Integrated estimation method and device for ecological water demand and water supplement

Abstract

The present application relates to the field of ecology, water resources management and remote sensing technology, and provides an integrated estimation method and device for ecological water demand and water supplement. The method comprises: performing deviation correction or fusion processing on precipitation, evapotranspiration and soil moisture content data respectively; determining a reference evapotranspiration based on the obtained consistent driving field and constructing a baseline vegetation correction coefficient, performing residual correction on the correction coefficient under the condition of no water stress, and determining the ecological water demand according to the corrected vegetation correction coefficient and the reference evapotranspiration; inputting the consistent driving field, the ecological water demand and the remote sensing vegetation observation information into multiple discrimination gates, performing logical operation on the discrimination results, and dividing each pixel into a multi-level water supplement priority label; determining the total loss term based on the consistent driving field, the ecological water demand and the water supplement priority label, and determining the ecological water supplement. The problem that the ecological water supplement accounting in the prior art is difficult to meet the water conservation constraint and is prone to systematic bias is solved.

Description

Technical Field

[0001] This invention relates to the fields of ecology, water resource management and remote sensing, and in particular to an integrated method and apparatus for estimating ecological water demand and water replenishment. Background Technology

[0002] In the field of ecological water use assessment, an estimation framework based on reference crop evapotranspiration and vegetation correction coefficients is commonly used. This framework first calculates reference evapotranspiration using a modified Penman-Montes method under the assumption of a "reference underlying surface and unrestricted water supply." Then, a vegetation correction coefficient is introduced to map the reference evapotranspiration to the potential evapotranspiration of a specific crop or natural vegetation under conditions without water stress, which is then used as the ecological water demand. While this type of method has a clear structure and well-defined physical meaning, its core parameters are mostly given by empirical tables or simplified rules, making it difficult to explicitly absorb the slow and continuous impacts of "human management practices and long-term structural evolution," such as the increase in leaf area index and roughness due to dense planting and variety improvement, and the enhancement of turbulent exchange through oasis management.

[0003] With the rapid development of Earth observation and meteorological products, the spatiotemporal continuity and coverage of regional-scale driving data such as precipitation, temperature, net radiation, and wind fields have significantly improved, leading to the widespread adoption of remote sensing-driven hydrological variable estimation. However, systematic biases generally exist between different data sources in terms of amplitude baselines, seasonal phases, and responses to extreme events, and the spatial heterogeneity of these biases is significantly correlated with topography. If these data are directly used as input for ecological water demand and replenishment assessments, the biases will be amplified in subsequent classification and water volume calculations, resulting in incomparability between regions or misjudgments during critical periods.

[0004] Current ecological water replenishment schemes often begin by obtaining stratified soil moisture content through remote sensing inversion, statistical learning, and spatial interpolation, and then use threshold or index methods to determine and classify "whether there is water shortage" and "water shortage level." This approach is feasible to some extent in areas with good data availability, but it fails to adequately reflect the combined logic of three types of evidence: phenological sensitivity, canopy structure evolution, and immediate physiological constraints. If only a single threshold or data source is relied upon for classification, false alarms or false alarms are likely to occur on complex underlying surfaces. In addition, current schemes simply approximate ecological water replenishment as "the difference between potential evapotranspiration and actual evapotranspiration," failing to take into account the combined effects of multiple channels such as changes in root zone water storage, deep seepage, lateral outflow, and capillary recharge. Therefore, it is difficult to meet the water conservation constraint, and the extrapolation performance is weak under different root depths and profile conditions. Summary of the Invention

[0005] This invention provides an integrated estimation method and apparatus for ecological water demand and replenishment, which solves the problems in the prior art where ecological water shortage classification relies on a single threshold or a single data source, which is prone to false alarms or omissions on complex underlying surfaces, and ecological replenishment calculation fails to take into account the merging effect of multiple water channels, making it difficult to meet water conservation constraints and easily generating systematic deviations.

[0006] This invention provides an integrated method for estimating ecological water demand and water replenishment, comprising: The precipitation data, evapotranspiration data, and soil moisture content data were respectively subjected to bias correction or fusion processing to obtain a consistent driving field; Based on the uniform driving field, a reference evapotranspiration is determined and a baseline vegetation correction coefficient is constructed. Under the condition of no water stress, the baseline vegetation correction coefficient is subjected to physically constrained residual correction to obtain the corrected vegetation correction coefficient. The ecological water requirement is determined based on the corrected vegetation correction coefficient and the reference evapotranspiration. The uniform driving field, the ecological water demand, and remote sensing vegetation observation information are input into multiple pre-constructed discrimination gates. The judgment results of the multiple discrimination gates are subjected to conjunctive logic operations to divide each pixel into multi-level water replenishment priority labels. Among them, the multiple discrimination gates respectively characterize phenological sensitivity, structural degradation degree, and immediate physiological limitation. Based on the unified driving field, the ecological water demand, and the water replenishment priority label, a total loss item is determined, and an ecological water replenishment amount is determined according to the ecological water demand and the total loss item; wherein, the total loss item is the combined amount of root zone water storage change, deep seepage, lateral outflow, and capillary recharge.

[0007] In one embodiment, the precipitation data refers to raw, unprocessed satellite precipitation product data, hereinafter referred to as satellite pixel precipitation data; the precipitation data is biased and corrected in the following manner: In the log-intensity space, the bias relationship between precipitation data from ground meteorological stations and precipitation data from satellite pixels is established, with topographic factors, intra-annual temporal phases, and satellite pixel intensity as covariates. Different regression kernel bandwidths are estimated for the dry and wet seasons respectively. Geographically weighted regression fitting is performed within a moving window using the regression kernel bandwidth corresponding to the season. The regression coefficients are mapped back to grid points to form a precipitation correction surface that distinguishes between the dry and wet seasons. The geographically weighted regression fitting simultaneously outputs pixel-level precipitation uncertainty and quality label. An adaptive threshold is set for the relative change of pixels in the precipitation correction surface that distinguishes between dry and wet seasons. When the relative change of pixels exceeds the adaptive threshold, the correction amplitude of the pixel is truncated to obtain the truncated precipitation correction surface. The truncated precipitation correction surface is weighted by the pixel-level precipitation uncertainty and quality label to form the precipitation data in the uniform driving field.

[0008] In one embodiment, the evaporation data includes at least three sets of evaporation products with different mechanisms and complementary error structures; the evaporation data is fused in the following manner: At least three sets of evapotranspiration products with different mechanisms and complementary error structures are selected. The random error and cross-correlation of each set of evapotranspiration products are jointly estimated. The optimal weight after correlation correction is obtained by Bayesian method. The optimal weight is adaptively learned in the subdomain of the same seasonal combination in the ecological zone. The Bayesian method outputs the pixel-level evapotranspiration uncertainty while obtaining the optimal weight. The fusion evaporation is obtained by weighting each set of evaporation products according to the optimal weight. It is then determined whether the fusion evaporation exceeds the net radiation supply limit. If it does, the weight of the evaporation product that caused the limit to be exceeded is reduced and the fusion evaporation is reweighted until the fusion evaporation does not exceed the net radiation supply limit, thus obtaining a uniform evaporation. The uniformized evaporation is the evaporation data in the uniformized driving field.

[0009] In one embodiment, the soil moisture content data includes surface soil water; the soil moisture content data is fused in the following manner: A two-stage inversion method was used to retrieve the root zone volumetric water content at multiple depths. In the first stage, machine learning was used to establish a nonlinear mapping of surface soil water, vegetation status indicators, energy status indicators, water memory indicators, and static surface factors to the target depth water content, and the mapping results for the first stage were obtained. The second segment calculates the prediction residual of the mapping result of the first segment at the site scale, and obtains a continuous residual surface through spatial interpolation. The continuous residual surface is then superimposed on the mapping result of the first segment to form a pixel-level stratified soil moisture content product. The pixel-level stratified soil moisture content product serves as the soil moisture content data in the uniform driving field.

[0010] In one embodiment, determining the reference evapotranspiration and constructing the baseline vegetation correction coefficient based on the uniform driving field includes: The reference evapotranspiration is determined based on the precipitation and meteorological data in the uniform driving field. The growing season is divided into multiple phenological stages: initial, development, middle and late. The vegetation correction coefficient anchor value for the middle stage is determined by the effective coverage and canopy height obtained by remote sensing. The vegetation correction coefficient anchor value for the middle stage is fine-tuned by combining wind speed and humidity to obtain the middle vegetation correction coefficient. Based on the evolution of phenological phase and coverage between different phenological stages, the vegetation correction coefficients for the initial, developmental, and later stages are continuously interpolated, and the vegetation correction coefficients for each phenological stage are spliced ​​together to form the baseline vegetation correction coefficients covering the entire growing season.

[0011] In one embodiment, the step of performing physically constrained residual correction on the baseline vegetation correction coefficient under water stress-free conditions to obtain a corrected vegetation correction coefficient, and determining the ecological water requirement based on the corrected vegetation correction coefficient and the reference evapotranspiration, includes: Under the no-moisture stress window, the ratio of actual evapotranspiration in the uniform driving field to the reference evapotranspiration is used as the management perception tag; wherein, the criterion for the no-moisture stress window is that the ratio of cumulative precipitation to reference evapotranspiration within a set period is not less than a preset threshold and that there are no fewer than a preset number of effective rainfalls during the period. A deep learning model is used to learn the residual between the management perception label and the baseline vegetation correction coefficient, and the residual is superimposed on the baseline vegetation correction coefficient to obtain the corrected vegetation correction coefficient. The ecological water requirement is obtained by multiplying the corrected vegetation correction coefficient by the reference evapotranspiration. The physical constraints applied in the physically constrained residual correction include: the corrected vegetation correction coefficient is limited to physically acceptable upper and lower bounds; the corrected vegetation correction coefficient maintains a non-negative monotonic response to the reference evapotranspiration and a non-negative monotonic response to the leaf area index; and the ecological water demand obtained after correction does not exceed the upper limit of evapotranspiration energy available from net radiation.

[0012] In one embodiment, the plurality of discrimination gates includes a discrimination gate characterizing phenological sensitivity, a discrimination gate characterizing the degree of structural degradation, and a discrimination gate characterizing immediate physiological limitation; The process involves inputting the unified driving field, the ecological water demand, and remote sensing vegetation observation information into multiple pre-constructed discrimination gates, performing a conjunction logic operation on the judgment results of the multiple discrimination gates, and dividing each pixel into multi-level water replenishment priority labels, including: The discrimination gate characterizing phenological sensitivity delineates the growing season phase based on the ratio of the current leaf area index of a pixel to the peak leaf area index of that pixel in the current year. A staged trigger threshold is smoothly generated based on the ratio. The leaf area index in the remote sensing vegetation observation information is input into the discrimination gate characterizing phenological sensitivity. When the ratio exceeds the staged trigger threshold, it is determined to be a phenological sensitive period. The discrimination gate characterizing the degree of structural degradation calculates the leaf area index anomaly and the solar induced fluorescence anomaly after removing the seasonal climate state. The leaf area index and solar induced fluorescence in the remote sensing vegetation observation information are input into the discrimination gate characterizing the degree of structural degradation. When the leaf area index anomaly is continuously negative for more than a preset number of days and the deviation of the solar induced fluorescence anomaly exceeds a preset standard deviation multiple, it is determined that the structure is weakening; otherwise, it is determined that the structure is stable or strengthening. The discrimination gate for characterizing immediate physiological limitation calculates the relative available water in the root zone based on the stratified soil moisture content in the uniform driving field, and calculates the evapotranspiration efficiency based on the actual evapotranspiration in the uniform driving field and the ecological water requirement. The relative available water in the root zone is characterized by the normalized position of the current volumetric water content relative to field capacity and wilting point water content, and the evapotranspiration efficiency is characterized by the ratio of actual evapotranspiration to ecological water requirement. When the relative available water in the root zone is lower than a preset available water threshold and the evapotranspiration efficiency is lower than a preset efficiency threshold, it is determined to be physiologically limited. When the relative available water in the root zone or the evapotranspiration efficiency is close to but does not fall below their respective preset thresholds, it is determined to be physiologically close to limited. The multi-level water replenishment priority label includes four levels: immediate water replenishment, priority water replenishment, observation and review, and water replenishment not recommended. When all three discrimination gates simultaneously determine that it is a phenological sensitive period, structural weakening, and physiological limitation, it is judged as immediate water replenishment. When it is judged as structural weakening and in a phenological sensitive period, and simultaneously judged as physiologically approaching limitation, it is judged as priority water replenishment. When it is judged as a phenological sensitive period, but the structure is judged as stable or strengthening and the evapotranspiration efficiency has not fallen below the efficiency threshold, it is judged as observation and review. If the above conditions are not met, water replenishment is not recommended.

