A method for estimating oasis water consumption using multi-source evapotranspiration fusion and dynamic weight optimization

Abstract

This invention provides a method for estimating oasis water consumption based on multi-source evapotranspiration fusion and dynamic weight optimization, relating to the fields of hydrological remote sensing and arid zone water resource management. The method involves evaluating multi-source evapotranspiration data to obtain a data quality index; dividing multi-temporal high-resolution remote sensing images into multiple functional zones to obtain a partitioning mask; calculating spatiotemporally adaptive dynamic weights based on the data quality index and multi-source evapotranspiration data; combining multi-source evapotranspiration data and the error covariance matrix to obtain a preliminary fused evapotranspiration field and a fused uncertainty field; and calculating the total water consumption of the oasis and its confidence interval. This invention addresses the problem of how to maximize the complementary advantages of multi-source evapotranspiration data under highly heterogeneous underlying surface conditions in oases through dynamic weight adjustment, uncertainty control, and physical constraint embedding, thereby improving the accuracy and reliability of regional water consumption estimation.

Description

Technical Field

[0001] This invention relates to the fields of hydrological remote sensing and water resources management in arid areas, and particularly to a method for estimating oasis water consumption by multi-source evapotranspiration fusion and dynamic weight optimization. Background Technology

[0002] Oases are core components of arid zone ecosystems, and accurate estimation of their water consumption plays a crucial supporting role in regional water resource allocation, ecosystem protection, and sustainable development decisions. Oasis ecosystems are highly heterogeneous, containing various land cover types such as farmland, natural vegetation, desert transition zones, water bodies, saline-alkali land, and urban built-up areas. The evapotranspiration processes of these different types differ significantly in their driving mechanisms, spatiotemporal variability, and responses to water stress. This makes oasis water consumption estimation a far more complex challenge than that for general watersheds or homogeneous underlying surfaces.

[0003] Currently, the mainstream methods for estimating regional evapotranspiration (ET) can be categorized as follows: energy balance models based on physical mechanisms (such as SEBS and METRIC), empirical and semi-empirical models based on vegetation index parameterization (such as PT-JPL), water balance models based on hydrological processes, and data-driven methods that have emerged in recent years (such as machine learning models). Each of these methods has its specific applicable conditions and data dependencies, and each has significant limitations in practical applications in oasis environments.

[0004] In the field of multi-source data fusion, Bayesian Model Averaging (BMA) is one of the most widely used frameworks. BMA achieves good results in hydrological forecasting and integrated climate prediction by weighting and averaging the posterior probabilities of each model. However, the BMA method has the following fundamental limitations in oasis ET fusion: First, the weights of BMA are fitted from historical training periods, which are static posterior weights and cannot respond to the real-time time-varying characteristics of data source quality. Second, the BMA framework assumes that the errors of each model follow a normal distribution and are spatially homogeneous, failing to distinguish the differentiated applicability of each model to different land cover types within the oasis. Third, BMA is essentially a purely statistical framework and does not embed physical conservation equations, which may produce fusion results that violate energy balance or water conservation under extreme water stress conditions. Fourth, the uncertainty quantification of classic BMA and its extended variants (such as EB-BMA) depends on parametric model assumptions; when multi-source data are fused under non-Gaussian error structures, the confidence interval estimation exhibits systematic bias.

[0005] The main shortcomings of existing research and technical solutions are reflected in the following four aspects: First, there is the issue of fixed weights in multi-source data fusion. When fusing remote sensing products (such as MOD16 and GLASS-ET), site observation data, and model outputs, existing methods (including BMA) generally use static weights based on statistical errors, which remain unchanged once calibrated. However, the accuracy of different ET data sources dynamically changes with seasons, meteorological conditions (such as cloud cover, alternation of wet and dry periods), and surface conditions. Fixed weights cannot reflect the time-varying characteristics of the relative reliability of each data source, resulting in a significant decrease in the accuracy of the fusion results during critical water-consuming periods (such as the spring greening period and peak irrigation periods).

[0006] Second, there is the problem of unquantified error propagation. Existing fusion frameworks typically only provide point estimates of ET, lacking a systematic characterization of the uncertainties of each data source and a quantification of their propagation patterns during the fusion process. When multiple data sources simultaneously exhibit high uncertainties at a certain time (such as continuous rain leading to a deterioration in the quality of optical remote sensing products), the errors in the fusion results will be amplified by superposition, but existing methods cannot identify and control this process.

[0007] Third, the issue of overlooking differences in land cover types. The evapotranspiration mechanisms of different land cover types within oases differ fundamentally: farmland evapotranspiration (ET) is regulated by irrigation systems, exhibiting strong anthropogenic characteristics; natural vegetation ET is constrained by both soil moisture and groundwater depth; soil evaporation and transpiration from sparse vegetation in desert transition zones need to be estimated separately; and ET in urban built-up areas is primarily driven by impermeable surfaces trapping evaporation and green vegetation transpiration, which differs significantly from natural land cover mechanisms. Existing methods use a uniform fusion framework for the entire region, neglecting this internal heterogeneity and leading to systemic bias.

[0008] Fourth, the problem of missing physical constraints. Most data fusion methods are based on a purely mathematical and statistical framework, failing to embed the energy balance equation (net radiation = latent heat + sensible heat + soil heat flux) and the water balance equation (precipitation + irrigation = ET + runoff + soil water change) as hard constraints into the fusion process. This can lead to physical inconsistencies in the fusion results, especially under extreme conditions of water stress, resulting in estimations that exceed the upper limit of energy supply or violate water conservation.

