content is uniformly classified and organized according to the trial zone and collection time sequence to form a complete field trial dataset.

[0076] Specifically, the field trial included three irrigation treatments and two planting density treatments, for a total of six treatments. Each treatment had three replicates, resulting in a total of 18 plots, each measuring 20m × 20m. The irrigation quota for each irrigation treatment was 2250 m³. 3 / hm 2 3100m 3 / hm 2 and 3700m 3 / hm 2 The planting densities were 80,000 plants / ha and 66,667 plants / ha, respectively. Two additional experimental plots were set up with sowing dates in mid-April and mid-May, with irrigation and planting densities consistent with the greywater treatment. Before sowing, five sampling points were selected in an S-shape in each experimental plot, and soil samples were collected from four layers: 0-20cm, 20-40cm, 40-60cm, and 60-100cm. Soil mechanical composition, bulk density, field water holding capacity, saturated hydraulic conductivity, available nitrogen, available phosphorus, and available potassium were measured. During the maize growth period, crop phenological information, plant height, stem diameter, and leaf area were collected every 10 days, and root distribution and biomass data were collected at each growth stage. At harvest, three 4m samples were randomly selected from each plot. 2 The quadrats were used to determine grain yield and thousand-grain weight. All experimental data were entered into a database and managed uniformly according to experimental zones and collection time.

[0077] For example, the field trial consisted of 20 plots, 18 of which were for a two-factor treatment of irrigation amount and planting density, and 2 for a treatment of sowing date. Soil samples collected before sowing showed that the bulk density of the 0-20cm soil layer in the experimental area was 1.35 g / cm³. 3 The field water holding capacity was 24.5%, and the saturated hydraulic conductivity was 0.8 cm / h. During the maize growth period, crop morphology data were collected 12 times, root distribution data 6 times, and biomass data 4 times. The yield range for each treatment, measured at harvest, was 9800 kg / hm².2 Up to 13200 kg / hm 2 The irrigation water use efficiency ranges from 1.8 kg / m³. 3 Up to 2.6 kg / m 3 All experimental data were entered into the field experiment database in a standardized format, resulting in over 12,000 valid data records.

[0078] Optionally, in step S2, the calculation formula used to screen the parameters of the MAIZSIM model through Morris global sensitivity analysis is as follows: Parameter basic effect calculation formula:

[0079]

[0080] In the formula, For the parameter variation step size, , For the parameter discrete level; For the first Group basic sample vector; For the first The unit vector corresponding to each parameter; For the sample The corresponding model output value; For the first Increase one parameter The model output value corresponding to the later sample; For the first Group of samples The basic effects of each model parameter;

[0081] The formula for calculating the mean parameter sensitivity is:

[0082]

[0083] In the formula, The global sensitivity mean of the parameter. This represents the number of trajectories analyzed for sensitivity.

[0084] Specifically, a literature review was first conducted to preliminarily determine 32 soil and crop parameters and their value ranges for the MAIZSIM model. The Morris method was then used for global sensitivity analysis, and the number of trajectories was set. Discrete level of each parameter The mean sensitivity of each parameter was calculated using leaf area index, yield, transpiration and evaporation, and average soil moisture content at a depth of 10 cm as output variables. with standard deviation The mean sensitivity Parameters with a value greater than 0.1 were identified as sensitive parameters, and a total of 12 sensitive parameters were selected, including maximum photosynthetic rate of leaves, root growth coefficient, soil saturated hydraulic conductivity, and field capacity. These parameters will be used for subsequent model calibration, while the remaining parameters will use the model's default values ​​or values ​​recommended in the literature.

[0085] For example, in the sensitivity analysis of the MAIZSIM model parameters, the mean sensitivity of the maximum photosynthetic rate of leaves for the yield output variable is... The mean sensitivity of the root growth coefficient is 0.86. The mean sensitivity of soil saturated hydraulic conductivity is 0.72. The values ​​were all greater than 0.1, and were therefore considered sensitive parameters; while the average sensitivity of the leaf senescence coefficient was 0.58. The value is 0.06, less than 0.1, and is therefore considered a non-sensitive parameter; the model default value is used. The mean sensitivity value of soil saturated hydraulic conductivity to the output variable of soil moisture content at a depth of 10cm is... The mean sensitivity of field water holding capacity is 0.92. The value is 0.81, which is also considered a sensitive parameter.