[0013] In one embodiment, determining the total loss item based on the consistent driving field, the ecological water demand, and the water replenishment priority label, and determining the ecological water replenishment amount based on the ecological water demand and the total loss item, includes: A supervisory quantity is constructed within a wet window, and a pre-constructed residual learning network is trained based on this supervisory quantity to obtain a trained residual learning network. The supervisory quantity is the precipitation data in the uniform driving field minus the ecological water demand. The residual learning network is a deep learning network that receives time-series data containing wet memory and canopy energy state and outputs residual corrections. A water conservation residual penalty term and a reservoir capacity upper limit constraint term are added to the training objective function. The water conservation residual penalty term constrains the predicted value of the total loss term to satisfy the root zone water balance closure, and the reservoir capacity upper limit constraint term constrains the root zone water content after water replenishment to not exceed the reservoir capacity upper limit corresponding to field capacity. The soil and topographic parameters in the uniform driving field are input into a pre-constructed physical prior model to obtain a first-order estimate of the total loss term; wherein, the physical prior model is established based on soil water holding capacity, water conductivity, topographic runoff characteristics, and infiltration limit. The time-series data containing wet memory and canopy energy state are input into the trained residual learning network to obtain the residual correction amount. The residual correction amount is then superimposed on the first-order estimation to obtain the predicted value of the total loss term. Pixels that are selected for immediate or priority water replenishment based on the water replenishment priority label are then added to the predicted value of the total loss item to obtain the ecological water replenishment amount for the selected pixels.

[0014] This invention also provides an integrated estimation device for ecological water demand and water replenishment, comprising the following modules: The data processing module is used to perform bias correction or fusion processing on precipitation data, evapotranspiration data and soil moisture content data respectively to obtain a consistent driving field; The water demand determination module is used to determine a reference evapotranspiration based on the uniform driving field and construct a baseline vegetation correction coefficient. Under the condition of no water stress, the baseline vegetation correction coefficient is subjected to physically constrained residual correction to obtain a corrected vegetation correction coefficient. The ecological water demand is determined based on the corrected vegetation correction coefficient and the reference evapotranspiration. The water shortage classification module is used to input the uniform driving field, the ecological water demand and remote sensing vegetation observation information into multiple pre-constructed discrimination gates, perform conjunctive logic operations on the judgment results of multiple discrimination gates, and divide each pixel into multi-level water replenishment priority labels; wherein, the multiple discrimination gates respectively characterize phenological sensitivity, structural degradation degree and immediate physiological limitation; The water replenishment determination module is used to determine the total loss item based on the unified driving field, the ecological water demand, and the water replenishment priority label, and to determine the ecological water replenishment item based on the ecological water demand and the total loss item; wherein, the total loss item is the combined amount of root zone water storage change, deep seepage, lateral outflow and capillary recharge.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the integrated estimation method for ecological water demand and water replenishment as described above.

[0016] This invention provides an integrated estimation method and apparatus for ecological water demand and recharge. It performs deviation correction or fusion processing on precipitation data, evapotranspiration data, and soil moisture content data to generate a consistent driving field. By eliminating systematic deviations between different data sources, it provides an accurate and spatiotemporally consistent input basis for subsequent estimations. Based on this driving field, a reference evapotranspiration is determined and a baseline vegetation correction coefficient is constructed. Then, under conditions without water stress, the baseline vegetation correction coefficient is subjected to physically constrained residual correction to obtain the corrected vegetation correction coefficient. This residual correction process ensures that the correction direction and magnitude conform to the physiological laws of vegetation through physical constraints, enabling the vegetation correction coefficient to overcome the static limitations of traditional empirical tables and dynamically adapt to the actual conditions of different underlying surfaces, thereby significantly improving the accuracy of ecological water demand estimation. This invention inputs the consistent driving field, ecological water demand, and remote sensing vegetation observation information into multiple discriminant gates that characterize phenological sensitivity, structural degradation degree, and immediate physiological constraints, respectively. By performing conjunctive logical operations on the judgment results of each discriminant gate, pixels are divided into multi-level recharge priority labels. This conjunctive logic requires that phenological, structural, and physiological evidence simultaneously meet corresponding conditions to trigger a water replenishment decision. This avoids misinterpretations of short-term meteorological noise or local anomalies by using a single threshold, and prevents missed water shortage assessments due to neglecting evidence from a single dimension. This significantly improves the accuracy and reliability of water shortage classification under complex underlying surface conditions. In the ecological water replenishment estimation process, this invention unifies multiple water inflow and outflow channels, such as root zone water storage changes, deep seepage, lateral outflow, and capillary recharge, into a total loss term. This ensures that water replenishment calculations strictly adhere to the root zone water balance relationship. By introducing this total loss term, water replenishment calculation comprehensively covers all important water loss and replenishment pathways, avoiding the water quantity non-closure problem caused by approximating demand solely by the difference between potential and actual evapotranspiration in existing technologies. This ensures that the estimation results meet water conservation constraints under different root depths and profile conditions, fundamentally eliminating systematic biases caused by neglecting the multi-channel merging effect, and improving the accuracy and applicability of water replenishment decisions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the integrated estimation method for ecological water demand and water replenishment provided by the present invention.

[0019] Figure 2 This is a spatial distribution map of precipitation in a certain watershed from 2016 to 2020 provided by the present invention.

[0020] Figure 3 This is a spatial distribution map of evapotranspiration in a certain watershed from 2016 to 2020 provided by the present invention.

[0021] Figure 4 This is a spatial distribution map of deep soil moisture in a certain watershed from 2016 to 2020, provided by the present invention.

[0022] Figure 5 This is a schematic diagram illustrating the verification results of precipitation, evapotranspiration, and soil moisture products for a certain watershed from 2016 to 2020, provided by this invention.

[0023] Figure 6 This is a spatial distribution map of ecological water demand in a certain watershed from 2016 to 2020, provided by the present invention.

[0024] Figure 7 This is a schematic diagram of the ecological water shortage classification results for a certain watershed from 2016 to 2020, provided by the present invention.

[0025] Figure 8 This is a spatial distribution map of ecological water replenishment in a certain watershed from 2016 to 2020, provided by the present invention.

[0026] Figure 9 This is a schematic diagram of the integrated estimation device for ecological water demand and water replenishment provided by the present invention.

[0027] Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] This embodiment provides an integrated estimation method for ecological water demand and water replenishment. This method achieves seamless connection from water demand calculation to water replenishment decision through multi-source data driving and multi-dimensional logical judgment. Figure 1 This is a flowchart illustrating the integrated estimation method for ecological water demand and water replenishment provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: S110. The precipitation data, evapotranspiration data and soil moisture content data are respectively subjected to bias correction or fusion processing to obtain a uniform driving field.

[0030] Acquire multi-source raw input data covering the target area, including precipitation, evapotranspiration, and soil moisture data. Since these data typically originate from different satellite sensors, ground observation stations, or assimilation models, they suffer from varying spatial resolutions, temporal frequencies, and systematic measurement errors. Therefore, it is necessary to perform bias correction or multi-scale fusion processing on these heterogeneous multi-source data to eliminate mismatches in physical magnitudes and spatiotemporal gaps, thereby constructing a uniform driving field that is fully aligned in time step and spatial grid. The uniform driving field refers to the spatially continuous and spatiotemporally consistent pixel-level data set containing the three key hydrological variables—precipitation, evapotranspiration, and soil moisture—obtained after bias correction or fusion processing.

[0031] S120. Based on the uniform driving field, determine the reference evapotranspiration and construct the baseline vegetation correction coefficient. Under the condition of no water stress, perform physical constraint residual correction on the baseline vegetation correction coefficient to obtain the corrected vegetation correction coefficient. Determine the ecological water demand based on the corrected vegetation correction coefficient and the reference evapotranspiration.

[0032] The baseline vegetation correction factor refers to the initial vegetation coefficient, constructed based on the FAO-56 framework under the assumption of no water stress, using phenological stage division, remote sensing parameters (effective cover, canopy height), and microclimate parameters (wind speed, humidity), without correction for management perception residuals. .

[0033] Physically constrained residual correction refers to the process of using the ratio of actual evapotranspiration to reference evapotranspiration as a management perception label in a water-free stress window, learning the residual between this label and the baseline vegetation coefficient through a time-series model, and then superimposing the learned residual onto the baseline vegetation coefficient to obtain the corrected vegetation coefficient. In this process, physical constraints are imposed on the correction results to ensure that the corrected vegetation coefficient is always within a physically acceptable range.

[0034] Based on the standardized driving field constructed above (such as extracting meteorological driving elements), a basic reference evapotranspiration is calculated. Simultaneously, an initial baseline vegetation correction coefficient is constructed by incorporating vegetation growth rhythms. To ensure this coefficient accurately reflects the water consumption characteristics of local vegetation, an ideal time window without water stress conditions is identified in the historical time series. Within this window, vegetation is in its optimal growth state with sufficient water supply. Using this as a benchmark, the baseline vegetation correction coefficient undergoes physically constrained residual correction to eliminate systematic biases introduced by empirical formulas, resulting in a corrected vegetation correction coefficient that highly reflects local realities. Multiplying this corrected vegetation correction coefficient by the aforementioned reference evapotranspiration yields the spatially continuous ecological water demand.

[0035] S130. Input the uniform driving field, the ecological water demand and remote sensing vegetation observation information into multiple pre-constructed discrimination gates, perform conjunctive logic operation on the judgment results of the multiple discrimination gates, and divide each pixel into multi-level water replenishment priority labels; wherein, the multiple discrimination gates respectively characterize phenological sensitivity, structural degradation degree and immediate physiological limitation.

[0036] After obtaining the water demand, the spatial diagnosis phase of water replenishment urgency begins. Multiple parallel discrimination gates are pre-constructed to characterize vegetation's water shortage crisis in different dimensions, specifically phenological sensitivity, structural degradation, and immediate physiological limitation. Phenological sensitivity assesses whether the current period is a critical growth stage where vegetation is most in need of water; structural degradation assesses whether vegetation has suffered physical or biomass damage due to water shortage; and immediate physiological limitation assesses whether the vegetation's current water metabolism has been inhibited. The uniform driving field, the calculated ecological water demand, and real-time remote sensing vegetation observation information are synchronously input into these discrimination gates. Subsequently, the Boolean values ​​or continuous evaluation results of these multiple discrimination gates are subjected to conjunctive logic operations to obtain the calculation result (i.e., comprehensively considering the concurrent triggering of various indicators). Based on the calculation result, each spatial pixel is precisely divided into multi-level water replenishment priority labels.

[0037] S140. Based on the unified driving field, the ecological water demand, and the water replenishment priority label, determine the total loss item, and determine the ecological water replenishment amount according to the ecological water demand and the total loss item; wherein, the total loss item is the combined amount of root zone water storage change, deep seepage, lateral outflow and capillary recharge.