[0009] Fifth, the limitations of distribution assumptions in uncertainty quantification methods. Existing methods (including BMA) generally assume that the errors of each data source follow a parameterized normal or t-distribution when quantifying ET fusion uncertainty. However, the error distribution of Oasis ET data sources often exhibits heavy-tailed, skewed, or even multimodal characteristics under conditions such as cloud pollution and irrigation abrupt changes. The parametric model assumptions lead to a serious underestimation of the confidence interval, which in turn affects the reliability of the decision. Summary of the Invention

[0010] To address the aforementioned shortcomings in existing technologies, this invention provides a multi-source evapotranspiration fusion and dynamic weight optimization method for estimating oasis water consumption. This method solves the problem of how to maximize the complementary advantages of multi-source evapotranspiration data under highly heterogeneous underlying surface conditions in oases through dynamic weight adjustment, uncertainty control, and physical constraint embedding, thereby improving the accuracy and reliability of regional water consumption estimation.

[0011] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a method for estimating oasis water consumption based on multi-source evapotranspiration fusion and dynamic weight optimization, comprising: S1: Based on the acquired multi-source evapotranspiration data, an evaluation mechanism is used to obtain the data quality index of each data source. S2: Based on the acquired multi-temporal high-resolution remote sensing images, the oasis study area is divided into multiple functional zones using a surface classification strategy to obtain a partition mask. S3: Based on multi-source evapotranspiration data and partitioning masks, the Riemannian manifold theory of information geometry is used to quantify the multi-source uncertainty of data from each data source in each functional area, and the error covariance matrix of each data source is obtained. S4: Based on the data quality index and multi-source evapotranspiration data, the spatiotemporal adaptive dynamic weights of each data source in each functional area are calculated using the physically enhanced variational Bayesian adaptive weight update mechanism. S5: Based on spatiotemporal adaptive dynamic weights, multi-source evapotranspiration data and error covariance matrix, combined with partitioning mask, hierarchical weighted fusion and covariance propagation are performed to obtain the preliminary fused evapotranspiration field and fused uncertainty field. S6: Based on the preliminary evapotranspiration field, the final evapotranspiration field of each functional area is obtained by using the energy balance model and the water balance model for dual physical constraint correction. S7: Based on the final evapotranspiration field and the fusion uncertainty field, spatial integration and error propagation calculations are performed to obtain the total water consumption of the oasis and its confidence interval, thus completing the oasis water consumption estimation.

[0012] The beneficial effects of this invention are as follows: This invention provides a method for estimating oasis water consumption by multi-source evapotranspiration fusion and dynamic weight optimization. By introducing the information geometry Riemannian manifold theory to quantify the uncertainty of each data source, the error covariance matrix is ​​obtained. Compared with the error estimation of traditional Euclidean space, it can handle complex error scenarios more robustly and retain the spatial correlation structure of the error.

[0013] By utilizing a physically enhanced variational Bayesian adaptive weight update mechanism to calculate spatiotemporally adaptive dynamic weights, we overcome the shortcomings of static weight fixation and reliance on subjective experience in traditional methods, and realize the adaptive response of weights to changes in data source quality.

[0014] By constructing a partitioned mask based on high-resolution remote sensing imagery and combining the mask with layered weighted fusion, the heterogeneity of the internal structure of the oasis surface is specifically addressed, allowing different data sources to maximize their advantages on the most suitable surface types.

[0015] By using energy balance and water balance models to perform dual physical constraint correction on the preliminary evapotranspiration field, it is ensured that the final evapotranspiration field does not violate physical laws in both the upper limit of energy supply and the macroscopic water balance, thus achieving physical self-consistency of the results.

[0016] Further, S1 includes: Based on the acquired multi-source evapotranspiration data and corresponding cloud cover data, temporal continuity indicators, irrigation pulse flags and crop coefficients, the cloud pollution self-enhancing penalty term, the cloud pollution modulation temporal continuity coupling term and the land surface type-specific nonlinear deviation penalty term are calculated respectively. By combining the cloud pollution self-enhancing penalty term, the cloud pollution modulation temporal continuity coupling term, and the surface type-specific nonlinear deviation penalty term, the data quality index of each data source is obtained.

[0017] The superlinear decay design of the contamination self-enhancing penalty term dramatically amplifies the penalty intensity when the cloud coverage probability is high, which is far more effective than linear penalty in suppressing the interference of severely cloud-contaminated data on the fusion results. The cloud contamination modulation time continuity coupling term, by introducing an irrigation pulse flag, distinguishes between "cloud-induced pseudo-mutations" and "irrigation-driven real step jumps," avoiding the misjudgment of reasonable irrigation responses as data anomalies and subsequent incorrect weighting, which is of great significance in irrigation district applications.

