A method and system for constructing a basic model of an arid irrigation area

Abstract

The application discloses a construction method of a basic model of a drought irrigation area, which comprises the following steps: acquiring multi-source data and dividing a target irrigation area into a plurality of calculation units; driving a hydrological model to perform simulation, and performing sensitivity analysis and batch calibration on hydrological model parameters to generate an ecological hydrological data set for each calculation unit; constructing a basic model of the drought irrigation area based on the ecological hydrological data set; learning general knowledge from the ecological hydrological data set by inputting an embedding module, a hybrid mask strategy and a self-supervised encoder-decoder structure based on a teacher-student mechanism in an upstream pre-training stage to obtain an upstream pre-training model; fixing main body parameters of the upstream pre-training model and constructing implicit hydrological representation in a downstream fine-tuning stage; fine-tuning a multi-task prediction module in combination with the constructed implicit hydrological representation and regional observation data of different tasks to obtain the basic model of the drought irrigation area. The application realizes multi-source data fusion and has adaptability to different drought irrigation areas.

Description

Technical Field

[0001] This invention relates to the field of irrigation district eco-hydrological process simulation technology, specifically to a method and system for constructing a basic model of an arid irrigation district. Background Technology

[0002] Arid regions possess abundant light, heat, and land resources, providing the natural conditions for expanding food production. However, they suffer from extreme water scarcity, fragile ecosystems, and a high dependence on irrigation for agriculture. With intensifying competition for water resources among humans, the scientific management of irrigation districts and the coordination of agricultural production with ecological protection have become core issues for regional sustainable development.

[0003] The eco-hydrological processes in irrigation areas are not only affected by natural climate and underlying surface conditions, but also strongly influenced by human activities such as crop planting, irrigation water diversion, and groundwater extraction. The complexity of these systems is far greater than that of natural watersheds. Currently, the modeling methods in this field are mainly divided into three categories, all of which have obvious limitations: (1) Mechanism models based on physical processes (such as SWAT, MODFLOW, etc.) have clear mechanism description capabilities, but the model structure has inherent uncertainties, numerous parameters, and poor regional applicability, making it difficult to directly extend to large-scale irrigation areas and also difficult to flexibly integrate increasingly rich satellite remote sensing information; (2) Pure data-driven models can capture complex nonlinear relationships, but generally lack physical interpretability, are overly dependent on data quality and quantity, and are mostly end-to-end models, making it difficult to systematically reflect the overall eco-hydrological behavior of multiple coupled processes; (3) Physical-data hybrid models attempt to combine the advantages of both, but in practice they often inherit the disadvantages of both, such as structural errors not being eliminated, dependence on explicit expression of partial differential equations, and difficulty in training in large-scale scenarios, and have not yet formed a general solution. Summary of the Invention

[0004] To overcome the shortcomings of existing eco-hydrological modeling methods for arid irrigation areas in terms of structural flexibility, parameter transferability, multi-source data fusion capability, and robustness under sparse data, this invention provides a method and system for constructing a basic model for arid irrigation areas. Based on inheriting physical mechanisms, it enhances the model's versatility, effectively integrates multi-source observation data such as satellite data, and adapts to the differentiated conditions of different arid irrigation areas, thereby supporting the precise implementation of strategies for efficient water resource utilization and sustainable agricultural ecological development in arid areas.

[0005] According to one aspect of the present invention, a method for constructing a basic model of an arid irrigation area is provided, comprising: acquiring multi-source data and dividing the target irrigation area into several computational units based on the multi-source data; using the multi-source data to drive a hydrological model for simulation, and performing sensitivity analysis and batch calibration on the hydrological model parameters, thereby generating an eco-hydrological dataset for each computational unit; constructing a basic model of the arid irrigation area based on the eco-hydrological dataset through upstream pre-training and downstream fine-tuning; wherein, in the upstream pre-training stage, general knowledge is learned from the eco-hydrological dataset through an input embedding module, a hybrid masking strategy, and a self-supervised encoder-decoder structure based on a teacher-student mechanism to obtain an upstream pre-trained model; in the downstream fine-tuning stage, the main parameters of the upstream pre-trained model are fixed and an implicit hydrological representation is constructed, and the multi-task prediction module is fine-tuned in combination with the constructed implicit hydrological representation and regional observation data of different tasks to obtain a basic model of the arid irrigation area suitable for the target irrigation area.

[0006] Furthermore, acquiring multi-source data and dividing the target irrigation area into several computational units based on the multi-source data includes:

[0007] Acquire multi-source data of the target irrigation area, including meteorological data, static attribute data, irrigation data, hydrological data, and satellite-aided data; perform resampling and reprojection processing on the multi-source data to obtain multi-source data with the same spatial resolution in the same target coordinate system; divide the target irrigation area into several computing units according to the spatial resolution of the multi-source data.

[0008] Furthermore, the multi-source data is used to drive the hydrological model for simulation, and sensitivity analysis and batch calibration of the hydrological model parameters are performed. This generates an eco-hydrological dataset for each computational unit, including: using meteorological data, static attribute data, irrigation data, and hydrological data to drive the hydrological model for simulation to obtain eco-hydrological element simulation prediction data; combining the simulation prediction data and actual observation data to perform sensitivity analysis and batch calibration of the hydrological model parameters to obtain calibrated deterministic parameters; using the hydrological model based on the deterministic parameters to generate eco-hydrological element simulation prediction data, thus generating an eco-hydrological dataset for each computational unit. The eco-hydrological dataset includes meteorological data, static attribute data, irrigation data, hydrological data, satellite-aided data, and eco-hydrological element simulation prediction data.

[0009] Furthermore, sensitivity analysis and batch calibration of hydrological model parameters are performed by combining simulated prediction data and actual observation data to obtain calibrated deterministic parameters. This includes: quantifying the influence of each parameter and its interaction on the hydrological model output through sensitivity analysis, and selecting a subset of parameters with higher sensitivity indices; and using an automatic optimization algorithm to perform batch calibration on the parameter subset, wherein the objective function of the calibration aims to minimize the difference between simulated prediction data and actual observation data to obtain calibrated deterministic parameters.