[0038] For areas deemed requiring intervention, the final water replenishment amount is calculated. Based on a consistent driving field, ecological water demand, and defined water replenishment priority labels, the physical consumption process after water input is comprehensively assessed to determine the total loss item. This total loss item encompasses the combined amounts of water changes in the root zone, deep seepage, lateral outflow, and capillary recharge occurring underground. Finally, the ecological water demand required by the plants themselves is added to the total loss item that must be consumed by environmental transport to calculate the actual amount of ecological water replenishment that needs to be allocated and delivered, thus forming a closed loop of the physical mechanism.

[0039] This invention first performs bias correction or fusion processing on precipitation, evapotranspiration, and soil moisture data to generate a consistent driving field. By eliminating systematic biases between different data sources, it provides an accurate and spatiotemporally consistent input basis for subsequent estimations. Based on this driving field, a reference evapotranspiration is determined and a baseline vegetation correction coefficient is constructed. Then, under conditions without water stress, the baseline vegetation correction coefficient is subjected to physically constrained residual correction to obtain the corrected vegetation correction coefficient. This residual correction process ensures that the correction direction and magnitude conform to the physiological laws of vegetation through physical constraints, enabling the vegetation correction coefficient to overcome the static limitations of traditional empirical tables and dynamically adapt to the actual conditions of different underlying surfaces, thereby significantly improving the accuracy of ecological water demand estimation.

[0040] To address the problem that existing technologies for ecological water shortage classification, which rely solely on a single threshold or data source, are prone to false alarms or missed alarms on complex underlying surfaces, this invention incorporates a unified driving field, ecological water demand, and remotely sensed vegetation observation information as inputs. These inputs represent multiple discriminant gates that characterize phenological sensitivity, structural degradation, and immediate physiological constraints, respectively. By performing a conjunctive logic operation on the judgment results of each discriminant gate, pixels are classified into multi-level water replenishment priority labels. This conjunctive logic requires that phenological, structural, and physiological evidence simultaneously meet corresponding conditions to trigger a water replenishment decision. This avoids false responses to short-term meteorological noise or local anomalies caused by a single threshold, and prevents missed water shortage judgments due to neglecting evidence from a single dimension. This significantly improves the accuracy and reliability of water shortage classification under complex underlying surface conditions.

[0041] To address the problems of existing technologies in ecological water replenishment calculations failing to account for the merging effects of multiple water channels, failing to meet water conservation constraints, and being prone to systematic biases, this invention unifies the estimation of ecological water replenishment by merging multiple water inflow and outflow channels, such as root zone water storage changes, deep seepage, lateral outflow, and capillary recharge, into a total loss term. This ensures that water replenishment calculations strictly adhere to the root zone water balance relationship. By introducing this total loss term, water replenishment calculation comprehensively covers all important water loss and replenishment pathways, avoiding the water volume non-closure problem caused by approximating demand solely by the difference between potential evapotranspiration and actual evapotranspiration in existing technologies. This ensures that the estimation results meet water conservation constraints under different root depths and profile conditions, fundamentally eliminating systematic biases caused by ignoring the merging effects of multiple channels, and improving the accuracy and applicability of water replenishment decisions.

[0042] In one embodiment, the precipitation data in S110 refers to raw, unprocessed satellite precipitation product data, including satellite pixel precipitation data; the precipitation data is biased and corrected in the following manner: In the log-intensity space, the bias relationship between precipitation data from ground meteorological stations and precipitation data from satellite pixels is established, with topographic factors, intra-annual temporal phases, and satellite pixel intensity as covariates. Different regression kernel bandwidths are estimated for the dry and wet seasons respectively. Geographically weighted regression fitting is performed within a moving window using the regression kernel bandwidth corresponding to the season. The regression coefficients are mapped back to grid points to form a precipitation correction surface that distinguishes between the dry and wet seasons. The geographically weighted regression fitting simultaneously outputs pixel-level precipitation uncertainty and quality label. An adaptive threshold is set for the relative change of pixels in the precipitation correction surface that distinguishes between dry and wet seasons. When the relative change of pixels exceeds the adaptive threshold, the correction amplitude of the pixel is truncated to obtain the truncated precipitation correction surface. The truncated precipitation correction surface is weighted by the pixel-level precipitation uncertainty and quality label to form the precipitation data in the uniform driving field.

[0043] Specifically, the bias correction process for precipitation data involves local bias correction of satellite precipitation. The specific steps are as follows: First, a consistent bias relationship is established between measured precipitation from ground weather stations and satellite pixel precipitation in logarithmic intensity space. When establishing this relationship, topographic factors (including altitude and slope), intra-annual temporal phase, and satellite pixel intensity (unit: millimeters) are used as covariates. The intra-annual temporal phase refers to the day of year. Its core relation is: ; In the formula, The logarithmic difference For the prediction function, Altitude The slope. This refers to precipitation data from surface meteorological stations, specifically the precipitation amount at the surface meteorological stations (unit: millimeters). ε represents satellite pixel precipitation data, i.e., satellite pixel precipitation (unit: millimeters); ε is a very small positive number (unit: millimeters) used to ensure the stability of the value on days with zero rain.

[0044] In one embodiment, the prediction function A locally linear prediction function based on geographic weighted regression is used. This applies to the target pixel (i.e., each spatial location). ,definition: in, For target pixel The local regression coefficients are obtained by fitting the data through geographically weighted regression. For precipitation (substitute) The time represents the precipitation recorded at the ground weather station, which is then substituted into the equation. (Precipitation data per satellite pixel).

[0045] To mitigate errors caused by seasonal phase mismatch, independent regression kernel bandwidths were estimated for both the dry and wet seasons. Specifically, the estimation of the regression kernel bandwidth employed adaptive optimization using cross-validation. Leave-one-out cross-validation (LOOCV) was used in the respective site sample sets for the dry and wet seasons. For each candidate bandwidth value, the following procedure was performed: a site sample was removed one by one; a geographically weighted regression was performed on the remaining sites at that candidate bandwidth; and the estimated logarithmic difference y at the removed sites was predicted using covariates (elevation, slope, DOY, and satellite pixel intensity). Then, the cross-validation residuals for that site are calculated, i.e., the difference between the true logarithmic difference y and the estimated value. The difference between (y-) After iterating through all eliminated sites, the squares of the cross-validation residuals for each site are summed to obtain the sum of squares of the cross-validation residuals for that candidate bandwidth. After iterating through all preset candidate bandwidth values, the bandwidth that minimizes the sum of squares of the cross-validation residuals is selected as the optimal regression kernel bandwidth for that season.

[0046] Because precipitation events are sparse in the dry season and the spatial correlation distance between stations is relatively long, its optimal kernel bandwidth is usually larger than that of the wet season. In contrast, wet season precipitation exhibits strong spatial variability and steep local gradients, resulting in a correspondingly smaller optimal kernel bandwidth. This allows for differentiated adaptation to the spatial structure of precipitation in different seasons. Using the optimal regression kernel bandwidth corresponding to the season, geographically weighted regression fitting is performed within a moving window centered on the target pixel. The fitted regression coefficients are then mapped back to spatial grid points, forming a precipitation correction surface that distinguishes between dry and wet seasons.

[0047] During the geographic weighted regression fitting process, the precipitation uncertainty and quality label at the pixel level are output simultaneously. The pixel-level precipitation uncertainty is characterized by the local regression prediction standard error at that pixel, calculated by combining the number of station samples participating in the regression within the pixel's neighborhood, the degree of local collinearity of the covariates, and the local variance of the regression residuals to obtain the prediction confidence interval width. A larger uncertainty value indicates lower correction reliability at that pixel. The quality label is a discrete level identifier assigned to each pixel according to preset rules. The determination rules for the quality label include a comprehensive criterion across multiple dimensions: first, whether the number of effective meteorological stations participating in the geographic weighted regression within the pixel's neighborhood is lower than a preset minimum station number threshold; second, whether the covariate combination corresponding to the pixel falls within the extrapolation region of the training samples (i.e., whether extrapolation regression has occurred); and third, the local regression fit goodness at that pixel (e.g., the local coefficient of determination R0). 2 Whether the quality label is below the preset goodness-of-fit threshold. When any dimensional criterion is triggered, the quality label of the pixel is downgraded. The higher the quality label level, the more reliable the correction result for that pixel.

[0048] To suppress overcorrection in areas of abrupt topographic changes, an adaptive threshold is set for the relative change of pixels in the precipitation correction surface that distinguishes between dry and wet seasons. The relative change of a pixel is defined as the ratio of the difference between the corrected precipitation and the original satellite precipitation before correction, divided by the original satellite precipitation before correction. Physically, it represents the relative change in precipitation caused by the correction operation. A positive relative change indicates that the correction increases precipitation; a negative value indicates that the correction decreases precipitation; and a larger absolute value indicates a more drastic correction.

[0049] The adaptive threshold is determined by statistically analyzing the distribution of relative changes in all pixels within the correction surface and calculating its quantiles (e.g., using the 2.5th and 97.5th percentiles as the lower and upper limits), or by dynamically determining the upper and lower bounds of the threshold using the mean plus or minus a certain number of standard deviations. Once the relative change of a pixel is detected to exceed the adaptive threshold, the correction amplitude of that pixel is immediately truncated to the threshold boundary, thus obtaining the truncated precipitation correction surface.

[0050] Finally, the truncated precipitation correction surface is weighted and fused using pixel-level precipitation uncertainty and quality labeling to form precipitation data in the unified driving field. The specific weighting method is as follows: for each pixel, the reciprocal of its precipitation uncertainty is used as the basic weight; pixels with lower uncertainty (i.e., higher correction confidence) are assigned a larger weight, while pixels with higher uncertainty (lower correction confidence) are assigned a smaller weight. Simultaneously, a modulation coefficient is applied to this basic weight according to the quality label level: pixels with high quality label levels retain or increase their weight, while the weight of pixels with low quality label levels (such as extrapolated areas and sparsely populated areas) is further attenuated. Based on this, the truncated precipitation correction value is weighted and mixed with the original satellite precipitation value: pixels with high quality and low uncertainty are dominated by the correction value; pixels with poor quality and high uncertainty retain more of the original satellite precipitation value to avoid low-quality correction introducing additional errors. After the above weighting processing, the final output pixel-level precipitation data serves as the precipitation driving data in the unified driving field for subsequent ecological water demand calculations.

[0051] In practical applications, taking a specific river basin as an example, this invention performs the complete deviation correction and calibration process described above on satellite precipitation products from 2016 to 2020, resulting in a high-precision calibrated product, such as... Figure 2 The spatial distribution of precipitation is shown. This method significantly eliminates the localized systematic biases of the original satellite products under complex terrain and alternating wet and dry seasons.

[0052] Existing precipitation correction technologies often employ global linear correction, which is highly susceptible to overcorrection or negative precipitation under complex terrain or extreme weather conditions. This invention introduces logarithmic intensity space and covariate regression, enabling it to keenly capture the reshaping effect of micro-topography on precipitation. Simultaneously, the dry-wet season adaptive kernel bandwidth and anomaly truncation mechanism provide robust dual protection, effectively preserving true precipitation extremes while suppressing non-physical distortions caused by regression extrapolation. Finally, the uncertainty weighting mechanism ensures the statistical robustness of the output results, giving the precipitation driving field extremely high fidelity.

[0053] In one embodiment, the evaporation data includes at least three sets of evaporation products with different mechanisms and complementary error structures; the evaporation data is fused in the following manner: At least three sets of evapotranspiration products with different mechanisms and complementary error structures are selected. The random error and cross-correlation of each set of evapotranspiration products are jointly estimated. The optimal weight after correlation correction is obtained by Bayesian method. The optimal weight is adaptively learned in the subdomain of the same seasonal combination in the ecological zone. The Bayesian method outputs the pixel-level evapotranspiration uncertainty while obtaining the optimal weight. The fusion evaporation is obtained by weighting each set of evaporation products according to the optimal weight. It is then determined whether the fusion evaporation exceeds the net radiation supply limit. If it does, the weight of the evaporation product that caused the limit to be exceeded is reduced and the fusion evaporation is reweighted until the fusion evaporation does not exceed the net radiation supply limit, thus obtaining a uniform evaporation. The uniformized evaporation is the evaporation data in the uniformized driving field.

