Multi-source satellite data driven ecological hydrological model construction method for irrigation district

Abstract

The present application relates to the technical field of ecological hydrological modeling and artificial intelligence, and particularly relates to a multi-source satellite data driven irrigation ecological hydrological basic model construction method, wherein the method comprises: obtaining multi-source satellite data and ecological hydrological related data of a target irrigation area; preprocessing the multi-source satellite data and ecological hydrological related data, and interpolating missing data to generate complete ecological hydrological data; fusing the time series satellite general representation generated from the multi-source satellite data and the physical variable embedding of the ecological hydrological data to generate a primary ecological hydrological general representation, and combining a self-supervised task to construct an irrigation ecological hydrological basic model, output a unified ecological hydrological general representation and perform ecological hydrological variable estimation. Thus, the problems of discontinuous results in time series or not conforming to physical laws due to reliance on statistical learning and lack of physical constraints, insufficient multi-source data fusion, and difficulty in achieving high-precision estimation of ecological hydrological elements in related technologies are solved.

Description

Technical Field

[0001] This invention relates to the field of eco-hydrological modeling and artificial intelligence, and in particular to a method for constructing a basic eco-hydrological model of an irrigation area driven by multi-source satellite data. Background Technology

[0002] Currently, ecohydrological research in irrigation districts typically employs models based on physical process equations to simulate the coupling relationships between surface water, groundwater, soil water, and vegetation processes, used to analyze water transport and energy balance. However, the application of such models in irrigation districts has the following limitations: First, parameter acquisition and calibration are difficult, with significant spatial differences in soil hydraulics and evapotranspiration-related parameters and a limited number of sampling points, leading to high uncertainty in the results; second, the model structure is based on idealized assumptions, making it difficult to accurately describe irrigation scheduling, shallow groundwater recharge, and heterogeneous soil conditions, resulting in decreased simulation accuracy; third, it is difficult to simultaneously utilize multi-source data such as satellite remote sensing and ground sensors, lacking data-driven feedback and constraints, resulting in insufficient performance for regional generalization and time-series prediction; fourth, a single model usually maintains high accuracy only within a small area, making it difficult to balance computational efficiency and accuracy at the regional scale.

[0003] Among related technologies, remote sensing data and deep learning methods are used to estimate eco-hydrological elements. Spatiotemporal estimation of surface variables can be achieved based on satellite observations. In order to make up for the lack of observations, some studies have also attempted to fuse multi-source data to improve spatiotemporal coverage by integrating different satellite data.

[0004] However, in related technologies, remote sensing and deep learning rely on statistical learning and a large number of ground samples, lacking physical constraints. This leads to results that are discontinuous in time or inconsistent with physical laws, making them difficult to apply to areas with scarce observations. They also have insufficient ability to estimate potential variables such as deep soil water and groundwater. Multi-source data fusion is prone to data mismatch or information bias, making it difficult to achieve spatiotemporal fusion consistent with physical processes. Consequently, it is difficult to obtain satisfactory eco-hydrological element estimation and dynamic prediction results under conditions of limited input data, which urgently needs improvement. Summary of the Invention

[0005] This invention provides a method for constructing a basic ecological hydrological model of irrigation districts driven by multi-source satellite data. This method addresses the problems in related technologies, such as the reliance on statistical learning and a large number of ground samples, the lack of physical constraints leading to discontinuous or non-physical results, difficulty in applying to areas with scarce observations, data mismatch or information bias in multi-source data fusion, difficulty in achieving spatiotemporal fusion consistent with physical processes, and difficulty in achieving high-precision estimation and dynamic prediction of ecological hydrological elements under limited input data.

[0006] The first aspect of this invention provides a method for constructing an eco-hydrological basic model of an irrigation area driven by multi-source satellite data, comprising the following steps: acquiring multi-satellite multi-temporal image data, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products of a target irrigation area to generate multi-source satellite data; performing spatial reprojection, resampling, and resolution unification processing on the multi-source satellite data to obtain standardized modeling units; and based on the standardized modeling units, interpolating the missing data of the target irrigation area using a pre-constructed eco-hydrological physical model to generate the target irrigation area's eco-hydrological basic model. The process involves: generating a time-series satellite general representation of the target irrigation area based on the multi-source satellite data; fusing the time-series satellite general representation with the physical variable embeddings of the eco-hydrological data to generate a primary time-series eco-hydrological general representation; establishing a basic eco-hydrological model of the irrigation area based on the primary time-series eco-hydrological general representation and pre-constructed self-supervised learning constraints; generating a unified eco-hydrological general representation based on the basic eco-hydrological model of the irrigation area and the eco-hydrological observation data of the target irrigation area; and outputting downstream eco-hydrological variable prediction results based on the unified eco-hydrological general representation.

[0007] Through the aforementioned technical means, this invention can standardize and physically consistent interpolate multi-source information, including multi-satellite images, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products from a target irrigation area, to generate standardized eco-hydrological data. It embeds and fuses temporal satellite observation features with the physical variables of the eco-hydrological data to construct a temporal eco-hydrological representation. Based on this temporal eco-hydrological representation and self-supervised learning constraints, it constructs a basic eco-hydrological model for the irrigation area and outputs prediction results. By deeply fusing multi-source satellite data, it achieves complementarity and unification of multi-modal observation information, including optical, radar, and ecological products, significantly improving the model's robustness and generalization ability under incomplete multi-source data conditions. This, in turn, enhances the accuracy and temporal stability of eco-hydrological element estimation, enabling collaborative estimation and dynamic prediction of multiple eco-hydrological elements from the surface to the ground.

[0008] Optionally, in one embodiment of the present invention, the step of interpolating the missing data of the target irrigation area using a pre-constructed eco-hydrological physical model to generate eco-hydrological data of the target irrigation area includes: calibrating the parameters of the pre-constructed eco-hydrological physical model using groundwater observation data, surface water process data, evapotranspiration data, leaf area index data, and aboveground biomass data of the target irrigation area to obtain a calibrated eco-hydrological physical model; running the calibrated eco-hydrological physical model on the standardized modeling unit to generate supplementary data of the target irrigation area; and interpolating the missing data of the target irrigation area based on the supplementary data to generate the eco-hydrological data.

[0009] Through the above-mentioned technical means, the embodiments of the present invention can use multi-dimensional measured data of the target irrigation area to calibrate the eco-hydrological physical model. Running this calibration model can generate reliable supplementary data on the standardized modeling unit, thereby effectively filling in the missing values ​​in the original multi-source data, overcoming the problem of data discontinuity, effectively restoring the spatiotemporal continuity of the data, and providing a high-quality input dataset that is spatiotemporally complete and physically consistent for subsequent modeling.

[0010] Optionally, in one embodiment of the present invention, the step of fusing the time-series satellite general representation and the physical variable embedding of the eco-hydrological data to generate a primary time-series eco-hydrological general representation includes: performing multidimensional encoding on the eco-hydrological data to generate the eco-hydrological data physical variable embedding; fusing the satellite general representation and the eco-hydrological data physical variable embedding to obtain a multimodal input matrix; extracting a preset number of latent representations based on the multimodal input matrix; and inputting the latent representations into a preset spatiotemporal feature fusion encoder to generate the primary time-series eco-hydrological general representation.

[0011] Through the above-mentioned technical means, the embodiments of the present invention can encode eco-hydrological data in multiple dimensions to generate physical variable embeddings. By using the cross-attention mechanism, potential representations are automatically extracted from the satellite general representation and physical variable embedding space, and input into the spatiotemporal feature fusion encoder to generate primary time-series eco-hydrological general representations. This avoids the limitations of relying solely on explicit physical variable coupling, enabling the model to identify key processes from complex spatiotemporal structures and characterize the temporal evolution law and multivariate coupling characteristics of eco-hydrological processes.

[0012] Optionally, in one embodiment of the present invention, the step of constructing a basic ecological hydrological model of an irrigation district using the primary time-series eco-hydrological general representation includes: determining the self-supervised learning constraints of the model; training the model using the self-supervised learning constraints and the primary time-series eco-hydrological general representation until the preset training conditions are met, thereby establishing the basic ecological hydrological model of the irrigation district.

[0013] Through the above-mentioned technical means, the embodiments of the present invention can train the model through self-supervised learning constraints, enabling the basic model of irrigation area eco-hydrology to aggregate eco-hydrological related features in a general representation and maintain the physical consistency of these features in time and space, thereby improving the accuracy and temporal stability of eco-hydrological element estimation.

[0014] Optionally, in one embodiment of the present invention, the step of outputting the eco-hydrological variable prediction results using the irrigation district eco-hydrological basic model includes: generating the unified eco-hydrological general representation based on the irrigation district eco-hydrological basic model and the eco-hydrological observation data of the target irrigation district; generating the general representation space of the target irrigation district based on the unified eco-hydrological general representation; and extracting the task features of the general representation space; constructing a downstream supervised learning lightweight model based on the task features, so as to output the eco-hydrological variable prediction results using the downstream supervised learning lightweight model.