[0086] Optionally, in step S2, the parameters of the MAIZSIM model are calibrated using a particle swarm optimization algorithm to obtain a localized MAIZSIM model.

[0087] Specifically, the particle swarm optimization algorithm was set to a population size of 50, an iteration count of 200, an inertia weight of 0.7, a cognitive factor of 1.5, and a social factor of 1.5. The objective function was to minimize the root mean square error between simulated and measured values, and the 12 selected sensitive parameters were calibrated. The objective function expression is:

[0088]

[0089] In the formula, These are the simulated values ​​from the model. These are measured values ​​on the ground. The sample size is used. Parameter calibration is performed using field trial data from the first year, and model validation is performed using field trial data from the second year. The coefficient of determination of the model on the validation set is... When the root mean square error (RMSE) is greater than 0.85 and less than 15% of the standard deviation of the measured value, the model is considered to be calibrated successfully, and the localized MAIZSIM model is obtained.

[0090] For example, after calibration using the particle swarm optimization algorithm, the coefficient of determination of the localized MAIZSIM model in leaf area index simulation is... The coefficient of determination for production simulation is 0.89, and the root mean square error (RMSE) is 0.35; The mean square error (RMSE) is 0.91, and the root mean square error (RMSE) is 420 kg / hm. 2 The coefficient of determination in the simulation of soil moisture content at a depth of 10 cm. The accuracy is 0.87, and the root mean square error (RMSE) is 1.2%, both of which meet the accuracy requirements of the model simulation.

[0091] Optionally, in step S3, at the spatial level, the two-dimensional grid data output by the localized MAIZSIM model is aggregated according to the independent irrigated fields within the farm, and the aggregated data and the basic agricultural observation data use the same spatial division unit; at the temporal level, the simulation step size of the localized MAIZSIM model is matched one by one with the collection time of the basic agricultural observation data; the localized MAIZSIM model starts preheating operation on a fixed date each year, and during the preheating stage, the measured soil moisture content data and groundwater level data of the entire farm are recorded; during the model backtracking operation stage, hourly meteorological data collected from field stations are accessed, and during the model prediction operation stage, long-term meteorological data and short-term meteorological data released by meteorological forecasting agencies are accessed, and the two types of meteorological data correspond to the 90-day agricultural sequence generation stage and the 15-day agricultural sequence generation stage, respectively.

[0092] Specifically, the MAIZSIM model divides each irrigated field into a 1m × 1m two-dimensional grid, simulating soil water and heat transport and crop growth processes within each grid. With a unified spatial scale, the average leaf area index and soil moisture content of all grids in each field are calculated to obtain simulated values ​​at the field scale, consistent with the spatial units of basic agricultural observation data. With a unified temporal scale, the MAIZSIM model uses a 1-hour simulation step, extracting simulated values ​​at each data collection time and matching them with observed values. The model begins its preheating run on March 1st each year, during which measured soil moisture content data and groundwater level data at a depth of 0-100cm are recorded across the entire farm, running until the sowing date. During the retrospective run, hourly data on temperature, humidity, solar radiation, wind speed, and precipitation collected from field weather stations are integrated, running up to the current time. During the forecasting phase, the system accesses daily meteorological data for the next 90 days and hourly meteorological data for the next 15 days released by meteorological forecasting agencies to generate a 90-day long-term agricultural condition sequence and a 15-day short-term agricultural condition sequence, respectively. The sequence content includes leaf area index, stratified soil moisture content, crop water requirements, etc. for each field.

[0093] For example, the MAIZSIM model divides each 5-hectare irrigated field into 50,000 1m×1m two-dimensional grids. During spatial aggregation, the average soil moisture content at a depth of 10cm is calculated across all grids in each field, yielding simulated soil moisture content values ​​at the field scale. For a unified time scale, the model's simulated values ​​at 13:00 daily are extracted and matched with image inversion values ​​collected by a drone at 13:00. The model begins its preheating run on March 1st each year, recording measured soil moisture content data (0-100cm depth) and groundwater level data across the entire farm, running until the sowing date of April 20th. During the retrospective run, hourly meteorological data from field weather stations is integrated, running up to the current date. During the forecasting run, meteorological data from weather forecasting agencies is integrated to generate daily agricultural condition sequences for the next 90 days and hourly agricultural condition sequences for the next 15 days, with the sequences updated daily. Figure 7 As shown.