[0054] Specifically, the fusion process of evapotranspiration data involves the consistent integration of multi-source evapotranspiration products. The specific steps are as follows: First, at least three sets of evapotranspiration products with fundamental differences in their physical inversion mechanisms and complementary error structures are rigorously selected. For example, products based on the surface energy balance residual method (such as inverting latent heat flux through remote sensing surface temperature), products based on the Penman-Montes formula (such as calculating evapotranspiration using meteorological driving data and vegetation impedance parameterization), and products based on machine learning or data-driven methods (such as extrapolating statistical models trained by flux station observations to the regional scale) can be selected. These three types of products each have advantages in radiation-driven accuracy, aerodynamic parameterization accuracy, and nonlinear pattern recognition capabilities, and their error sources are independent or weakly correlated, providing a statistical basis for subsequent joint fusion.

[0055] Before fusion, instead of simply assuming that the errors of each product are independent, the random error variances of each set of evaporation products and the cross-correlation of errors between products are jointly estimated. Specifically, an error covariance matrix is ​​constructed using the deviation between the estimated values ​​of each product at the same pixel and time step and the mean of the multi-product set. The diagonal elements of this matrix represent the random error variance of each product itself, while the off-diagonal elements represent the degree of error correlation between any two sets of products. When two sets of products have a positive correlation in error due to sharing similar input driving data or similar physical parameterization schemes, this correlation information will be used in subsequent weight calculations to suppress the joint weights of the two sets, avoiding double counting of systematic bias.

[0056] After obtaining the error covariance matrix, the optimal weights are solved using the Bayesian method. Specifically, the prior error statistical characteristics of each product are used as the prior distribution, and the joint likelihood of the observations of multiple products is used as the data likelihood function. The posterior distribution of the weights of each product is derived using Bayes' theorem, and the posterior expectation is taken as the optimal weight after correlation correction. This weight simultaneously satisfies the non-negativity constraint and the normalization constraint (i.e., the sum of the weights of each product is 1).

[0057] To avoid insufficient spatial adaptability caused by a one-size-fits-all approach, the learning process for the optimal weights is not conducted uniformly across the entire study area, but rather adaptively within subdomains combining "ecological zones + seasons." Specifically, the study area is divided into multiple spatial subdomains based on vegetation type or ecological zone, and different seasons are distinguished along the temporal dimension. The Bayesian weight optimization described above is then performed independently within each "ecological zone × season" subdomain unit. This strategy allows for completely different product weight combinations in wet seasons in forest areas and dry seasons in grassland areas, fully adapting to the spatial and temporal heterogeneity of product accuracy under different ecological types and hydrothermal conditions.

[0058] While solving for the optimal weights, the posterior distribution of the Bayesian method also outputs the pixel-level evaporation uncertainty of each pixel at each time step. This uncertainty is characterized by the standard deviation or confidence interval width of the Bayesian posterior distribution, reflecting the credibility of the fusion result at that pixel, and serving as a confidence factor for subsequent threshold triggering and loss weighting.

[0059] As an example, the pixel-level evaporation uncertainty Characterized by the standard deviation of the Bayesian posterior distribution: ; In the formula, This represents the actual evaporation. For observation vectors of multiple evaporation products, This represents the variance of the posterior distribution; The pixel-level evapotranspiration uncertainty serves as a confidence factor for the discrimination gate characterizing the degree of structural degradation and / or the discrimination gate characterizing immediate physiological limitations in subsequent water shortage assessments. It is used to dynamically adjust the corresponding discrimination thresholds. The adjustment method is as follows: multiply the base threshold by an adjustment factor that monotonically increases with evapotranspiration uncertainty, i.e.: ; In the formula, The adjusted discrimination threshold, Based on the threshold, This is the preset adjustment coefficient.

[0060] Specifically, the pixel-level evapotranspiration uncertainty, as a confidence factor for the discrimination gate characterizing the degree of structural degradation in water shortage assessment, is used to adjust the threshold for determining solar-induced fluorescence anomalies from the original threshold -1.5σ to... σ is the historical standard deviation of the solar-induced fluorescence anomaly in that pixel; and / or, as a confidence factor for a discriminant gate characterizing immediate physiological limitation, it is used to adjust the evapotranspiration efficiency threshold from the original threshold of 0.7 to , where α and β are preset adjustment coefficients.

[0061] The pixel-level evapotranspiration uncertainty is also used as a loss-weighted confidence factor in the replenishment estimation, and is applied to the training of the residual learning network with a 1 / Construct a weighted loss function for the weights.

[0062] The evapotranspiration products are linearly weighted and summed according to the optimal weights obtained through adaptive learning in each subdomain to obtain a preliminary fused evapotranspiration value. Subsequently, an energy consistency gatekeeper mechanism is introduced: the upper limit of the available evapotranspiration energy from net surface radiation for each pixel at the current time step (i.e., the available radiation energy after deducting soil heat flux converted into equivalent evapotranspiration) is calculated, and it is determined whether the preliminary fused evapotranspiration value exceeds this upper limit. If it does not exceed this upper limit, the fused evapotranspiration value is directly retained; if it does exceed this upper limit, it indicates that at least one evapotranspiration product is overestimated at that pixel, the weight of the evapotranspiration product causing the exceedance is reduced, and then the weighted fusion is re-performed based on the adjusted weights. This judgment-adjustment-reweighting process is executed iteratively until the fused evapotranspiration value does not exceed the upper limit of available net radiation, ultimately obtaining a consistent evapotranspiration that satisfies the thermodynamic energy conservation constraint. This consistent evapotranspiration, as evapotranspiration data in the consistent driving field, is used for subsequent ecological water demand calculations.

[0063] In practical applications, taking a certain river basin as an example, this invention performs the above-described complete and consistent integration process on no fewer than three evapotranspiration products from 2016 to 2020 to obtain integrated products, such as... Figure 3The spatial distribution map of evapotranspiration is shown. This method significantly eliminates the problem of unstable accuracy of single evapotranspiration products under specific ecological or seasonal conditions, while ensuring that the fusion results strictly adhere to the principle of energy conservation.

[0064] This invention achieves high-precision consistency processing of evapotranspiration data by fusing Bayesian optimal weights from multi-source evapotranspiration products and applying energy consistency constraints. First, at least three sets of evapotranspiration products with different mechanisms and complementary error structures are selected. By jointly estimating the random errors and cross-correlation of each product, a Bayesian method is used to obtain the optimal weights after correlation correction. These weights adaptively learn within subdomains of the same seasonal combination in the ecological zone, capturing the differences in evapotranspiration characteristics across different underlying surface types and phenological stages, thus avoiding the insufficient adaptability of globally uniform weights in spatially heterogeneous regions. The Bayesian method generates pixel-level uncertainty while outputting the fused evapotranspiration, providing a confidence basis for subsequent water shortage assessment and water quantity calculation. After weighting each set of products according to the optimal weights to obtain the fused evapotranspiration, a net available radiation upper limit is further introduced as an energy conservation constraint. By dynamically detecting and reducing the weights of evapotranspiration products that exceed the limit, it is ensured that the fused evapotranspiration always remains within the physically acceptable energy upper limit. This energy consistency gatekeeping mechanism effectively avoids the defect of violating energy closure due to overestimation of evapotranspiration from a single product. The resulting consistent evapotranspiration not only integrates the complementary advantages of multiple product sources but also strictly satisfies the physical constraints of surface energy balance. Through the above fusion process, the evapotranspiration data constructed by this invention has higher accuracy and reliability on complex underlying surfaces in arid and semi-arid regions, providing a solid data foundation for ecological water demand estimation and water scarcity assessment.

[0065] In one embodiment, the soil moisture content data includes surface soil water, which is the input to the first machine learning model. After a two-stage inversion, the final output is a pixel-level stratified soil moisture content product. The soil moisture content data is fused in the following manner: A two-stage inversion method was used to retrieve the root zone volumetric water content at multiple depths. In the first stage, machine learning was used to establish a nonlinear mapping of surface soil water, vegetation status indicators, energy status indicators, water memory indicators, and static surface factors to the target depth water content, and the mapping results for the first stage were obtained. The second segment calculates the prediction residual of the mapping result of the first segment at the site scale, and obtains a continuous residual surface through spatial interpolation. The continuous residual surface is then superimposed on the mapping result of the first segment to form a pixel-level stratified soil moisture content product. The pixel-level stratified soil moisture content product serves as the soil moisture content data in the uniform driving field.

[0066] Specifically, the fusion processing of soil moisture content data involves a two-stage inversion of stratified soil water and local recalibration based on dense aerial and ground observations. The specific steps are as follows: Stratified inversion of root zone volumetric water content at three key depths of 30 cm, 60 cm, and 100 cm is performed. The first stage employs machine learning, using surface soil water, vegetation status indicators (normalized difference in vegetation index, leaf area index), energy status indicators (surface temperature, albedo), water memory indicators (cumulative precipitation and cumulative evapotranspiration over multiple time windows, in millimeters), and static surface factors (soil texture, digital elevation, slope) as inputs. Ground-based observation station measured profile water content is used as the label to establish a nonlinear mapping of the above multidimensional features to the target depth water content, outputting the predicted water content value for each pixel at each target depth—the first stage mapping result.

[0067] In this process, a gated monotonicity mechanism is introduced to ensure the physical rationality of the inverted profile. Considering the objective existence of the phenomenon of "wetter in shallow layers and drier in deep layers" in riparian zones, shallow groundwater areas, and secondary saline areas, this invention does not adopt a one-size-fits-all hard monotonicity constraint across layers. When the probability of shallow groundwater occurrence is high, the topographic confluence index is large, or the historical profile statistics of the pixel show a significant "shallow wet, deep dry" pattern, the monotonicity constraint is automatically relaxed to allow the unconventional profile structure to be reflected accurately. In conventional areas, a soft monotonicity penalty is used, adding a regularization penalty term to the loss function that shows the monotonic change of profile water content with depth to avoid anti-physical profile jumps.

[0068] The second stage calculates the predicted residuals between the mapping results of the first stage and the actual observations at the station scale, and obtains a continuous residual surface covering the entire study area through ordinary kriging interpolation. This continuous residual surface is then superimposed pixel by pixel onto the mapping results of the first stage to form a pixel-level stratified soil moisture content product after station anchoring correction, which serves as the soil moisture content data in the uniform driving field.

[0069] In areas where conditions permit, high-precision aerial and ground-based dense observation data are further introduced to perform local recalibration of three key hydrological variables: precipitation, evapotranspiration, and soil water. Three-dimensional vegetation structure information and fine topographic data are acquired using airborne lidar to correct local biases in vegetation parameterization in evapotranspiration products and update topographic factors in soil water inversion. High spatiotemporal resolution near-surface meteorological elements and eddy covariance flux data are acquired using mobile weather towers to perform independent local cross-validation and bias correction on precipitation and evapotranspiration products. Multi-depth measured moisture content data are obtained using densified soil profile monitoring equipment to supplement the sample density for station residual calculations in the two-stage inversion.

[0070] After the above-mentioned dense observation data is locally recalibrated, the weights of the calibrated samples are written back into the base data of the uniform driving field. In the precipitation bias correction stage, the weights of the regression training samples of dense stations are added. In the evapotranspiration fusion stage, flux observations are incorporated into the prior constraints of Bayesian weights. In the soil water inversion stage, dense profile observations are added to the station set of Kriging interpolation to further reduce the pixel-level uncertainty of each variable.