[0015] Through the above-mentioned technical means, the embodiments of the present invention can achieve spatiotemporal estimation and scenario prediction of eco-hydrological elements without relying on complete ground monitoring or groundwater data during the application stage, relying only on multi-source satellites and meteorological drives. The task query mechanism ensures flexible retrieval and interpretable inference of different variables and depths. The time series prediction driven by the drive enables the model to have forward-looking simulation and trend forecasting capabilities, providing reliable data support for hydrological management and ecological monitoring at the irrigation district scale.

[0016] A second aspect of this invention provides a device for constructing a basic eco-hydrological model of an irrigation area driven by multi-source satellite data, comprising: an acquisition module for acquiring multi-satellite multi-temporal image data, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products of a target irrigation area to generate multi-source satellite data; and a generation module for performing spatial reprojection, resampling, and resolution unification processing on the multi-source satellite data to obtain standardized modeling units, and interpolating the missing data of the target irrigation area based on the standardized modeling units using a pre-constructed eco-hydrological physical model to generate the eco-hydrological model of the target irrigation area. The system comprises: a data fusion module, used to generate a time-series satellite general representation of the target irrigation area based on the multi-source satellite data, and to fuse the time-series satellite general representation with the physical variable embedding of the eco-hydrological data to generate a primary time-series eco-hydrological general representation; and a construction module, used to combine the primary time-series eco-hydrological general representation with pre-constructed self-supervised learning constraints to establish a basic eco-hydrological model of the irrigation area, generate a unified eco-hydrological general representation based on the basic eco-hydrological model of the irrigation area and the eco-hydrological observation data of the target irrigation area, and output downstream eco-hydrological variable prediction results based on the unified eco-hydrological general representation.

[0017] Through the aforementioned technical means, this invention can standardize and physically consistent interpolate multi-source information, including multi-satellite images, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products from a target irrigation area, to generate standardized eco-hydrological data. It embeds and fuses temporal satellite observation features with the physical variables of the eco-hydrological data to construct a temporal eco-hydrological representation. Based on this temporal eco-hydrological representation and self-supervised learning constraints, it constructs a basic eco-hydrological model for the irrigation area and outputs prediction results. By deeply fusing multi-source satellite data, it achieves complementarity and unification of multi-modal observation information, including optical, radar, and ecological products, significantly improving the model's robustness and generalization ability under incomplete multi-source data conditions. This, in turn, enhances the accuracy and temporal stability of eco-hydrological element estimation, enabling collaborative estimation and dynamic prediction of multiple eco-hydrological elements from the surface to the ground.

[0018] Optionally, in one embodiment of the present invention, the generation module includes: a calibration unit, used to calibrate the parameters of the pre-constructed eco-hydrological physical model using groundwater observation data, surface water process data, evapotranspiration data, leaf area index data, and aboveground biomass data of the target irrigation area, to obtain a calibrated eco-hydrological physical model; a supplementation unit, used to run the calibrated eco-hydrological physical model on the standardized modeling unit to generate supplementary data for the target irrigation area; and an interpolation unit, used to interpolate the missing data of the target irrigation area based on the supplementary data to generate the eco-hydrological data.

[0019] Through the above-mentioned technical means, the embodiments of the present invention can use multi-dimensional measured data of the target irrigation area to calibrate the eco-hydrological physical model. Running this calibration model can generate reliable supplementary data on the standardized modeling unit, thereby effectively filling in the missing values ​​in the original multi-source data, overcoming the problem of data discontinuity, effectively restoring the spatiotemporal continuity of the data, and providing high-quality input data that is spatiotemporally complete and physically consistent for subsequent modeling.

[0020] Optionally, in one embodiment of the present invention, the fusion module includes: an encoding unit for multidimensional encoding of the eco-hydrological data to generate the eco-hydrological data physical variable embedding; a multimodal fusion unit for fusing the satellite general representation with the eco-hydrological data physical variable embedding to obtain a multimodal input matrix; a representation extraction unit for extracting a preset number of potential representations based on the multimodal input matrix; and a generation unit for inputting the potential representations into a preset spatiotemporal feature fusion encoder to generate the primary time-series eco-hydrological general representation.

[0021] Through the above-mentioned technical means, the embodiments of the present invention can encode eco-hydrological data in multiple dimensions to generate physical variable embeddings. By using the cross-attention mechanism, potential representations are automatically extracted from the satellite general representation and physical variable embedding space, and input into the spatiotemporal feature fusion encoder to generate primary time-series eco-hydrological general representations. This avoids the limitations of relying solely on explicit physical variable coupling, enabling the model to identify key processes from complex spatiotemporal structures and characterize the temporal evolution law and multivariate coupling characteristics of eco-hydrological processes.

[0022] Optionally, in one embodiment of the present invention, the construction module includes: a determination unit for determining the self-supervised learning constraints of the model; and a training unit for training the model using the self-supervised learning constraints and the primary time-series eco-hydrological general representation until the preset training conditions are met, thereby establishing the basic eco-hydrological model of the irrigation area.

[0023] Through the above-mentioned technical means, the embodiments of the present invention can train the model through self-supervised learning constraints, enabling the basic model of irrigation area eco-hydrology to aggregate eco-hydrological related features in a general representation and maintain the physical consistency of these features in time and space, thereby improving the accuracy and temporal stability of eco-hydrological element estimation.

[0024] Optionally, in one embodiment of the present invention, the construction module includes: a feature extraction unit, configured to generate the unified eco-hydrological general representation based on the irrigation district eco-hydrological basic model and the eco-hydrological observation data of the target irrigation district, generate a general representation space of the target irrigation district based on the unified eco-hydrological general representation, and extract task features from the general representation space; and a prediction unit, configured to construct a downstream supervised learning lightweight model based on the task features, so as to output the eco-hydrological variable prediction results using the downstream supervised learning lightweight model.

[0025] Through the above-mentioned technical means, the embodiments of the present invention can achieve spatiotemporal estimation and scenario prediction of eco-hydrological elements without relying on complete ground monitoring or groundwater data during the application stage, relying only on multi-source satellites and meteorological drives. The task query mechanism ensures flexible retrieval and interpretable inference of different variables and depths. The time series prediction driven by the drive enables the model to have forward-looking simulation and trend forecasting capabilities, providing reliable data support for hydrological management and ecological monitoring at the irrigation district scale.

[0026] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for constructing a basic ecological hydrological model of an irrigation area driven by multi-source satellite data as described in the above embodiments.

[0027] A fourth aspect of the present invention provides a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for constructing a multi-source satellite data-driven ecological hydrological basic model for irrigation districts.

[0028] A fifth aspect of the present invention provides a computer program product, which stores a computer program that, when executed by a processor, implements the above-described method for constructing a basic ecological hydrological model of an irrigation district driven by multi-source satellite data.

[0029] This invention can integrate multi-source information from multiple satellites, including multi-temporal images, driving data, attribute data, groundwater observation data, and eco-hydrological satellite products, to generate standardized eco-hydrological data through standardization and physical consistency interpolation. It embeds and fuses temporal satellite observation features with the physical variables of the eco-hydrological data to construct a temporal eco-hydrological representation. Based on this temporal eco-hydrological representation and self-supervised learning constraints, it builds a basic eco-hydrological model for the irrigation area and outputs prediction results. By deeply fusing multi-source satellite data, it achieves complementarity and unification of multi-modal observation information such as optical, radar, and ecological products, significantly improving the model's robustness and generalization ability under incomplete multi-source data conditions. This, in turn, improves the accuracy and temporal stability of eco-hydrological element estimation, enabling collaborative estimation and dynamic prediction of multiple eco-hydrological elements from the surface to the ground. This solves the problems in related technologies, such as the lack of physical constraints due to reliance on statistical learning and a large number of ground samples, which leads to discontinuous results or non-compliance with physical laws, making it difficult to apply to areas with scarce observations, data mismatch or information bias in multi-source data fusion, difficulty in achieving spatiotemporal fusion consistent with physical processes, and difficulty in achieving high-precision estimation and dynamic prediction of eco-hydrological elements under the condition of limited input data.

[0030] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0031] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0032] Figure 1 A flowchart illustrating a method for constructing a basic eco-hydrological model of an irrigation district driven by multi-source satellite data, according to an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of a self-looping supervision module based on a weak Markov model according to an embodiment of the present invention;

[0034] Figure 3A flowchart of a method for constructing an irrigation district eco-hydrological basic model driven by multi-source satellite data according to an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the basic architecture of an irrigation district eco-hydrological model according to an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of a multi-source satellite data-driven irrigation district eco-hydrological basic model construction device according to an embodiment of the present invention;

[0037] Figure 6 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of the present invention.