[0018] Furthermore, the expression for the cloud pollution self-enhancing penalty term is: ; The expression for the time continuity coupling term of the cloud pollution modulation is: ; The expression for the surface type-specific nonlinear deviation penalty term is: ; ; in, This indicates a self-enhancing penalty for cloud pollution. This represents the cloud pollution index of the S-th data source. This represents the empirical value coefficient of optical remote sensing in arid areas. This represents the time-continuity coupling term of cloud contamination modulation. This represents the time continuity index of the S-th data source. This represents the cloud pollution index of the S-th data source. Represents the attenuation factor constant. Indicates the adjustment parameter. Indicates irrigation status. This indicates the irrigation pulse flag, where t represents the time. This represents a nonlinear deviation penalty term specific to land surface type. This represents the decay rate of the k-th functional region. Indicates the penalty index. This indicates the deviation under the penalty index. Indicates the deviation amount. This represents the crop coefficient of the S-th data source. This represents the upper limit of the optimal crop coefficient range for the k-th functional zone. denoted by , represents the lower limit of the optimal crop coefficient interval for the k-th functional zone, and x represents the spatial pixel coordinates.

[0019] The three penalty factors are weighted by adjustable empirical weights to form a unified data quality index, which is then transmitted to the downstream weight update module in the form of a single scalar, avoiding decision conflicts caused by the parallel transmission of multiple indicators. At the same time, the default parameter values ​​are derived from optical remote sensing experience in arid areas, which has a certain degree of plug-and-play applicability and reduces user calibration costs.

[0020] Further, S2 includes: Based on the acquired multi-temporal high-resolution remote sensing images, combined with vegetation phenology information, impermeable surface extraction results and auxiliary geographic data, the oasis study area is divided into six functional zones: irrigated farmland, natural vegetation, water and wetland, oasis-to-desert transition zone, bare soil and saline-alkali area, and urban built-up area. A zoning mask is constructed and obtained.

[0021] The division of the six functional zones covers the complete spatial gradient within the oasis, from strong anthropogenic disturbance (irrigated farmland, urban built-up areas) to weak anthropogenic disturbance (natural vegetation, desert transition zone), which is highly consistent with the classification of the main driving mechanisms of oasis evapotranspiration. By introducing vegetation phenological information, the zoning boundaries can reflect seasonal changes in the surface condition, avoiding the problem of ignoring phenological dynamics with fixed boundaries. The inclusion of auxiliary geographic data (digital elevation, soil type, etc.) further enhances the reliability of functional zone division in areas with complex terrain.

[0022] Further, S3 includes: Based on the multi-source evapotranspiration data and the partition mask, mutually independent data sources are selected to construct the error covariance estimation matrix within the time window; The error covariance estimation matrix is ​​mapped to a symmetric positive definite matrix space, and the geodesic distance between the covariance matrices of each data source is obtained by using the affine invariant Riemannian metric. Based on the geodesic distance, the error covariance matrix of each data source is obtained by solving the Riemann trigonometric calibration equations and combining them with the physical reliability modulation factor determined by the spatial variance of the cloud pollution index and vegetation index for triple weighted aggregation.

[0023] By mapping the error covariance estimation matrix to a symmetric positive definite matrix space and employing an affine invariant Riemannian metric, the positive definiteness constraint of the covariance matrix in geometric operations is fundamentally guaranteed, avoiding the non-positive definite covariance matrix that may be generated by traditional Euclidean mean operations. Riemann trigonometric calibration based on geodesic distance has smaller bias under non-Gaussian error structures compared to the direct triple difference method, and the error covariance estimation results are more robust.

[0024] Furthermore, the expression for the physical reliability modulation factor is: ; in, Indicates the physical reliability modulation factor. Indicates the cloud pollution suppression coefficient. Represents a triplet The average cloud pollution index at the current moment for the three data sources. Indicates the heterogeneity adjustment coefficient. Represents the x-neighborhood window of spatial cell coordinates Spatial variance of normalized vegetation index within the region denoted as the mean of the spatial variance of the normalized vegetation index for the entire region, t represents time, and x represents the spatial pixel coordinates.

[0025] The physical reliability modulation factor couples two independent physical information channels, cloud pollution suppression and vegetation heterogeneity perception, in a multiplicative manner, achieving dual physical correction of the aggregation weights of pure statistical errors. Specifically, when a data source of a certain triplet simultaneously exhibits high cloud pollution, the factor automatically reduces its aggregation contribution, preventing systematic errors from the same source from being incorrectly canceled out during triangulation. At the same time, the spatial variance term of the local vegetation index provides additional error suppression for areas with high surface heterogeneity, improving the estimation reliability in mixed pixel scenarios.

[0026] Furthermore, when weighting triples, if singularities occur in the triple estimation, an adaptive regularization perturbation is used to correct the aggregation weights. The expression for the corrected aggregation weights is as follows: ; ; in, This represents the corrected aggregate weights. Represents the original aggregate weights. This represents the adaptive regularization perturbation. Represents a triplet The determinant of the error covariance estimation matrix, This represents the basic disturbance quantity. The trace of the covariance matrix. Indicates the heterogeneity sensitivity parameter. This represents the spatial variance of the normalized vegetation index within the x-neighborhood window of the spatial cell coordinates. This represents the mean spatial variance of the normalized vegetation index for the entire region. Let represent the basic error covariance matrix of the triplet, and x represent the spatial pixel coordinates.

[0027] The adaptive regularization perturbation is positively correlated with the spatial variance of the local normalized vegetation index. It automatically enhances the regularization intensity in areas with highly complex (heterogeneous) land cover, while maintaining a small perturbation in homogeneous areas, thus achieving the effect of "regularization on demand". Compared with fixed perturbation, this design avoids the unnecessary loss of accuracy in homogeneous areas caused by global uniform regularization, while ensuring numerical stability in areas with high incidence of singularities.