[0094] Optionally, in step S3, the calculation formula used by the iterative set smoother is as follows:

[0095] The formula for the model's input-output relationship is:

[0096]

[0097] In the formula, For the observation vector, For the MAIZSIM forward simulation model, For model parameter vectors, Let be the observation error vector, and let be the observation error covariance. ;

[0098] The formula for generating the initial parameter set is:

[0099]

[0100] In the formula, For the initial parameter set, For the parameter sample size, to The initial parameter samples for each group are listed in order.

[0101] The parameter iterative update formula is:

[0102]

[0103] The formula for calculating the set sensitivity matrix is:

[0104]

[0105] In the formula, For the first The set of parameters for the next iteration The parameter set from the previous iteration. This is the iteration step size coefficient; Let the prior covariance matrix be the parameter. , For the initial parameter set The deviation matrix from its mean; For the first The set approximation of the Jacobian matrix output by the model in the next iteration is derived from the parameter set. The corresponding model output set is calculated; The observation error covariance matrix; This is the set sensitivity matrix.

[0106] Specifically, the number of parameter samples for the iterative set smoother. Set to 100, the initial parameter set is randomly generated within the calibrated parameter value range using a log-normal distribution. Iteration step size coefficient. Set to version 1.0, the iteration process employs Gauss-Newton calculations, integrating all time-series basic agricultural observation data. The iteration process terminates and model calibration is completed when the relative change in the mean of the parameter sets between two adjacent iterations is less than 1e-4. This algorithm can simultaneously utilize all available historical observation information, effectively suppressing interference from anomalous observations and improving the robustness and accuracy of model parameter calibration.

[0107] For example, in a model calibration process, the initial parameter set contained 100 samples. After 8 iterations, the relative change in the mean of the parameter set was 8.6e-5, which was less than the preset convergence threshold of 1e-4, and the iteration ended. The calibrated model reduced the root mean square error in leaf area index simulation from 0.35 to 0.28, and the root mean square error in soil moisture content simulation from 1.2% to 0.9%, further improving the model simulation accuracy.

[0108] Optionally, in step S4, the constraints include the farm's total annual irrigation water volume, time-series data of cross-regional water diversion, maximum and minimum storage capacity of the reservoir, daily water delivery volume of the irrigation system, range of single-time irrigation water volume for a single field, and soil moisture values ​​corresponding to different growth stages of maize. The calculation is divided into long-term planning calculation and short-term real-time calculation. The long-term planning uses 15 days as a time unit, while the short-term real-time calculation uses a natural day as a time unit. The long-term planning generates phased water allocation data and reservoir water storage data, while the short-term real-time calculation generates irrigation water volume data and irrigation time period data for each field. After integration, a complete field irrigation execution instruction is formed.

[0109] Specifically, the long-term programming operation is performed one day before the start of each irrigation cycle. Based on the agricultural forecast sequence and weather forecast data for the next 90 days, and using the cumulative irrigation water volume every 15 days and the water storage capacity of the reservoir at the end of each period as decision variables, a multi-objective optimization model is constructed with the objectives of minimizing irrigation water volume and maximizing economic benefits. Constraints include: the farm's total annual irrigation water volume shall not exceed 370,000 m³. 3 The maximum capacity of the reservoir is 100,000 m³. 3 The minimum capacity is 10,000 m³. 3 The daily water delivery volume of the irrigation system shall not exceed 5000 m³. 3 The irrigation volume for a single plot is between 10-30 mm. Soil moisture content should be no less than 60% of field capacity during the corn seedling stage, no less than 70% during the jointing stage, and no less than 75% during the grain-filling stage. Short-cycle real-time calculations are performed daily at 00:00, based on a 15-day agricultural forecast sequence and hourly weather forecast data. Using the daily irrigation volume for each plot as the decision variable, and under the constraint of available water allocated through long-cycle planning, the irrigation plan for each plot is optimized for that day. The preferred irrigation period is from 20:00 to 6:00 the following day to reduce evaporation losses. Figure 2 As shown, Figure 2 'a' represents the distribution map of crop water stress index in various plots. Figure 2 b represents a schematic diagram showing the distribution of irrigation periods for different plots.

[0110] Optionally, in step S4, the formula for calculating the net crop yield used in the field irrigation process is as follows:

[0111]

[0112] In the formula, For net income from planting, This refers to the market price per unit of corn. For the farm's planting area, The yield of corn per unit area, For unit water supply cost, For the first Irrigation water volume for each plot of land This is the surface water utilization coefficient.

