A Method and System for Estimating Evapotranspiration in Rice Irrigation Areas Based on Multi-Source Remote Sensing and Tower-Based Data Constraints

By using multi-source remote sensing and base data constraints, mixed pixels are decomposed and the evapotranspiration estimation of rice irrigation areas is corrected. This solves the problem that the physiological state of a single crop is difficult to reflect under complex terrain, and improves the accuracy and adaptability of the estimation.

CN123115708APending Publication Date: 2026-11-17NANJING HYDRAULIC RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611626194.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing regional evapotranspiration estimation methods are difficult to accurately reflect the physiological state of a single crop under complex land cover, and the uniform physical model parameters are difficult to adapt to rapidly changing phenological periods, resulting in overfitting when the model is applied across regions and years.

Method used

By acquiring multi-source remote sensing data and tower base data, the mixed pixels are decomposed, fluorescent endmembers per unit area for different rice cultivation types are extracted, and combined with vegetation parameters and meteorological data, the total primary productivity is corrected using a residual mapping model. A carbon-water coupling model is constructed to estimate canopy transpiration and field surface evaporation, and the total evapotranspiration of paddy fields is estimated.

Benefits of technology

It reduces the systematic bias caused by uniform model parameters, improves the accuracy of water consumption estimation in complex rice-growing areas, adapts to rapid phenological changes, and reduces the risk of overfitting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115708A_ABST
    Figure CN123115708A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for estimating evapotranspiration in rice-growing irrigation areas based on multi-source remote sensing and data constraints. The method acquires fluorescent mixed pixels, vegetation parameters, meteorological data, and rice planting structure of the target irrigation area; decomposes the mixed pixels based on the planting structure to extract fluorescent endmembers per unit area for different rice cultivation types; calculates the continuous total primary productivity baseline for the region based on vegetation parameters and meteorological data; extracts temporal state characteristics and uses a residual mapping model to predict regional biases, compensating them to the baseline to obtain corrected total primary productivity; inputs the corrected results into a carbon-water coupling model to estimate canopy transpiration, and combines this with independently estimated field evaporation to synthesize total evapotranspiration of the paddy field. This invention can reduce the transmission of systemic biases caused by uniform model parameters and improve the accuracy of water consumption estimation in complex rice-growing areas.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing and farmland moisture process monitoring technology, and in particular to a method and system for estimating evapotranspiration in rice irrigation areas based on multi-source remote sensing and data constraints. Background Technology

[0002] In complex agricultural ecosystems such as rice-irrigated areas in southern China, single-season rice, double-season rice, and non-rice background features are interspersed, with rapid phenological transitions and frequent changes in field surface moisture. Evapotranspiration is one of the key indicators characterizing farmland water consumption and agricultural water resource utilization. Accurately estimating evapotranspiration under complex surface conditions is of great significance for analyzing water consumption patterns during crop growth stages and optimizing regional water and carbon cycle models.

[0003] Existing methods for estimating regional evapotranspiration typically rely on surface energy balance models or solar energy utilization process models. These methods utilize low- to medium-resolution satellite remote sensing vegetation parameters combined with meteorological driving data, employing a relatively uniform parameterization scheme for large-scale continuous calculations. Meanwhile, some data-driven methods utilize multi-source remote sensing observation variables, using limited carbon and water flux data provided by ground eddy covariance systems as absolute value labels to establish machine learning supervised mapping models to fit regional absolute fluxes.

[0004] However, under complex land cover, coarse-resolution remote sensing pixels contain a mixture of crops with different canopy structures, making it difficult for the input signal to accurately reflect the physiological state of a single crop. At the same time, uniform physical model parameters are difficult to adapt to rapidly changing phenological periods. Directly using a small number of ground flux samples to forcibly fit high-dimensional remote sensing features can easily lead to overfitting when the model is applied across regions and years. Summary of the Invention

[0005] Purpose of the invention: To provide a method and system for estimating evapotranspiration in rice-growing irrigation areas based on multi-source remote sensing and tower-based data constraints, in order to solve the above-mentioned problems in the prior art.

[0006] Technical solution: Firstly, a method for estimating evapotranspiration in rice-irrigated areas based on multi-source remote sensing and data constraints, comprising: Acquire sunlight-induced chlorophyll-fluorescence mixed pixels, vegetation parameters, meteorological data, and rice planting structure information of the target irrigation area; Based on rice planting structure information, the solar-induced chlorophyll fluorescence mixed pixels were decomposed, and the unit area fluorescence endmembers of different rice planting types were extracted. Calculate regional continuous total primary productivity benchmarks based on vegetation parameters and meteorological data for different rice cultivation types; Based on fluorescent endmembers per unit area, vegetation parameters and meteorological data, a time-series state characterization is constructed. A pre-trained residual mapping model is used to predict regional bias based on the time-series state characterization. The regional bias is then compensated to the regional continuous total primary productivity benchmark to obtain the corrected total primary productivity. The corrected total primary productivity was used as a physiological driver input into a pre-defined carbon-water coupling model to estimate canopy transpiration. Evaporation on the field surface is estimated based on vegetation parameters and meteorological data. The evaporation on the field surface is combined with the canopy transpiration to obtain the total evapotranspiration of paddy fields at the field scale in the irrigation area.

[0007] The second aspect is a rice irrigation area evapotranspiration estimation system based on multi-source remote sensing and tower-based data constraints. The system includes a processor coupled with a memory. The memory stores instructions, which are executed by the processor to implement any one of the methods in the first aspect.

[0008] Beneficial effects: It reduces the transmission of systematic bias caused by uniform model parameters and improves the accuracy of water consumption estimation in complex rice-growing areas. Attached Figure Description

[0009] Figure 1 This is a flowchart of the method for estimating evapotranspiration in rice-growing irrigation areas based on multi-source remote sensing and tower-based data constraints, as proposed in this invention.

[0010] Figure 2 This is a flowchart of the process for extracting unit area fluorescent endmembers for different rice cultivation types in this invention.

[0011] Figure 3 This is a flowchart of the pre-trained decomposition model obtained in this invention.

[0012] Figure 4 This is a flowchart illustrating the construction of artificial mixed samples and the contribution of real fluorescence in this invention.

[0013] Figure 5 This is a comparison and verification result diagram of the calibrated GPP and the tower-based EC-GPP in this invention.

[0014] Figure 6 This is a comparison and verification result between the total evapotranspiration estimation results of the model paddy field in this invention and the ET observed at the tower base. Detailed Implementation

[0015] In order to solve the problems mentioned in the background art, combined with Figures 1 to 6 The present invention will be specifically described through the following embodiments.

[0016] Example 1 According to one aspect of this application, this embodiment provides a method for estimating evapotranspiration in rice-irrigated areas based on multi-source remote sensing and tower-based data constraints. The method includes the following steps: Step 101: Obtain the solar-induced chlorophyll fluorescence mixed pixels, vegetation parameters, meteorological data, and rice planting structure information of the target irrigation area.

[0017] In agricultural ecosystems such as rice-irrigated areas in southern China, single-season rice, double-season rice, and non-rice background features are typically interspersed. To describe the water consumption in this region, this embodiment establishes a unified spatiotemporal computation grid. Specifically, the grid containing the low-spatial-resolution solar-induced chlorophyll-fluorescence mixing pixels is used as the baseline template. Rice planting structure information and vegetation parameters extracted from high-spatial-resolution satellite remote sensing images, as well as meteorological data obtained from the reanalysis dataset, are aligned in terms of spatial scale and temporal step for subsequent pixel decomposition and physical process simulation.

[0018] Step 102: Based on rice planting structure information, decompose the solar-induced chlorophyll fluorescence mixed pixels and extract the unit area fluorescence endmembers for different rice planting types.

[0019] Because the mixed pixels in remote sensing images are typically larger than the area of ​​a single farmland plot, a single pixel may contain multiple land features with different phenological characteristics and canopy structures. If unprocessed mixed image data is used, it will be difficult to reflect the true photosynthetic physiological state of a single rice species.

[0020] In this embodiment, by combining the area proportions of various land features within a pixel, a pre-constructed decomposition network is used to strip away the mixed signal, obtaining fluorescent endmembers per unit area. Here, the fluorescent endmembers refer to remotely sensed sunlight-induced chlorophyll fluorescence endmembers, i.e., SIF endmembers. This process weakens the interference of non-target land features on the physiological signals of the target rice, obtaining pure physiological characterization information consistent with the phenological rhythms of single-season or double-season rice.

[0021] Step 103: Calculate the regional continuous total primary productivity baseline based on vegetation parameters and meteorological data for different rice cultivation types.