[0071] In practical applications, taking a certain watershed as an example, this invention retrieved deep soil moisture products from 2016 to 2020, and the results are as follows: Figure 4 The spatial distribution map of deep soil moisture is shown. 100°0'0"E, 110°0'0"E, and 120°0'0"E are longitude coordinates, representing 100°, 110°, and 120°E, respectively. 40°0'0"N and 50°0'0"N are latitude coordinates, representing 40° and 50°N, respectively. N, W, S, and E represent the four directions: north, west, south, and east, respectively. Soil volumetric water content represents the volume percentage of water in a unit volume of soil. 30 cm, 60 cm, and 100 cm represent the soil moisture inversion results at three depth layers of the soil profile, i.e., 30 cm, 60 cm, and 100 cm below the surface. Figure 4 The spatial distribution of soil volumetric water content at three depths of 30 cm, 60 cm, and 100 cm is shown. Figure 4 As can be seen, the spatial patterns of soil moisture at different depths are generally similar. Humid areas are mainly distributed in the central and eastern parts of the basin and near the riverbanks, while arid areas are concentrated in the northwest and high-altitude mountainous areas. With increasing depth, the spatial heterogeneity of soil moisture gradually decreases. Moisture distribution at the top 30 cm depth is significantly influenced by precipitation and evapotranspiration, resulting in a high degree of patch fragmentation. Moisture distribution at 60 cm depth tends to be continuous, reflecting the combined effects of root uptake and lateral recharge. The spatial variation of moisture at 100 cm depth is more gradual, with deep moisture mainly controlled by soil texture and groundwater depth. The simultaneous provision of differentiated information across the three depths indicates that the stratified soil moisture product constructed in this invention can effectively characterize the spatial patterns of water content in different soil layers, providing a reliable background field of root zone moisture for ecological water demand estimation and water shortage identification. Furthermore, precipitation, evapotranspiration, and soil moisture products from 2016 to 2020 were validated, and the validation results are as follows: Figure 5 As shown. Figure 5 The comparisons shown in the figures are: corrected precipitation versus observed precipitation, fused evapotranspiration versus observed evapotranspiration, and simulated soil moisture content versus observed soil moisture content at three depths of 30 cm, 60 cm, and 100 cm. Figure 5Data shows that the corrected precipitation and observed precipitation have good consistency; the trends of fused evapotranspiration and observed evapotranspiration are highly consistent; the simulated soil moisture content at all three depths can effectively track the dynamic changes of observed soil moisture content, and different depths exhibit differentiated moisture distribution characteristics. The inverted values ​​of each component show good correlation and consistency with ground observations. The soil moisture results can simultaneously provide differentiated information at three depths of 30 cm, 60 cm, and 100 cm, indicating that this invention can accurately characterize key processes of the regional water cycle and provide reliable data support for ecological water demand estimation and water replenishment decisions.

[0072] This invention employs a two-stage inversion framework to fuse stratified soil moisture data, significantly improving the accuracy and spatial continuity of root zone water retrieval. The first stage utilizes a machine learning model to integrate multi-dimensional features such as surface soil water, vegetation state indicators, energy state indicators, water memory indicators, and static surface factors to establish a nonlinear mapping of water content at the target depth. This design fully leverages the implicit vertical water transport information in remote sensing and reanalysis data, enabling the model to capture the complex hydrological connections between the surface and deep layers, avoiding the systematic biases of traditional linear interpolation or empirical formulas on heterogeneous underlying surfaces. The second stage calculates the predicted residuals of the first stage's mapping results at the station scale and uses spatial interpolation to generate a continuous residual surface, extending local ground truth information to the regional scale. This effectively corrects the spatial heterogeneity errors in the first stage results caused by insufficient training samples or model generalization limitations. Superimposing the residual surface onto the first stage's mapping results preserves the machine learning's ability to fit multivariate relationships while incorporating the precise constraints of station observations on local hydrology. By organically combining the above two-stage inversion framework, a high-precision pixel-level stratified soil moisture content product can be constructed.

[0073] In one embodiment, determining the reference evapotranspiration and constructing the baseline vegetation correction coefficient based on the uniform driving field includes: The reference evapotranspiration is determined based on the precipitation and meteorological data in the uniform driving field. The growing season is divided into multiple phenological stages: initial, development, middle and late. The vegetation correction coefficient anchor value for the middle stage is determined by the effective coverage and canopy height obtained by remote sensing. The vegetation correction coefficient anchor value for the middle stage is fine-tuned by combining wind speed and humidity to obtain the middle vegetation correction coefficient. Based on the evolution of phenological phase and coverage between different phenological stages, the vegetation correction coefficients for the initial, developmental, and later stages are continuously interpolated, and the vegetation correction coefficients for each phenological stage are spliced ​​together to form the baseline vegetation correction coefficients covering the entire growing season.

[0074] Specifically, based on bias-corrected precipitation and meteorological data in the uniform driving field, the reference evapotranspiration (denoted as ) is calculated according to the FAO-56 Penman-Montes equation. The precipitation input directly uses the bias-corrected precipitation product from the uniformized driving field, rather than the original satellite precipitation, to ensure the consistency of the driving source. For short-term missing measurements in meteorological elements, a phenological conservation interpolation method is used to fill the gaps. This method maintains seasonal phase stability during interpolation, avoiding deviations from the current phenological stage's climatological characteristics due to interpolation values, thereby preventing anchor point drift in subsequent vegetation correction coefficient reconstruction.

[0075] The growing season is divided into four phenological stages: initial, developmental, mid-term, and late-term. The vegetation correction coefficient (denoted as ) is reconstructed using a combination of phenological anchor points and microclimate fine-tuning. The time trajectory of ). For the most vigorous intermediate stage of growth, the effective coverage obtained by remote sensing (denoted as ). ) and canopy height (denoted as The anchor value for the vegetation correction factor in the mid-term stage is determined by the following formula: In the formula, This is the mid-term vegetation correction factor. This is the near-lower limit coefficient for bare land. This is the vegetation correction factor for full coverage. For pixel geometric coverage, The effective coverage is calculated after adjustments based on the solar altitude angle and sensor geometry. The canopy height is represented by this formula. This formula ensures that the medium-term anchor point value will not exceed the physically reasonable range due to overestimation of remote sensing coverage by minimizing both geometric and effective coverage. The bare land near-lower limit coefficient refers to the minimum vegetation coefficient value representing bare land or a surface with extremely low cover.

[0076] After determining the intermediate anchor point, the wind speed at a height of two meters (denoted as...) is taken into account. ) and minimum relative humidity (denoted as By fine-tuning the microclimate of the total vegetation cover coefficient, the mid-term vegetation correction coefficient is obtained, and its adjustment formula is as follows: In the formula, This is the full coverage factor for a given canopy height under standard wind and humidity conditions. This formula reflects the effect of increased evapotranspiration due to enhanced canopy turbulence exchange when the actual wind speed is 2 m / s higher than the standard conditions, and the effect of reduced atmospheric evapotranspiration demand when the minimum relative humidity is 45% higher than the standard conditions. The fine-tuning range increases with increasing canopy height.

[0077] Based on the evolution of phenological phase and vegetation cover across different phenological stages, vegetation correction coefficients were continuously interpolated for the initial, developmental, and late-stage phases. The coefficients for the developmental stage smoothly transitioned from the late initial stage value to the mid-stage anchor value, while the coefficients for the late-stage stage smoothly decreased from the mid-stage anchor value to the end of the growing season value. The interpolation curves continuously evolved with the actual changes in vegetation cover and phenological phase. The vegetation correction coefficients for each phenological stage were then stitched together to form a baseline vegetation correction coefficient curve covering the entire growing season, serving as the basis for subsequent residual correction.

[0078] Ultimately, potential evaporation (denoted as...) (Based on vegetation correction factor) Compared with reference evaporation The product is determined as follows: In practical applications, taking a certain watershed as an example, this invention, based on consistent driving field data from 2016 to 2020, completed the daily calculation of reference evapotranspiration and the phenological reconstruction of baseline vegetation correction coefficients.

[0079] Existing technologies often directly refer to static table values ​​recommended by the FAO for vegetation correction coefficients, ignoring the dynamic changes in regional microclimate conditions and specific vegetation growth. This leads to a systematic underestimation of evapotranspiration demand in semi-arid regions with high wind speeds or dry air, and potentially an overestimation in humid regions. This invention achieves high-precision dynamic reconstruction of baseline vegetation correction coefficients through phenological stage division and the synergistic use of multi-source remote sensing parameters. First, effective cover and canopy height are used to determine mid-term anchor points, enabling vegetation coefficients to respond to actual changes in vegetation canopy structure and avoiding static biases caused by traditional empirical table assignments. Second, wind speed and humidity are introduced to perform microclimate correction on mid-term coefficients, improving the adaptability of vegetation coefficients to local meteorological conditions and making potential evapotranspiration estimates closer to the actual energy exchange process of the underlying surface. Finally, continuous interpolation is performed on the initial, developmental, and late-stages based on the evolution of phenological phases and cover, ensuring the temporal continuity and phenological consistency of vegetation coefficients throughout the growing season. The baseline vegetation correction coefficient constructed in this way not only retains the physical interpretability of the FAO framework, but also has the dynamic response capability driven by remote sensing, providing a high-quality reference field for subsequent residual correction based on management perception, thereby significantly improving the accuracy and reliability of ecological water demand estimation.

[0080] In one embodiment, the step of performing physically constrained residual correction on the baseline vegetation correction coefficient under water stress-free conditions to obtain a corrected vegetation correction coefficient, and determining the ecological water requirement based on the corrected vegetation correction coefficient and the reference evapotranspiration, includes: Under the no-moisture stress window, the ratio of actual evapotranspiration in the uniform driving field to the reference evapotranspiration is used as the management perception tag; wherein, the criterion for the no-moisture stress window is that the ratio of cumulative precipitation to reference evapotranspiration within a set period is not less than a preset threshold and that there are no fewer than a preset number of effective rainfalls during the period. A deep learning model is used to learn the residual between the management perception label and the baseline vegetation correction coefficient, and the residual is superimposed on the baseline vegetation correction coefficient to obtain the corrected vegetation correction coefficient. The ecological water requirement is obtained by multiplying the corrected vegetation correction coefficient by the reference evapotranspiration. The physical constraints applied in the physically constrained residual correction include: the corrected vegetation correction coefficient is limited to physically acceptable upper and lower bounds; the corrected vegetation correction coefficient maintains a non-negative monotonic response to the reference evapotranspiration and a non-negative monotonic response to the leaf area index; and the ecological water demand obtained after correction does not exceed the upper limit of evapotranspiration energy available from net radiation.

[0081] Specifically, the aforementioned effective rainfall refers to a rainfall event in which the amount of a single rainfall event exceeds a certain preset threshold. This threshold is usually set as the minimum amount of rainfall that can be effectively absorbed and utilized by the root system of vegetation or that can significantly affect soil moisture conditions. The preset threshold can be set to 5 mm, or to 0.5 times the daily reference evapotranspiration.

[0082] To incorporate the shaping effects of long-term management practices such as community density and variety renewal on canopy roughness, leaf area index, and rewetting behavior, a water-stress-free window was first automatically identified in the historical time series of a uniform driving field. The criteria for this window were a four-day cumulative precipitation to reference evapotranspiration ratio of not less than 0.8, with at least two effective rainfall events occurring during this period. Within this window, the vegetation is in its optimal growth state with sufficient water supply, and evapotranspiration is not inhibited by soil moisture deficit; therefore, actual evapotranspiration accurately reflects the potential water consumption capacity of the vegetation under current management conditions.