[0113] Specifically, the market price per unit of corn The unit water supply cost is calculated using the local average purchase price for that year. Including water fees, electricity fees, equipment maintenance fees, etc., surface water utilization coefficient Take 0.92. Production Irrigation water volume was obtained from simulations using the localized MAIZSIM model. These are the decision variables. The NSGA-III multi-objective optimization algorithm is used to solve the problem, setting the population size to 100, the number of iterations to 100, the crossover probability to 0.9, and the mutation probability to 0.1, obtaining the Pareto optimal solution set. Then, based on the actual needs of the farm, by introducing objective weights, the multi-objective problem is transformed into a single-objective problem to determine the final irrigation scheme, such as... Figure 8 As shown.

[0114] For example, the farm has a planting area of ​​120 hectares, the market price of corn is 2.4 yuan / kg, and the unit water supply cost is 0.6 yuan / m². 3 The surface water utilization coefficient is 0.92. After long-term planning calculations, the available irrigation water volume for a certain 15-day period was determined to be 45,000 m³. 3 The system generates a daily irrigation plan using short-cycle real-time calculations. For example, the irrigation water volume for field 3 is 20mm, corresponding to an irrigation volume of 1000m³. 3 The irrigation period is from 22:00 to 0:00 the next day; the irrigation water volume for plot 6 is 15mm, corresponding to an irrigation volume of 750m³. 3 Irrigation will be carried out from 0:00 to 2:00 the following day; the remaining fields will not be irrigated. According to the formula for calculating net planting income, this irrigation plan is expected to increase the farm's net income by 12,600 yuan on that day.

[0115] Optionally, in step S5, the combined simulation scenario includes a hydrological type scenario, an observation error scenario, an external water diversion scenario, and a field management scenario. The hydrological type is divided into three categories: abundant water, normal water, and low water. The observation error is divided into three levels. The external water diversion is configured with multiple sets of random time-series data. The field management is matched with different plant densities and different sowing times. The simulation process uses the localized MAIZSIM model as the simulation carrier and simultaneously sets the traditional irrigation mode as the comparison group. Multiple sets of parallel simulations are carried out based on the independent irrigated fields inside the farm as the basic computing unit. After the simulation is completed, the yield data and irrigation water volume data of each group of simulations are extracted in batches to complete the corresponding probability distribution statistics.

[0116] Specifically, the hydrological scenario is divided into three categories based on precipitation data from the study area over the past 30 years: wet years, normal years, and dry years. The observation error scenario is set at three levels: low, medium, and high, corresponding to relative errors of 5%, 10%, and 15%, respectively. The external water diversion scenario randomly generates 20 different sets of water diversion time-series data to reflect the uncertainty of water inflow. The field management scenario sets two planting densities and two sowing dates, for a total of four combinations. By combining the above four scenarios, 3 × 3 × 20 × 4 = 720 simulation scenarios are generated. Each scenario is tested using the irrigation decision-making method of this invention and traditional empirical irrigation methods to simulate the growth and irrigation processes of maize throughout its entire growth period. After the simulation is completed, the yield, irrigation volume, and net planting income under each scenario are statistically analyzed, and the probability distributions of relative yield loss, irrigation volume increase, and economic benefit loss are calculated.

[0117] Optionally, in step S5, the calculation formula used for uncertainty quantification is as follows:

[0118] The formula for calculating relative production loss is:

[0119]

[0120] In the formula, For relative output loss, This represents the total number of simulations. For the first Yield data under conditions where the agricultural situation is completely known in this simulation. For the first Yield data under conditions of uncertainty in observation information in this simulation; the formula for calculating the increase in irrigation volume is:

[0121]

[0122] In the formula, To increase the irrigation amount, For the first Irrigation data under the condition that the agricultural situation is completely known in this simulation. For the first Irrigation data under conditions of uncertainty in observation information in this simulation; the formula for calculating economic benefit loss is:

[0123]

[0124] In the formula, For economic losses, For the first The net planting profit under the condition that the agricultural conditions are completely known in this simulation. For the first Net planting income under conditions of uncertainty in observation information in this simulation.