[0022] Specifically, after obtaining pure physiological information for different rice cultivation types, a reference benchmark with clear physical meaning and continuous spatial coverage is constructed. Using a conventional light energy utilization process model, combined with the photosynthetically active absorption ratio represented by vegetation indices, and environmental stress factors such as temperature and water, the corresponding productivity estimates are calculated pixel by pixel.

[0023] Step 104: Extract time-series state characteristics based on unit area fluorescence endmembers, vegetation parameters and meteorological data. Use a pre-trained residual mapping model to predict regional bias based on time-series state characteristics. Compensate the regional bias to the regional continuous total primary productivity benchmark to obtain corrected total primary productivity.

[0024] This process model provides continuous computational capabilities at a regional scale. To meet the demands of large-scale computation, the model typically employs a relatively uniform parameterization scheme, leading to systematic biases in the output results during the rapid phenological transition phase of rice. This embodiment utilizes a constraint mechanism that integrates data-driven and mechanistic models. In the time series, fluorescent endmembers per unit area are combined with corresponding meteorological and vegetation environmental states, and compressed into a low-dimensional temporal state representation through a feature extraction network. Furthermore, a residual mapping model, pre-trained based on limited high-quality ground-based observation data, is used to predict the amount of error to be compensated in the mechanistic model under the current temporal state representation. The predicted regional bias is then superimposed onto the regional continuous total primary productivity baseline. This approach reduces the risk of overfitting caused by fitting a high-dimensional absolute flux with a small number of ground samples, ensuring the spatial continuity of the process model in the regional estimation results and achieving dynamic bias compensation for different physiological and environmental states.

[0025] Step 105: Input the corrected total primary productivity as a physiological driver into the preset carbon-water coupling model to estimate canopy transpiration.

[0026] During the absorption of carbon by vegetation through photosynthesis, water loss to the atmosphere is usually accompanied by the opening of stomata. Based on this principle, a carbon-water coupling model calculates vegetation water consumption by establishing a coupling relationship between stomatal conductance and photosynthesis. By substituting the error-compensated corrected total primary productivity into this coupling relationship, the estimation bias of the upstream mechanism model during key growth stages is prevented from being transmitted to the water consumption calculation stage, thereby improving the accuracy of canopy transpiration estimation.

[0027] Step 106: Estimate field surface evaporation based on vegetation parameters and meteorological data, and combine field surface evaporation with canopy transpiration to obtain total evapotranspiration of paddy fields at the irrigation area plot scale.

[0028] In this embodiment, the total water consumption of the paddy field ecosystem comprises two parts: vegetation-driven transpiration and physical-environment-driven surface evaporation. Surface evaporation is primarily controlled by the available net radiative energy reaching the surface, the degree of crop canopy obstruction, and the wet / dry constraints between the surface and the near-surface atmosphere. By independently estimating this physical evaporation, it is summed with physiologically driven canopy transpiration to form a complete dual-source evapotranspiration estimation framework. This framework allows canopy water consumption and non-canopy water consumption to be constrained by corresponding state information, generating regional continuous data products suitable for water resource allocation and irrigation demand assessment in complex paddy irrigation areas.

[0029] Example 2 Based on the above embodiments, this embodiment describes the spatiotemporal unification and parameter extraction process of multi-source data. In one possible implementation, acquiring meteorological data for the target irrigation area includes the following steps: Step 201: Obtain raw meteorological grid data.

[0030] Specifically, a meteorological reanalysis dataset covering the target irrigation area and its surrounding buffer zone is acquired. The required meteorological variables, such as air temperature, dew point temperature, surface air pressure, surface downward shortwave radiation, and net radiation component, are extracted from the meteorological reanalysis dataset. Meteorological conversion methods, such as the Tetens empirical formula, are used to calculate the saturated vapor pressure, actual vapor pressure, and vapor pressure deficit. The physical dimensions of each meteorological variable are then converted to a pre-defined unified standard.

[0031] Step 202: Using the target pixels within the target irrigation area as a reference, determine the nearest effective meteorological grid points.

[0032] Specifically, a unified spatial computing grid is established, and the spatial grid of the solar-induced chlorophyll-fluorescence mixed pixels is used as the target spatial template. The geometric center of each target pixel in the target spatial template is used as the spatial interpolation reference position, and several effective meteorological sampling grid points closest to the geometric center of the target pixels are selected from the original meteorological grid data.

[0033] Step 203: Calculate the distance weight based on the spatial distance between the target pixel and the effective meteorological grid points.

[0034] Accordingly, the Euclidean distance between the geometric center of the target pixel and each effective meteorological sampling grid point is calculated. The inverse distance weighting method is used to determine the spatial interpolation weights, i.e., a non-linear negative correlation is established between the distance weights and the square of the Euclidean distance. When the geometric center of the target pixel coincides with the position of a certain effective meteorological sampling grid point, the data corresponding to that grid point is taken as the meteorological data for that target pixel.

[0035] Step 204: Use distance weights to perform spatial interpolation on the original meteorological grid data to obtain meteorological data that is consistent with the spatial grid of the target pixel.

[0036] Specifically, the values ​​of various meteorological variables at the effective meteorological sampling grid points are obtained, multiplied by their corresponding distance weights and summed, and then divided by the sum of the distance weights involved in the calculation. This process downscales the coarser resolution original meteorological grid data and maps it to the target spatial template, outputting meteorological data consistent with the target pixel spatial grid.

[0037] This embodiment also specifies the acquisition and processing procedures for vegetation parameters. Multispectral satellite remote sensing products are acquired, and surface reflectance data for each band are read. When performing data quality control based on scene classification information, clouds, cloud shadows, and invalid data pixels are removed, while vegetation, bare land, and water body category pixels are retained. Specifically, water body category pixels are retained to prevent incorrect removal by conventional filtering algorithms during the paddy field flooding and transplanting periods, when the surface appears wet or has water surface characteristics, thus ensuring the integrity of observation data in the early stages of rice growth.

[0038] Next, based on the quality-controlled surface reflectance data, the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), and Land Surface Moisture Index (LDMA) are calculated using existing vegetation index calculation rules. In the temporal dimension, multi-period satellite remote sensing products are time-series aggregated using a preset time window as the step size. For multiple valid observations within the same time window, a representative value for that time window is generated by calculating the median of pixels. For missing phases in the time series, an initial value sequence is generated by linear interpolation using adjacent valid phases. Subsequently, a second-order difference weighted smoothing algorithm is used to denoise and smooth this initial value sequence, outputting smoothed multi-temporal vegetation parameters.

[0039] Furthermore, for the leaf area index in vegetation parameters, instead of using single-band surface reflectance for algebraic calculation, the complete organizational structure of the satellite remote sensing product is preserved and input into the preset biophysical parameter inversion module for analytical inversion, extracting leaf area index data that matches the target spatial resolution.

[0040] As an optional implementation method, when there are densely distributed ground meteorological observation stations inside and around the target irrigation area, the original meteorological grid data can be replaced with the measured meteorological data of the ground meteorological observation stations, and combined with the Kriging spatial interpolation algorithm, meteorological data consistent with the target pixel spatial grid can be generated.

[0041] Example 3 Based on Example 1, this example uses the online inference and demixing process of mixed pixels based on phenological priors to solve the problem that coarse-resolution remote sensing signals are difficult to characterize the physiological state of a single crop under complex rice cultivation systems.

[0042] In one possible implementation, extracting fluorescent endmembers per unit area for different rice crop types includes the following steps: Step 301: Based on rice planting structure information, obtain the area ratio of each component in the solar-induced chlorophyll fluorescence mixed pixel.

[0043] In this embodiment, high-resolution rice planting structure data matching the target spatial grid is acquired. Within each target cell, the surface components are divided into single-season rice, double-season rice, and background components. The background component is used to uniformly represent other non-rice land cover within the target cell, excluding single-season and double-season rice, to account for their contribution to the fluorescence signal of the mixed cell. The area proportions of the three components within the target cell are statistically analyzed. When spatial resampling or classification aggregation causes the sum of the area proportions of the three components to deviate from 1, normalization is performed on the area proportions of single-season rice, double-season rice, and background components to ensure that the sum of the area proportions of the three components is always equal to 1.

[0044] As an alternative implementation, if computing resources allow and there is sufficient high-resolution land cover classification data, the background components can be further subdivided into independent sub-components such as water bodies, dryland crops, and forest land, and the area ratio of each sub-component can be calculated separately to adapt to complex land cover scenarios.

[0045] Step 302: Input the solar-induced chlorophyll fluorescence mixed pixel, the area ratio of each component, and the auxiliary environmental variables corresponding to the mixed pixel into the pre-trained decomposition model to obtain the predicted fluorescence contribution of each component.