[0028] Further, S4 includes: Based on the data quality index and the multi-source evapotranspiration data, an energy balance penalty term representing the upper limit of available energy and a water balance penalty term representing the monthly water mismatch are embedded in the lower bound objective function of variational inference to construct a physical enhancement evidence lower bound under the physical enhancement variational Bayesian adaptive weight update mechanism. By maximizing the lower bound of the physical augmentation evidence, the weight parameters are iteratively updated to obtain the spatiotemporally adaptive dynamic weights of each data source within each functional area.

[0029] Furthermore, the expressions for the lower bound of the physical enhancement evidence and the included energy balance penalty term and water balance penalty term are as follows: ; ; ; The expression for the spatiotemporal adaptive dynamic weight is: ; ; ; in, Let q denote the lower bound of the physical enhancement evidence, q denote the variational approximation distribution, and p denote the prior distribution. Let q represent the expectation under distribution q. Represents the relative entropy divergence. This represents observed evapotranspiration data. Indicates the surface type adaptability coefficient. This represents the adaptive penalty coefficient for energy balance at the current time t. This represents the adaptive penalty coefficient for water balance in the k-th functional zone at time t. Represents the energy balance residual term. This represents the water balance residual term. Indicates the amount of fusion evaporation. This indicates the evaporation ratio at the current moment. G represents the net surface radiation flux, and G represents the soil heat flux. This represents the joint error variance of available energy observations. This represents the reference evapotranspiration for water balance. This represents the joint error variance of the observations of each component of the water balance. An indicator function representing the availability of water balance data. and These represent the posterior accuracy parameters before and after the iterative update, respectively. and Let represent the posterior mean parameters before and after the iterative update, respectively. Indicates the lower bound of physically enhanced evidence. This represents a truncation operator that projects a value onto the interval [0.5, 2.0]. This represents spatiotemporal adaptive dynamic weights. This represents the base weight of each data source, and N represents the total number of data sources.

[0030] Further, S6 includes: Based on the preliminary evapotranspiration field, the evaporation ratio is independently retrieved from the surface temperature and albedo characteristic space using an energy balance model to determine the upper limit of available energy. The energy balance upper limit constraint correction is then applied to the preliminary evapotranspiration field to obtain the corrected evapotranspiration field. Based on the changes in precipitation, irrigation water diversion, runoff and soil water storage, the water balance model is used to calculate the reference evapotranspiration for water balance. The water balance reference evapotranspiration is used to perform water balance verification and linear scale correction on the corrected evapotranspiration field on a monthly scale to obtain the final evapotranspiration field of each functional area. Attached Figure Description

[0031] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart of an oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization, as shown in some embodiments of this specification. Detailed Implementation

[0032] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0033] Example Figure 1 This is an exemplary flowchart illustrating a multi-source evapotranspiration fusion and dynamic weight optimization method for estimating oasis water consumption, based on some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0034] S1: Based on the acquired multi-source evapotranspiration data, an evaluation mechanism is used to obtain the data quality index of each data source.

[0035] Multi-source evapotranspiration data is a collection of data characterizing surface evapotranspiration from different sources and estimation mechanisms. For example, multi-source evapotranspiration data can include remote sensing product data, simulation results based on energy balance models, eddy covariance station observation data, and estimation results based on meteorological drivers.

[0036] In some embodiments, the processor can acquire multi-source evapotranspiration data by receiving external database transmissions or real-time sensor uploads.

[0037] Data quality index is a comprehensive evaluation indicator that quantifies the reliability of different evapotranspiration data sources under specific spatiotemporal conditions. For example, data quality index may include cloud pollution self-enhancing penalty components, cloud pollution modulation temporal continuity components, and surface type-specific nonlinear deviation penalty components.

[0038] In some embodiments, the processor can obtain a data quality index by performing nonlinear activation synthesis calculations on multi-source evapotranspiration data and its auxiliary indicators.

[0039] In some embodiments, the processor can calculate, based on the acquired multi-source evapotranspiration data and corresponding cloud cover data, temporal continuity index, irrigation pulse flag, and crop coefficient, a cloud pollution self-reinforcing penalty term, a cloud pollution modulation temporal continuity coupling term, and a land type-specific nonlinear deviation penalty term; and combine the cloud pollution self-reinforcing penalty term, the cloud pollution modulation temporal continuity coupling term, and the land type-specific nonlinear deviation penalty term to obtain the data quality index of each data source.

[0040] Cloud cover data is a numerical value that reflects the probability or degree to which a specific pixel in a remotely sensed image is obscured by clouds. For example, cloud cover data may include normalized cloud contamination probability values ​​or cloud mask coverage ratios.

[0041] In some embodiments, the processor can extract cloud coverage data by parsing the quality control bands of satellite imagery or meteorological cloud maps.

[0042] Time continuity metrics are indicators that measure the stationarity of fluctuations in a single data source over time. For example, time continuity metrics may include variance or smoothness scores calculated based on the differences between data from different time periods.

[0043] In some embodiments, the processor can obtain time continuity indicators by performing statistical analysis on historical time-series evapotranspiration data.

[0044] An irrigation pulse flag is a discrete state variable that indicates whether artificial irrigation has occurred at a specific time point. For example, an irrigation pulse flag can include a Boolean value of 0 or 1.

[0045] In some embodiments, the processor can obtain the irrigation pulse flag by reading the operation record of the irrigation district management system or the pump start / stop log.