[0010] Furthermore, the upstream pre-training stage learns general knowledge from the eco-hydrological dataset through an input embedding module, a hybrid masking strategy, and a self-supervised encoder-decoder structure based on a teacher-student mechanism. This includes: inputting the eco-hydrological dataset into the input embedding module to obtain a complete embedding sequence; applying a hybrid masking strategy to the complete embedding sequence to obtain a masked embedding sequence; inputting the complete embedding sequence and the masked embedding sequence into the self-supervised encoder-decoder structure based on the teacher-student mechanism, so that the teacher encoder generates a high-quality target representation based on the high-dimensional embedding vector, and the student encoder generates a fragmented data representation based on the masked embedding sequence; inputting the fragmented data representation into the decoder and the projection layer, so that the decoder maps the fragmented data representation back to the original variable space and reconstructs the masked variable values, and the projection layer obtains the projected representation based on the fragmented data representation.

[0011] Furthermore, in the downstream fine-tuning stage, the fixed parameters of the upstream pre-trained model are frozen and an implicit hydrological representation is constructed. The multi-task prediction module is then fine-tuned using the constructed implicit hydrological representation and regional observation data from different tasks to obtain a basic model of arid irrigation areas suitable for the target irrigation district. This includes: freezing the input embedding module and decoder; partially freezing and fine-tuning the student encoder to fix the main parameters of the upstream pre-trained model; constructing an implicit hydrological representation based on the upstream pre-trained model with fixed main parameters; constructing a conditional vector for each prediction location by combining variable embedding vectors and time embedding vectors; mapping the conditional vectors to modulation parameters based on the concept of linear modulation, and applying the modulation parameters to the implicit hydrological representation for conditional modulation to obtain task-specific features; integrating the task-specific features into the multi-task prediction module of the hybrid expert system; and then fine-tuning the multi-task prediction module specifically using regional observation data from different tasks to form a basic model of arid irrigation areas suitable for the target irrigation district.

[0012] Furthermore, the downstream fine-tuning stage also includes: constructing a multi-task prediction joint loss function, wherein the loss function is designed as follows:

[0013] ,

[0014] In the formula, This represents the downstream monitoring loss; b, t, and v represent the sample number index, time index, and variable index, respectively. The weights represent the relative importance of the b-th sample, time t, and v-th variable in the loss. This represents the supervision mask, used to indicate whether there are true observations for the b-th sample, time t, and the v-th variable used to calculate the loss; , , Let B represent the observed value of the b-th sample, the time t-th time, the observed value of the v-th variable, the model predicted mean, and the model predicted variance, respectively; B represents the total number of batches, T represents the number of time periods, and V represents the number of variables.

[0015] According to one aspect of the present invention, a system for constructing a basic model of an arid irrigation area is provided, comprising: a target irrigation area division module, used to acquire multi-source data and divide the target irrigation area into several computational units based on the multi-source data; an eco-hydrological dataset generation module, used to drive a hydrological model to perform simulation using the multi-source data, and to perform sensitivity analysis and batch calibration of the hydrological model parameters, thereby generating an eco-hydrological dataset for each computational unit; and a basic model training module, used to construct a basic model of the arid irrigation area based on the eco-hydrological dataset through upstream pre-training and downstream fine-tuning; wherein, in the upstream pre-training stage, general knowledge is learned from the eco-hydrological dataset through an input embedding module, a hybrid masking strategy, and a self-supervised encoder-decoder structure based on a teacher-student mechanism to obtain an upstream pre-trained model; in the downstream fine-tuning stage, the main parameters of the upstream pre-trained model are fixed and an implicit hydrological representation is constructed, and the multi-task prediction module is fine-tuned in combination with the constructed implicit hydrological representation and regional observation data of different tasks to obtain a basic model of the arid irrigation area suitable for the target irrigation area.

[0016] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to execute the method for constructing a basic model for arid irrigation areas.

[0017] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the method for constructing a basic model of an arid irrigation area.

[0018] The above technical solution first acquires multi-source data from the irrigation area and divides it into computational units; then, it uses a traditional hydrological model to generate a physically consistent training dataset (i.e., an eco-hydrological dataset); subsequently, in the upstream pre-training stage, it constructs an implicit hydrological representation that integrates prior knowledge of mechanisms and satellite observations through an input embedding module and a hybrid masking strategy, and performs pre-training based on a teacher-student self-supervised architecture; finally, in the downstream fine-tuning stage, it achieves multi-task adaptation by fine-tuning with only a small amount of measured data. This invention innovatively realizes a modeling paradigm of "general basic model + a small amount of measured fine-tuning," overcoming the shortcomings of traditional physical model parameter calibration being cumbersome, structure being rigid, and the lack of physical consistency of data-driven models. It can simultaneously complete the reconstruction of key hydrological variables and multi-objective joint prediction with a unified model, providing efficient technical support for the coordinated regulation and refined management of water resources in arid irrigation areas.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] (1) In terms of model construction method, this invention constructs a physical consistency hydrological implicit representation based on mechanism prior and satellite observation. Through the joint learning mechanism of "simulation prior - observation constraint", physical laws are deeply embedded in the model, forming a high-dimensional knowledge core that combines physical rationality and data adaptability, providing a powerful and universal transferable foundation for downstream tasks.

[0021] (2) In terms of model application architecture, this invention designs a unified pre-training-fine-tuning framework for multiple tasks. This framework enables a single model to be quickly fine-tuned with a small amount of regional data, which can be efficiently adapted to multiple prediction tasks. It realizes collaborative simulation from reconstruction of difficult-to-observe variables (such as deep soil water) to joint prediction of multiple variables (such as water consumption and soil moisture), which significantly improves the practicality and transferability of the model.

[0022] (3) Regarding the generation of training data, this invention creates a data augmentation method that uses a validated hydrological model and multi-source satellite observations to synthesize a massive amount of training samples. This method effectively supplements the deep hydrological process data that are difficult to obtain in field observations through model simulation, and integrates satellite remote sensing indicator variables that are closely related to eco-hydrological processes, thereby significantly alleviating the constraints on model training caused by the scarcity of observation data. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying 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.