[0125] For example, 100 randomized simulations were conducted under scenarios of a dry year, moderate observation error, average external water diversion, and conventional planting density and sowing date. The average yield under conditions of completely known crop conditions was 12,500 kg / hm². 2 The average irrigation volume is 3100m³. 3 / hm 2 The average net income is 25,200 yuan / hm² 2 The average yield under conditions of uncertainty in observational information is 12100 kg / hm². 2 The average irrigation volume is 3250m³. 3 / hm 2 The average net income is 23,800 yuan / hm² 2 Based on the uncertainty quantification formula, the relative yield loss is 3.2%, the irrigation volume increase is 4.8%, and the economic benefit loss is 5.6%. Compared with traditional experience-based irrigation models, the irrigation decision-making method of this invention can reduce irrigation volume by 12.5%, increase yield by 8.3%, and improve economic benefits by 15.2% in this scenario. Figure 9 As shown.

[0126] Based on the same inventive concept, such as Figure 3 As shown, one embodiment of the present invention also proposes a precision irrigation decision-making system for arid farms under water resource constraints. This system is used to execute the aforementioned precision irrigation decision-making method for arid farms under water resource constraints. The system includes:

[0127] Multi-source data acquisition module: This module is used to carry multispectral drones, thermal infrared drones, soil moisture sensors, and meteorological sensors to simultaneously acquire multi-source images of cornfields, stratified soil moisture content, field meteorological data, and data from the entire field experiment cycle, and then upload the acquired data to the data processing and storage module.

[0128] Data processing and storage module: Connected to the multi-source data acquisition module, it has a built-in Swin-ConvNeXt deep learning model and inverse distance weight interpolation algorithm. It is used to calculate leaf area index based on UAV imagery, generate stratified soil moisture content based on ground sensor data and UAV imagery data, and organize, classify and store the collected data to form basic agricultural observation datasets and field test datasets.

[0129] Model localization module: Connected to the data processing and storage module, it integrates the Morris global sensitivity analysis algorithm and the particle swarm optimization algorithm. It is used to screen and calibrate the MAIZSIM model parameters based on basic agricultural observation data and field test data to obtain a localized MAIZSIM model.

[0130] Agricultural sequence generation module: Connected to the model localization module and external meteorological data sources, it is used to unify the spatiotemporal scale of the localized MAIZSIM model and basic agricultural observation data, call the iterative ensemble smoother to complete model calibration, and input long-term and short-term meteorological data to generate 90-day long-term agricultural sequence and 15-day short-term agricultural sequence.

[0131] Irrigation Decision Solving Module: Connected to the crop condition sequence generation module, it takes water availability, irrigation infrastructure capacity, and crop growth stage as constraints, receives long-term and short-term crop condition sequences output by the crop condition sequence generation module, and uses the NSGA-Ⅲ multi-objective optimization algorithm to solve for field irrigation execution instructions.

[0132] Simulation Risk Assessment Module: Connected to the model localization module, it builds a virtual farmland simulation scenario based on the localized MAIZSIM model, configures simulation scenarios for hydrological types, observation errors, external water diversion, and field management combinations, conducts batch simulations, and statistically analyzes indicators related to yield, irrigation volume, and economic benefits to complete the risk assessment of irrigation decisions.

[0133] Command issuance module: It is connected to the irrigation decision-solving module and the field irrigation execution equipment respectively. It is used to receive the field irrigation execution commands output by the irrigation decision-solving module and issue the commands to the irrigation facilities of each field in the farm to complete the irrigation operation scheduling.

[0134] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0135] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention after reading the specification and implementing it. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for precision irrigation decision-making for farms in arid regions under water resource constraints, characterized in that, Specifically, it includes: S1. Multi-source images and soil moisture content data of cornfields are collected using UAVs and ground sensors. The leaf area index is calculated from the images using the Swin-ConvNeXt deep learning model. Soil moisture content information is obtained by combining UAV images and ground sensor data. Spatial correction is performed using the inverse distance weighting method to form basic agricultural observation data. S2. Using field test data and obtained basic agricultural observation data as input, the parameters of the MAIZSIM model are screened through Morris global sensitivity analysis, and the parameters of the MAIZSIM model are calibrated through particle swarm optimization algorithm to obtain the localized MAIZSIM model. S3. Unify the spatiotemporal scale of the localized MAIZSIM model and basic agricultural observation data, use an iterative ensemble smoother to fuse the basic agricultural observation data to calibrate the model, and connect to the multi-source meteorological data-driven model to generate a 90-day long-term agricultural sequence and a 15-day short-term agricultural sequence. S4. Using the generated long-term and short-term agricultural conditions sequences as inputs, and constrained by water availability, irrigation infrastructure capacity and crop growth stage, the NSGA-Ⅲ multi-objective optimization algorithm is used to solve for the field irrigation execution instructions. S5. Based on the obtained localized MAIZSIM model, construct a virtual farmland simulation scenario, set up combined simulation scenarios and carry out batch simulations, and use uncertainty quantification algorithm to statistically analyze the relative yield loss, irrigation volume increase and economic benefit loss between yield and irrigation water use efficiency, and complete the irrigation decision risk assessment.