[0083] Under the moisture-free stress window, the actual evapotranspiration in the uniform driving field (denoted as...) ) and reference evapotranspiration (denoted as The ratio of ) is defined as the management perception label (denoted as ) ): This invention does not directly learn the complete vegetation correction coefficient, but only learns the management perception tags. Baseline vegetation correction factor The small residuals between (denoted as) This is to maintain the inheritance of baseline phenological structure and improve learning stability. The residual is defined as: In the formula, represents the input time series sequence within the backtracking window from the tw-th time step to the t-th time step, including time series features such as reference evapotranspiration component and leaf area index, w is the length of the backtracking window, and θ is the model parameter. This represents a time step. The present invention employs a lightweight deep learning model to learn the above mapping relationship. This deep learning model uses a two-layer multi-head self-attention encoder as its backbone to map the multi-dimensional temporal input within the backtracking window to the residual correction amount of the current time step.

[0084] residual Overlay on baseline vegetation correction factor The corrected vegetation correction factor (denoted as ) is obtained above. ): During the above residual correction process, the following physical constraints were applied to ensure the scientific rationality of the results. Corrected vegetation correction coefficient. The vegetation correction coefficient is constrained within physically acceptable upper and lower bounds to prevent negative values ​​or outliers exceeding the physiological limits of vegetation. The corrected vegetation correction coefficient maintains a non-negative monotonic response to the reference evapotranspiration and also a non-negative monotonic response to the leaf area index; that is, the coefficient will not decrease in the opposite direction as evapotranspiration potential increases and vegetation growth improves. The corrected ecological water requirement does not exceed the upper limit of evapotranspiration energy available from net radiation, ensuring energy conservation from a thermodynamic perspective.

[0085] The corrected vegetation correction coefficient Multiply by reference evaporation rate The ecological water demand after taking into account management effects (denoted as ) is obtained. This ecological water demand comprehensively reflects the actual potential water consumption demand of vegetation under the current phenological stage, current management status, and current meteorological conditions.

[0086] In the formula, This represents the ecological water demand at time step t (e.g., day t). This represents the vegetation correction factor after correction at time step t. This represents the reference evaporation at time step t.

[0087] In practical applications, this invention, taking a certain watershed as an example, calculated the ecological water demand from 2016 to 2020, and the results are as follows: Figure 6 The spatial distribution map of ecological water demand is shown.

[0088] In existing technologies, the determination of vegetation correction coefficients typically relies on static table lookups or pure empirical regression, which fails to reflect the continuous shaping effect of long-term management factors such as community density adjustments and variety replacement on canopy evapotranspiration characteristics, leading to significant biases in water demand estimation in actively managed areas. This invention effectively separates management effects from complex water stress signals by extracting management-aware labels within a strictly selected water-stress-free window. The strategy of learning only small residuals of baseline coefficients rather than reconstructing complete coefficients preserves the structural stability of the FAO phenological framework while providing the model with the flexibility to capture local management differences. The simultaneous application of multiple physical constraints transforms the deep learning model from an uninterpretable black box into a gray box constrained by physical laws, effectively preventing abnormal coefficient fluctuations and anti-physical predictions caused by data noise, and significantly improving the robustness and scientific credibility of ecological water demand estimation in cross-regional and cross-year applications.

[0089] In one embodiment, the present invention proposes a three-gate, four-level pixel-level discrimination system that organically combines phenological, structural, and physiological evidence to overcome the shortcomings of single thresholds, which are prone to false alarms and false negatives. The multiple discrimination gates include a gate characterizing phenological sensitivity, a gate characterizing the degree of structural degradation, and a gate characterizing immediate physiological limitations. The process involves inputting the unified driving field, the ecological water demand, and remote sensing vegetation observation information into multiple pre-constructed discrimination gates, performing a conjunction logic operation on the judgment results of the multiple discrimination gates, and dividing each pixel into multi-level water replenishment priority labels, including: The discrimination gate characterizing phenological sensitivity delineates the growing season phase based on the ratio of the current leaf area index of a pixel to the peak leaf area index of that pixel in the current year. A staged trigger threshold is smoothly generated based on the ratio. The leaf area index in the remote sensing vegetation observation information is input into the discrimination gate characterizing phenological sensitivity. When the ratio exceeds the staged trigger threshold, it is determined to be a phenological sensitive period. The discrimination gate characterizing the degree of structural degradation calculates the leaf area index anomaly and the solar induced fluorescence anomaly after removing the seasonal climate state. The leaf area index and solar induced fluorescence in the remote sensing vegetation observation information are input into the discrimination gate characterizing the degree of structural degradation. When the leaf area index anomaly is continuously negative for more than a preset number of days and the deviation of the solar induced fluorescence anomaly exceeds a preset standard deviation multiple, it is determined that the structure is weakening; otherwise, it is determined that the structure is stable or strengthening. The discrimination gate for characterizing immediate physiological limitation calculates the relative available water in the root zone based on the stratified soil moisture content in the uniform driving field, and calculates the evapotranspiration efficiency based on the actual evapotranspiration in the uniform driving field and the ecological water requirement. The relative available water in the root zone is characterized by the normalized position of the current volumetric water content relative to field capacity and wilting point water content, and the evapotranspiration efficiency is characterized by the ratio of actual evapotranspiration to ecological water requirement. When the relative available water in the root zone is lower than a preset available water threshold and the evapotranspiration efficiency is lower than a preset efficiency threshold, it is determined to be physiologically limited. When the relative available water in the root zone or the evapotranspiration efficiency is close to but does not fall below their respective preset thresholds, it is determined to be physiologically close to limited. The multi-level water replenishment priority label includes four levels: immediate water replenishment, priority water replenishment, observation and review, and water replenishment not recommended. When all three discrimination gates simultaneously determine that it is a phenological sensitive period, structural weakening, and physiological limitation, it is judged as immediate water replenishment. When it is judged as structural weakening and in a phenological sensitive period, and simultaneously judged as physiologically approaching limitation, it is judged as priority water replenishment. When it is judged as a phenological sensitive period, but the structure is judged as stable or strengthening and the evapotranspiration efficiency has not fallen below the efficiency threshold, it is judged as observation and review. If the above conditions are not met, water replenishment is not recommended.

[0090] Specifically, the discrimination gate characterizing phenological sensitivity uses the relative leaf area index (LAI) to depict the phase of a pixel within its growing season. The LAI from remotely sensed vegetation observations is input into this gate, and the ratio of the pixel's current LAI to the peak LAI for that pixel in the current year is calculated; this ratio is the relative LAI (denoted as LAI). ): In the formula, t is the time step, and LAI(t) is the leaf area index at the t-th time step. This represents the peak leaf area index for that pixel in that year. Based on... The generation of phased trigger thresholds is smoothed, making the greening and senescence stages more prone to triggering (because vegetation is most sensitive to water stress during these two stages), while being more cautious during the peak stage (because vegetation has a well-developed root system and stronger drought resistance during the peak stage). When the threshold for triggering this stage is exceeded, the pixel is determined to be in a phenological sensitive period.

[0091] After removing seasonal climatology from the discriminant gate characterizing the degree of structural degradation, the anomalous amounts of leaf area index (LAI) and solar-induced fluorescence (SEL) were observed simultaneously. LAI and SEL from remotely sensed vegetation observations were input into the discriminant gate, and the following calculations were performed: In the formula, This refers to the leaf area index anomaly, which is the deviation between the current leaf area index and its corresponding annual climatological mean. This refers to the solar-induced fluorescence anomaly, which is the deviation between the current solar-induced chlorophyll fluorescence and its corresponding annual climatological mean. Leaf area index, and These represent the spatial-temporal climatological mean values ​​of leaf area index and solar-induced chlorophyll fluorescence at the corresponding intra-annual date, respectively. Sun-induced chlorophyll fluorescence. When A negative value for at least 8 consecutive days indicates a continuous shrinkage of the physical structure of the vegetation canopy, and If the deviation is no greater than -1.5σ (σ is the historical standard deviation of the solar-induced fluorescence anomaly of the pixel), it indicates that the photosynthetic activity has declined significantly and synchronously. Both pieces of evidence jointly determine that the pixel is structurally weakened; otherwise, it is determined that the structure is stable or strengthening.

[0092] A diagnostic method combining a discriminant gorge characterizing immediate physiological limitation with relative root zone water availability and evapotranspiration efficiency is employed. Relative root zone water availability (denoted as ) is calculated based on stratified soil moisture content within a uniform driving field. Evapotranspiration efficiency is calculated based on actual evapotranspiration and ecological water demand in a uniformly driven field. The relative available water in the root zone is characterized by the normalized position of the current volumetric water content relative to field capacity and wilting point water content. In the formula, This represents the current volumetric water content of the root zone. The moisture content at the wilting point. This refers to field holding capacity. Evapotranspiration efficiency is calculated from actual evapotranspiration. Ecological water demand The ratio is represented by this. When Not greater than 0.45 and A value not exceeding 0.7 indicates a severe shortage of available water in the root zone and significant inhibition of vegetation transpiration, classifying it as physiological limitation. or When a value approaches but does not fall below its respective preset threshold, it is determined to be physiologically restricted.

[0093] The results of the three discrimination gates were combined using logical operations to classify each pixel into four levels of water replenishment priority labels. When all three discrimination gates simultaneously determine that the pixel is in a phenologically sensitive period, structurally weakened, and physiologically restricted, the triple evidence indicates that the pixel is in the most critical ecological state, and water replenishment is deemed necessary immediately. When the pixel is determined to be structurally weakened and in a phenologically sensitive period, and simultaneously determined to be physiologically restricted, it indicates that although the pixel has not completely fallen below the physiological red line, the degradation trend is clear and it is in a critical phenological window, and water replenishment is given priority. When the pixel is determined to be in a phenologically sensitive period but structurally stable or strengthening, and the evapotranspiration efficiency has not fallen below the threshold, it indicates that although the pixel is in a sensitive period, the vegetation itself is still in good condition, and observation and verification are required. Pixels that do not meet the above conditions are deemed not to require water replenishment.

[0094] In practical applications, this invention, taking a certain watershed as an example, conducted ecological water shortage classification from 2016 to 2020. The classification results are as follows: Figure 7 As shown.

[0095] Existing technologies for ecological water replenishment decisions typically rely on a single soil moisture threshold or empirical calendar cycles, which can easily lead to widespread false alarms in drought-stricken areas when the actual vegetation can still withstand the conditions, or to cases where severely damaged vegetation is masked by mean values ​​and goes unreported. This invention constructs a three-gate, four-level discrimination system that uses cross-validation diagnosis from three independent dimensions: phenological timing sensitivity, canopy structure degradation, and immediate physiological limitation, effectively overcoming the limitations of single-indicator discrimination. The phenological gate's staged trigger threshold design allows for differentiation of the varying degrees of urgency despite the same water shortage at different growth stages. The structural gate simultaneously introduces two anomalies, LAI and SIF, and sets a time persistence threshold, effectively filtering out instantaneous false signals caused by cloud pollution or sensor noise. The physiological gate jointly determines soil moisture supply and vegetation evapotranspiration demand, avoiding the drawback of relying solely on soil moisture as a single indicator, which cannot distinguish between the fundamentally different situations of dry soil with drought-resistant vegetation and dry soil with damaged vegetation. The four-tiered classification mechanism enables differentiated and precise allocation of limited water resources, prioritizing the allocation of the most scarce water replenishment resources to the pixels with the highest ecological benefits.