[0024] Figure 1A flowchart illustrating a method for constructing a basic model of an arid irrigation area, as provided in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of an embedded module provided in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the upstream teacher-student self-supervised architecture provided in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the downstream multi-task joint prediction structure provided in an embodiment of the present invention. Detailed Implementation

[0028] It should be noted that:

[0029] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0031] 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. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0032] Please refer to the appendix. Figure 1 This invention provides a method for constructing a basic model of an arid irrigation area, specifically including the following steps:

[0033] Step S1: Acquire multi-source data and divide the target irrigation area into several calculation units based on the multi-source data.

[0034] Step S11: Obtain multi-source data for the target irrigation area. This multi-source data includes meteorological data, static attribute data, irrigation data, hydrological data, and satellite-aided data.

[0035] In step S11, the multi-source data includes, but is not limited to, meteorological data, hydrological data, vegetation data, topographic data, soil texture data, location information, irrigation data, land use data, and surface temperature remote sensing observation data. Meteorological data includes, but is not limited to, rainfall, average air pressure, maximum temperature, minimum temperature, average temperature, average relative humidity, sunshine duration, average wind speed, and net radiation in the target irrigation area. Irrigation data includes, but is not limited to, the actual irrigation volume in the target irrigation area. Hydrological data includes, but is not limited to, the potential evapotranspiration, actual evapotranspiration, and groundwater depth in the target irrigation area. Vegetation data includes, but is not limited to, the NDVI index, EVI index, leaf area index (LAI), and yield data of the target irrigation area. Topographic data includes, but is not limited to, the elevation and slope of the target irrigation area. Soil texture data includes, but is not limited to, the proportion of sand, clay, and silt in the soil of the target irrigation area. Location information includes, but is not limited to, the latitude and longitude of the target irrigation area. Surface temperature remote sensing observation data includes, but is not limited to, the daytime and nighttime surface temperatures of the target irrigation area. The static attribute data mainly includes topographic data, soil texture data, location information, and land use data, while the satellite-aided data includes satellite-observed vegetation indices (NDVI, EVI) and land surface temperature (LST).

[0036] Step S12 involves resampling and reprojecting the multi-source data to obtain multi-source data with the same spatial resolution in the same target coordinate system.

[0037] It should be noted that the multi-source data of the target irrigation area obtained in step S11 may be station data in the form of stations or remote sensing raster data obtained through remote sensing (such as surface temperature remote sensing observation data). For station data covering the entire area, interpolation processing needs to be performed in advance, followed by resampling and reprojection. For discrete station data or remote sensing raster data, no interpolation processing is performed; they are only used as observation data for fine-tuning, and resampling and reprojection are performed directly. Based on this, step S12 obtains multi-source data with the same spatial resolution in the same target coordinate system.

[0038] Step S13: Divide the target irrigation area into several calculation units based on the spatial resolution of the multi-source data.

[0039] It should be noted that step S12 has unified all the multi-source data into the same target coordinate system and the same spatial resolution, and has also performed pixel alignment. Step S13 divides the target irrigation area according to the spatial resolution of the multi-source data. In fact, it divides each pixel into a computing unit. Each computing unit has complete meteorological data, hydrological data, vegetation data, topographic data, soil texture data, irrigation data, location information, land use data, and surface temperature remote sensing observation data.

[0040] Step S2 involves using multi-source data to drive the hydrological model for simulation, and performing sensitivity analysis and batch calibration of the hydrological model parameters to generate an eco-hydrological dataset for each computational unit.

[0041] In step S2, the hydrological model uses meteorological data, static attribute data, irrigation data, and hydrological data as input data for simulation, and outputs simulated and predicted data containing eco-hydrological elements. To ensure that the simulated and predicted data output by the hydrological model reflects the actual situation, this invention calibrates the parameters of the hydrological model using actual observation data. Understandably, the parameters of a hydrological model are uncertain, and there are many parameters (each module in the hydrological model has corresponding formulas and parameters). Therefore, before calibrating the model parameters, parameters must be pre-screened (i.e., parameter sensitivity analysis is performed) to reduce the parameter range. Then, by minimizing the difference between the actual observation data and the simulated and predicted data output by the hydrological model, a set of deterministic model parameters is calibrated.

[0042] Step S21: Use meteorological data, static attribute data, hydrological data and irrigation data to drive the hydrological model to perform simulation, so as to obtain simulated and predicted data of eco-hydrological elements.

[0043] It should be noted that before driving the hydrological model for simulation, the hydrological model needs to be selected. Specifically, based on the geographical location of the target irrigation area, the distribution of input data and actual observation data, a hydrological model that best describes the physical processes of the irrigation area should be selected. Input data refers to the data input to the hydrological model, including meteorological data, static attribute data, hydrological data, and irrigation data. Actual observation data includes field sampling observation data and satellite observation data (these two types of actual observation data are mainly used for model calibration). Field sampling observation data mainly refers to data obtained from field experiments in the target irrigation area (such as measurements of soil moisture content, soil salinity, groundwater level, and drainage volume). Satellite observation data mainly refers to soil moisture content and LAI (leaf area index) data. In this embodiment, the selected hydrological model is the SWAP model. Understandably, in some feasible embodiments, other hydrological models that conform to the actual conditions of the irrigation area can be selected, and this is not limited here.

[0044] Step S22: Combine simulated prediction data and actual observation data to perform sensitivity analysis and batch calibration of the parameters of the hydrological model in order to obtain the calibrated deterministic parameters.

[0045] In step S22, sensitivity analysis is used to quantify the impact of each parameter and its interaction on the hydrological model output, and a subset of parameters with high sensitivity indices is selected. Subsequently, an automatic optimization algorithm is used to batch calibrate this parameter subset. The objective function aims to minimize the difference between the simulated prediction data and the actual observation data output by the hydrological model, thereby obtaining a set of deterministic model parameters (i.e., deterministic parameters). It should be noted that hydrological models typically have many parameters; therefore, sensitivity analysis is needed to select key and important parameters (i.e., key model parameters, such as highly sensitive soil hydraulic parameters, crop parameters, drainage parameters, salinity parameters, etc. Theoretically, each module should have some parameters that need calibration, but this is not limited here) for calibration. Other parameters can remain at their default values. Minimizing the difference between the model's simulated values ​​and the actual observation values ​​is essentially achieved by changing the values ​​of the selected key model parameters to alter the simulated prediction data output by the hydrological model until the difference between the simulated prediction data and the actual observation data is minimized. At this point, the values ​​of the key model parameters constitute the required set of deterministic parameters after calibration.