[0046] In this embodiment, a multidimensional input feature vector is constructed for model inference. Specifically, observed values ​​of sunlight-induced chlorophyll-fluorescence mixed pixels, the area proportion of single-season rice, the area proportion of double-season rice, the area proportion of background components, the total area proportion of rice, time-cycle encoded variables, and variables such as vegetation indices and photosynthetically active radiation for different rice crop types are extracted. These variables are combined and concatenated to form an input vector containing multidimensional features. The median imputation method is used to process missing values ​​in the input vector, and the Z-score method is used for standardization. In this embodiment, auxiliary environmental variables include at least one of the following: total rice area proportion, time-cycle encoded variables, vegetation indices for different rice crop types, and photosynthetically active radiation.

[0047] Next, the standardized input vector is fed into the pre-trained decomposition model. The decomposition model adopts a multilayer perceptron (MLP) structure. Through the nonlinear forward propagation network of the decomposition model, mapping calculations are performed, and the preliminary predicted values ​​of the contribution of single-season rice to the fluorescence of mixed pixels, the preliminary predicted values ​​of the contribution of double-season rice to the fluorescence of mixed pixels, and the preliminary predicted values ​​of the contribution of background components to the fluorescence of mixed pixels are extracted at the output end.

[0048] Step 303: Extract the relative activity of vegetation index based on vegetation parameters, and use the relative activity of vegetation index to impose phenological prior constraints on the fluorescence contribution to obtain the adjusted fluorescence contribution.

[0049] To prevent purely data-driven neural network outputs from violating the physical laws of crop growth, this embodiment introduces phenological prior information based on vegetation parameters. Within the corresponding annual cycle, high values ​​of vegetation indices for single-season and double-season rice within the target pixel are extracted, such as high values ​​of the Annual Enhanced Vegetation Index (EVI) or the Normalized Difference Vegetation Index (NDVI). The ratio of the vegetation index at the current time phase to the high value of the annual vegetation index is calculated to characterize the relative activity of vegetation indices in the target rice canopy. The relative activity of vegetation indices refers to the relative annual level of vegetation indices for different rice crop types.

[0050] Next, a soft conservation constraint is applied to the preliminary contribution prediction value using the relative activity of vegetation indices. A nonlinear mapping relationship is established between relative activity and contribution penalty weights. When the relative activity of the target rice type is in a low-activity state, the retention weight of the corresponding preliminary contribution prediction value is reduced to suppress abnormal fluorescence contribution predictions generated by the target crop in the low-activity phase; when the relative activity is in a high-activity state, the original preliminary contribution prediction value is retained. Through this phenological prior constraint processing, an adjusted fluorescence contribution value that conforms to the crop growth rhythm is output.

[0051] Considering that this embodiment extracts physiological signals from rice components, after completing the phenological constraints on rice components, the residual amount between the observed fluorescence value of the mixed pixel and the fluorescence contribution amount of each rice component after phenological constraints is used to determine the closing fluorescence contribution amount of the background component. The closing fluorescence contribution amount of the background component is mainly used to maintain the closure of the mixed pixel signal and is not used as the main output of the rice unit area fluorescence endmember in this embodiment.

[0052] Step 304: Calculate the fluorescent endmembers per unit area for different rice cultivation types based on the adjusted fluorescence contribution and the corresponding area ratio.

[0053] Specifically, the adjusted fluorescence contribution of single-season and double-season rice is obtained. A lower limit threshold for area proportion is set. When the area proportion of the target rice crop type is lower than the lower limit threshold, the endmember of that phase is marked as invalid to prevent the calculation results from diverging due to a very small divisor. When the area proportion of the target rice crop type is greater than or equal to the lower limit threshold, the ratio of the adjusted fluorescence contribution to the corresponding area proportion is calculated to obtain the fluorescence endmember per unit area for that rice crop type. The corresponding calculation formula is as follows: ; in, Indicates the first Fluorescent endmembers per unit area in rice-like crops This indicates the first phenological event after processing with prior phenological constraints. Adjusted fluorescence contribution for rice-like crop types Indicates the first The proportion of rice-like crops in the area.

[0054] Next, the unit area fluorescence endmember time series of rice crop types that match the target spatiotemporal grid are output as information inputs characterizing the physiological states of single-season rice and double-season rice, and then passed to the subsequent regional total primary productivity benchmark correction process.

[0055] Example 4 To address the problem that the fluorescence contribution of each component within a real coarse-resolution mixed pixel lacks a true label, making it difficult to directly train a supervised model, this embodiment provides a mechanism for constructing artificial mixed samples with true contribution labels, and performs supervised training and spatial generalization verification of the neural network based on this mechanism. In one possible implementation, the process of constructing artificial mixed samples and true fluorescence contributions includes the following steps: Step 401: Select high-purity candidate pixels from historical mixed pixels based on historical area ratio, and calculate the initial value of candidate fluorescent endmembers.

[0056] In this embodiment, area ratio thresholds for various components are set. When the area ratio of a certain component is greater than a preset high purity threshold, and the area ratios of other target components are less than preset mixing thresholds, it is selected as a high purity candidate pixel for the corresponding component. For example, if the area ratio of double-cropping rice is greater than 0.7 and the area ratio of single-cropping rice is less than 0.15, the pixel in question is selected as a high purity double-cropping rice candidate pixel.

[0057] Next, the mixed fluorescence observation value of the candidate pixel is obtained, and the ratio of the mixed fluorescence observation value to the area ratio of the component is calculated. A small constant is added to the denominator to prevent division by zero errors, thus obtaining the initial value of the candidate fluorescence endmember. The quantile truncation method is used to control outliers in the initial value of the candidate fluorescence endmember, limiting it to a preset physical range.

[0058] Step 402: Extract the relative activity of historical vegetation index using historical vegetation parameters, and implement a soft gating mechanism on the initial value of candidate fluorescent endmembers using the relative activity of historical vegetation index to remove abnormal endmembers in low-activity states and obtain effective candidate endmembers.

[0059] In this embodiment, the high annual vegetation index value at the location of the candidate pixel is extracted and used as a reference benchmark. The ratio of the current vegetation index to this reference benchmark is calculated to obtain the historical vegetation index relative activity. Introducing the historical vegetation index relative activity, a soft-gating mechanism is applied to the initial values ​​of candidate fluorescent endmembers, thereby establishing a positive correlation between relative activity and the effective strength of candidate endmembers as training labels. When the target crop is in a low-activity state, such as in the early transplanting stage or late harvesting stage, the weight of the corresponding candidate endmembers as training labels is reduced; when the vegetation index is in a high-activity state, the original candidate endmember values ​​are retained as the main training information. Through this mechanism, endmember contamination caused by high area proportions but actually in an inactive period is suppressed, and effective candidate endmembers that conform to physical laws are extracted.

[0060] Step 403: Randomly generate non-negative area weights that satisfy the summation constraint as the simulated area ratio.

[0061] Specifically, effective candidate endmember sample libraries for each type of component are established according to year and time step. Endmember combinations are randomly selected from the sample libraries of single-season rice, double-season rice, and background components in the same time phase. For each selection, a set of non-negative values ​​is randomly generated as the simulated area weight. This set of values ​​is normalized, and the sum of the simulated area weights for single-season rice, double-season rice, and background components is constrained to always be equal to 1.

[0062] Step 404: Based on the non-negative area weights and the corresponding valid candidate endmembers, construct artificial mixed fluorescence and the contribution of real fluorescence; obtain vegetation parameters and meteorological data corresponding to the time phase of artificial mixed fluorescence as auxiliary environmental variables, and combine artificial mixed fluorescence, simulated area ratio and auxiliary environmental variables to form artificial mixed sample.

[0063] In this embodiment, the effective candidate endmembers of double-cropping rice, single-cropping rice, and background components are multiplied by their respective non-negative area weights to obtain the true fluorescence contribution of each component. The artificial mixed fluorescence is calculated by summing the true fluorescence contributions of each component, using the following formula: ; in, This indicates the artificially synthesized mixed fluorescence. This represents the simulated non-negative area weights for randomly generated double-cropping rice. This indicates effective candidate endmembers for double-cropping rice. This represents the non-negative area weights for randomly generated single-season rice simulations. This indicates the effective candidate endmembers for single-season rice. This indicates that the randomly generated background components simulate non-negative area weights. Indicates valid candidate endmembers for background components.

[0064] Furthermore, the corresponding tag generation formula for the actual fluorescence contribution of each component is as follows: ; in, Indicates the first The true fluorescence contribution of the component class Indicates the first Simulated non-negative area weights for class components Indicates the first Valid candidate endmembers for class components.