[0038] Figure label:

[0039] 10-Multi-source satellite data-driven irrigation area eco-hydrological basic model construction device; 100-Acquisition module, 200-Generation module, 300-Fusion module, 400-Construction module; 601-Memory, 602-Processor, 603-Communication interface. Detailed Implementation

[0040] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0041] The following describes, with reference to the accompanying drawings, a method for constructing a basic eco-hydrological model of an irrigation area driven by multi-source satellite data according to an embodiment of the present invention. Addressing the issues raised in the background section regarding related technologies, which rely on statistical learning and a large number of ground samples, lack physical constraints, and result in discontinuous or non-physically consistent results, making them difficult to apply to areas with scarce observations, and prone to data mismatch or information bias in multi-source data fusion, making it difficult to achieve spatiotemporal fusion consistent with physical processes, and consequently, struggling to achieve high-precision estimation and dynamic prediction of eco-hydrological elements under limited input data conditions, the present invention provides a method for constructing a basic eco-hydrological model of an irrigation area driven by multi-source satellite data. This method integrates multi-satellite multi-temporal images, driving data, attribute data, groundwater observation data, and eco-hydrological phase data of the target irrigation area. This method utilizes multi-source information, including satellite products, to generate standardized eco-hydrological data through standardization and physical consistency interpolation. It embeds and fuses temporal satellite observation characteristics with physical variables from eco-hydrological data to construct a temporal eco-hydrological representation. Based on this representation and self-supervised learning constraints, a basic eco-hydrological model for irrigation districts is built, and prediction results are output. Deep fusion of multi-source satellite data achieves complementarity and unification of multi-modal observation information, including optical, radar, and ecological products. This significantly improves the model's robustness and generalization ability under conditions of incomplete multi-source data, thereby enhancing the accuracy and temporal stability of eco-hydrological element estimation. It enables collaborative estimation and dynamic prediction of multiple eco-hydrological elements from the surface to the subsurface. This addresses problems in related technologies, such as reliance on statistical learning and large ground samples, lack of physical constraints leading to discontinuous or non-physically consistent results, difficulty in applying to areas with scarce observations, data mismatch or information bias in multi-source data fusion, difficulty in achieving spatiotemporal fusion consistent with physical processes, and difficulty in achieving high-precision estimation and dynamic prediction of eco-hydrological elements under limited input data.

[0042] Specifically, Figure 1 This is a flowchart illustrating a method for constructing an irrigation district eco-hydrological basic model driven by multi-source satellite data, as provided in an embodiment of the present invention.

[0043] like Figure 1 As shown, the method for constructing a basic eco-hydrological model of an irrigation district driven by multi-source satellite data includes the following steps:

[0044] In step S101, multi-satellite multi-phase image data, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products of the target irrigation area are acquired to generate multi-source satellite data.

[0045] It is understood that the eco-hydrological related satellite products in the embodiments of the present invention may include satellite products such as surface soil water, evapotranspiration, leaf area index and aboveground biomass of crops.

[0046] In practical implementation, this invention can address the needs of eco-hydrological modeling in the Hetao Irrigation District by acquiring multi-satellite, multi-temporal imagery data covering the entire area. This includes Sentinel-1 SAR radar data, Sentinel-2 MSI reflectivity data, Landsat 8 / 9 OLI / TIRS optical data, MODIS MOD09 / MYD09 surface reflectivity products, GF series high-resolution optical images, and GEDI lidar data. Simultaneously, eco-hydrological satellite products are acquired, including surface soil water (e.g., SMAP L3 products), evapotranspiration (e.g., MOD16A2), leaf area index (e.g., MCD15A2H), and crop aboveground biomass (AGB). These multi-source data complement each other temporally and spatially, providing multimodal observational support for subsequent spatiotemporal fusion and eco-hydrological feature extraction.

[0047] Furthermore, meteorological driving data, soil property data, groundwater observation data, and irrigation management data for the irrigation area were collected. Meteorological data, including elements such as temperature, precipitation, radiation, wind speed, and humidity, mainly came from observations at regional meteorological stations, supplemented by reanalysis data (such as ERA5-Land or CMADS). Soil property data, including indicators such as soil texture, bulk density, organic matter, field capacity, and saturated hydraulic conductivity, were multi-layered and integrated with field surveys in the Hetao Irrigation Area and the global SoilGrids dataset. Groundwater data came from monitoring well water level records, canal recharge information, and pumping data. Irrigation data came from canal flow and field water use records from local water conservancy departments. The inverse distance weighting method or Kriging interpolation method was used to transform discrete observation data into continuous grids for meteorological, groundwater, and irrigation point data, obtaining the spatial distribution characteristics of driving, property, and groundwater elements.

[0048] This invention can systematically integrate various types of information, such as multi-satellite multi-temporal images, driving data, and attribute data, to generate multi-source satellite data, providing rich and complementary information sources for subsequent fusion modeling. This effectively solves the problem of incomplete information or insufficient observation frequency from a single data source, laying a solid data foundation for achieving high spatiotemporal resolution and physically consistent eco-hydrological simulation.

[0049] In step S102, the multi-source satellite data is subjected to spatial reprojection, resampling and resolution unification processing to obtain standardized modeling units. Based on the standardized modeling units, the missing data of the target irrigation area is interpolated using a pre-constructed eco-hydrological physical model to generate eco-hydrological data of the target irrigation area.

[0050] It is understood that the standardized modeling unit in the embodiments of the present invention can be understood as a regular gridded spatial data layer formed after spatial alignment and resolution unification, and all data are calculated under this grid framework; the pre-built eco-hydrological physical model can be SWAP or other models suitable for water-energy-crop process coupling simulation in arid irrigation areas, and can be used to fill in missing data values.

[0051] In practical implementation, embodiments of this invention can spatially reproject and resample multi-source satellite imagery and ground observation data, unifying them to the UTM WGS 1984 Zone 49N projection coordinate system, using a consistent spatial resolution (e.g., 1000 m) and time step (daily). Embodiments of this invention can align data from different sources and at different resolutions to the same spatiotemporal grid. In the temporal dimension, linear interpolation or moving average methods are used to unify multi-temporal images and meteorological data to the same time step. Spatially, bilinear or cubic convolution resampling is used for continuous variables, and nearest-neighbor resampling is used for categorical variables; in the temporal dimension, linear interpolation or moving averages are used to achieve temporal alignment between multi-temporal images and meteorological driving data. After processing, a standardized modeling unit is formed, containing multi-source satellite features, meteorological driving data, soil properties, groundwater level, irrigation amount, and quality control marker layers, providing a unified input basis for subsequent physical model interpolation and multimodal feature fusion.

[0052] Furthermore, a pre-built eco-hydrological physical model is run to generate eco-hydrological data such as multi-layer soil moisture, water flux, evapotranspiration, and crop phenotypes aligned with a unified grid. The results are directly used as interpolation data for sparsely observed areas to generate eco-hydrological data for the target irrigation area that has physical consistency and contains information such as multi-layer soil moisture, water flux, and vegetation phenotypes.

[0053] This invention can standardize multi-source data through spatial reprojection and resampling, and use eco-hydrological physical models to supplement missing data to generate eco-hydrological data. Thus, the spatial heterogeneity of multi-source data is eliminated through standardization, which solves the modeling error caused by spatiotemporal misalignment of data in traditional models, effectively restores the spatiotemporal continuity of data, avoids the loss of modeling information due to missing data, and the generated eco-hydrological data has physical consistency, providing high-quality input for subsequent feature fusion and model construction.

[0054] Optionally, in one embodiment of the present invention, a pre-constructed eco-hydrological physical model is used to interpolate missing data of the target irrigation area to generate eco-hydrological data of the target irrigation area. This includes: calibrating the parameters of the pre-constructed eco-hydrological physical model using groundwater observation data, surface water process data, evapotranspiration data, leaf area index data, and aboveground biomass data of the target irrigation area to obtain a calibrated eco-hydrological physical model; running the calibrated eco-hydrological physical model on a standardized modeling unit to generate supplementary data of the target irrigation area; and interpolating the missing data of the target irrigation area based on the supplementary data to generate eco-hydrological data.

[0055] It is understood that parameter calibration in the embodiments of the present invention can be understood as the process of adjusting the key parameters in the eco-hydrological physical model using limited field observation data of the target irrigation area, so that the model simulation results and local measured values ​​achieve the best fit; supplementary data can be understood as the estimated values ​​of missing data output by the model after calibration.

[0056] In practical implementation, embodiments of this invention can utilize groundwater observation data, surface water process data, evapotranspiration data, leaf area index (LAI) data, and aboveground biomass (AGB) data to calibrate the parameters of the eco-hydrological model. The physical model used for calibration can be SWAP or other models suitable for coupled simulation of water-energy-crop processes in arid irrigation areas. Model inputs include meteorological drivers, soil properties, groundwater boundaries, and crop physiological parameters. Parameter calibration is based on existing parameters from similar studies of the target irrigation area. By adjusting soil hydraulic parameters (such as saturated hydraulic conductivity, field capacity, and porosity), crop root parameters (such as root depth and root water absorption distribution), and evapotranspiration coefficients (such as crop coefficient Kc and maximum transpiration rate), the simulation results are made consistent with the observed data across multiple time scales. The calibration process employs a strategy combining stepwise calibration and sensitivity analysis to establish physical constraints between observable variables (surface moisture, evapotranspiration, and crop phenotype) and latent variables (deep soil water and groundwater), ensuring that the model's temporal response matches the actual regional processes.

[0057] After parameter calibration, the calibrated physical model was run on standardized modeling units. Model-driven data included daily meteorological factors (temperature, precipitation, radiation, humidity, wind speed), irrigation amount, and soil property parameters. Output variables covered soil moisture, water flux, evapotranspiration, crop transpiration, and crop physiological variables such as leaf area index (LAI) and aboveground biomass (AGB) at different depths. To ensure spatial consistency, all simulation results were aligned with the unified modeling grid (UTMWGS 1984 Zone 49N) and maintained consistency with the satellite observation time step in the temporal dimension. Model-driven calculations were performed in areas with missing data (such as sparse groundwater areas and unmonitored irrigation areas) to supplement surface and subsurface eco-hydrological data, achieving continuous spatial and temporal interpolation. The water and energy fluxes output by the model satisfied mass conservation and energy balance constraints, thus ensuring the physical consistency of the generated data.