[0046] In this embodiment, the global EFAST algorithm is used to analyze the impact of a single parameter acting alone and multiple parameters interacting on the physical model results. By decomposing the model variance, the contribution of each parameter acting alone and interacting with each other to the total variance is calculated, thereby obtaining the first-order sensitivity index of each parameter and the total sensitivity index, as shown below.

[0047] (1)

[0048] (2)

[0049] (3)

[0050] (4)

[0051] in, This represents each parameter to be analyzed. Indicates parameters The characteristic frequencies are used to uniquely identify the influence of each parameter in subsequent analyses. s is a scalar representing the search variable. This represents a transformation function, the purpose of which is to transform... Oscillation mapped to parameters On the actual probability distribution. D represents the total variance of the model. and These are the Fourier coefficients corresponding to frequency p, where M represents the number of samples and p represents the harmonic index in the Fourier series. Indicates parameters The first-order sensitivity index, and It is a harmonic The corresponding Fourier coefficients. Indicates parameters Total sensitivity index, Indicates all that are not included A set of frequencies.

[0052] Then, a subset of parameters with higher sensitivity indices is selected, and batch calibration of all computational units in the target irrigation area is performed using the PEST algorithm. This involves minimizing the difference between simulated and observed data using the objective function, thereby obtaining a set of deterministic model parameters. The objective function equation is... This can be expressed by the formula:

[0053] (5)

[0054] Where n represents the number of observations, This represents the weighting coefficients for different observations. Indicates time The observed value at that time, This indicates that the model performs in time when using the Params series of parameters. The simulated values ​​below.

[0055] Step S23: Using the hydrological model, simulated and predicted eco-hydrological elements are generated based on deterministic parameters, and then an eco-hydrological dataset is generated for each computing unit; the eco-hydrological dataset includes meteorological data, static attribute data, irrigation data, hydrological data, satellite-aided data, and simulated and predicted eco-hydrological elements.

[0056] In step S23, the deterministic parameters obtained in step S22 are applied to the hydrological model, and the hydrological model is run again. The hydrological model can then generate output results of eco-hydrological data that are more consistent with the actual situation. This output result, combined with the input data of each calculation unit in step S13 (including meteorological data and time-series data from satellite-assisted observations), generates a seamless and continuous eco-hydrological dataset for each calculation unit that is consistent with the actual situation of the irrigation area (including key eco-hydrological elements such as meteorological variables, actual irrigation volume, stratified soil moisture content, stratified soil salinity, drainage volume, lower boundary water flow, actual transpiration, actual evaporation, groundwater depth, LAI, dry matter accumulation, plant height, vegetation indices (NDVI, EVI), and land surface temperature (LST).

[0057] Step S3: Based on the eco-hydrological dataset, construct a basic model for arid irrigation areas through upstream pre-training and downstream fine-tuning.

[0058] In step S3, the basic model for arid irrigation areas includes an input embedding module, a teacher-student-based self-supervised encoder-decoder structure, and a multi-task prediction module. The basic model undergoes two distinct training phases (steps S1 and S2 are essentially about preparing the eco-hydrological dataset): an upstream pre-training phase and a downstream fine-tuning phase. In the upstream pre-training phase, general implicit hydrological features are learned from the eco-hydrological dataset using the input embedding module, a hybrid masking strategy, and a teacher-student-based self-supervised encoder-decoder structure to obtain the upstream pre-trained model. In the downstream fine-tuning phase, the main parameters of the upstream pre-trained model are fixed, and implicit hydrological representations are constructed. The multi-task prediction module is then fine-tuned based on these implicit hydrological representations and regional observation data for different tasks to obtain a basic model for arid irrigation areas suitable for the target irrigation district. The specific steps of step S3 will be further described below:

[0059] Step S31: In the upstream pre-training stage, general knowledge is learned from the eco-hydrological dataset through input embedding module, hybrid masking strategy and self-supervised encoder-decoder structure based on teacher-student mechanism to obtain upstream pre-trained model.

[0060] In step S31, based on the eco-hydrological dataset generated in step S2, and combined with the time-series data from satellite-assisted observations in step S1, the upstream pre-trained model is trained using embedding encoding, a hybrid masking strategy, and a self-supervised learning architecture based on a teacher-student mechanism. During training, general knowledge is continuously learned from the eco-hydrological dataset to obtain a well-trained upstream pre-trained model. This well-trained upstream pre-trained model is used to output an implicit hydrological representation that deeply integrates prior mechanistic knowledge with remote sensing signals. Furthermore, step S31 specifically includes the following sub-steps:

[0061] Step S311: Input the eco-hydrological dataset into the input embedding module to obtain the complete embedding sequence.

[0062] In step S311, the eco-hydrological dataset (original multivariate time-series data) is projected into a high-dimensional space through the input embedding module (i.e., the embedding layer) to obtain a complete embedding sequence (i.e., a high-dimensional embedding vector). The original multivariate time-series data refers to a seamless and continuous eco-hydrological dataset generated in step S2 (including key eco-hydrological elements such as meteorological variables, actual irrigation amount, stratified soil moisture content, stratified soil salinity, drainage amount, lower boundary water flow, actual transpiration, actual evaporation, groundwater depth, LAI, dry matter accumulation, plant height, vegetation index, and surface temperature).

[0063] It should be noted that the input embedding module comprehensively considers static attributes, time information, variable type, the numerical value itself, and the depth position of the variable on the vertical profile (e.g., Figure 2 (As shown). The deep embedding is designed based on the relative positions of different variable types in the vertical profile. Variable type identifiers include: atmospheric variables, soil variables, groundwater variables, satellite biophysical parameter variables, and crop variables. For atmospheric variables, relative position identifiers include surface, 2m, and 10m; for soil variables, continuous encoding is used for absolute positions; for groundwater variables, learnable position encoding is used; for satellite biophysical parameter variables, relative position identifiers include canopy and surface; for crop variables, relative position identifiers include aboveground parts, roots, and whole plant.