[0065] In another alternative embodiment, the process of artificially mixing samples and constructing the actual fluorescence contribution includes: Based on the area ratio of each component in the historical fluorescence mixed pixels, high-purity pixels of single-season rice, double-season rice and background components were screened respectively; based on the fluorescence observation value of high-purity pixels and the area ratio of the corresponding components, the SIF endmember value per unit area of ​​each component was calculated, and outlier control was performed on the fluorescence endmember value per unit area. Based on the time series of vegetation indices for different rice crop types, the relative status of the current vegetation index with respect to the annual reference level is determined. According to the relative status, the initial value of the fluorescence endmember per unit area for the corresponding rice crop type is subjected to soft gating to reduce the influence of endmember samples with lower vegetation index levels on the construction of training samples, thereby obtaining effective fluorescence endmember samples per unit area for model training. Establish unit area fluorescence endmember sample libraries for each component according to year and time phase, and extract endmember samples from the corresponding sample libraries of each component in the same year and time. Generate non-negative area weights that satisfy the summation constraint as simulated area ratios. Construct artificial mixed fluorescence based on endmember samples and corresponding simulated area ratios, and calculate the fluorescence contribution of each component.

[0066] In a further embodiment, the decomposition model is pre-trained through the following steps: Step 405: Obtain artificially mixed samples, which include artificially mixed fluorescence, simulated area ratio, and auxiliary environmental variables; obtain the actual fluorescence contribution corresponding to the artificially mixed samples.

[0067] Artificially mixed fluorescence, simulated area ratio, and features such as vegetation parameters, photosynthetically active radiation, and time period encoding are combined to construct a multi-dimensional input feature vector, forming artificially mixed samples. Correspondingly, the actual fluorescence contributions of single-season rice, double-season rice, and background components are used as supervision labels for the corresponding artificially mixed samples.

[0068] Step 406: Using artificially mixed samples as input and the corresponding real fluorescence contribution as supervision labels, supervise the training of the initial neural network to obtain the decomposition model.

[0069] Specifically, a multilayer perceptron with multiple hidden layers is constructed as the initial neural network. Median imputation is used to handle missing values ​​in the input artificial mixture sample features, and Z-score standardization is performed. The processed artificial mixture samples are then input into the initial neural network, outputting the predicted fluorescence contribution of each component. The mean squared error loss function is used to calculate the loss between the predicted and actual fluorescence contributions. Based on this loss value, the adaptive moment estimation (Adam) algorithm is used to update the network parameters. During training, L2 regularization constraints are introduced to limit weight fluctuations, and an early stopping mechanism is enabled based on the loss changes observed in the internal validation samples to obtain a pre-trained decomposition model.

[0070] Furthermore, the supervised training process for the initial neural network also includes: Step 4061: Divide the target irrigation area into multiple spatial blocks; reserve some spatial blocks as independent spatial reserved areas; use artificial mixed samples outside the independent spatial reserved areas to update parameters; use artificial mixed samples within the independent spatial reserved areas to verify the spatial applicability of the decomposition model and obtain the spatial verification error index; when the spatial verification error index meets the preset generalization performance conditions, confirm that the decomposition model training is complete.

[0071] In another possible implementation, the process of artificially mixing samples and constructing the actual fluorescence contribution includes the following steps: Obtain fluorescent endmember samples per unit area for each component and generate simulated area ratios for each component that satisfy the total area ratio constraint; construct artificial mixed samples based on the fluorescent endmember samples per unit area and the corresponding simulated area ratios and determine the fluorescence contribution of each component; combine the artificial mixed fluorescence, the simulated area ratios of each component, and auxiliary environmental variables to form training samples; use the training samples as input and the fluorescence contribution of each component as supervision labels to train the initial neural network and obtain the decomposition model.

[0072] In terms of geospatial dimension, the study area is divided into multiple regular spatial blocks. Spatial blocks with a predetermined proportion are randomly selected and designated as independent spatial reserve areas. During parameter updates, only artificially mixed samples outside these independent spatial reserve areas are used. After model training, the artificially mixed samples within the independent spatial reserve areas are input into the decomposed model to calculate the prediction error, thereby verifying the model's applicability and stability in different spatial locations. This spatial independence verification mechanism is independent of conventional random sample partitioning, enhancing the reliability of the model's cross-regional generalization.

[0073] This embodiment constructs artificial mixed samples and introduces a soft-gating mechanism for vegetation activity to train the decomposition network. It then performs fine decomposition on coarse-resolution fluorescent mixed pixels, reduces background noise from non-target ground features, and extracts physiological information consistent with the phenological rhythm of the preset rice crop type.

[0074] Example 5 Based on Example 1, this example constructs a spatially continuous reference benchmark for total primary productivity with clear physical meaning, based on the light energy utilization process model.

[0075] In one possible implementation, a regional continuous total primary productivity baseline is calculated based on vegetation parameters and meteorological data, including the following steps: Step 501: Obtain the vegetation parameters corresponding to single-season rice and double-season rice respectively, and extract the photosynthetically effective vegetation absorption characterization quantity and water state characterization quantity from the vegetation parameters.

[0076] Multi-temporal vegetation parameters for different rice cultivation types were obtained after spatiotemporal unification and smoothing. An enhanced vegetation index was extracted and used as a measure of photosynthetically active vegetation absorption, reflecting the absorption capacity of the rice canopy for photosynthetically active radiation. A land surface moisture index was extracted as a measure of water status, reflecting changes in the moisture content of the vegetation canopy and background soil.

[0077] Step 502: Extract photosynthetically active radiation data and temperature data from meteorological data.

[0078] From the spatially interpolated meteorological dataset, photosynthetically active radiation (PANR) data and near-surface air temperature data, matched with the target pixels and corresponding time steps, are extracted. PANR data characterizes the effective photosynthetic energy input reaching the vegetation canopy, while air temperature data is used to subsequently assess the degree to which environmental heat conditions limit the activity of rice photosynthetic enzymes.

[0079] Step 503: Calculate the temperature regulation factor based on the air temperature data, and calculate the moisture regulation factor based on the moisture state characterization quantity.

[0080] In this embodiment, preset physiological parameters such as the minimum, maximum, and optimum temperatures required for rice photosynthesis are obtained. Based on the extracted air temperature data and preset physiological parameters, a temperature regulation factor is calculated using a conventional quadratic function relationship of temperature stress to describe the inhibitory effect on photosynthesis when the current air temperature deviates from the optimum temperature.

[0081] Furthermore, the highest time-series values ​​of the water state characterization (land surface water index) for the target pixels during the rice growing season are used as the reference value for the maximum water index. A ratio mapping relationship is constructed using the water state characterization at the current time step and the reference value for the maximum water index to calculate the water regulation factor, which describes the degree to which the current water deficit restricts stomatal conductance and photosynthetic capacity.

[0082] In one alternative implementation, the phenological adjustment factor can be set to a constant of 1, so that the annual changes in photosynthetic growth of rice are mainly driven by the coordinated dynamic changes in photosynthetically effective vegetation absorption, water state, air temperature, and photosynthetically effective radiation.

[0083] In another alternative implementation, a segmented phenological regulation function based on key phenological nodes can be introduced, so that the changes in photosynthetic growth of rice throughout the year are mainly driven by the coordinated dynamic changes of photosynthetically effective vegetation absorption, water state, air temperature data, and photosynthetically effective radiation data.

[0084] Step 504: Calculate the regional continuous total primary productivity baseline for rice cultivation type based on the preset maximum light energy utilization rate, photosynthetically effective vegetation absorption characterization, photosynthetically effective radiation, temperature regulation factor, and water regulation factor.

[0085] In this embodiment, maximum light energy utilization parameters are set for single-season and double-season rice. A conventional vegetation photosynthetic process model, such as the vegetation photosynthesis (VPM) model, is used. The preset maximum light energy utilization, photosynthetically effective vegetation absorption, photosynthetically effective radiation data, temperature regulation factor, and water regulation factor are multiplied together to calculate the total primary productivity of the target pixel at the current time step. The value of the maximum light energy utilization parameter can be determined by referring to publicly available literature values ​​for C3 crop light energy utilization. Those skilled in the art can obtain appropriate values ​​from existing rice ecosystem flux observation studies without any inventive effort.

[0086] For all target pixels and time series within the target irrigation area, the above calculation process is performed to generate a spatially continuous total primary productivity baseline covering both single-season and double-season rice areas. This baseline adopts a unified process structure and parameterization scheme and is not directly used as the final absolute estimation result, but rather provides a basic reference field for subsequent state-dependent bias correction based on the base residuals.

[0087] Example 6 This embodiment is based on the online inference process of regional total primary productivity correction based on time-series state representation and residual mapping. It takes fluorescent endmembers and productivity benchmarks as inputs and dynamically compensates the benchmark model through prediction bias.