[0058] The eco-hydrological data output from the physical model is integrated with driving data such as meteorological and irrigation data, as well as attribute data such as soil and topography. Based on data functional characteristics, it is divided into three categories: driving data (meteorological factors and irrigation inputs), attribute data (soil physicochemical properties, topographic features), and state / flux data (soil moisture, evapotranspiration, groundwater level, crop phenotypes, etc.). Each type of data corresponds one-to-one with a modeling unit spatially and is unified to the same time step temporally. The integrated eco-hydrological dataset not only contains observable variables but also potential eco-hydrological elements obtained through model interpolation, providing continuous and physically consistent input constraints for subsequent spatiotemporal feature fusion and deep model training.

[0059] The embodiments of the present invention can use multi-dimensional measured data of the target irrigation area to calibrate the eco-hydrological physical model. Running this calibration model can generate reliable supplementary data on the standardized modeling unit, thereby effectively filling in the missing values ​​in the original multi-source data, overcoming the problem of data discontinuity, effectively restoring the spatiotemporal continuity of the data, and providing high-quality input data that is spatiotemporally complete and physically consistent for subsequent modeling.

[0060] In step S103, a time-series satellite general characterization of the target irrigation area is generated based on multi-source satellite data, and the time-series satellite general characterization and the physical variable embedding of eco-hydrological data are fused to generate a primary time-series eco-hydrological general characterization.

[0061] It is understood that the temporal satellite general representation in the embodiments of the present invention can be understood as spatial features automatically extracted from multi-source satellite data through a deep learning network, and the joint fusion of multi-source features is achieved through a temporal attention mechanism and a spatial attention mechanism; physical variable embedding can be understood as the physical features in eco-hydrological data being mapped into feature vectors by an encoder.

[0062] In actual implementation, the embodiments of the present invention can first perform temporal registration and spatial alignment on each data source to address the differences in spatial resolution, temporal resolution and observation frequency of different satellite images, so as to ensure that the multi-source satellite data within the same modeling unit correspond in the temporal dimension and match in the spatial dimension, thus providing a consistent input basis for feature fusion.

[0063] Based on this, embodiments of the present invention can add time encoding and satellite attribute encoding to satellite imagery data from different sources to enhance the model's perception of observation time, sensor type, and observation conditions. Spatial features are extracted using convolutional neural networks, and multi-source features are jointly fused through temporal and spatial attention mechanisms to generate a general satellite representation with temporal continuity and spatial consistency.

[0064] Specifically, the multi-source satellite fusion encoder conducts self-supervised training by introducing a mask reconstruction task and teacher-student model representation consistency constraints. By reconstructing the features of the masked images and maintaining the consistency of representations among models, a unified time-series satellite general representation is obtained.

[0065] Furthermore, embodiments of the present invention can encode the driving, attribute, and state / flux data by time, location, variable name, and numerical value to obtain physical variable embeddings; after fusing the physical variable embeddings of eco-hydrological data with time-series satellite general representations, latent representations are extracted through a cross-attention mechanism, and multimodal spatiotemporal dependencies are modeled through a spatiotemporal feature fusion encoder (self-attention). The high-dimensional time-series feature set output by the encoder is the primary time-series eco-hydrological general representation, which carries cross-modal, cross-variable, and cross-spatiotemporal coupling information for subsequent self-supervised training and task querying.

[0066] The embodiments of the present invention can extract time-series satellite general characteristics from multi-source satellite data, integrate these characteristics with the physical variable embedding of eco-hydrological data, and generate a primary time-series eco-hydrological general characteristic, which can accurately reflect the dynamic change process of the irrigation area surface, retain the core physical attributes of eco-hydrological data, and lay the foundation for improving the scientificity and accuracy of model prediction.

[0067] Optionally, in one embodiment of the present invention, fusing time-series satellite general representations and physical variable embeddings of eco-hydrological data to generate a primary time-series eco-hydrological general representation includes: performing multidimensional encoding on eco-hydrological data to generate eco-hydrological data physical variable embeddings; fusing satellite general representations and eco-hydrological data physical variable embeddings to obtain a multimodal input matrix; extracting a preset number of latent representations based on the multimodal input matrix; and inputting the latent representations into a preset spatiotemporal feature fusion encoder to generate a primary time-series eco-hydrological general representation.

[0068] It is understood that, in the embodiments of the present invention, multidimensional encoding can be understood as the process of converting multiple types of features of eco-hydrological data into a unified format through various encoding methods (such as time encoding, location encoding, variable name encoding, and numerical encoding); the multimodal input matrix can be understood as a data structure that integrates features from different sources (modalities); the preset spatiotemporal feature fusion encoder can be a learning module based on a self-attention mechanism, used to learn and extract unified spatiotemporal features from multimodal inputs. The preset spatiotemporal feature fusion encoder can be set by those skilled in the art according to the actual situation, and no specific restrictions are imposed here.

[0069] For example, embodiments of the present invention can perform time encoding, location encoding, variable name encoding, and numerical encoding on driving data, attribute data, and state / flux data, respectively, to generate physical variable embeddings. Time encoding reflects seasonality and diurnal variation patterns; location encoding distinguishes the spatial distribution characteristics of different modeling units; variable name encoding distinguishes different physical elements (such as temperature, precipitation, soil moisture content, evapotranspiration, etc.) through learnable vectors; and numerical encoding normalizes physical quantity values ​​and converts them into trainable vector representations. After multiple encodings, various physical variables are mapped to an embedding space of the same dimension, forming a unified physical variable embedding.

[0070] Numerical encoding is used to represent quantitative information about physical variables. First, the original variable values ​​are normalized to convert them into dimensionless numerical values. Then, it is mapped to a fixed-dimensional feature vector through a linear encoder, and its calculation formula is:

[0071] ，

[0072] in, This represents the learnable linear transformation weight matrix. As the bias vector, the output is... This is a 64-dimensional numerical feature vector used to characterize the quantitative magnitude information of the variable. Through this mapping, different physical quantities (such as temperature, precipitation, evapotranspiration, etc.) can be expressed on a unified numerical scale, providing a basis for the model to learn the relative changes across variables.

[0073] Location encoding is used to represent the spatial relationships of variables with vertical structure across different depth layers. For stratified variables (such as soil moisture content at different depths), a simple, learnable neural network function is used for each depth. Generate a deep embedding vector using the following formula:

[0074] ,

[0075] in, For a learnable, simple neural network function, the input is a depth value. (e.g., 10 cm, 20 cm, 40 cm, etc.), the output is a 64-dimensional vector. This encoding can learn the correlations and differences between different depth layers, thus reflecting the vertical distribution characteristics of groundwater. For variables without vertical layers (such as meteorological factors or satellite indices), a default constant vector is assigned. To maintain consistency in the dimensions of the input structure.

[0076] Variable name embedding is used to introduce semantic and identity information for physical variables. For each physical variable type... (e.g., temperature, precipitation, evapotranspiration, soil moisture, LAI, etc.) are assigned a globally shared, learnable embedding vector that remains consistent across all samples and time steps to characterize the semantic role of the variable in ecohydrological processes.

[0077] Formally, define the variable embedding matrix:

[0078]

[0079] in This represents the total number of physical variables involved in the modeling. For the ... Each variable has a corresponding semantic vector obtained by looking up a table:

[0080]

[0081] During model training, the variable embedding vectors and network parameters are optimized together through backpropagation, enabling the model to automatically learn the intrinsic relationships between different physical variables at the level of statistical characteristics and physical processes, such as the positive correlation between temperature and evapotranspiration, and the lagged effect of precipitation on soil moisture, without the need to explicitly introduce artificial rules or prior assumptions.

[0082] Temporal encoding is used to capture temporal features and periodic patterns. It employs a periodic temporal embedding form based on sine and cosine functions, projected through a learnable matrix. Its computational form is as follows:

[0083] ，

[0084] in, Indicates a time step (such as date or birth period number). For the first k A time frequency is used to control the periodic scale of time changes. This is a learnable projection matrix. Through this encoding, the model can identify time-dependent features such as seasonal variations, crop growth cycles, and climate fluctuations.

[0085] Finally, the four types of encoding results are summed and fused to obtain a unified 64-dimensional embedding vector:

[0086] ,

[0087] in, The final physical variable embedding comprehensively incorporates the variable's numerical attributes, spatial location, physical semantics, and temporal features. Physical variable embedding provides a unified and physically semantically consistent input foundation for subsequent latent feature extraction and spatiotemporal coupling learning.

[0088] Encoded physical variables are embedded and aligned with time-series satellite general representations in the time dimension to construct a multimodal time-series input matrix, enabling a joint representation of physical variables and satellite observation information. To improve the model's robustness and generalization ability under data gaps, a masking mechanism is introduced during the fusion process, randomly generating two types of samples for the input features: complete input and masked input. Complete input samples are fed into the teacher model to generate the target representation; masked input samples are fed into the student model to perform feature reconstruction and representation consistency learning. Through this teacher-student structure of masked self-supervised training, the model can maintain feature reconstruction ability and cross-modal consistency even with partial feature loss, thereby enhancing the model's stability in fusing multi-source asynchronous data.