[0064] ① Static attribute embedding is shown below:

[0065] (6)

[0066] In the formula, s is a static feature vector (soil texture, elevation, land use, etc.). The projection matrix is ​​for static features. Embed vectors for static attributes.

[0067] ②The time embedding is as follows:

[0068] (7)

[0069] In the formula, T is the learnable temporal location encoding matrix, and t is the time step index. This is the temporal embedding vector.

[0070] ③ Variable embedding is shown below:

[0071] (8)

[0072] In the formula, V is the learnable variable embedding matrix, and i is the i-th variable. Embed the vector for the variable.

[0073] ④ Numerical embedding is shown below:

[0074] (9)

[0075] In the formula, , These are the learnable weights and bias matrices, respectively. This represents the observed value of variable i at time t. It is a numerical embedding vector.

[0076] ⑤ The deep embedding is designed according to the relative position of the variables in the vertical profile. The variable type is identified as shown in formula (10), specifically including: atmospheric variables, soil variables, groundwater variables, satellite biophysical parameter variables, and crop variables, as shown in formulas (11)-(15). The total deep embedding is shown in formula (16).

[0077] (10)

[0078] (11)

[0079] (12)

[0080] (13)

[0081] (14)

[0082] (15)

[0083] (16)

[0084] In the formula, i represents different variables. Represents a deep embedding vector. Indicates the type identifier of variable i (such as atmosphere, soil, groundwater, satellite biophysics, crop). The location code of variable i is represented; type(i)∈{atm, soil, gw, sur_bio, crp}, which represent atmospheric variables, soil variables, groundwater variables, satellite biophysical parameter variables, and crop variables, respectively. Indicates altitude level (e.g., surface, 2m from surface, 10m from surface, etc.); Indicates soil depth; This indicates a feature engineering process that enables continuous soil depth. Indicates the closest pre-trained soil layer index; This represents a learnable groundwater depth embedding; Indicates the relative location of satellite biophysical parameter variables (such as canopy, surface); Indicates the level of crop variables (such as aboveground parts, roots, whole plant, etc.). Represents the embedding of different variables, i∈{atm,soil, gw, sur_bio, crp}; onehot is the one-hot encoding of different categories.

[0085] ⑥ The final complete embedding formula is shown below:

[0086] (17)

[0087] Let represent the total embedding vector, and the remaining variables are the same as in the formula above.

[0088] Step S312: Apply a hybrid masking strategy to the complete embedding sequence to obtain a masked embedding sequence.

[0089] In step S312, the masked embedding sequence is a fragmented sequence after hybrid masking. The hybrid masking strategy includes constant masking of meteorological variables, local time window masking of candidate variables, whole-sequence masking of candidate variables, and block masking for specified future time periods. Variables subject to local time window masking include surface soil moisture content, LAI, LST, NDVI, and EVI, to simulate the loss caused by satellite obstruction due to clouds and fog in actual conditions. Variables subject to whole-sequence masking include deep soil moisture content, soil salinity in different soil layers, groundwater depth, crop dry matter accumulation, and crop height—variables difficult to observe over large areas. Block masking typically masks all non-meteorological variables for the next 1-7 days. It is important to note that multiple masks can be applied simultaneously to a single sample, and meteorological variables must be forcibly preserved.

[0090] Step S313: Input the complete embedding sequence and the masked embedding sequence into the teacher-student-based self-supervised encoder-decoder structure, so that the teacher encoder generates a high-quality target representation based on the high-dimensional embedding vector, and the student encoder generates a fragmented data representation based on the masked embedding sequence.

[0091] In step S313, the complete embedding sequence output from step S311 and the incomplete sequence after mixing and masking output from step S312 are input into a teacher-student-based self-supervised encoder-decoder structure. This teacher-student-based self-supervised encoder-decoder structure includes a teacher encoder and a student encoder with the same structure but different parameter update methods. The teacher encoder always processes the complete embedding sequence to output a high-quality target representation (i.e., the teacher's original high-dimensional representation); the student encoder processes the masked embedding sequence and learns the ability to infer the complete state from partial information (e.g., ...). Figure 3 As shown), the output is a partial data representation (i.e., the student's original high-dimensional representation), which is the implicit hydrological representation output by the upstream pre-trained model.

[0092] Furthermore, the parameters of the teacher encoder are obtained from the parameters of the student encoder using a momentum update method, and do not participate in gradient backpropagation. The update formula is shown below:

[0093] (18)

[0094] In the formula, The parameters represent the teacher encoder. The parameters representing the student encoder, This represents the momentum coefficient.

[0095] It should be noted that both the teacher encoder and the student encoder of this invention are encoders designed using Transformer technology, that is, Transformer encoders. Each Transformer encoder has N identical Transformer blocks stacked together. In this embodiment, each Transformer encoder has 8 Transformer blocks, each Transformer block has 8 attention heads, the hidden layer dimension is 384, and the dropout rate is 0.1.

[0096] Step S314: Input the incomplete data representation to the decoder and the projection layer so that the decoder maps the incomplete data representation back to the original variable space and reconstructs the masked variable values, and the projection layer obtains the projected representation based on the incomplete data representation.

[0097] In step S314, after the student encoder extracts the incomplete data representation, it inputs it into the projection layer and the decoder (corresponding to the contrastive learning branch and the reconstruction task branch, respectively). In the contrastive learning branch, the cosine similarity (representational contrastive loss) between the projected representation obtained from the incomplete data representation through the projection layer and the high-quality target representation output by the teacher encoder is calculated. Simultaneously, in the reconstruction task branch, the incomplete data representation obtained by the student encoder is reconstructed by the decoder to obtain the masked values ​​(i.e., the values ​​at the masked positions identified by the mask matrix, i.e., the missing parts are filled in), to obtain physically meaningful variable values. These values ​​are then compared with the true variable values ​​to calculate the reconstruction loss. The representational contrastive loss and the reconstruction loss are combined to obtain the total loss. Furthermore, the decoder is designed as a lightweight multilayer perceptron to avoid overburdening the reconstruction task on the decoder.