[0088] In one possible implementation, obtaining the corrected total primary productivity includes the following steps: Step 601: Based on the crop type and the relative position of the time step within the growing season in the rice planting structure information, obtain the phenological cycle information corresponding to the rice crop type.

[0089] Specifically, the crop type code, crop growing season code, relative position variable within the growing season at the current time step, and the intra-year date identifier for the corresponding time step of rice within the target irrigation area are obtained. To reduce the interference of abrupt changes in time spanning years on sequence feature extraction, trigonometric functions are used to encode the intra-year date identifier, and intra-year periodic features composed of sine and cosine components are calculated. The above crop code, relative position variable, and intra-year periodic features are combined to form phenological periodic information reflecting the growth rhythm of rice.

[0090] Step 602: Match the fluorescent endmembers per unit area, vegetation parameters, meteorological data, and phenological cycle information in the time dimension to construct multivariate time series for different rice cultivation types.

[0091] We acquired unit-area fluorescence endmembers for different rice crop types, along with processed vegetation parameters (enhanced vegetation index) and meteorological data, such as photosynthetically active radiation, air temperature, and vapor pressure deficit. Using a preset time step, we performed pixel-by-pixel and hour-by-hour alignment of the unit-area fluorescence endmembers, vegetation parameters, meteorological data, and phenological cycle information. These multi-source variables were then combined and assembled according to preset dimensions to construct an annual unlabeled multivariate time series reflecting the coordinated changes in rice physiology, canopy status, meteorological environment, and phenological background.

[0092] Step 603: Input the multivariate time series into the pre-trained time encoder to obtain a low-dimensional time series state representation, which is used to represent the physiological, environmental and phenological states of rice.

[0093] Specifically, the constructed multivariate time series is input into a pre-trained temporal encoder. Using the temporal encoder, high-dimensional multi-source information is nonlinearly compressed and features are extracted, outputting a 6-dimensional latent feature vector corresponding to each time step, forming a low-dimensional temporal state representation. The 6-dimensional latent feature vector at the t-th time step is used as a low-dimensional representation of the comprehensive physiological, environmental, and phenological state of rice at that time step, connecting the unlabeled temporal series with finite base-supervised information, reducing the risk of overfitting caused by high-dimensional feature mapping. It should be noted that the offline training process of the temporal encoder in this embodiment is described in detail in subsequent embodiment 7.

[0094] Step 604: Input the time-series state representation into the residual mapping model to obtain the regional bias, whereby the regional bias is the predicted residual corresponding to the regional continuous total primary productivity benchmark.

[0095] A 6-dimensional latent feature vector is input into a pre-trained residual mapping model. The residual mapping model establishes a mapping relationship between the low-dimensional temporal state representation and the output bias of the light energy utilization process model. Forward computation is performed using this model to obtain the prediction residual for the current pixel at the current time step; this prediction residual is then used as the region bias that needs to be compensated. The offline training process of the residual mapping model in this embodiment is described in detail in subsequent embodiment 7.

[0096] Step 605: Superimpose the regional deviation onto the regional continuous total primary productivity benchmark, perform non-negative constraint processing, and obtain the corrected total primary productivity.

[0097] In this embodiment, the regional continuous total primary productivity baseline is added to the predicted regional deviation. Considering the actual physical meaning of plant physiological processes, the total primary productivity cannot be negative. Therefore, a non-negativity constraint is applied to the sum, and the maximum value between the sum and 0 is extracted to obtain the final corrected total primary productivity. The corresponding calculation formula is as follows: ; in, Indicates the first The pixel in the first Corrected total primary productivity at each time step This is the function for finding the maximum value. Indicates the first The pixel in the first A regional continuous total primary productivity benchmark at a given time step. Indicates the first The pixel in the first Regional bias obtained from predictions at each time step This is the zero value used for the lower limit of nonnegativity constraints.

[0098] Example 7 Based on Example 6, this example addresses the problem of scarce base observation samples and difficulty in supporting high-dimensional multi-source variable regional state representation learning. It provides a model training mechanism based on a combination of label-free temporal self-supervised learning and base residual constraints.

[0099] In one possible implementation, the pre-trained temporal encoder is obtained through training in the following unlabeled self-supervised manner: Step 701: Obtain historical unlabeled multivariate time series data, which includes historical fluorescent endmembers per unit area.

[0100] In this embodiment, a large amount of unlabeled remote sensing and meteorological matching data of the target irrigation area in historical years is obtained. Using the same data construction method as in Embodiment 6, historical unit area fluorescence endmembers, vegetation parameters, meteorological data and phenological cycle information are combined to construct a historical unlabeled multivariate time series reflecting the comprehensive status of rice, so as to cover rice pixels in the target irrigation area that lack ground flux observation labels.

[0101] Step 702: Randomly mask the effective observations in the historical unit area fluorescent endmembers to obtain the masked multivariate time series.

[0102] In historical unlabeled multivariate time series, the effective observation time steps of fluorescence endmembers per unit area are located. Fluorescence values ​​within the effective observation time steps are randomly masked according to a preset masking ratio, replacing them with mask identifiers. Vegetation parameters, meteorological data, and phenological cycle information within the same time step are preserved unchanged, thus obtaining the masked multivariate time series. This masking operation is performed only on the original effective observations, avoiding the introduction of missing data as erroneous supervisory information into the subsequent training process.

[0103] Step 703: Input the masked multivariate time series into the initial time encoder to predict the masked historical fluorescence endmembers per unit area. Extract fluorescence phase change features, crop growth stage features, and intra-annual cycle features based on the unmasked original observations in the historical unlabeled multivariate time series. Combine the prediction results with the fluorescence phase change features, crop growth stage features, and intra-annual cycle features to construct a multi-task joint loss function.

[0104] In this embodiment, an initial temporal encoder consisting of one-dimensional temporal convolutional layers is constructed, with its channel number decreasing layer by layer, for example, sequentially passing through one-dimensional convolutional layers with 32 channels, 16 channels, and 6 channels, to progressively compress the high-dimensional time series into a low-dimensional latent representation. The masked multivariate time series is then input into the initial temporal encoder to predict the fluorescence endmember values ​​at the masked locations.

[0105] To enable the time-series encoder to learn the comprehensive characteristics of rice physiological states, environmental changes, and phenological processes, this embodiment sets multiple training objectives. Among them, fluorescence reconstruction loss constrains the encoder to recover the fluorescent endmembers per unit area of ​​the masked structure, ensuring that the low-dimensional time-series representation retains the main fluorescence physiological information; adjacent change loss constrains the recovery of fluorescence changes between adjacent time phases, ensuring that the low-dimensional time-series representation retains fluorescence temporal change characteristics; cropping position loss, growth stage loss, and cycle loss are used to ensure that the low-dimensional time-series representation retains information on cropping progress, growth stages, and intra-annual cycles, respectively. Based on the differences between the predicted or reconstructed results corresponding to each training objective and the known information, the corresponding losses are calculated and combined according to preset weights to form a multi-task joint loss function.

[0106] ; in, This represents the total value of the constructed multi-task joint loss function. This represents the preset weighting coefficients for the fluorescence reconstruction task. This represents the fluorescence reconstruction loss used to recover artificially masked fluorescence values. The preset weighting coefficients represent the fluorescence phase change task. This represents the adjacent change loss used to preserve information about the direction and magnitude of fluorescence changes. The adjacent change loss is calculated only for samples where the previous valid observation time step can be determined. This represents the preset weighting coefficient for the seasonal position task. This represents the seasonal position loss, used to characterize the relative position of the current time step within the seasonal period. This represents the preset weighting coefficients for tasks in the growth phase. This indicates the growth stage losses used to constrain different growth stages of rice. This represents the preset weighting coefficient for annual periodic tasks. This represents the periodic loss used to preserve the periodic characteristics of dates within the year.

[0107] In some optional embodiments, the formula for calculating fluorescence reconstruction loss is as follows: ; Where M represents the set of masked time steps, and |M| represents the number of masked time steps. This represents the predicted fluorescence endmember value at the t-th masked time step. This represents the corresponding original true fluorescent endmember value.

[0108] For example, the formula for calculating the loss due to adjacent changes is as follows: ; in, This represents the predicted change in the fluorescent endmember between the t-th time step and the previous effective time step. This represents the corresponding actual change.

[0109] For example, the formula for calculating the seasonal location loss is: ; in, This represents the encoder's predicted relative position within the season at time step t. This indicates the actual relative position label within the season at that time step.

[0110] Furthermore, the growth stage loss employs the cross-entropy loss function to constrain the encoder's prediction of the discrete growth stage category at the current time step. The periodic loss is calculated as follows: ; in, and These represent the sine and cosine components of the dates predicted by the encoder within the year. and This represents the corresponding actual value.