[0096] In one embodiment, the present invention merges multiple water channels that have been stably decomposed into a total loss term, and uses a hybrid architecture of physical kernel + residual learning for accurate estimation. The step of determining the total loss term based on the consistent driving field, the ecological water demand, and the water replenishment priority label, and determining the ecological water replenishment amount based on the ecological water demand and the total loss term, includes: A supervisory quantity is constructed within a wet window, and a pre-constructed residual learning network is trained based on this supervisory quantity to obtain a trained residual learning network. The supervisory quantity is the precipitation data in the uniform driving field minus the ecological water demand. The residual learning network is a deep learning network that receives time-series data containing wet memory and canopy energy state and outputs residual corrections. A water conservation residual penalty term and a reservoir capacity upper limit constraint term are added to the training objective function. The water conservation residual penalty term constrains the predicted value of the total loss term to satisfy the root zone water balance closure, and the reservoir capacity upper limit constraint term constrains the root zone water content after water replenishment to not exceed the reservoir capacity upper limit corresponding to field capacity. The soil and topographic parameters in the uniform driving field are input into a pre-constructed physical prior model to obtain a first-order estimate of the total loss term; wherein, the physical prior model is established based on soil water holding capacity, water conductivity, topographic runoff characteristics and infiltration limit; The time-series data containing wet memory and canopy energy state are input into the trained residual learning network to obtain the residual correction amount. The residual correction amount is then superimposed on the first-order estimation to obtain the predicted value of the total loss term. Pixels that are selected for immediate or priority water replenishment based on the water replenishment priority label are then added to the predicted value of the total loss item to obtain the ecological water replenishment amount for the selected pixels.

[0097] Specifically, Establish the root zone water balance equation at the root zone scale: In the formula, For changes in root zone water storage, For precipitation, For ecological water replenishment, For shallow groundwater capillary recharge, For ecological water demand, For deep leakage, It is a lateral outward discharge. Based on this, the total loss term is defined. : This total loss term combines four water channels—root zone water storage changes, deep seepage, lateral outflow, and capillary recharge—into a unified physical quantity, characterizing all water transport losses during recharge, excluding vegetation evapotranspiration. Under the recharge window, the core relationship of ecological recharge is: The wet window refers to a period of sufficient rainfall, good soil moisture conditions, and minimal water stress on vegetation. During this period, the combined effects of changes in root zone water storage, deep seepage, lateral drainage, and capillary recharge (total loss item) can be considered. The difference between precipitation and ecological water demand can be approximated for monitoring. Monitoring parameters are constructed during historical wet seasons (i.e., periods of abundant precipitation and significant water balance signals). This monitoring quantity is used to standardize precipitation data in the driving field. Subtract ecological water demand , This supervisory quantity is approximately equal to the true value of the total loss term under the wetted window, providing indirect training labels for the residual learning network that do not rely on direct measurements.

[0098] The residual learning network is a deep learning network that receives time-series data including wetting memory and canopy energy state and outputs residual corrections. Wetting memory refers to time-series variables reflecting the previous water accumulation status of the soil and vegetation system, including cumulative precipitation, cumulative evapotranspiration, temporal changes in volumetric water content of each layer in the root zone over multiple time windows, and the number of days between precipitation events and the current time step. These variables characterize the historical dynamic process of soil water retention and dissipation from previous precipitation or recharge events, providing the network with historical information on the water supply side. Canopy energy state refers to time-series variables reflecting the current energy balance and physiological activity level of the vegetation canopy, including net surface radiation, surface temperature, reference evapotranspiration, leaf area index, and solar-induced fluorescence. These variables characterize the canopy's evapotranspiration potential and photosynthetic activity level driven by radiation, providing the network with real-time status information on the water demand side. Both types of features are considered in the backtracking window. arrive The internal time-step organization is a multivariate time series, which are then input into the encoder of the residual learning network.

[0099] The backbone of the network employs a two-layer multi-head self-attention encoder, with channel width and the number of attention heads controlled within a feasible deployment range. Location representation uses a hybrid form of learnable absolute location vectors and annual periodic sine and cosine functions to ensure accurate phenological phase identification even under short-sequence input conditions. To strike a balance between intra-class sharing and sample diversity, the network introduces a sample-level lightweight adaptation mechanism, injecting land cover type, ecological zone embedding, and pixel adaptation vectors into each layer through Adaptive Layer Normalization (AdaLN). In the formula, Indicates that under a given condition vector Under the condition of, for intra-layer features The output features after conditional adaptive layer normalization. Features within the layer For layer normalization, It is a conditional vector (obtained by concatenating the land cover type, ecological zone and cell adaptation vector). and These are learnable scaling and translation parameters. This conditionalization mechanism allows the same network to automatically adjust its internal feature distribution when processing pixels in different ecological zones and different land cover types, without needing to train a separate model for each land cover.

[0100] The training objective function consists of three parts. The first part is the principal regression error term, which is the sum of the predicted values ​​and the supervised values ​​of the residual learning network. The mean squared error between the two values ​​drives the network to fit the true magnitude of the total loss term. The second part is the water conservation residual penalty term, which applies the predicted value of the total loss term. Substitute into the root zone water balance equation and calculate the closure residual ε. close The closure residual is defined as the difference between the sum of all input terms (precipitation, ecological recharge, and shallow groundwater capillary recharge) and the sum of all output terms (ecological water demand, root zone water storage changes, deep seepage, and lateral discharge) in the root zone water balance equation. Ideally, this difference should be strictly zero, indicating a complete closure of the water balance. The water conservation residual penalty term uses the square of the closure residual as the penalty amount in the objective function, automatically increasing the penalty for samples with poor closure, driving the network to actively reduce the closure residual during training so that the prediction results satisfy the water balance closure as much as possible. The third part is the reservoir capacity upper limit constraint term. When the predicted ecological recharge causes the root zone water content to exceed the reservoir capacity upper limit corresponding to the field capacity, a penalty is generated. In the network output layer, the total predicted loss term is limited to the range of a single window at most up to the field capacity, physically preventing excessive recharge.

[0101] By inputting the soil and topographic parameters from the uniform driving field into a pre-built physical prior model, a first-order estimate of the total loss term is obtained (denoted as...). This physical prior model is established based on soil water-holding capacity, hydraulic conductivity, topographic runoff characteristics, and infiltration limit, using a set of static parameters. (Including soil texture, topographical classification, etc.) and current hydrothermal forcing The model takes as input (including reference evapotranspiration, net radiation, wind speed, and humidity) and outputs a physical baseline estimate of the total loss term. Internally, the physical prior model decomposes the total loss term into multiple physical components, calculates them separately, and then superimposes them. The deep percolation component is based on Darcy's law, calculating gravity-driven downward percolation using soil saturated hydraulic conductivity, root zone bottom hydraulic gradient, and soil texture parameters. The lateral outflow component estimates horizontal water loss based on topographic runoff index and surface slope, combined with soil lateral hydraulic conductivity. The capillary recharge component estimates bottom-up water replenishment based on the capillary rise height between shallow groundwater depth and root zone bottom, combined with soil capillary hydraulic conductivity. The root zone water storage change component is calculated as the product of the difference in root zone volumetric water content between different time steps and the effective root zone depth. The superposition of these components constitutes a first-order estimate of the total loss term, capturing the main characteristics of water transport determined by soil and topographic background properties.

[0102] Time-series data including wet memory and canopy energy state Input the trained residual learning network to obtain the residual correction amount. This correction compensates for the nonlinear response and spatial heterogeneity that the physical prior model failed to characterize. The residual correction is then superimposed onto the first-order estimate to obtain the predicted value of the total loss term: In the formula, This is the predicted value for the total loss item. The first-order order estimate given for the physical kernel. For static parameter set, For the current hydrothermal forcing, This refers to the residual correction value output by the residual learning network. The input sequence within the backtracking window, For time steps, This is the length of the backtracking window.

[0103] Pixels are selected based on their water replenishment priority tags, determining whether they require immediate or priority water replenishment. The ecological water demand of these selected pixels is then calculated. Add the predicted value of the total loss item Ecological water replenishment : Pixels deemed to be under observation and not recommended for water replenishment are excluded from the current water replenishment calculation, thereby concentrating limited water resources on areas with the highest ecological benefits and the greatest risk of degradation.

[0104] Meanwhile, this deep learning framework, along with the aforementioned method for vegetation coefficient residuals, The learning network shares a unified architecture. When learning from vegetation coefficient residuals, the network monitors the backpropagation gradient in the feedforward layer, and when... or When a trend is observed, the regularization strength of the sample is automatically increased and a small parameter backoff is performed to ensure that the corrected vegetation correction coefficient maintains a non-negative monotonic response to the reference evapotranspiration and leaf area index. For total loss estimation, the network imposes a hard constraint on the upper limit of the storage capacity at the output layer, while simultaneously penalizing samples with poor closure driven by water conservation residuals. The two tasks share the encoder backbone and conditionalization mechanism, with task-specific configurations only applied to the output head and constraint policies.

[0105] In practical applications, this invention, taking a certain watershed as an example, calculated the ecological water replenishment from 2016 to 2020, and the results are as follows. Figure 8 The spatial distribution map of ecological water replenishment is shown.

[0106] Existing technologies for estimating ecological water replenishment typically focus only on vegetation evapotranspiration demand, neglecting water transport losses in the soil profile due to infiltration, lateral drainage, and capillary recharge. This results in insufficient actual replenishment to maintain the target moisture content in the rhizosphere. While some schemes consider water loss, they rely on purely empirical coefficients, failing to adapt to the spatial heterogeneity of loss channels under different soil textures and topographic conditions. This invention establishes a complete water replenishment calculation framework based on physical mechanisms by merging multiple water channels in the rhizosphere water conservation equation into a unified total loss term. The hybrid estimation architecture of physical kernel + residual learning allows the physical prior model to capture the main loss magnitude determined by the background properties of soil and topography, while the residual learning network only compensates for fine-grained nonlinear biases not characterized by the physical model, avoiding the overfitting risk of purely data-driven methods when training samples are insufficient. The dual-track embedding of the water conservation penalty term and the reservoir capacity upper limit constraint ensures physical consistency from within the network rather than at the post-processing level. The conditional adaptive normalization mechanism enables a single network to adapt to multiple ecological zones and land cover types, avoiding the high cost of modeling each type of land surface separately. Pre-screening with water replenishment priority labels ensures that water replenishment calculations are performed only on pixels that truly require intervention, achieving precise water allocation and efficient utilization.

[0107] The integrated estimation device for ecological water demand and water replenishment provided by the present invention is described below. The integrated estimation device for ecological water demand and water replenishment described below and the integrated estimation method for ecological water demand and water replenishment described above can be referred to in correspondence.

[0108] The integrated estimation device for ecological water demand and water replenishment provided by this invention refers to... Figure 9 As shown, it includes the following modules: Data processing module 210 is used to perform deviation correction or fusion processing on precipitation data, evapotranspiration data and soil moisture content data respectively to obtain a consistent driving field; The water demand determination module 220 is used to determine a reference evapotranspiration based on the uniform driving field and construct a baseline vegetation correction coefficient. Under the condition of no water stress, the baseline vegetation correction coefficient is subjected to physically constrained residual correction to obtain a corrected vegetation correction coefficient. The ecological water demand is determined based on the corrected vegetation correction coefficient and the reference evapotranspiration. The water shortage classification module 230 is used to input the uniform driving field, the ecological water demand and remote sensing vegetation observation information into multiple pre-constructed discrimination gates, perform conjunctive logic operation on the judgment results of multiple discrimination gates, and divide each pixel into multi-level water replenishment priority labels; wherein, the multiple discrimination gates respectively characterize phenological sensitivity, structural degradation degree and immediate physiological limitation; The water replenishment determination module 240 is used to determine the total loss item based on the unified driving field, the ecological water demand and the water replenishment priority label, and to determine the ecological water replenishment item according to the ecological water demand and the total loss item; wherein, the total loss item is the combined amount of root zone water storage change, deep seepage, lateral outflow and capillary recharge.

[0109] Figure 10 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 10 As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communication bus 340. The processor 310, communications interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions from the memory 330 to execute an integrated estimation method for ecological water demand and replenishment.