[0098] Furthermore, the formula for calculating the joint loss function is as follows:

[0099] (19)

[0100] (20)

[0101] (twenty one)

[0102] (twenty two)

[0103] (twenty three)

[0104] (twenty four)

[0105] In the formula, Indicates the total loss. This represents the contrast loss between the teacher encoder and the student encoder. This represents the reconstruction loss of the encoder-decoder structure by reconstructing the masked variables. It is to prevent the loss of characterization collapse. It is a feature decorrelation loss. N is the number of samples. and These represent the masked input sample (i.e., the masked embedding sequence input to the student encoder) and the complete input sample (i.e., the complete embedding sequence input to the teacher encoder), respectively. and Let represent the latent representations output by the student encoder and the teacher encoder, respectively, namely, the incomplete data representation and the high-quality target representation. M represents the total number of all masked spatiotemporal points. This represents the decoder's prediction of the mask portion. This represents the actual value of the masked variable. D is the dimension of the hidden feature. This represents the value of the d-th hidden dimension across all samples in the representation matrix output by the encoder. This represents the variance of the d-th feature. This represents the minimum variance threshold. For the stability term, we take 0.0001. C represents the covariance matrix of the encoder representation, and H is the encoder representation. The mean vector for each dimension, Let i be the covariance between the i and j-th dimensions. This is the sum of squares of all off-diagonal elements. , , , These refer to the weights of contrast loss, reconstruction loss, variance regularization, and decorrelation regularization, respectively. In this embodiment, , , , The values ​​are 0.5, 1.0, 0.005, and 0.005 respectively. Understandably, the weights of each loss can be set according to actual needs, and are not limited here.

[0106] It should be noted that different losses have different characteristics, enabling the arid irrigation area basic model of this invention to exhibit excellent properties in various aspects. Among them, the representation contrast loss requires the basic model to capture the consistency of similar content in the representation space, that is, the consistency of the semantic core, without caring about the specific value of a particular variable. The reconstruction loss, on the other hand, forces the encoder to generate representations containing sufficient information so that the decoder can reconstruct the original input from the representations, focusing on values ​​or details. The loss to prevent representation collapse aims to ensure the uniformity and diversity of representation distribution, ensuring that the basic model does not output useless constants. The feature decorrelation loss focuses on the internal structure of the representation, preventing dimensional redundancy and avoiding a situation where all dimensions describe the same feature, thus allowing more information to be conveyed with fewer dimensions.

[0107] Step S32: In the downstream fine-tuning stage, the main parameters of the upstream pre-trained model are fixed and an implicit hydrological representation is constructed. The multi-task prediction module is fine-tuned by combining the constructed implicit hydrological representation with regional observation data of different tasks to obtain a basic model of arid irrigation area suitable for the target irrigation area.

[0108] In step S32, the implicit hydrological representation obtained in the upstream pre-training stage is a general knowledge integration. Downstream, a multi-task prediction module is needed to achieve predictions for different tasks. Using the implicit hydrological representation output from the upstream pre-training stage, combined with a small amount of regional observation data corresponding to different downstream tasks (such as predictions of specific eco-hydrological variables), the multi-task prediction module is fine-tuned to form a basic model for arid irrigation areas suitable for the target irrigation district. The accurate multi-task prediction includes the inversion of key hydrological variables and multi-objective collaborative simulation, specifically comprising the following sub-steps:

[0109] Step S321: Freeze the input embedding module and decoder, and perform partial freeze fine-tuning on the student encoder to fix the main parameters of the upstream pre-trained model.

[0110] In step S321, the main parameters of the upstream pre-trained model are fixed by freezing or partially freezing. Specifically, the input embedding module is completely frozen to ensure the consistency of the variable semantic space; the student encoder is partially frozen, with only a small number of high-level parameters unfrozen and updated at a low learning rate to adapt to the downstream task, while preserving the latent space structure of the upstream representation; and the decoder is completely frozen to ensure the stability of its auxiliary reconstruction of missing variables (i.e., incomplete mask embedding sequences).

[0111] Step S322: Construct an implicit hydrological representation based on the upstream pre-trained model with fixed main parameters.

[0112] In step S322, the student encoder obtained from upstream pre-training is used as the frozen feature extraction module to maintain the physical consistency of its representation space. The complete variables obtained in the upstream pre-training stage (i.e., the incomplete data representation obtained by the student encoder through the incomplete mask embedding sequence and then input into the decoder to obtain the missing parts, so as to achieve low-weight completion of the missing variables, wherein the incomplete mask embedding sequence + the missing parts together constitute the complete variables) are then input into the frozen student encoder to obtain a high-dimensional latent representation (i.e., implicit hydrological representation), so as to use the high-dimensional latent representation to assist in the downstream multi-task fine-tuning.

[0113] Step S323: Perform multi-task feature aggregation and prediction based on implicit hydrological representation.

[0114] In step S323, conditional vectors are constructed for each prediction location downstream by combining the variable embedding vectors and time embedding vectors from the upstream pre-training stage. Based on the concept of linear modulation (FiLM), the conditional vectors are mapped to modulation parameters. These modulation parameters are then applied to the implicit hydrological representation obtained in step S322 for conditional modulation to obtain task-specific features. Finally, these task-specific features are integrated into the multi-task prediction module of a hybrid expert system (MoE). The multi-task prediction module is then fine-tuned using regional observation data from different tasks to map the implicit vectors to the predicted mean and uncertainty (i.e., the multi-task prediction results), thereby obtaining a basic model for arid irrigation areas suitable for the target irrigation district (e.g., [missing data]). Figure 4 (As shown).

[0115] In this embodiment, the specific formulas for the above steps are as follows:

[0116] (25)

[0117] (26)

[0118] (27)

[0119] (28)

[0120] In the formula, This is the condition vector for each predicted location downstream; , These are the variable embedding vector and the time embedding vector, respectively; For a multilayer perceptron, the condition vector is mapped to modulation parameters; , These are the scaling factor and offset in the modulation parameters, respectively; h represents the downstream features after conditional modulation; h is the high-dimensional latent representation obtained in step S321. , These are the predicted mean and variance, respectively. It is a hybrid expert system.