[0089] Among them, the masking method is used to randomly mask some features or time segments of the input sample, thereby simulating common missing scenarios of multi-source data in actual observations during self-supervised training, such as cloud and rain obstruction, data loss, or sensor inconsistency. The masking method can include one or a combination of the following three random masking strategies: Feature Dimension Masking: Masking all feature dimensions of the input sample with a preset probability to simulate the situation where features are completely missing, used to examine the model's reconstruction ability when key features are unavailable; Temporal Random Masking: Randomly selecting several time steps in the input temporal feature sequence with a preset probability for masking, simulating the situation of missing data in individual time phases or abnormal satellite observations, thereby enhancing the model's robustness to missing inputs in different time phases; Segmented Masking: Dividing the input temporal sequence into several continuous time periods, and randomly selecting one segment for overall masking, to simulate the situation of continuous observation loss, cloud and rain obstruction, or long-term sensor failure.

[0090] During model training, the three masking strategies mentioned above can be used independently or in combination, and the masking probability can be dynamically adjusted according to task complexity or data integrity. By introducing diverse masking methods during the training phase, the model can learn more robust spatiotemporal feature representations even when features are missing, thereby improving the learning performance of subsequent mask reconstruction tasks and teacher-student consistency constraints, and achieving temporal continuity and physical consistency of representations.

[0091] Furthermore, the multimodal input matrix is ​​fed into the cross-attention module, and a fixed number of latent temporal feature vectors are extracted through attention weight calculation. This process establishes a dynamic weight relationship between the time and modal dimensions, enabling the model to automatically aggregate key features from different physical variables and satellite information. Ultimately, a model containing... N ( N =128) latent features, each of which comprehensively reflects the coupling characteristics and spatiotemporal evolution information of multiple elements in the eco-hydrological system. The resulting latent representations not only achieve deep fusion between different physical modes, but also provide a unified representation space for subsequent self-supervised optimization and task feature extraction.

[0092] Specifically, after obtaining the multimodal time-series input matrix (which includes physical variable encoding embeddings and satellite general representations), let this matrix be:

[0093] ，

[0094] in, The number of multimodal input variables, Indicates the first Embedded input variables.

[0095] Initialize a set of learnable latent representations:

[0096] ，

[0097] in, N =128 indicates the number of latent representations, meaning the model can learn 128 latent representations, each representing a different type of information subspace. These latent representations are learned automatically during training and can be viewed as containers of latent information within the model.

[0098] Using latent representation as a query ( Query ) vector, multi-source input features as keys ( Key Vector and value ( Value Vectors are used to perform cross-attention operations to achieve information aggregation:

[0099] ，

[0100] ，

[0101] ，

[0102] in, These are learnable weight matrices, used to generate feature vectors for queries, keys, and values, respectively.

[0103] Calculate the attention weight matrix:

[0104] ,

[0105] in, d =64 represents the feature dimension. This weight matrix indicates the degree of correlation between each potential representation and the satellite general representation or physical variable embedding.

[0106] Update the latent representation based on attention weights:

[0107] ,

[0108] That is, each latent representation is weighted and aggregated from all input features according to the attention distribution to form an updated latent representation. Through parameter updates during the training of the entire base model, a stable latent representation extractor is obtained, ultimately leading to a stable set of latent features.

[0109]

[0110] Each of them This represents the potential features generated after fusing different physical variables and satellite characteristics.

[0111] This set constitutes a unified potential representation that can characterize the coupling relationships and spatiotemporal evolution features of multiple elements in the eco-hydrological system at a high-dimensional semantic level.

[0112] In this embodiment of the invention, for potential characterization A spatiotemporal fusion encoder was designed to further model the spatiotemporal dependencies between potential representations and generate a primary time-series eco-hydrological general representation.

[0113] The spatiotemporal fusion encoder employs an 8-layer stacked structure, with each layer containing two main sub-blocks: a temporal fusion block and a spatial fusion block. A feedforward network and a gated fusion mechanism are connected after each sub-block. Layers are concatenated via residual connections and layer normalization to ensure gradient stability and continuous information transmission. The calculation process is as follows:

[0114] Multi-Head Self-Attention (MHSA) is applied to the latent representation along the time dimension to capture the dependencies and patterns of change between different time steps. The computational form is as follows:

[0115] ,

[0116] in, Indicates the first n A potential representation inT The potential representation of each time step This represents a multi-head self-attention operation along the time dimension (with optional relative position bias to enhance temporal sequence awareness). Through this module, the model can aggregate dynamic information across time steps, thereby identifying temporal evolution patterns in eco-hydrological processes.

[0117] After temporal fusion, multi-head self-attention is applied to the features of each time step along the space (i.e., the latent variable dimension) to model the interactions and cooperative relationships between different latent physical variables. The computational form is as follows:

[0118] ,

[0119] in, Indicates the first t The potential representation of each time step This represents a multi-head self-attention operation along the latent representation dimension. This module enables global dependency modeling among latent physical variables, allowing the model to capture the coupling characteristics between different eco-hydrological elements.

[0120] Each sub-block output is connected to a feedforward network, and the GELU activation function is used to enhance nonlinear expressiveness. To facilitate information exchange between time blocks and variable blocks, a gating fusion mechanism is designed:

[0121] In the l In the layer, a portion of the information from the time block output is passed to the variable block input via a gating function, and vice versa, thereby achieving bidirectional modulation of time features and variable features, which helps reduce feature redundancy and prevent overfitting.

[0122] After 8 layers of spatiotemporal alternation stacking and gating fusion, the encoder outputs: . This invention aims to integrate the primary temporal-series eco-hydrological general representation with the interaction of time dynamics and variables. The primary temporal-series eco-hydrological general representation simultaneously characterizes the temporal evolution and multivariate coupling characteristics of eco-hydrological processes within a unified space, providing semantically complete input representations for subsequent self-supervised learning and downstream tasks.

[0123] The embodiments of the present invention can encode eco-hydrological data in multiple dimensions to generate physical variable embeddings. By using a cross-attention mechanism, potential representations are automatically extracted from satellite general representations and physical variable embedding spaces, and input into a spatiotemporal feature fusion encoder to generate a primary time-series eco-hydrological general representation. This avoids the limitations of relying solely on explicit physical variable coupling, enabling the model to identify key processes from complex spatiotemporal structures and characterize the temporal evolution and multivariate coupling characteristics of eco-hydrological processes.

[0124] In step S104, a basic eco-hydrological model of the irrigation area is established by combining the primary time-series eco-hydrological general representation and the pre-constructed self-supervised learning constraints. Based on the basic eco-hydrological model of the irrigation area and the eco-hydrological observation data of the target irrigation area, a unified eco-hydrological general representation is generated, and the prediction results of downstream eco-hydrological variables are output based on the unified eco-hydrological general representation.

[0125] It is understood that the pre-built self-supervised learning constraints in the embodiments of the present invention may include: conventional self-supervised tasks and customized self-supervised tasks. Conventional self-supervised tasks include: reconstruction tasks, teacher-student representation consistency constraints, and representation anti-collapse constraints. Customized self-supervised tasks include: comparative learning self-supervision based on relevant satellite products and weak Markov self-circulation supervision. The irrigation district eco-hydrological basic model can be understood as a model built based on the primary time-series eco-hydrological general representation for predicting eco-hydrological variables. Its core is to learn the mapping relationship between the representation and the target variable.

[0126] For example, embodiments of the present invention can select satellite products (including but not limited to soil moisture products, evapotranspiration products, leaf area index products, photosynthetically active radiation absorption ratio products, solar induced fluorescence products, etc.) that are related to the target variable based on the physical correlation of time-series eco-hydrological elements. The similarity of the general representation generated based on the similarity constraints of satellite products of different pixels makes the general representation generated by the basic model more closely resemble the eco-hydrological element characteristics in the feature aggregation direction.

[0127] Specifically, by implementing self-supervised learning constraints and temporal physical consistency optimization between the predicted general representation and the actual general representation, and combining conventional and customized self-supervised tasks, the eco-hydrological relevance and evolutionary continuity of the model-generated representation are enhanced. This enables the irrigation district eco-hydrological basic model to learn the evolutionary continuity and causal dependence of the primary temporal eco-hydrological general representation during the training process, thereby embedding the dynamic evolution law of eco-hydrological processes into the generated unified eco-hydrological general representation and constructing the irrigation district eco-hydrological basic model.

[0128] Furthermore, in this embodiment of the invention, the basic model of irrigation area eco-hydrology is used to generate a general eco-hydrological representation, and task-related features are extracted through learnable task queries to drive downstream models to estimate eco-hydrological variables. In the application stage, under the condition of inputting only satellite and driving data, eco-hydrological elements such as evapotranspiration, soil moisture, groundwater level and crop phenotype can be estimated.