2. The method according to claim 1, wherein, In step S1, the drones are divided into multispectral drones and thermal infrared drones. Both types of drones carry out routine flight operations throughout the entire growth cycle of corn, and the flight range covers all independent irrigated fields in the farm. Multiple fixed time periods are set every day to complete image acquisition. The types of images acquired include RGB images, multispectral images and thermal infrared images. Ground soil moisture sensors are arranged in a grid pattern inside each irrigated field, and meteorological sensors are simultaneously deployed in the field area. The meteorological sensors continuously collect field environmental data, and the soil moisture sensors continuously collect soil moisture content data at different soil layers. The drone image acquisition work was carried out simultaneously with the ground sensor acquisition work.

3. The method of claim 1, wherein, In step S1, the Swin-ConvNeXt deep learning model is composed of an input adaptation layer, a dual-branch feature extraction layer, a feature fusion layer, a global average pooling layer, and a single-value regression layer connected in sequence. The model internally sets up a dual-branch feature extraction structure with parallel SwinTransformer and ConvNeXt branches. The image samples input to the model are first processed uniformly and then enter the input adaptation layer to integrate and transform heterogeneous image channels. The two independent branches extract features from the image data respectively. All the extracted feature data are sent to the feature fusion layer for integration. The integrated feature vector is passed to the global average pooling layer to achieve spatial dimension compression. The processed feature data is finally input to the single-value regression layer to complete the corresponding numerical output.

4. The method of claim 1, wherein, In step S2, the field trial data comes from the full-cycle collection of multi-group zonal trials. The trial area is divided into trial areas with different irrigation treatment combinations, trial areas with different planting density combinations, and dedicated trial areas distinguished by sowing date. During the sowing stage, soil samples are collected in layers and indoor tests are completed to form soil-related data. Throughout the entire growth cycle of maize, crop phenological information, plant morphology information, root distribution information, and aboveground and belowground biomass information are collected and recorded sequentially according to different growth stages. During the crop harvest stage, grain yield-related information is collected uniformly. Irrigation operation information of each zone is recorded synchronously throughout the entire trial. All collected content is uniformly classified and organized according to the trial zone and collection time sequence to form a complete field trial dataset.

5. The method of claim 1, wherein, In step S2, MAIZSIM is a maize-specific mechanism model. It adopts a two-dimensional spatial architecture and performs simulation calculations with a fixed time step. The model is equipped with soil hydrothermal coupling calculation units, solute transport units, crop growth and development units, and root growth calculation units. Soil-related calculations rely on spatial discrete grids to carry out global numerical analysis. The root growth part adopts a calculation architecture corresponding to biomass allocation. The model synchronously integrates plant morphology units, growth process units, and dry matter accumulation units. Data communication and synchronous calculations are realized between various functional units to form an integrated model calculation system.

6. The method of claim 1, wherein, In step S3, at the spatial level, the two-dimensional grid data output by the localized MAIZSIM model is aggregated according to the independent irrigated fields within the farm. The aggregated data and the basic agricultural observation data use the same spatial division unit. At the temporal level, the simulation step size of the localized MAIZSIM model is matched one by one with the collection time of the basic agricultural observation data. The localized MAIZSIM model starts its preheating operation on a fixed date each year. During the preheating phase, measured soil moisture content data and groundwater level data of the entire farm are recorded. During the model retrospective operation phase, hourly meteorological data collected from field stations are accessed. During the model prediction operation phase, long-term and short-term meteorological data released by meteorological forecasting agencies are accessed. These two types of meteorological data correspond to the 90-day agricultural sequence generation stage and the 15-day agricultural sequence generation stage, respectively.