[0121] Furthermore, a joint loss function for multi-task prediction is constructed to simultaneously constrain the prediction accuracy, feature decoupling, and uncertainty representation of multiple tasks, thereby achieving highly robust and interpretable spatiotemporal prediction. The design of the loss function is shown below:

[0122] (29)

[0123] In the formula, This represents the downstream monitoring loss; b, t, and v represent the sample number index, time index, and variable index, respectively. The weights represent the relative importance of the b-th sample, time t, and v-th variable in the loss. This represents the supervision mask, used to indicate whether there are true observations for the b-th sample, time t, and the v-th variable used to calculate the loss; , , Let B represent the observed value of the b-th sample, the time t-th time, the observed value of the v-th variable, the model predicted mean, and the model predicted variance, respectively; B represents the total number of batches, T represents the number of time periods, and V represents the number of variables.

[0124] Based on the same technical concept as the aforementioned embodiments, this invention also provides a system for constructing a basic model for arid irrigation areas, comprising: a target irrigation area division module, used to acquire multi-source data and divide the target irrigation area into several computational units based on the multi-source data; an eco-hydrological dataset generation module, used to drive the hydrological model to perform simulation using multi-source data, and to perform sensitivity analysis and batch calibration of the hydrological model parameters, thereby generating a set of eco-hydrological datasets for each computational unit; and a basic model training module, used to construct a basic model for arid irrigation areas based on the eco-hydrological datasets through upstream pre-training and downstream fine-tuning; wherein, in the upstream pre-training stage, general knowledge is learned from the eco-hydrological datasets through an input embedding module, a hybrid masking strategy, and a self-supervised encoder-decoder structure based on a teacher-student mechanism to obtain an upstream pre-trained model; in the downstream fine-tuning stage, the main parameters of the upstream pre-trained model are fixed and an implicit hydrological representation is constructed, and the multi-task prediction module is fine-tuned in combination with the constructed implicit hydrological representations and regional observation data of different tasks to obtain a basic model for arid irrigation areas suitable for the target irrigation area.

[0125] Based on the same technical concept as the foregoing embodiments, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor calls the program instructions to execute the method for constructing a basic model of an arid irrigation area.

[0126] Based on the same technical concept as the foregoing embodiments, the present invention also provides a non-transitory computer-readable storage medium that stores computer instructions that cause the computer to execute the method for constructing a basic model of an arid irrigation area.

[0127] In summary, firstly, this invention constructs a physically consistent implicit hydrological representation based on mechanistic priors and satellite observations. This implicit hydrological representation, through the introduction of a joint learning mechanism of "simulation priors-observation constraints" in the pre-training stage, deeply embeds the physical laws and expert knowledge inherent in the hydrological mechanism equations into the basic model of arid irrigation areas in a data-driven manner. Simultaneously, the basic model's forced learning maintains consistency with multi-source satellite remote sensing observations within the representation space. The resulting implicit hydrological representation not only embeds rich hydrological process priors but also profoundly reveals the intrinsic physical correlation between observational data and hydrological state variables, thus forming a high-dimensional knowledge core that combines physical rationality and observational adaptability, providing a powerful and transferable general foundation for various downstream tasks. Secondly, this invention innovatively designs a pre-training-fine-tuning framework oriented towards multiple tasks, enabling the collaborative simulation and prediction of core hydrological elements in irrigation areas using a single basic model. This framework constructs a general physical knowledge base in the upstream pre-training stage, and in the downstream fine-tuning stage, only a small amount of region-specific data is needed to fine-tune the basic model, allowing for rapid adaptation to different prediction tasks. Based on this, the basic model can reconstruct key hydrological variables that are difficult to observe directly with high precision (such as deep soil moisture content, groundwater level dynamics, and deep seepage), and can also achieve joint prediction of multiple variables (such as crop evapotranspiration water consumption estimation and soil moisture diagnosis). This fundamentally overcomes the limitations of traditional methods with their complex parameters, providing a unified and efficient technical tool for the systematic regulation and refined management of water resources in irrigation areas. Furthermore, this invention creates a data augmentation method that synthesizes massive training samples using a widely validated hydrological model and multi-source satellite data. This method effectively supplements deep soil process data that are difficult to obtain in field observations and incorporates satellite observation variables closely related to eco-hydrological processes, thus significantly alleviating the constraints on model training caused by sparse observation data. This invention has significant practical value, providing key state variable inversion and process prediction support for the optimal allocation of water resources in arid irrigation areas, the formulation of precision irrigation programs, and the construction of sustainable groundwater utilization strategies. It effectively improves the pertinence and reliability of management decisions, thereby systematically enhancing the scientific foundation and practical effectiveness of comprehensive water resource management in arid areas.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a basic model of an arid irrigation area, characterized in that, include: Acquire multi-source data and divide the target irrigation area into several calculation units based on the multi-source data; The multi-source data is used to drive the hydrological model for simulation, and sensitivity analysis and batch calibration of the hydrological model parameters are performed. This generates an eco-hydrological dataset for each computational unit, including: using meteorological data, static attribute data, irrigation data, and hydrological data to drive the hydrological model for simulation to obtain simulated and predicted eco-hydrological elements; combining the simulated and predicted data with actual observation data to perform sensitivity analysis and batch calibration of the hydrological model parameters to obtain calibrated deterministic parameters; and using the hydrological model to generate simulated and predicted eco-hydrological elements based on the deterministic parameters, thus generating an eco-hydrological dataset for each computational unit. The eco-hydrological dataset includes meteorological data, static attribute data, irrigation data, hydrological data, satellite-aided data, and simulated and predicted eco-hydrological elements. Based on the aforementioned eco-hydrological dataset, a basic model for arid irrigation areas is constructed through upstream pre-training and downstream fine-tuning. In the upstream pre-training stage, general knowledge is learned from the eco-hydrological dataset using an input embedding module, a hybrid masking strategy, and a self-supervised encoder-decoder structure based on a teacher-student mechanism to obtain the upstream pre-trained model. In the downstream fine-tuning stage, the main parameters of the upstream pre-trained model are fixed, and an implicit hydrological representation is constructed. This implicit hydrological representation, combined with regional observation data from different tasks, is used to fine-tune the multi-task prediction module to obtain a basic model for arid irrigation areas suitable for the target irrigation district.