[0129] This invention can combine primary time-series eco-hydrological general representations with pre-constructed self-supervised learning constraints to establish a basic eco-hydrological model for irrigation districts. By outputting the predicted eco-hydrological variables through the model and inputting high-quality features from the primary time-series eco-hydrological general representations, the model can quickly learn the evolution patterns of eco-hydrological variables, improve the model's fitting ability, effectively reflect the spatiotemporal variation characteristics of eco-hydrological variables in irrigation districts, and provide scientific decision support for irrigation district water resource allocation, ecological protection, and other work.

[0130] Optionally, in one embodiment of the present invention, a basic model of irrigation district eco-hydrology is established by combining primary time-series eco-hydrological general representations and pre-constructed self-supervised learning constraints, including: determining the self-supervised learning constraints of the model; training the model using the self-supervised learning constraints and primary time-series eco-hydrological general representations until the preset training conditions are met, thereby establishing a basic model of irrigation district eco-hydrology.

[0131] It is understood that the self-supervised learning constraints in the embodiments of the present invention can be understood as training tasks that do not require manual annotation, which can be used to constrain the model training process so that the model output conforms to objective physical laws; the preset training conditions can be that the loss of the model on the validation set reaches a preset threshold or the number of iterations meets the requirements, to ensure that the model is trained sufficiently and does not overfit.

[0132] In actual implementation, the embodiments of the present invention can perform self-supervised learning constraints and temporal physical consistency optimization, and combine conventional and customized self-supervised tasks to enhance the eco-hydrological relevance and evolutionary continuity of the model-generated representations.

[0133] Specifically, the routine self-supervised tasks include mask reconstruction, teacher-student consistency constraints, and representation collapse prevention tasks, with corresponding losses of: , , In addition, it includes two specially designed self-supervised tasks, which are explained in detail below:

[0134] In this embodiment of the invention, in order to preserve the eco-hydrological semantic structure in the high-dimensional space of the shared representation generated by the model, for any two spatiotemporal samples and The similarity between the eco-hydrological representations and satellite eco-hydrological product vectors was calculated and ensured to be consistent. First, the two time-series eco-hydrological representations... and The dimensionality is reduced to two one-dimensional vectors using a trainable two-layer neural network. and Combine the M satellite products (such as SM, ET, LAI, FPAR, SIF, etc., each product being pre-normalized) from the two corresponding spatiotemporal samples into two one-dimensional vectors. and The cosine similarity of a sample pair in the two spaces is defined as:

[0135]

[0136] To achieve "product vector proximity" Similar representations; large differences in product vectors To address the objective of "significant differences in representation," the similarity preservation loss (similarity regression) is constructed as follows:

[0137]

[0138] in, This is the set of sample indices for the current training batch. By minimizing... The model aligns with the "paired geometric relationships" consistent with eco-hydrological products in the representation space, thereby avoiding reliance on binary classification comparisons of hard positive / negative samples and continuously preserving physical semantic similarity.

[0139] To further constrain the physical consistency of the model over time, an autoregressive predictor is set up in the eco-hydrological general representation space. The principle is as follows Figure 2 As shown.

[0140] The predictor is a summary characterization of the previous 7 time steps. Meteorological driving data for the next 7 time steps Given the input, the predicted general representation for the subsequent iterations is:

[0141] ,

[0142] By imposing a consistency constraint on the predicted general representation and the actual general representation, the formula is as follows:

[0143] ,

[0144] The model can learn the temporal evolution continuity and causal dependence of general eco-hydrological representations, thus embedding the dynamic evolution law of eco-hydrological processes into the general representation space. This constraint belongs to the "weak Markov property" paradigm, that is, it does not enforce strict state transition equations, but allows a certain degree of process randomness and environmental disturbances while maintaining temporal dependence, which is more in line with the non-stationary evolution characteristics of eco-hydrological systems in actual arid irrigation areas.

[0145] The total loss is:

[0146] ,

[0147] in, λ 1~ λ5 represents the weighting coefficient for each loss term, used to balance the importance of different self-supervised tasks.

[0148] Through joint training with the above five types of self-supervised constraints, the model achieves eco-hydrological semantic aggregation and temporal physical continuity enhancement in a unified latent space. This enables the generated general representation to reflect the ecological semantics contained in multi-source satellite data and maintain dynamic consistency in the time dimension, providing a reliable foundation for subsequent variable estimation and prediction.

[0149] The embodiments of the present invention can train the model through self-supervised learning constraints, enabling the basic model of irrigation area eco-hydrology to aggregate eco-hydrological features in a general representation and maintain the physical consistency of these features in time and space, thereby improving the accuracy and temporal stability of eco-hydrological element estimation.

[0150] Optionally, in one embodiment of the present invention, the output of eco-hydrological variable prediction results using the irrigation district eco-hydrological basic model includes: generating a unified eco-hydrological general representation based on the irrigation district eco-hydrological basic model and the eco-hydrological observation data of the target irrigation district; generating a general representation space of the target irrigation district based on the unified eco-hydrological general representation; and extracting task features from the general representation space; constructing a downstream supervised learning lightweight model based on the task features, so as to output eco-hydrological variable prediction results using the downstream supervised learning lightweight model.

[0151] It is understood that the eco-hydrological observation data of the target irrigation area in the embodiments of the present invention can be multi-source satellite time-series data and sparse observation data; the task features can be understood as the subset of features selected from them that are most relevant to a specific prediction target; the downstream supervised learning lightweight model can be understood as a supervised learning model with a simple structure and high computational efficiency, which does not require a complex feature extraction process and can be used to quickly map task features to specific prediction results, adapting to the diverse prediction needs of the irrigation area.

[0152] In actual implementation, the embodiments of the present invention can perform task feature extraction and eco-hydrological variable estimation and prediction, generate a general representation space using the trained model, extract task features through a task query mechanism, and realize the estimation and time series prediction of eco-hydrological variables.

[0153] Specifically, after training, the irrigation district eco-hydrological basic model performs forward inference under the condition of inputting only multi-source satellite time-series data and meteorological driving data: First, the time-series satellite general representation and driving data are mapped to a unified dimension according to the preprocessing and encoding process. Then, the general eco-hydrological representation is obtained by sequentially passing through the cross-attention aggregation module and the spatiotemporal feature fusion encoder. This characterization can still be generated even with little or no surface attributes / groundwater observations, and is used to depict the spatiotemporal variations and multi-element coupling characteristics of ecohydrological processes, serving as a unified representation space for downstream tasks.

[0154] For the target eco-hydrological task (such as soil moisture content at different depths, evapotranspiration, groundwater level, crop physiological parameters, etc.), based on the name of the physical variable v and layer depth / level information d Construct the task query header. Preferably, variable name embedding is used. With deep embedding After concatenation, the query vector is obtained through learnable projection. Using the query vector as Query With latent representation set for Key / Value Perform attention retrieval to obtain task-related potential features.

[0155] Using observed samples (measured or reliable products) as supervisory signals, a lightweight regression / classification head (linear layer or shallow MLP) is configured on top of the task features to complete the estimation of target variables; for multivariate, multi-depth or multi-time tasks, multiple sets of query vectors can be generated in parallel for batch retrieval and modeling.

[0156] In the prediction phase, meteorological driving data for future time periods are input into a trained time-series predictor along with current or historical general eco-hydrological representations. Based on historical evolutionary characteristics and future driving conditions, the predictor recursively generates general eco-hydrological representations for future time steps. The model then utilizes a task query mechanism to extract task-related features from the generated general eco-hydrological representations and infers future eco-hydrological variables using a pre-trained downstream model, achieving time-series prediction and dynamic evolution simulation of ecological processes. To improve prediction stability and physical plausibility, the model can employ a rolling recursion strategy in long-term series predictions, combining teacher-mandated and autoregressive approaches to progressively use predicted representations as input for the next step. Simultaneously, physical constraints or smoothing regularization terms can suppress prediction bias and unreasonable fluctuations, ensuring that the prediction results maintain physical continuity and ecological interpretability over time.

[0157] The embodiments of the present invention can achieve spatiotemporal estimation and scenario prediction of eco-hydrological elements without relying on complete ground monitoring or groundwater data during the application stage, relying only on multi-source satellites and meteorological drives. The task query mechanism ensures flexible retrieval and interpretable inference of different variables and depths. The time series prediction driven by the drive enables the model to have forward-looking simulation and trend forecasting capabilities, providing reliable data support for hydrological management and ecological monitoring at the irrigation district scale.

[0158] Specifically, it can be combined with Figure 3 and Figure 4As shown, a specific embodiment of the present invention is used to elaborate in detail the working principle of the multi-source satellite data-driven irrigation district eco-hydrological basic model construction method.

[0159] like Figure 3 As shown, embodiments of the present invention may include the following steps:

[0160] Step S301: Multi-source data acquisition and modeling unit construction.

[0161] In particular, the embodiments of the present invention can unify the spatial scale of multi-source satellite, meteorological and ground data to form a standardized input unit.

[0162] Step S302: Physical model interpolation and eco-hydrological data generation.

[0163] In this embodiment of the invention, a calibrated physical model can be used to fill in missing measurement areas and generate physically consistent eco-hydrological process data.

[0164] Step S303: Multi-source satellite spatiotemporal feature fusion.

[0165] In this embodiment of the invention, multi-source satellite observation information can be integrated to form a continuous and unified time-series satellite general representation.

[0166] Step S304: Multimodal feature encoding and latent representation generation.