[0046] The crop coefficient is an empirical parameter that reflects the proportion of water consumption characteristics of a specific vegetation at a specific growth stage to the evapotranspiration of a reference crop. For example, the crop coefficient can include the basic crop coefficient and the soil evaporation coefficient.

[0047] In some embodiments, the processor can obtain crop coefficients by consulting agricultural meteorological standard tables or by inverting based on empirical formulas for vegetation indices.

[0048] The cloud pollution self-reinforcing penalty term is a penalty weight that is superlinearly decaying when assessing data quality for scenarios with high cloud coverage. For example, the cloud pollution self-reinforcing penalty term may include a basic linear complement term and an exponential amplification term that decays faster with the probability of cloud coverage.

[0049] In some embodiments, the processor can obtain a cloud pollution self-enhancing penalty term by substituting cloud coverage data into a self-enhancing exponential function.

[0050] The temporal continuity coupling term is a time series quality score that considers the dual effects of cloud-induced interpolation spurious mutations and irrigation-driven real step jumps. For example, the temporal continuity coupling term may include a cloud contamination attenuation factor, an original temporal continuity index, and a mutation judgment tolerance parameter controlled by an irrigation pulse flag.

[0051] In some embodiments, the processor can obtain the time continuity coupling term by modulating the time continuity index by coupling the cloud pollution index and the irrigation pulse flag bit.

[0052] The land surface type-specific nonlinear deviation penalty term is a negative evaluation indicator that quantifies the degree to which the crop coefficient of the data source deviates from the reasonable physiological constraint range of each land surface functional zone. For example, the land surface type-specific nonlinear deviation penalty term may include a cloud pollution self-reinforcing penalty term, a temporal continuity coupling term, deviation amount, land type decay rate, and a step penalty index under extreme deviation.

[0053] In some embodiments, the processor can obtain a land type-specific nonlinear deviation penalty term by comparing the crop coefficient of the data source with the optimal crop coefficient range of the functional area and calculating exponential decay.

[0054] In some embodiments, the expression for the cloud pollution self-enhancing penalty term is: ; The expression for the time continuity coupling term of cloud pollution modulation is: ; The expression for the surface type-specific nonlinear deviation penalty term is: ; ; in, This indicates a self-enhancing penalty for cloud pollution. This represents the cloud pollution index of the S-th data source. This represents the empirical value coefficient of optical remote sensing in arid areas. This represents the time-continuity coupling term of cloud contamination modulation. This represents the time continuity index of the S-th data source. This represents the cloud pollution index of the S-th data source. Represents the attenuation factor constant. Indicates the adjustment parameter. Indicates irrigation status. This indicates the irrigation pulse flag, where t represents the time. This represents a nonlinear deviation penalty term specific to land surface type. This represents the decay rate of the k-th functional region. Indicates the penalty index. This indicates the deviation under the penalty index. Indicates the deviation amount. This represents the crop coefficient of the S-th data source. This represents the upper limit of the optimal crop coefficient range for the k-th functional zone. This represents the lower limit of the optimal crop coefficient interval for the k-th functional zone.

[0055] In some embodiments, the expression for the data quality index of each data source is: ; in, Indicates the data quality index. , and All are empirical weights (default values ​​are 0.4, 0.3, and 0.3 respectively).

[0056] In some embodiments, the land cover type with stronger ET process constraints takes a larger value (e.g., 2.5 for water bodies and 1.0 for bare soil). The moderate deviation is taken as 2, when Exceeding the land category threshold It automatically switches to 3 to achieve a sharp suppression of extreme deviations.

[0057]

[0058] S2: Based on the acquired multi-temporal high-resolution remote sensing images, the oasis study area is divided into multiple functional zones using a surface classification strategy, resulting in a partition mask.

[0059] Multi-temporal high-resolution remote sensing images are high-spatial-resolution records of electromagnetic wave reflection or radiation from the Earth's surface, acquired at different points in time. For example, multi-temporal high-resolution remote sensing images can include Sentinel-2 satellite imagery or Landsat-8 satellite multispectral imagery.

[0060] In some embodiments, the processor can acquire multi-temporal high-resolution remote sensing images by sending a request to a commercial satellite data provider.

[0061] Functional zones are spatial units with relatively homogeneous eco-hydrological characteristics that divide a study area based on land cover type, water-driving mechanisms, and human activity characteristics. For example, functional zones may include irrigated farmland areas, natural vegetation areas, water and wetland areas, oasis-to-desert transition zones, bare soil and saline-alkali areas, and urban built-up areas.

[0062] In some embodiments, the processor can obtain functional areas by performing supervised classification analysis on multi-temporal high-resolution remote sensing images.

[0063] A zoning mask is a spatial matrix or raster layer that corresponds to the spatial extent of the study area and is used to identify the functional zone category to which each cell belongs. For example, a zoning mask may include a two-dimensional array storing classification codes or a raster file with geographic coordinates.

[0064] In some embodiments, the processor can obtain a partition mask by rasterizing the functional area partitioning results.

[0065] In some embodiments, the processor can divide the oasis study area into six functional zones based on the acquired multi-temporal high-resolution remote sensing images, combined with vegetation phenology information, impermeable surface extraction results and auxiliary geographic data, including irrigated farmland, natural vegetation, water and wetland, oasis-to-desert transition zone, bare soil and saline-alkali area, and urban built-up area, and construct and obtain a partition mask.