2. The method for constructing a basic model of an arid irrigation area as described in claim 1, characterized in that, Acquiring multi-source data and dividing the target irrigation area into several computational units based on the multi-source data includes: Acquire multi-source data for the target irrigation area, including meteorological data, static attribute data, irrigation data, hydrological data, and satellite-aided data. The multi-source data is resampled and reprojected to obtain multi-source data with the same spatial resolution in the same target coordinate system. The target irrigation area is divided into several calculation units based on the spatial resolution of the multi-source data.

3. The method for constructing a basic model of an arid irrigation area as described in claim 1, characterized in that, Sensitivity analysis and batch calibration of hydrological model parameters were performed by combining simulated prediction data and actual observation data to obtain calibrated deterministic parameters, including: Sensitivity analysis was used to quantify the impact of each parameter and its interaction on the hydrological model output, and a subset of parameters with higher sensitivity indices was selected. An automatic optimization algorithm is used to perform batch calibration on the parameter subset, wherein the objective function of the calibration aims to minimize the difference between the simulated prediction data and the actual observed data in order to obtain the calibrated deterministic parameters.

4. The method for constructing a basic model of an arid irrigation area as described in claim 1, characterized in that, The upstream pre-training phase learns general knowledge from the ecohydrological dataset through an input embedding module, a hybrid masking strategy, and a self-supervised encoder-decoder structure based on a teacher-student mechanism, including: The eco-hydrological dataset is input into the input embedding module to obtain the complete embedding sequence; A hybrid masking strategy is applied to the complete embedding sequence to obtain a masked embedding sequence; The complete embedding sequence and the masked embedding sequence are input into the teacher-student self-supervised encoder-decoder structure, so that the teacher encoder generates a high-quality target representation based on the high-dimensional embedding vector, and the student encoder generates a fragmented data representation based on the masked embedding sequence. The incomplete data representation is input to the decoder and the projection layer, so that the decoder maps the incomplete data representation back to the original variable space and reconstructs the masked variable values, and the projection layer obtains the projected representation based on the incomplete data representation.

5. The method for constructing a basic model of an arid irrigation area as described in claim 4, characterized in that, In the downstream fine-tuning stage, the fixed parameters of the upstream pre-trained model are frozen and an implicit hydrological representation is constructed. The multi-task prediction module is then fine-tuned using the constructed implicit hydrological representation and regional observation data from different tasks to obtain a basic model for arid irrigation areas suitable for the target irrigation district, including: The input embedding module and decoder are frozen, and the student encoder is partially frozen and fine-tuned to fix the main parameters of the upstream pre-trained model. Implicit hydrological representations are constructed based on upstream pre-trained models with fixed main parameters; A conditional vector is constructed for each prediction location by combining the variable embedding vector and the temporal embedding vector; Based on the concept of linear modulation, conditional vectors are mapped to modulation parameters, and the modulation parameters are applied to implicit hydrological characterization for conditional modulation in order to obtain task-specific features; Task-specific features are integrated into the multi-task prediction module of the hybrid expert system. The multi-task prediction module is then fine-tuned based on regional observation data for different tasks to form a basic model of arid irrigation areas suitable for the target irrigation area.

6. The method for constructing a basic model of an arid irrigation area as described in claim 5, characterized in that, The downstream fine-tuning stage also includes: constructing a joint loss function for multi-task prediction, wherein the design of the loss function is as follows: ， In the formula, This represents the downstream monitoring loss; b, t, and v represent the sample number index, time index, and variable index, respectively. The weights represent the relative importance of the b-th sample, time t, and v-th variable in the loss. This represents the supervision mask, used to indicate whether there are true observations for the b-th sample, time t, and the v-th variable used to calculate the loss; , , Let B represent the observed value of the b-th sample, the time t-th time, the observed value of the v-th variable, the model predicted mean, and the model predicted variance, respectively; B represents the total number of batches, T represents the number of time periods, and V represents the number of variables.

7. A system for constructing a basic model of an arid irrigation area, characterized in that, include: The target irrigation district division module is used to acquire multi-source data and divide the target irrigation district into several calculation units based on the multi-source data. An eco-hydrological dataset generation module is used to drive a hydrological model to perform simulations using the multi-source data, and to conduct sensitivity analysis and batch calibration of the hydrological model parameters, thereby generating an eco-hydrological dataset for each computing unit. This includes: driving a hydrological model to perform simulations using meteorological data, static attribute data, irrigation data, and hydrological data to obtain simulated and predicted eco-hydrological elements; combining the simulated and predicted data with actual observation data to perform sensitivity analysis and batch calibration of the hydrological model parameters to obtain calibrated deterministic parameters; and using the hydrological model to generate simulated and predicted eco-hydrological elements based on the deterministic parameters, thereby generating an eco-hydrological dataset for each computing unit. The eco-hydrological dataset includes meteorological data, static attribute data, irrigation data, hydrological data, satellite-aided data, and simulated and predicted eco-hydrological elements. The basic model training module is used to construct a basic model of arid irrigation areas based on the eco-hydrological dataset through upstream pre-training and downstream fine-tuning. In the upstream pre-training stage, general knowledge is learned from the eco-hydrological dataset through an input embedding module, a hybrid masking strategy, and a self-supervised encoder-decoder structure based on a teacher-student mechanism to obtain an upstream pre-trained model. In the downstream fine-tuning stage, the main parameters of the upstream pre-trained model are fixed and an implicit hydrological representation is constructed. The multi-task prediction module is then fine-tuned by combining the constructed implicit hydrological representation with regional observation data for different tasks to obtain a basic model of arid irrigation areas suitable for the target irrigation area.

8. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute a method for constructing a basic model of an arid irrigation area as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute a method for constructing a basic model of an arid irrigation area as described in any one of claims 1 to 6.

Images