[0167] In this embodiment of the invention, physical variables can be encoded in multiple dimensions and fused with satellite general representations, and temporal latent representations can be extracted through cross-attention.

[0168] Step S305: Self-supervised learning constraints and temporal-physical consistency optimization.

[0169] In particular, embodiments of the present invention can combine conventional and customized self-supervised tasks to enhance the eco-hydrological relevance and evolutionary continuity of the model.

[0170] Step S306: Task feature extraction and eco-hydrological variable estimation and prediction.

[0171] In this embodiment of the invention, a trained basic model can be used to generate a general eco-hydrological representation, and to perform task queries and downstream variable estimation and prediction.

[0172] like Figure 4 As shown, the basic model architecture of the irrigation area's eco-hydrology includes (a) a multi-source satellite fusion module, (b) a pre-training module, and (c) a downstream task module.

[0173] Specifically, the multi-source satellite fusion module addresses the differences in spatial resolution, temporal resolution, and observation frequency among various satellite imagery sources. First, it performs temporal registration and spatial alignment of each data source to ensure that multi-source satellite data within the same modeling unit correspond in the temporal dimension and match in the spatial dimension, providing a consistent input foundation for feature fusion. Using different types of satellite data, such as Sentinel, LandSat, and GEDI, as input, the "multi-source satellite fusion encoder" integrates heterogeneous satellite observation information, ultimately generating a unified, universal satellite representation and achieving standardized feature extraction from multi-source satellite data.

[0174] The pre-training module, based on the satellite general representation output by the multi-source satellite fusion module, performs time encoding, location encoding, variable name encoding, and numerical encoding on the driving data, attribute data, and state / flux data, respectively, to generate physical variable embeddings. Simultaneously, feedback constraints from the physical model are introduced, and a spatiotemporal feature fusion encoder completes the spatiotemporal coupling of satellite observation features and eco-hydrological physical features, ultimately generating a general eco-hydrological representation.

[0175] After training, using general eco-hydrological representations as input and learningable task queries, task-related representations are extracted through a cross-attention mechanism, then input into a lightweight task network, and with the help of a small number of labeled supervision signals, the estimation and prediction of eco-hydrological variables are finally completed, realizing the transformation from general representations to specific task outputs.

[0176] Specifically, the basic eco-hydrological model for irrigation districts can perform forward inference with a minimum input of multi-source satellite time-series data and meteorological driving data: First, the general representation of time-series satellite data and driving data are mapped to a unified dimension according to a preprocessing and encoding process. Then, through a cross-attention aggregation module and a spatiotemporal feature fusion encoder, a general eco-hydrological representation corresponding to each modeling unit is obtained. This representation can still be generated even with no or few surface attributes / groundwater observations missing, and is used to characterize the spatiotemporal variations and multi-element coupling features of eco-hydrological processes, serving as a unified representation space for downstream tasks.

[0177] The multi-source satellite data-driven method for constructing an ecological hydrological basic model for irrigation areas, as proposed in this embodiment, can standardize and physically consistent interpolate multi-source information such as multi-satellite multi-temporal images, driving data, attribute data, groundwater observation data, and ecological hydrological related satellite products from multiple satellites in the target irrigation area to generate standardized ecological hydrological data. It embeds and fuses temporal satellite observation features with physical variables of the ecological hydrological data to construct a temporal ecological hydrological representation. Based on this temporal ecological hydrological representation and self-supervised learning constraints, it constructs an ecological hydrological basic model for the irrigation area and outputs prediction results. Through deep fusion of multi-source satellite data, it achieves complementarity and unification of multi-modal observation information such as optical, radar satellite, and ecological hydrological related satellite products, significantly improving the model's robustness and generalization ability under incomplete multi-source data conditions. This, in turn, improves the accuracy and temporal stability of ecological hydrological element estimation, enabling collaborative estimation and dynamic prediction of multiple ecological hydrological elements from the surface to the ground. This solves the problems in related technologies, such as the lack of physical constraints due to reliance on statistical learning and a large number of ground samples, resulting in discontinuous or non-physical results, making it difficult to apply to areas with scarce observations, data mismatch or information bias in multi-source data fusion, difficulty in achieving spatiotemporal fusion consistent with physical processes, and difficulty in achieving high-precision estimation and dynamic prediction of eco-hydrological elements under the condition of limited input data.

[0178] Next, referring to the accompanying drawings, a device for constructing a basic ecological hydrological model of an irrigation area driven by multi-source satellite data according to an embodiment of the present invention is described.

[0179] Figure 5 This is a schematic diagram of the structure of the irrigation district eco-hydrological basic model construction device driven by multi-source satellite data according to an embodiment of the present invention.

[0180] like Figure 5 As shown, the multi-source satellite data-driven irrigation area eco-hydrological basic model construction device 10 includes: acquisition module 100, generation module 200, fusion module 300 and construction module 400.

[0181] The acquisition module 100 is used to acquire multi-satellite multi-temporal image data, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products of the target irrigation area to generate multi-source satellite data.

[0182] The generation module 200 is used to perform spatial reprojection, resampling and resolution unification processing on multi-source satellite data to obtain standardized modeling units. Based on the standardized modeling units, the missing data of the target irrigation area are interpolated using a pre-built eco-hydrological physical model to generate eco-hydrological data of the target irrigation area.

[0183] The fusion module 300 is used to generate a time-series satellite general characterization of the target irrigation area based on multi-source satellite data, and to fuse the time-series satellite general characterization with the physical variable embedding of eco-hydrological data to generate a primary time-series eco-hydrological general characterization.

[0184] Module 400 is used to build a basic eco-hydrological model for irrigation districts by combining the primary time-series eco-hydrological general representation with pre-built self-supervised learning constraints. Based on the basic eco-hydrological model of irrigation districts and the eco-hydrological observation data of the target irrigation district, a unified eco-hydrological general representation is generated, and the prediction results of downstream eco-hydrological variables are output based on the unified eco-hydrological general representation.

[0185] Optionally, in one embodiment of the present invention, the generation module 200 includes: a calibration unit, a supplement unit, and an interpolation unit.

[0186] The calibration unit is used to calibrate the parameters of a pre-constructed eco-hydrological physical model using groundwater observation data, surface water process data, evapotranspiration data, leaf area index data, and aboveground biomass data of the target irrigation area, so as to obtain a calibrated eco-hydrological physical model.

[0187] The supplementary unit is used to run the calibrated eco-hydrological physics model on the standardized modeling unit to generate supplementary data for the target irrigation district.

[0188] The interpolation unit is used to interpolate missing data in the target irrigation area based on supplementary data to generate eco-hydrological data.

[0189] Optionally, in one embodiment of the present invention, the fusion module 300 includes: an encoding unit, a multimodal fusion unit, and a generation unit.

[0190] The coding unit is used to perform multidimensional coding on eco-hydrological data to generate eco-hydrological data physical variable embeddings.

[0191] The multimodal fusion unit is used to fuse satellite general representations with the embedding of physical variables from eco-hydrological data to obtain a multimodal input matrix.

[0192] The characterization extraction unit is used to extract a preset number of latent characterizations based on the multimodal input matrix.

[0193] The generation unit is used to input the potential representation into a preset spatiotemporal feature fusion encoder to generate a primary temporal eco-hydrological general representation.

[0194] Optionally, in one embodiment of the present invention, the construction module 400 includes a determination unit and a training unit.

[0195] Among them, the determining unit is used to determine the self-supervised learning constraints of the model.

[0196] The training unit is used to train the model using self-supervised learning constraints and basic temporal eco-hydrological general representations until the preset training conditions are met, thus establishing a basic eco-hydrological model for the irrigation area.

[0197] Optionally, in one embodiment of the present invention, the construction module 400 includes a feature extraction unit and a prediction unit.

[0198] The feature extraction unit is used to generate a unified eco-hydrological general representation based on the basic eco-hydrological model of the irrigation area and the eco-hydrological observation data of the target irrigation area, generate a general representation space of the target irrigation area based on the unified eco-hydrological general representation, and extract the task features of the general representation space.

[0199] The prediction unit is used to build a lightweight downstream supervised learning model based on task features, and to use the lightweight downstream supervised learning model to output the prediction results of eco-hydrological variables.

[0200] It should be noted that the explanation of the above-mentioned embodiment of the method for constructing an ecological and hydrological basic model of an irrigation area driven by multi-source satellite data also applies to the device for constructing an ecological and hydrological basic model of an irrigation area driven by multi-source satellite data in this embodiment, and will not be repeated here.

[0201] The multi-source satellite data-driven irrigation area eco-hydrological basic model construction device proposed in this embodiment of the invention can standardize and physically consistent interpolate standardized eco-hydrological data by integrating multi-source information such as multi-satellite multi-temporal images, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products from multiple satellites in the target irrigation area. It embeds and fuses temporal satellite observation features with physical variables of eco-hydrological data to construct temporal eco-hydrological representations. Based on the temporal eco-hydrological representations and self-supervised learning constraints, it constructs an irrigation area eco-hydrological basic model and outputs prediction results. By deeply fusing multi-source satellite data, it achieves complementarity and unification of multi-modal observation information such as optical, radar, and ecological products, significantly improving the robustness and generalization ability of the model under incomplete multi-source data conditions. This, in turn, improves the accuracy and temporal stability of eco-hydrological element estimation, and realizes collaborative estimation and dynamic prediction of eco-hydrological multi-element from the surface to the ground. This solves the problems in related technologies, such as the lack of physical constraints due to reliance on statistical learning and a large number of ground samples, resulting in discontinuous or non-physical results, making it difficult to apply to areas with scarce observations, data mismatch or information bias in multi-source data fusion, difficulty in achieving spatiotemporal fusion consistent with physical processes, and difficulty in achieving high-precision estimation and dynamic prediction of eco-hydrological elements under the condition of limited input data.