[0066] Vegetation phenological information reflects the temporal characteristics of vegetation growth and development stages as the seasons change. For example, vegetation phenological information can include specific time points such as the greening-up stage, the heading stage, and the yellowing-out stage.

[0067] In some embodiments, the processor can obtain vegetation phenological information by analyzing feature points of vegetation index curves over a long time series.

[0068] Impermeable surface extraction results are spatial distribution data that identify the extent of surface cover, such as man-made structures and roads, within the study area where water cannot directly infiltrate. For example, impermeable surface extraction results can include an impermeable surface percentage map or a binarized distribution map.

[0069] In some embodiments, the processor can obtain impermeable surface extraction results by applying the impermeability index algorithm image segmentation technique.

[0070] Ancillary geographic data is the basic spatial context information used to assist in land surface classification and hydrological analysis. For example, ancillary geographic data may include digital elevation models, soil type distribution maps, or groundwater depth data.

[0071] In some embodiments, the processor can obtain auxiliary geographic data by retrieving a local database of the geographic information system.

[0072] S3: Based on multi-source evapotranspiration data and partitioning masks, the Riemannian manifold theory of information geometry is used to quantify the multi-source uncertainty of data from each data source in each functional area, and the error covariance matrix of each data source is obtained.

[0073] The error covariance matrix is ​​a mathematical matrix that describes the individual error variances of each data source in multi-source data and the correlation of errors between different data sources. For example, the error covariance matrix can include the source variance components on the diagonal and the covariance components off the diagonal.

[0074] In some embodiments, the processor can infer the error covariance matrix from time series data from multiple independent data sources using information geometry Riemannian manifold theory.

[0075] In some embodiments, the processor can select independent data sources based on the multi-source evapotranspiration data and the partition mask, and construct an error covariance estimation matrix within a time window; map the error covariance estimation matrix to a symmetric positive definite matrix space, and calculate the geodesic distance between the covariance matrices of each data source using an affine invariant Riemannian metric; based on the geodesic distance, solve the Riemann trigonometric calibration equations, and combine them with a physical reliability modulation factor determined by the spatial variance of the cloud pollution index and vegetation index for triplet weighted aggregation to obtain the error covariance matrix of each data source.

[0076] The error covariance estimation matrix is ​​an empirical covariance matrix initially calculated based on sample data within a finite time window. For example, the error covariance estimation matrix can include the time average of the product of the deviations of the sample means.

[0077] In some embodiments, the processor can obtain the error covariance estimation matrix by performing statistical operations on the residual sequence of the paired data source within a time window.

[0078] In some embodiments, for a data source pair (X,Y), the formula for constructing the error covariance estimation matrix is ​​as follows: ; in, This represents the error covariance estimation matrix of data sources X and Y within the time window T; Indicates the length of the time window; and These represent data sources X and Y at different times. Observed values; and These represent data sources X and Y within the time window, respectively. The mean within.

[0079] A symmetric positive definite matrix space is a mathematical manifold space consisting of real symmetric matrices with all positive eigenvalues. It is used to guarantee the physical validity of the covariance matrix in geometric operations. For example, a symmetric positive definite matrix space can include the set of all local error covariance estimates that satisfy non-negative definite constraints.

[0080] In some embodiments, the processor can obtain a data structure located in a symmetric positive definite matrix space by applying a small regularization perturbation constraint to the initial covariance matrix.

[0081] Geodesic distance is the shortest path length connecting two positive definite matrix points on a Riemannian manifold surface, used as an alternative to straight-line distance in Euclidean space to measure differences in non-Gaussian error distributions. For example, geodesic distance can include the Frobenius norm of the matrix logarithm calculated based on affine-invariant Riemannian metrics.

[0082] In some embodiments, the processor can obtain the geodesic distance by calling the Riemannian manifold geometry computation library to perform logarithmic mapping and norm calculation on the error covariance estimation matrix.

[0083] In some embodiments, the expression for geodesic distance is: ; in, This represents the error covariance estimation matrix. and Geodesic distance between them; Represents the matrix logarithm operation; This represents the Frobenius norm.

[0084] The cloud contamination index is a numerical indicator that quantifies the severity of cloud interference in remote sensing data at a given moment. For example, the cloud contamination index may include derived cloud probability thresholds.

[0085] In some embodiments, the processor can obtain the cloud pollution index by reading the probability quality identifier attached to the cloud mask of the remote sensing product.

[0086] The spatial variance of vegetation indices is a statistic that describes the degree of unevenness in vegetation cover distribution within a local neighborhood around a specific pixel. For example, the spatial variance of vegetation indices can include the dispersion of the normalized vegetation index within a fixed radius window of the central pixel.

[0087] In some embodiments, the processor can obtain the spatial variance of vegetation indices by applying a local window sliding variance statistical algorithm to the raster image.

[0088] The physical reliability modulation factor is an adjustment coefficient that corrects the weighted aggregate error covariance derived from pure statistics by incorporating external physical observation information. For example, the physical reliability modulation factor may include an exponential term to suppress covariant errors caused by synchronous cloud contamination and a heterogeneity penalty term to perceive mixed pixel effects.

[0089] In some embodiments, the processor can obtain the physical credibility modulation factor by multiplying the mean cloud pollution index at the current moment with the spatial variance of the vegetation index.