[0110] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0112] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An integrated estimation method for ecological water demand and water replenishment, characterized in that, include: The precipitation data, evapotranspiration data, and soil moisture content data were respectively subjected to bias correction or fusion processing to obtain a consistent driving field; Based on the uniform driving field, a reference evapotranspiration is determined and a baseline vegetation correction coefficient is constructed. Under the condition of no water stress, the baseline vegetation correction coefficient is subjected to physically constrained residual correction to obtain the corrected vegetation correction coefficient. The ecological water requirement is determined based on the corrected vegetation correction coefficient and the reference evapotranspiration. The uniform driving field, the ecological water demand, and remote sensing vegetation observation information are input into multiple pre-constructed discrimination gates. The judgment results of the multiple discrimination gates are subjected to conjunctive logic operations to divide each pixel into multi-level water replenishment priority labels. Among them, the multiple discrimination gates respectively characterize phenological sensitivity, structural degradation degree, and immediate physiological limitation. Based on the unified driving field, the ecological water demand, and the water replenishment priority label, a total loss item is determined, and an ecological water replenishment amount is determined according to the ecological water demand and the total loss item; wherein, the total loss item is the combined amount of root zone water storage change, deep seepage, lateral outflow, and capillary recharge.

2. The integrated estimation method for ecological water demand and water replenishment according to claim 1, characterized in that, The precipitation data includes satellite pixel precipitation data; bias corrections are performed on the precipitation data using the following methods: In the log-intensity space, the bias relationship between precipitation data from ground meteorological stations and precipitation data from satellite pixels is established, with topographic factors, intra-annual temporal phases, and satellite pixel intensity as covariates. Different regression kernel bandwidths are estimated for the dry and wet seasons respectively. Geographically weighted regression fitting is performed within a moving window using the regression kernel bandwidth corresponding to the season. The regression coefficients are mapped back to grid points to form a precipitation correction surface that distinguishes between the dry and wet seasons. The geographically weighted regression fitting simultaneously outputs pixel-level precipitation uncertainty and quality label. An adaptive threshold is set for the relative change of pixels in the precipitation correction surface that distinguishes between dry and wet seasons. When the relative change of pixels exceeds the adaptive threshold, the correction amplitude of the pixel is truncated to obtain the truncated precipitation correction surface. The truncated precipitation correction surface is weighted by the pixel-level precipitation uncertainty and quality label to form the precipitation data in the uniform driving field.

3. The integrated estimation method for ecological water demand and water replenishment according to claim 1, characterized in that, The evaporation data includes at least three sets of evaporation products with different mechanisms and complementary error structures; the evaporation data is fused using the following method: At least three sets of evapotranspiration products with different mechanisms and complementary error structures are selected. The random error and cross-correlation of each set of evapotranspiration products are jointly estimated. The optimal weight after correlation correction is obtained by Bayesian method. The optimal weight is adaptively learned in the subdomain of the same seasonal combination in the ecological zone. The Bayesian method outputs the pixel-level evapotranspiration uncertainty while obtaining the optimal weight. The fusion evaporation is obtained by weighting each set of evaporation products according to the optimal weight. It is then determined whether the fusion evaporation exceeds the net radiation supply limit. If it does, the weight of the evaporation product that caused the limit to be exceeded is reduced and the fusion evaporation is reweighted until the fusion evaporation does not exceed the net radiation supply limit, thus obtaining a uniform evaporation. The uniformized evaporation is the evaporation data in the uniformized driving field.

4. The integrated estimation method for ecological water demand and water replenishment according to claim 1, characterized in that, The soil moisture content data includes surface soil water; the soil moisture content data is fused using the following method: A two-stage inversion method was used to retrieve the root zone volumetric water content at multiple depths. In the first stage, machine learning was used to establish a nonlinear mapping of surface soil water, vegetation status indicators, energy status indicators, water memory indicators, and static surface factors to the target depth water content, and the mapping results for the first stage were obtained. The second segment calculates the prediction residual of the mapping result of the first segment at the site scale, and obtains a continuous residual surface through spatial interpolation. The continuous residual surface is then superimposed on the mapping result of the first segment to form a pixel-level stratified soil moisture content product. The pixel-level stratified soil moisture content product serves as the soil moisture content data in the uniform driving field.

5. The integrated estimation method for ecological water demand and water replenishment according to claim 1, characterized in that, The process of determining the reference evapotranspiration and constructing the baseline vegetation correction coefficient based on the uniform driving field includes: The reference evapotranspiration is determined based on the precipitation and meteorological data in the uniform driving field. The growing season is divided into multiple phenological stages: initial, development, middle and late. The vegetation correction coefficient anchor value for the middle stage is determined by the effective coverage and canopy height obtained by remote sensing. The vegetation correction coefficient anchor value for the middle stage is fine-tuned by combining wind speed and humidity to obtain the middle vegetation correction coefficient. Based on the evolution of phenological phase and coverage between different phenological stages, the vegetation correction coefficients for the initial, developmental, and later stages are continuously interpolated, and the vegetation correction coefficients for each phenological stage are spliced ​​together to form the baseline vegetation correction coefficients covering the entire growing season.

6. The integrated estimation method for ecological water demand and water replenishment according to claim 1, characterized in that, The process of performing physically constrained residual correction on the baseline vegetation correction coefficient under water stress-free conditions to obtain the corrected vegetation correction coefficient, and determining the ecological water requirement based on the corrected vegetation correction coefficient and the reference evapotranspiration, includes: Under the no-moisture stress window, the ratio of actual evapotranspiration in the uniform driving field to the reference evapotranspiration is used as the management perception tag; wherein, the criterion for the no-moisture stress window is that the ratio of cumulative precipitation to reference evapotranspiration within a set period is not less than a preset threshold and that there are no fewer than a preset number of effective rainfalls during the period. A deep learning model is used to learn the residual between the management perception label and the baseline vegetation correction coefficient, and the residual is superimposed on the baseline vegetation correction coefficient to obtain the corrected vegetation correction coefficient. The ecological water requirement is obtained by multiplying the corrected vegetation correction coefficient by the reference evapotranspiration. The physical constraints applied in the physically constrained residual correction include: the corrected vegetation correction coefficient is limited to physically acceptable upper and lower bounds; the corrected vegetation correction coefficient maintains a non-negative monotonic response to the reference evapotranspiration and a non-negative monotonic response to the leaf area index; and the ecological water demand obtained after correction does not exceed the upper limit of evapotranspiration energy available from net radiation.

7. The integrated estimation method for ecological water demand and water replenishment according to claim 1, characterized in that, The multiple discrimination gates include a discrimination gate characterizing phenological sensitivity, a discrimination gate characterizing the degree of structural degradation, and a discrimination gate characterizing immediate physiological limitation; The process involves inputting the unified driving field, the ecological water demand, and remote sensing vegetation observation information into multiple pre-constructed discrimination gates, performing a conjunction logic operation on the judgment results of the multiple discrimination gates, and dividing each pixel into multi-level water replenishment priority labels, including: The discrimination gate characterizing phenological sensitivity delineates the growing season phase based on the ratio of the current leaf area index of a pixel to the peak leaf area index of that pixel in the current year. A staged trigger threshold is smoothly generated based on the ratio. The leaf area index in the remote sensing vegetation observation information is input into the discrimination gate characterizing phenological sensitivity. When the ratio exceeds the staged trigger threshold, it is determined to be a phenological sensitive period. The discrimination gate characterizing the degree of structural degradation calculates the leaf area index anomaly and the solar induced fluorescence anomaly after removing the seasonal climate state. The leaf area index and solar induced fluorescence in the remote sensing vegetation observation information are input into the discrimination gate characterizing the degree of structural degradation. When the leaf area index anomaly is continuously negative for more than a preset number of days and the deviation of the solar induced fluorescence anomaly exceeds a preset standard deviation multiple, it is determined that the structure is weakening; otherwise, it is determined that the structure is stable or strengthening. The discrimination gate for characterizing immediate physiological limitation calculates the relative available water in the root zone based on the stratified soil moisture content in the uniform driving field, and calculates the evapotranspiration efficiency based on the actual evapotranspiration in the uniform driving field and the ecological water requirement. The relative available water in the root zone is characterized by the normalized position of the current volumetric water content relative to field capacity and wilting point water content, and the evapotranspiration efficiency is characterized by the ratio of actual evapotranspiration to ecological water requirement. When the relative available water in the root zone is lower than a preset available water threshold and the evapotranspiration efficiency is lower than a preset efficiency threshold, it is determined to be physiologically limited. When the relative available water in the root zone or the evapotranspiration efficiency is close to but does not fall below their respective preset thresholds, it is determined to be physiologically close to limited. The multi-level water replenishment priority label includes four levels: immediate water replenishment, priority water replenishment, observation and review, and water replenishment not recommended. When all three discrimination gates simultaneously determine that it is a phenological sensitive period, structural weakening, and physiological limitation, it is judged as immediate water replenishment. When it is judged as structural weakening and in a phenological sensitive period, and simultaneously judged as physiologically approaching limitation, it is judged as priority water replenishment. When it is judged as a phenological sensitive period, but the structure is judged as stable or strengthening and the evapotranspiration efficiency has not fallen below the efficiency threshold, it is judged as observation and review. If the above conditions are not met, water replenishment is not recommended.

8. The integrated estimation method for ecological water demand and water replenishment according to claim 1, characterized in that, The process of determining the total loss item based on the unified driving field, the ecological water demand, and the water replenishment priority label, and determining the ecological water replenishment amount based on the ecological water demand and the total loss item, includes: A supervisory quantity is constructed within a wet window, and a pre-constructed residual learning network is trained based on this supervisory quantity to obtain a trained residual learning network. The supervisory quantity is the precipitation data in the uniform driving field minus the ecological water demand. The residual learning network is a deep learning network that receives time-series data containing wet memory and canopy energy state and outputs residual corrections. A water conservation residual penalty term and a reservoir capacity upper limit constraint term are added to the training objective function. The water conservation residual penalty term constrains the predicted value of the total loss term to satisfy the root zone water balance closure, and the reservoir capacity upper limit constraint term constrains the root zone water content after water replenishment to not exceed the reservoir capacity upper limit corresponding to field capacity. The soil and topographic parameters in the uniform driving field are input into a pre-constructed physical prior model to obtain a first-order estimate of the total loss term; wherein, the physical prior model is established based on soil water holding capacity, water conductivity, topographic runoff characteristics and infiltration limit; The time-series data containing wet memory and canopy energy state are input into the trained residual learning network to obtain the residual correction amount. The residual correction amount is then superimposed on the first-order estimation to obtain the predicted value of the total loss term. Pixels that are selected for immediate or priority water replenishment based on the water replenishment priority label are then added to the predicted value of the total loss item to obtain the ecological water replenishment amount for the selected pixels.

9. An integrated estimation device for ecological water demand and water replenishment, characterized in that, include: The data processing module is used to perform bias correction or fusion processing on precipitation data, evapotranspiration data and soil moisture content data respectively to obtain a consistent driving field; The water demand determination module is used to determine a reference evapotranspiration based on the uniform driving field and construct a baseline vegetation correction coefficient. Under the condition of no water stress, the baseline vegetation correction coefficient is subjected to physically constrained residual correction to obtain a corrected vegetation correction coefficient. The ecological water demand is determined based on the corrected vegetation correction coefficient and the reference evapotranspiration. The water shortage classification module is used to input the uniform driving field, the ecological water demand and remote sensing vegetation observation information into multiple pre-constructed discrimination gates, perform conjunctive logic operations on the judgment results of multiple discrimination gates, and divide each pixel into multi-level water replenishment priority labels; wherein, the multiple discrimination gates respectively characterize phenological sensitivity, structural degradation degree and immediate physiological limitation; The water replenishment determination module is used to determine the total loss item based on the unified driving field, the ecological water demand, and the water replenishment priority label, and to determine the ecological water replenishment item based on the ecological water demand and the total loss item; wherein, the total loss item is the combined amount of root zone water storage change, deep seepage, lateral outflow and capillary recharge.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the integrated estimation method for ecological water demand and water replenishment as described in any one of claims 1 to 8.

Images