[0202] Figure 6 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. The electronic device may include:

[0203] The memory 601, the processor 602, and the computer program stored on the memory 601 and capable of running on the processor 602.

[0204] When the processor 602 executes the program, it implements the method for constructing a basic ecological hydrological model of irrigation area driven by multi-source satellite data provided in the above embodiments.

[0205] Furthermore, electronic devices also include:

[0206] Communication interface 603 is used for communication between memory 601 and processor 602.

[0207] The memory 601 is used to store computer programs that can run on the processor 602.

[0208] The memory 601 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0209] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0210] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0211] Processor 602 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0212] This invention also provides a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for constructing an irrigation district eco-hydrological basic model driven by multi-source satellite data.

[0213] This invention also provides a computer program product storing a computer program that, when executed by a processor, implements the above-described method for constructing a basic ecological hydrological model of an irrigation district driven by multi-source satellite data.

[0214] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0215] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0216] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0217] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0218] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0219] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0220] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0221] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for constructing a basic eco-hydrological model of an irrigation district driven by multi-source satellite data, characterized in that, Includes the following steps: The system acquires multi-satellite multi-temporal image data, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products of the target irrigation area to generate multi-source satellite data. The driving data is meteorological driving data, which includes air temperature, precipitation, radiation, wind speed, and humidity. The attribute data is soil attribute data, which includes soil texture, bulk density, organic matter, field water holding capacity, and saturated hydraulic conductivity. The multi-source satellite data is subjected to spatial reprojection, resampling, and resolution unification processing to obtain standardized modeling units. Based on the standardized modeling units, the missing data of the target irrigation area is interpolated using a pre-constructed eco-hydrological physical model to generate eco-hydrological data of the target irrigation area. The time-series satellite representation of the target irrigation area is generated based on the multi-source satellite data, and the physical variable embedding of the time-series satellite representation and the eco-hydrological data is fused to generate a primary time-series eco-hydrological representation. The time-series satellite representation is a feature automatically extracted from the multi-source satellite data by a deep learning network, and the feature is obtained by the joint fusion of multi-source features through a time attention mechanism and a spatial attention mechanism. Combining the primary time-series eco-hydrological representation and pre-constructed self-supervised learning constraints, a basic eco-hydrological model for the irrigation district is established. Based on the basic eco-hydrological model and the eco-hydrological observation data of the target irrigation district, a unified eco-hydrological representation is generated, and downstream eco-hydrological variable prediction results are output based on the unified eco-hydrological representation. The establishment of the basic eco-hydrological model by combining the primary time-series eco-hydrological representation and pre-constructed self-supervised learning constraints includes: determining the self-supervised learning constraints of the model; training the model using the self-supervised learning constraints and the primary time-series eco-hydrological representation until preset training conditions are met, thus establishing the basic eco-hydrological model for the irrigation district. The pre-constructed self-supervised learning constraints include: conventional self-supervised tasks and customized self-supervised tasks. The conventional self-supervised tasks include: reconstruction tasks, teacher-student representation consistency constraints, and representation anti-collapse constraints. The customized self-supervised tasks include: comparative learning self-supervision based on relevant satellite products and weak Markov self-circulation supervision.

2. The method for constructing a basic eco-hydrological model of an irrigation district driven by multi-source satellite data according to claim 1, characterized in that, The process of interpolating missing data from the target irrigation area using a pre-constructed eco-hydrological physical model to generate eco-hydrological data for the target irrigation area includes: The parameters of the pre-constructed eco-hydrological physical model are calibrated using groundwater observation data, surface water process data, evapotranspiration data, leaf area index data, and aboveground biomass data of the target irrigation area to obtain the calibrated eco-hydrological physical model. The calibrated eco-hydrological physical model is run on the standardized modeling unit to generate supplementary data for the target irrigation area; The missing data of the target irrigation area are interpolated based on the supplementary data to generate the eco-hydrological data.

3. The method for constructing a basic eco-hydrological model of an irrigation district driven by multi-source satellite data according to claim 1, characterized in that, The embedding of physical variables that integrates the time-series satellite representation and the eco-hydrological data to generate a primary time-series eco-hydrological representation includes: The eco-hydrological data is multidimensionally encoded to generate the eco-hydrological data physical variable embedding; The satellite representations are fused with the embedded physical variables of the eco-hydrological data to obtain a multimodal input matrix; A predetermined number of latent representations are extracted based on the multimodal input matrix; The potential representation is input into a preset spatiotemporal feature fusion encoder to generate the primary temporal eco-hydrological representation.

4. The method for constructing a basic eco-hydrological model of an irrigation district driven by multi-source satellite data according to claim 1, characterized in that, The process of generating a unified eco-hydrological characterization based on the basic eco-hydrological model of the irrigation district and the eco-hydrological observation data of the target irrigation district, and outputting downstream eco-hydrological variable prediction results based on the unified eco-hydrological characterization, includes: Based on the basic eco-hydrological model of the irrigation area and the eco-hydrological observation data of the target irrigation area, a unified eco-hydrological representation is generated. Based on the unified eco-hydrological representation, a representation space of the target irrigation area is generated, and the task features of the representation space are extracted. A lightweight downstream supervised learning model is constructed based on the task characteristics, and the prediction results of the eco-hydrological variables are output using the lightweight downstream supervised learning model.

5. A device for constructing a basic eco-hydrological model of an irrigation area driven by multi-source satellite data, characterized in that, include: The acquisition module is used to acquire multi-satellite multi-temporal image data, driving data, attribute data, groundwater observation data, and eco-hydrological related satellite products of the target irrigation area to generate multi-source satellite data. The driving data is meteorological driving data, which includes air temperature, precipitation, radiation, wind speed, and humidity. The attribute data is soil attribute data, which includes soil texture, bulk density, organic matter, field water holding capacity, and saturated hydraulic conductivity. The generation module is used to perform spatial reprojection, resampling and resolution unification processing on the multi-source satellite data to obtain standardized modeling units, and based on the standardized modeling units, to interpolate the missing data of the target irrigation area using a pre-built eco-hydrological physical model to generate eco-hydrological data of the target irrigation area. The fusion module is used to generate a time-series satellite representation of the target irrigation area based on the multi-source satellite data, and to fuse the time-series satellite representation and the physical variable embedding of the eco-hydrological data to generate a primary time-series eco-hydrological representation. The time-series satellite representation is a feature obtained by automatically extracting spatial features from the multi-source satellite data through a deep learning network and then fusing the multi-source features through a time attention mechanism and a spatial attention mechanism. A construction module is used to establish a basic eco-hydrological model of the irrigation area by combining the primary time-series eco-hydrological representation and pre-constructed self-supervised learning constraints. Based on the basic eco-hydrological model of the irrigation area and the eco-hydrological observation data of the target irrigation area, a unified eco-hydrological representation is generated, and the prediction results of downstream eco-hydrological variables are output based on the unified eco-hydrological representation. The step of establishing the basic eco-hydrological model of the irrigation area by combining the primary time-series eco-hydrological representation and pre-constructed self-supervised learning constraints includes: determining the self-supervised learning constraints of the model; training the model using the self-supervised learning constraints and the primary time-series eco-hydrological representation until preset training conditions are met, thus establishing the basic eco-hydrological model of the irrigation area. The pre-constructed self-supervised learning constraints include: conventional self-supervised tasks and customized self-supervised tasks. The conventional self-supervised tasks include: reconstruction tasks, teacher-student representation consistency constraints, and representation anti-collapse constraints. The customized self-supervised tasks include: comparative learning self-supervision based on relevant satellite products and weak Markov self-circulation supervision.

6. The device for constructing a basic eco-hydrological model of an irrigation area driven by multi-source satellite data according to claim 5, characterized in that, The generation module includes: The calibration unit is used to calibrate the parameters of the pre-constructed eco-hydrological physical model using groundwater observation data, surface water process data, evapotranspiration data, leaf area index data and aboveground biomass data of the target irrigation area, so as to obtain the calibrated eco-hydrological physical model. A supplementary unit is used to run the calibrated eco-hydrological physical model on the standardized modeling unit to generate supplementary data for the target irrigation area; An interpolation unit is used to interpolate the missing data of the target irrigation area based on the supplementary data to generate the eco-hydrological data.

7. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for constructing a basic ecological hydrological model of an irrigation district driven by multi-source satellite data as described in any one of claims 1-5.

8. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method for constructing a basic ecological hydrological model of irrigation area driven by multi-source satellite data as described in any one of claims 1-4.

9. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement the method for constructing an irrigation district eco-hydrological basic model driven by multi-source satellite data as described in any one of claims 1-4.

Images