[0090] In some embodiments, the expression for the physical reliability modulation factor is: ; in, Indicates the physical reliability modulation factor. Indicates the cloud pollution suppression coefficient. Represents a triplet The average cloud pollution index at the current moment for the three data sources. Indicates the heterogeneity adjustment coefficient. Represents the x-neighborhood window of spatial cell coordinates Spatial variance of normalized vegetation index within the region denoted as the mean of the spatial variance of the normalized vegetation index for the entire region, t represents time, and x represents the spatial pixel coordinates.

[0091] In some embodiments, when weighting triples, if singularities occur in the triple estimation, an adaptive regularization perturbation is used to correct the aggregation weights. The expression for the corrected aggregation weights is as follows: ; ; in, This represents the corrected aggregate weights. Represents the original aggregate weights. This represents the adaptive regularization perturbation. Represents a triplet The determinant of the error covariance estimation matrix, This represents the basic disturbance quantity. The trace of the covariance matrix. Indicates the heterogeneity sensitivity parameter. This represents the spatial variance of the normalized vegetation index within the x-neighborhood window of the spatial cell coordinates. This represents the mean spatial variance of the normalized vegetation index for the entire region. Let represent the basic error covariance matrix of the triplet, and x represent the spatial pixel coordinates.

[0092] In some embodiments, the estimation of the covariance of each source error is solved by the Riemann trigonometric calibration equations. Taking data source X as an example, the specific formula is as follows: ; in, This represents the estimation result of the error covariance matrix of data source X; This represents the Riemann exponent mapping at the identity matrix; This represents the logarithmic mapping at the identity matrix; This represents the parallel translation operator along a geodesic line from point A to point B. and These represent addition and subtraction operations in the tangent space, respectively. , , and These represent the empirical error covariance estimation matrices for the corresponding data source pairs.

[0093] In some embodiments, the formula for deriving the scalar uncertainty coefficient of the data source from the trace of the error covariance matrix is ​​as follows: ; in, Represents the scalar uncertainty coefficient of the S-th data source; This represents the trace of the error covariance matrix of the S-th data source; This represents the spatiotemporal mean of the data source within the corresponding functional area.

[0094] S4: Based on the data quality index and multi-source evapotranspiration data, the spatiotemporal adaptive dynamic weights of each data source in each functional area are calculated using a physically enhanced variational Bayesian adaptive weight update mechanism.

[0095] The Physics-Enhanced Variational Bayesian Adaptive Weight Update Mechanism is an iterative optimization algorithm module that forcibly embeds energy and water conservation constraints into a Bayesian probabilistic inference framework. For example, the Physics-Enhanced Variational Bayesian Adaptive Weight Update Mechanism may include truncated normal conjugate priors, mean-field variational approximation calculation steps, and a gradient descent operator containing a residual penalty term.

[0096] In some embodiments, the processor can obtain the results of the physical-enhanced variational Bayesian adaptive weight update mechanism by executing a code program that contains variational lower bound maximization logic and loads the physical equations.

[0097] Spatiotemporally adaptive dynamic weights are weighting coefficients that dynamically change with time and spatial pixel location, reflecting the relative reliability of each data source under the current spatiotemporal state. For example, spatiotemporally adaptive dynamic weights can include the normalized product of the base weights and the mean parameter of land surface type adaptability updated by physical augmentation.

[0098] In some embodiments, the processor can obtain spatiotemporally adaptive dynamic weights through the iterative convergence output of a physically enhanced variational Bayesian adaptive weight update mechanism.

[0099] In some embodiments, the processor can, based on the data quality index and the multi-source evapotranspiration data, embed an energy balance penalty term representing exceeding the upper limit of available energy and a water balance penalty term representing monthly-scale water mismatch into the lower bound objective function of variational inference, to construct a physical enhancement evidence lower bound under a physical enhancement variational Bayesian adaptive weight update mechanism; by maximizing the physical enhancement evidence lower bound, the weight parameters are iteratively updated to obtain the spatiotemporal adaptive dynamic weights of each data source in each functional area.

[0100] The lower bound objective function is a lower bound function used in variational inference to approximate complex posterior distributions. For example, the lower bound objective function can include the expected log-likelihood term of the observed data minus the dispersion term between the prior and the variational distribution.

[0101] In some embodiments, the processor can obtain the lower bound objective function of evidence by constructing a mathematical expectation formula based on the observation dataset and probability model assumptions.

[0102] The energy balance penalty is a one-sided numerical penalty imposed on the objective function when the merged evapotranspiration exceeds the upper limit of available surface energy. For example, the energy balance penalty may include the ratio of the squared expected value of the excess available energy to the variance of the joint error of energy observations.

[0103] In some embodiments, the processor can obtain an energy balance penalty term by calculating the one-sided expectation of the difference between the current fusion evapotranspiration and the inverted upper limit of available energy.

[0104] The water balance penalty is a constraint imposed on the objective function when the merged evapotranspiration deviates from the monthly watershed water budget balance. For example, the water balance penalty may include the ratio of the expected square of the difference between the merged evapotranspiration and the water balance reference value to the variance of the joint error of the water component observations.

[0105] In some embodiments, the processor can obtain a water balance penalty term by comparing the fused evapotranspiration with a reference value calculated based on the water balance equation.

[0106] The physical enhancement evidence lower bound is a modified objective function that superimposes the residuals of dual physical constraints (energy and water quantity) onto the traditional variational evidence lower bound. For example, the physical enhancement evidence lower bound can include the standard evidence lower bound minus the energy balance penalty term and the water balance penalty term with adaptive coefficients.