[0111] Step 704: Update the parameters of the initial temporal encoder by minimizing the multi-task joint loss function to obtain the pre-trained temporal encoder.

[0112] A conventional gradient descent optimization algorithm is employed, with the goal of minimizing the joint loss function across multiple tasks. Backpropagation is used to calculate the gradient, and the network weights of the initial temporal encoder are iteratively updated. Training stops when the preset maximum number of iterations is reached, or when the early stopping mechanism is triggered, the network parameters are fixed, and the pre-trained temporal encoder is output.

[0113] In another alternative embodiment, the pre-trained temporal encoder is obtained through the following self-supervised learning training: Obtain multivariate time series data of historical rice crop types; Randomly mask the effective observations in fluorescent endmembers per unit area; The masked multivariate time series is input into the initial time encoder to predict the fluorescent endmembers per unit area of ​​the masked area, and a multi-task joint training objective is constructed by combining the fluorescence phase change characteristics, crop growth stage characteristics and intra-annual cycle characteristics. The initial temporal encoder is updated based on the joint training objective to obtain the temporal encoder.

[0114] Furthermore, the pre-trained residual mapping model is obtained through training in the following manner: Step 705: Obtain the total primary productivity of the tower base observation and obtain the continuous total primary productivity benchmark of the tower base location in the historical region of the corresponding time phase.

[0115] Carbon flux observation data from the eddy covariance system within the target irrigation district were acquired over a historical period. After routine quality control and time aggregation processing, the total primary productivity observed at the tower base was obtained. Based on the geographic coordinates of the eddy covariance system, a historical continuous total primary productivity benchmark for the corresponding observation phase was extracted for the tower base location.

[0116] Step 706: Calculate the base observation residual between the total primary productivity observed at the base and the historical region continuous total primary productivity benchmark.

[0117] The difference between the total primary productivity observed at the base station and the historical continuous total primary productivity benchmark is calculated to obtain the base station observation residual. This processing method ensures that the limited base station observation data is used only to calibrate the deviation of the physical process model under different states, rather than to fit high-dimensional absolute productivity values.

[0118] Step 707: Input the historical multivariate time series of the tower base location in the corresponding time phase into the pre-trained time encoder for feature extraction to obtain the historical time series state representation of the tower base location in the corresponding time phase.

[0119] The historical multivariate time series of the base position in the corresponding time phase is extracted and input into a pre-trained time encoder. After forward computation, the historical time series state representation of the corresponding dimension is extracted.

[0120] Step 708: Using historical time series state representation as input features and base observation residuals as supervision labels, the initial mapping model is fitted with parameters by combining regularization constraints to obtain a pre-trained residual mapping model.

[0121] In this embodiment, a ridge regression mapping model with an L2 regularization term is constructed. Using historical time-series state representations as independent variables and base case observation residuals as dependent variables, the regression coefficients are solved by minimizing an objective function that includes mean squared error and a regularization penalty term. The regularization penalty term is used to limit excessive fluctuations in the regression coefficients and reduce the risk of overfitting. The formula for calculating the parameter fitting is as follows: ; in, This represents the estimated value of the optimal intercept term obtained by solving the problem. This represents the estimated value of the optimal regression coefficient vector obtained by solving the problem. This represents the parameter operation that finds the minimum value of the objective function. This represents the intercept variable to be solved. This represents the vector variable of regression coefficients to be solved. This represents the total number of observation samples from the base of the tower that participated in the model fitting. Represents the time step index variable. Indicates the first The residuals of the tower base observations at each time step Indicates the first A historical temporal state representation vector at each time step This represents the preset L2 regularization coefficient. This represents the L2 norm squared penalty term of the regression coefficient vector.

[0122] Furthermore, to evaluate the cross-year generalization capability of the mapping model, a leave-one-out validation mechanism is introduced. Specifically, a specific year from the historical dataset is selected as an independent leave-out year. The data from this leave-out year is not used in the training of the temporal encoder or the parameter fitting of the initial mapping model; it is only used to independently calculate the error metric between the predicted residual and the true residual. After completing leave-one-out validation for all years and confirming that the cross-year stability meets the requirements, the parameter fitting is re-executed using samples from all valid years, outputting the final pre-trained residual mapping model.

[0123] This embodiment uses a self-supervised encoder to mine the temporal state representation of massive unlabeled regions. Taking the continuous output of the light energy utilization process model as a benchmark, it establishes a residual mapping using limited ground observation data, preserves the spatial continuity of the mechanism model, and reduces the risk of overfitting caused by directly fitting high-dimensional features with small samples, thus completing the dynamic compensation for estimation bias under different environmental conditions.

[0124] As an example, to illustrate the effectiveness of the above-mentioned time-series state characterization and residual mapping methods, the corrected total primary productivity (GPP) is compared with the total primary productivity (EC-GPP) obtained from observations based on the flux tower eddy covariance method (EC) during the same period. The results are as follows: Figure 5 As shown, the corrected GPP exhibits high consistency with the base station observations, indicating that residual compensation based on low-dimensional time-series state representation can effectively reduce the systematic bias of the baseline model. To quantitatively evaluate the relationship between the model estimation results and the base station observation data, this embodiment uses the linear regression coefficient of determination (R²). 2 The root mean square error (RMSE), mean absolute error (MAE), and mean bias (Bias) are used as evaluation indicators. Among them, R... 2 RMSE is used to characterize the degree of agreement between the estimated and observed values ​​in terms of linear trends. RMSE reflects the overall error level and is sensitive to larger errors. MAE represents the average magnitude of the absolute error between the estimated and observed values. Bias is used to determine whether the model as a whole systematically overestimates or underestimates.

[0125] Example 8 Based on Examples 1 and 6, this example calculates canopy transpiration and field evaporation based on corrected total primary productivity, combined with the carbon-water coupling mechanism and the principle of surface energy balance, and generates the final spatiotemporal data products.

[0126] In one possible implementation, the corrected total primary productivity is used as a physiological driver input into a pre-defined carbon-water coupling model to estimate canopy transpiration, including the following steps: Step 801: Extract air temperature, water vapor pressure deficit, and atmospheric pressure from meteorological data.

[0127] Meteorological data that has undergone spatial interpolation is acquired, and the air temperature (near-surface air temperature), vapor pressure deficit, and atmospheric pressure corresponding to the target pixel and the current time step are extracted from it. To ensure the consistency of units for the pressure term in subsequent stomatal-photosynthetic coupling relationship calculations, the dimensions of vapor pressure deficit and atmospheric pressure are uniformly constrained.

[0128] Step 802: Calculate the CO2 compensation point based on air temperature. During photosynthesis, crops absorb CO2 through their stomata and simultaneously release water vapor into the atmosphere. This embodiment uses the corrected total primary productivity to characterize crop carbon absorption intensity and calculates canopy transpiration by combining air temperature, vapor pressure deficit (VPD), atmospheric pressure (Pa), and annual atmospheric CO2 concentration. The corresponding CO2 compensation point calculation formula is as follows: ;in, This refers to the mole fraction of atmospheric oxygen. This refers to the air temperature.

[0129] Step 803: Input the corrected total primary productivity, water vapor pressure deficit, atmospheric pressure, carbon dioxide compensation point, and preset atmospheric carbon dioxide concentration into the stomatal-photosynthetic coupling relationship in the carbon-water coupling model to calculate canopy transpiration.

[0130] Specifically, corrected total primary productivity is substituted into the calculation as a physiological driver to prevent the transmission of systematic bias from the upstream baseline model to the water consumption stage. The reciprocal of marginal water use efficiency is obtained, multiplied by the vapor pressure deficit, and the square root is calculated using a constant coefficient to obtain the humidity demand term. The reciprocal of marginal water use efficiency can be determined by regression analysis of baseline carbon and water flux observation data based on the species physiological characteristics of the target rice type, or by directly using recommended parameter values ​​for rice crop types already published in this field.