7. The method of claim 1, wherein, In step S3, the iterative set smoother generates multiple sets of initial parameter samples based on the distribution characteristics of the model parameters. The iterative calculation process adopts the Gauss-Newton calculation method. The calculation process integrates all time-series basic agricultural observation data. The model parameter set is continuously updated during the iteration process. When the numerical change generated by adjacent iterations reaches the preset judgment standard, the iteration process ends and the model calibration is completed.

8. The method of claim 1, wherein, In step S4, the constraints include the farm's total annual irrigation water volume, time-series data of cross-regional water diversion, maximum and minimum storage capacity of the reservoir, daily water delivery volume of the irrigation system, range of single irrigation water volume for a single field, and lower limit threshold of soil moisture content corresponding to different growth stages of maize. The calculation is divided into long-cycle planning calculation and short-cycle real-time calculation. The long-cycle planning uses 15 days as a time unit, while the short-cycle real-time calculation uses a natural day as a time unit. The long-cycle planning generates phased water allocation data and reservoir water storage data, while the short-cycle real-time calculation generates irrigation water volume data and irrigation time period data for each field. After integration, a complete field irrigation execution instruction is formed.

9. The method of claim 1, wherein, In step S5, the combined simulation scenario includes hydrological type scenario, observation error scenario, external water diversion scenario, and field management scenario. The hydrological type is divided into three categories: abundant water, normal water, and low water. The observation error is divided into three levels. The external water diversion is configured with multiple sets of random time series data. The field management is matched with different plant densities and different sowing times. The simulation process uses the localized MAIZSIM model as the simulation carrier and simultaneously sets the traditional irrigation mode as the comparison group. Multiple sets of parallel simulations are carried out based on the independent irrigated fields inside the farm as the basic computing unit. After the simulation is completed, the yield data and irrigation water volume data of each group of simulations are extracted in batches, and the corresponding probability distribution statistics are completed.

10. A precision irrigation decision system for arid zone farms under water resource constraints, characterized in that, The system is used for a precision irrigation decision-making method for arid farms under water resource constraints as described in any one of claims 1-9. The system includes a multi-source data acquisition module, a data processing and storage module, a model localization module, an agricultural condition sequence generation module, an irrigation decision solving module, a simulation risk assessment module, and an instruction issuance module. The multi-source data acquisition module is used to carry a multispectral drone, a thermal infrared drone, a soil moisture sensor and a meteorological sensor to simultaneously acquire multi-source images of cornfields, stratified soil moisture content, field meteorology and full-cycle data of field experiments, and upload the acquired data to the data processing and storage module in real time. The data processing and storage module is equipped with a built-in Swin-ConvNeXt deep learning model and an inverse distance weight interpolation algorithm. It is used to calculate the leaf area index based on UAV images, generate layered soil moisture content based on UAV images and ground sensor data, complete data sorting, classification and storage, and construct and store basic agricultural observation datasets and field test datasets. The model localization module interfaces with the data processing and storage module, and integrates the Morris global sensitivity analysis algorithm and the particle swarm optimization algorithm to screen and calibrate the MAIZSIM model parameters and output the localized MAIZSIM model. The agricultural sequence generation module is connected to the model localization module and the external meteorological data source respectively. It is used to unify the spatiotemporal scale of the localized MAIZSIM model and the observation data, call the iterative ensemble smoother to complete the model calibration, and input long-term and short-term meteorological data to generate a 90-day long-term agricultural sequence and a 15-day short-term agricultural sequence. The irrigation decision-solving module, with constraints such as water availability, irrigation infrastructure capacity, and crop growth stage, integrates the NSGA-Ⅲ multi-objective optimization algorithm, receives the crop condition sequence output by the crop condition sequence generation module, solves and outputs field irrigation execution instructions; The simulation risk assessment module is connected to the model localization module. It builds a virtual farmland simulation scene based on the localized MAIZSIM model, configures multiple combined simulation scenarios such as hydrology, observation error, external water diversion, and field management, and conducts batch simulations in parallel. It calls the uncertainty quantification algorithm to complete the statistics and assessment of risk indicators related to yield, irrigation volume, and economic benefits. The instruction issuing module is connected to the irrigation decision solving module and the field irrigation execution equipment, respectively. It is used to receive irrigation execution instructions and issue the instructions to the irrigation facilities in each field of the farm to complete the automated irrigation operation scheduling.

Images