[0131] Accordingly, the difference between the atmospheric carbon dioxide concentration and the carbon dioxide compensation point is calculated, multiplied by atmospheric pressure, and the square root is calculated to obtain the pressure supply term. The corrected total primary productivity is then multiplied by a constant coefficient, multiplied by the humidity demand term, and finally divided by the pressure supply term to calculate canopy transpiration. The corresponding calculation formula is as follows: ; in, This represents the calculated canopy transpiration. This represents corrected gross primary productivity, expressed in gC·m³. - ²·d - ¹, This represents the reciprocal of marginal water use efficiency, and is a dimensionless coefficient whose value implicitly includes unit conversion. This indicates a water vapor pressure deficit, expressed in kPa. Atmospheric pressure is expressed in kPa. This indicates the preset atmospheric carbon dioxide concentration, expressed in parts per million (ppm). The calculated carbon dioxide compensation point is represented in parts per million (ppm). 1.5 is the unit conversion coefficient formed under the total primary productivity, carbon dioxide concentration, and canopy transpiration unit system used in this embodiment. 1.6 is the conversion factor in the relationship between water vapor and carbon dioxide stomatal diffusion. Further, obtaining the total evapotranspiration of the paddy field specifically includes the following steps: Step 804: Extract net radiation and relative humidity from meteorological data, with relative humidity expressed as a decimal ratio from 0 to 1; extract the leaf area index from vegetation parameters. Extract net radiation reflecting surface energy input, relative humidity reflecting atmospheric humidity, and leaf area index reflecting the degree of canopy shading from vegetation parameters. Step 805: Calculate the potential evaporation of the field surface based on net radiation, leaf area index, and a preset canopy extinction coefficient.

[0132] In this embodiment, conventional physical meteorological relationships are employed, and the slope of the saturated vapor pressure curve and the ecliptic constant are calculated based on air temperature and atmospheric pressure. Surface soil heat flux is obtained, based on net radiation according to a preset proportionality coefficient, which can be determined by those skilled in the art based on the characteristics of the paddy field surface. In this embodiment, considering the time scale and paddy field conditions, the contribution of surface heat flux to net radiation is relatively small, and the surface heat flux is approximated as 0. The effective energy reaching the field surface after net radiation attenuates within the canopy is calculated. This attenuation process can be performed exponentially using a natural constant, with the exponent being the negative product of the canopy extinction coefficient and the leaf area index. Combining the ratio of the saturated vapor pressure curve slope to the ecliptic constant, and the difference between the effective energy and the surface soil heat flux, the potential evaporation of the field surface is calculated. The preset canopy extinction coefficient can be determined based on the structural characteristics of the rice canopy and radiative transfer theory; those skilled in the art can obtain typical values ​​from existing studies on rice canopy radiation interception.

[0133] Step 806: Calculate the wet and dry constraints based on relative humidity and water vapor pressure deficit.

[0134] In this embodiment, an index term reflecting the surface wetness is constructed using relative humidity, and this term is defined as the surface wetness term. An index term reflecting the air dryness is constructed using relative humidity and water vapor pressure deficit, and this term is defined as the air dryness limitation term. Next, the fourth power of relative humidity is used as the surface wetness term, and a power function with relative humidity as the base and water vapor pressure deficit as the exponent is calculated as the air dryness limitation term. The corresponding calculation formula is as follows: ; ; in, This represents the surface wetness term constructed based on relative humidity. Represents relative humidity, in decimal form, ranging from 0 to 1. This represents the air humidity and dryness constraint term constructed based on relative humidity and water vapor pressure deficit. This indicates a vapor pressure deficit. It should be noted that the vapor pressure deficit is represented by a pure numerical value in kPa as a dimensionless exponent, which is a technical convention in remote sensing estimation of evapotranspiration in this field. This is a preset constant exponent.

[0135] Step 807: Calculate field surface evaporation based on potential evaporation and wet / dry constraints.

[0136] Calculate the difference between constant 1 and the surface moisture term, and multiply it by the air humidity limitation term to obtain the adjustment term. Add the surface moisture term and the adjustment term to obtain the comprehensive humidity limitation factor. Multiply the comprehensive humidity limitation factor by the potential evaporation of the field surface to calculate the field surface evaporation, using the following formula: ; in, This represents the calculated evaporation from the field surface. Indicates surface moisture. Indicates the air humidity limit. This indicates potential evaporation on the field surface.

[0137] Step 808: Add the evaporation from the field surface to the evaporation from the canopy to obtain the total evapotranspiration of the paddy field.

[0138] Under the same spatial pixel and time step, canopy transpiration and field evaporation are superimposed and synthesized from two sources to obtain the total evapotranspiration of the paddy field at the current pixel and the current time step. This dual-source structure allows canopy transpiration and non-transpiration water consumption to be constrained by state information corresponding to their physical processes.

[0139] Furthermore, after obtaining the total evapotranspiration of the paddy fields, it also includes: Step 809: Serialize the total evapotranspiration of the paddy field according to the preset time step.

[0140] Obtain the total evapotranspiration of all rice pixels within the target irrigation area at different time points. Arrange the total evapotranspiration results in chronological order according to a uniform preset time interval, forming a time series covering the target analysis period.

[0141] Step 810: Map the serialized total evapotranspiration of paddy fields to a unified preset spatial grid to generate and output spatiotemporal data products of continuous paddy field evapotranspiration in the region.

[0142] The serialized total evapotranspiration data of paddy fields is mapped to a unified spatial computing grid according to the geographic coordinate index of the target pixels. The mapped data is then formatted and encapsulated to generate a spatiotemporal data product of continuous regional paddy field evapotranspiration containing spatial coordinates, time markers, and evapotranspiration values. This product is then output to provide data support for crop water consumption analysis and regional water resource allocation in complex paddy irrigation areas.

[0143] This embodiment uses residual-corrected productivity data as a physiological driver and substitutes it into the stomatal-photosynthetic coupling relationship. Combined with surface wet and dry constraints, it independently calculates field surface evaporation, blocking the transmission of upstream system deviations to the water consumption stage, thereby improving the stability and reliability of regional water and carbon flux estimation in complex agricultural scenarios.

[0144] As an example, to verify the reliability of the obtained total evapotranspiration results for paddy fields, the evapotranspiration (ET) data of the base cell was extracted and compared with the evapotranspiration (EC-ET) data observed by the base eddy correlation method during the same period. The results are as follows: Figure 6 As shown, the model-estimated ET and the tower-based observed ET have good consistency, indicating that the total evapotranspiration change in paddy fields can be characterized by the dual-source synthesis of canopy transpiration and field surface evaporation.

[0145] Example 9 This embodiment provides a method for decomposing daylight-induced chlorophyll-fluorescence hybrid pixels based on refined background components. When computational resources allow and there is sufficient high-resolution land classification data, this embodiment no longer refers to all non-rice land cover as a single background component. Instead, it reduces the interference of highly heterogeneous land surfaces on the physiological signals of the target rice by subdividing non-rice land cover categories. One possible implementation includes the following steps: Step 901: Obtain the area ratio of multiple background sub-components based on refined surface classification data.

[0146] Specifically, high-resolution, refined land classification data of the target irrigation area was acquired. Within each target cell, in addition to identifying single-season rice and double-season rice, non-rice land cover was subdivided into multiple independent background sub-components, such as forest land, dryland crops, water bodies, and built-up land. The area proportions of single-season rice, double-season rice, and each background sub-component within the target cell were calculated separately. Conventional normalization methods were used to process the area proportions of each component to ensure that the sum of the area proportions of single-season rice, double-season rice, and all background sub-components is always equal to 1.

[0147] Step 902: Screen high-purity candidate pixels of each category of background sub-components and construct multi-component artificial mixed fluorescent samples.

[0148] In this embodiment, independent high-purity area ratio thresholds and infiltration restriction thresholds are set for each identified background sub-component. Candidate pixels for each background sub-component that meet the threshold conditions are selected from historical mixed pixels, and the initial values ​​of the corresponding candidate fluorescent endmembers are calculated. For background sub-components with vegetation characteristics, such as woodlands and dryland crops, a soft-gating mechanism is implemented using the relative activity of historical vegetation indices to remove abnormal endmembers in a low-activity state, thus obtaining effective candidate endmembers for each background sub-component.

[0149] Next, to construct artificially mixed samples suitable for training multi-class decomposition models, a set of non-negative area weights with a sum limited to 1 is randomly generated. This set contains simulated weights corresponding to single-season rice, double-season rice, and each background sub-component. The effective candidate endmembers of double-season rice, single-season rice, and each background sub-component are multiplied by their respective non-negative area weights and summed to calculate the expanded artificially mixed fluorescence value. The corresponding calculation formula is as follows: ; In the formula, This indicates the extended artificially mixed fluorescence value obtained through construction. This represents the simulated non-negative area weights for randomly generated double-cropping rice. This indicates effective candidate endmembers for double-cropping rice. This represents the non-negative area weights for randomly generated single-season rice simulations. This indicates the effective candidate endmembers for single-season rice. This represents the total number of background sub-components after subdivision. Indicates the index number of the background subcomponent. Represents the randomly generated first... Simulated non-negative area weights for each background subcomponent Indicates the first Valid candidate endmembers for each background subcomponent.

[0150] Accordingly, the product of the effective candidate endmembers of each component and their corresponding simulated non-negative area weights is used as the label of the true fluorescence contribution of the corresponding component.

[0151] Step 903: Based on the decomposition model of multi-class input, predict the refined fluorescence contribution and calculate the fluorescence endmember per unit area.

[0152] In this embodiment, an expanded input feature vector is constructed, and the observed values ​​of sunlight-induced chlorophyll fluorescence mixed pixels, the area ratio of single-season rice, the area ratio of double-season rice, the area ratio of each background sub-component, and related auxiliary environmental variables are extracted and combined to form a multi-class inference input vector. This vector is then input into a multi-output multilayer perceptron model pre-trained based on artificial samples constructed in step 902, and the predicted fluorescence contribution of single-season rice, the fluorescence contribution of double-season rice, and the fluorescence contribution of each background sub-component are output.

[0153] After obtaining the predicted fluorescence contribution, phenological prior constraints are applied in conjunction with the relative activity of vegetation indices to suppress anomalous contributions. The fluorescence endmembers per unit area for different rice crop types are calculated by dividing the adjusted fluorescence contribution of single-season or double-season rice by the corresponding area ratio. This scheme provides clean input variables for subsequent temporal state characterization by removing the independent influence of preset non-rice land cover (forest land with high fluorescence contribution and water bodies with zero fluorescence contribution).

[0154] Based on the above embodiments, this application also provides a rice irrigation area evapotranspiration estimation system based on multi-source remote sensing and tower-based data constraints. The system includes a processor coupled to a memory. The memory stores instructions, which are executed by the processor to implement the methods described in the above embodiments.

[0155] This application also provides an electronic device, including: one or more processors and a memory, wherein the one or more processors are coupled to the memory, the memory is used to store computer program code, the computer program code including computer instructions, and when the one or more processors execute the computer instructions, the electronic device executes the methods executed in the above-described method embodiments. This application also provides a chip, including: a processor and a communication interface, wherein the processor calls a computer program or instructions through the communication interface to execute the methods executed in the above-described method embodiments. This application also provides a computer program product, wherein when the computer program product runs on a computer, the computer executes the methods executed in the above-described method embodiments. This application also provides a computer-readable storage medium storing a computer program, wherein when the computer program runs on a computer, the computer executes the methods executed in the above-described method embodiments.

[0156] It should be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A method for estimating evapotranspiration in rice-irrigated areas based on multi-source remote sensing and data constraints, characterized in that... include: Acquire sunlight-induced chlorophyll-fluorescence mixed pixels, vegetation parameters, meteorological data, and rice planting structure information of the target irrigation area; Based on rice planting structure information, the solar-induced chlorophyll fluorescence mixed pixels were decomposed, and the unit area fluorescence endmembers of different rice planting types were extracted. Calculate regional continuous total primary productivity benchmarks based on vegetation parameters and meteorological data for different rice cultivation types; Based on fluorescent endmembers per unit area, vegetation parameters and meteorological data, a time-series state characterization is constructed. A pre-trained residual mapping model is used to predict regional bias based on the time-series state characterization. The regional bias is then compensated to the regional continuous total primary productivity benchmark to obtain the corrected total primary productivity. The corrected total primary productivity was used as a physiological driver input into a pre-defined carbon-water coupling model to estimate canopy transpiration. Evaporation on the field surface is estimated based on vegetation parameters and meteorological data. The evaporation on the field surface is combined with the canopy transpiration to obtain the total evapotranspiration of paddy fields at the field scale in the irrigation area.

2. The method according to claim 1, characterized in that, Extracting fluorescent endmembers per unit area for different rice cultivation types, specifically including: Based on rice planting structure information, the area ratio of each component in the solar-induced chlorophyll fluorescence mixed pixel was obtained; The fluorescence contribution of each component was obtained by inputting the mixed pixels of sunlight-induced chlorophyll fluorescence, the area ratio of each component, and the auxiliary environmental variables corresponding to the mixed pixels into the decomposition model. The relative activity of vegetation indices is extracted based on vegetation parameters, and the relative activity of vegetation indices is used to impose phenological prior constraints on the fluorescence contribution to obtain the adjusted fluorescence contribution. Based on the adjusted fluorescence contribution and the corresponding area ratio, the fluorescent endmembers per unit area for different rice cultivation types are calculated.

3. The method according to claim 2, characterized in that, The decomposition model is obtained through pre-training using the following steps: Artificially mixed samples were obtained, which included artificially mixed fluorescence, simulated area ratios, and auxiliary environmental variables. Obtain the true fluorescence contribution corresponding to the artificially mixed samples; Using artificially mixed samples as input and the corresponding real fluorescence contribution as supervision labels, the initial neural network is trained in a supervised manner to obtain the decomposition model.

4. The method according to claim 3, characterized in that, The process of constructing artificially mixed samples and real fluorescence contributions specifically includes: High-purity candidate pixels are screened from historical mixed pixels based on historical area ratios, and the initial values ​​of candidate fluorescent endmembers are calculated. Historical vegetation parameters were used to extract the relative activity of historical vegetation indices. Based on this, a soft gating mechanism was implemented on the initial values ​​of candidate fluorescent endmembers to remove abnormal endmembers in low-activity states and obtain effective candidate endmembers. Randomly generate non-negative area weights that satisfy the summation constraint as the simulated area ratio; Based on the non-negative area weights and the corresponding valid candidate endmembers, artificial mixed fluorescence and the contribution of real fluorescence are constructed. Vegetation parameters and meteorological data corresponding to the time phase of artificial mixed fluorescence were obtained as auxiliary environmental variables. The artificial mixed fluorescence, simulated area ratio, and auxiliary environmental variables were combined to form an artificial mixed sample.

5. The method according to claim 3, characterized in that, The supervised training process for the initial neural network also includes: The target irrigation area is divided into multiple spatial blocks; Reserve some space blocks as independent spatial areas; The parameters are updated using artificially mixed samples outside the independent spatial reserved area. The spatial applicability of the decomposition model is verified using artificially mixed samples within the independent spatial reserved area to obtain the spatial verification error index. When the spatial verification error index meets the preset generalization performance conditions, the decomposition model is confirmed.

6. The method according to claim 1, characterized in that, Based on vegetation parameters and meteorological data, regional continuous gross primary productivity benchmarks for rice cultivation types were calculated, including: Extract photosynthetically effective vegetation absorption and water status characteristics from vegetation parameters; Extract photosynthetically active radiation data and temperature data from meteorological data; Temperature regulation factor is calculated based on air temperature data, and moisture regulation factor is calculated based on moisture state characterization. Based on maximum light energy utilization, photosynthetically effective vegetation absorption, photosynthetically effective radiation, temperature regulation factor, and water regulation factor, the regional continuous total primary productivity baselines for different rice cultivation types were calculated.

7. The method according to claim 1, characterized in that, To obtain corrected total primary productivity, including: Based on the crop type and the relative position of the time step within the growing season in the rice planting structure information, phenological cycle information corresponding to different rice crop types is obtained. By matching fluorescent endmembers per unit area, vegetation parameters, meteorological data, and phenological cycle information in the time dimension, multivariate time series of rice cultivation types are constructed. By inputting multivariate time series data into a pre-trained time encoder, a low-dimensional temporal state representation can be obtained. The temporal state representation is input into the residual mapping model to obtain the regional bias; The regional bias is superimposed onto the regional continuous total primary productivity benchmark and subjected to non-negative constraint processing to obtain the corrected total primary productivity.

8. The method according to claim 1, characterized in that, The canopy transpiration is estimated by inputting the corrected total primary productivity as a physiological driver into a pre-defined carbon-water coupling model, specifically including: Extract air temperature, water vapor pressure deficit, and atmospheric pressure from meteorological data; Calculate the carbon dioxide compensation point based on air temperature; The total primary productivity, water vapor pressure deficit, atmospheric pressure, carbon dioxide compensation point, and preset atmospheric carbon dioxide concentration are corrected and input into the stomatal-photosynthetic coupling relationship in the carbon-water coupling model to calculate canopy transpiration.

9. The method according to claim 1, characterized in that, Obtain meteorological data for the target irrigation area, specifically including: Obtain raw meteorological grid data; Based on the target pixels within the target irrigation area, determine the nearest effective meteorological grid points; Based on the spatial distance between the target pixel and the effective meteorological grid points, the distance weight is calculated. Spatial interpolation of the original meteorological grid data is performed using distance weights to obtain meteorological data consistent with the spatial grid of the target pixel.

10. A rice irrigation area evapotranspiration estimation system based on multi-source remote sensing and tower-based data constraints, the system comprising a processor coupled to a memory, the memory storing instructions which, when executed by the processor, implement the method as described in any one of claims 1 to 9.