[0107] In some embodiments, the processor can obtain the lower bound of physically enhanced evidence by linearly combining the variational inference lower bound with the physical conservation penalty function.

[0108] In some embodiments, the expressions for the lower bound of physical enhancement evidence and the included energy balance penalty term and water balance penalty term are as follows: ; ; ; The expression for the spatiotemporal adaptive dynamic weight is: ; ; ; in, Let q denote the lower bound of the physical enhancement evidence, q denote the variational approximation distribution, and p denote the prior distribution. Let q represent the expectation under distribution q. Represents the relative entropy divergence. This represents observed evapotranspiration data. Indicates the surface type adaptability coefficient. This represents the adaptive penalty coefficient for energy balance at the current time t. This represents the adaptive penalty coefficient for water balance in the k-th functional zone at time t. Represents the energy balance residual term. This represents the water balance residual term. Indicates the amount of fusion evaporation. This indicates the evaporation ratio at the current moment. G represents the net surface radiation flux, and G represents the soil heat flux. This represents the joint error variance of available energy observations. This represents the reference evapotranspiration for water balance. This represents the joint error variance of the observations of each component of the water balance. An indicator function representing the availability of water balance data. and These represent the posterior accuracy parameters before and after the iterative update, respectively. and Let represent the posterior mean parameters before and after the iterative update, respectively. Indicates the lower bound of physically enhanced evidence. This represents a truncation operator that projects a value onto the interval [0.5, 2.0]. This represents spatiotemporal adaptive dynamic weights. This represents the base weight of each data source, and N represents the total number of data sources.

[0109] In some embodiments, the formulas for the adaptive physical penalty coefficients of the energy balance constraint and the water balance constraint are as follows: ; ; in, This represents the adaptive penalty coefficient for energy balance at the current time t; Indicates the intensity of the energy balance baseline penalty; This represents the hot start rate parameter; Indicates the warm start time window; This represents the data quality index of the net radiation data at the current moment; This represents the adaptive penalty coefficient for water balance in the k-th functional zone at the current time t. Indicates the baseline penalty intensity for water balance; An indicator function that indicates the availability of water balance data; This represents the joint error variance of the observations of each component of the water balance in the kth functional zone. This represents the maximum tolerance threshold for the variance of the water balance error.

[0110] In some embodiments, the gradient term formula expanded during the variational parameter update process is specifically as follows: ; in, This represents the partial derivative of the lower bound of physical enhancement evidence with respect to the surface type fit coefficient; This represents the partial derivative of the log-likelihood of the observed data with respect to the surface type fit coefficient; The partial derivative of the energy balance residual term; This represents the partial derivative of the water balance residual term.

[0111] In some embodiments, when there are no verification observations at a certain time, the formula for the posterior precision parameter retention precision decay mechanism is as follows: ; in, This represents the posterior accuracy parameter updated after the observation time; Represents the posterior precision parameter of the previous time step; The forgetting factor is set.

[0112] In some embodiments, when there are no verification observations at a certain time, the formula for shrinking the mean to the safety fit coefficient is as follows: ; in, This represents the posterior mean parameter updated after the time of no observation. This represents the posterior mean parameter of the previous time step; This represents the safety adaptation coefficient of the k-th functional area, determined based on long-term physical closure statistics. This represents a truncation operator that projects a value onto the interval [0.5, 2.0].

[0113] S5: Based on spatiotemporal adaptive dynamic weights, multi-source evapotranspiration data and error covariance matrix, combined with partitioning mask, hierarchical weighted fusion and covariance propagation are performed to obtain the preliminary fused evapotranspiration field and fused uncertainty field.

[0114] The preliminary fused evapotranspiration field is the spatial distribution data of regional evapotranspiration based on multi-source data after weighted summation and spatial smoothing using spatiotemporal adaptive dynamic weights, but before final physical forced correction. For example, the preliminary fused evapotranspiration field may include pixel-level preliminary evapotranspiration estimates stored in a geographic information matrix.

[0115] In some embodiments, the processor can obtain a preliminary fused evaporation field by executing a hierarchical weighted fusion algorithm and an adaptive Gaussian filter.

[0116] In some embodiments, the formula for calculating the preliminary fusion evaporation in each functional area is as follows: ; in, This represents the initial fusion evapotranspiration of the k-th functional area at time t, representing the spatial pixel coordinate x. Indicates the final fusion weight; This represents the evapotranspiration observation value of the S-th data source at time t with spatial pixel coordinate x; This indicates the total number of data sources.

[0117] The fusion uncertainty field is a spatial distribution matrix characterizing the error variance and inter-pixel covariance of the preliminary fusion result across all pixels. For example, the fusion uncertainty field can include the covariance propagation result after weighting the error covariance matrices of each data source by squaring them according to the final fusion weights.

[0118] In some embodiments, the processor can obtain the fused uncertainty field by applying the complete covariance propagation law to the original error covariance matrix.

[0119] In some embodiments, the scalar uncertainty field formula derived from the trace of the covariance matrix is ​​specifically as follows: ; in, The scalar fusion uncertainty field represents the spatial pixel coordinate x at time t; This represents the trace of the covariance matrix of the fused uncertainty field at that pixel.

[0120] S6: Based on the preliminary evapotranspiration field, the final evapotranspiration field of each functional zone is obtained by using the energy balance model and the water balance model for dual physical constraint correction.

[0121] The energy balance model is a physical model that describes the energy budget conservation relationship between net surface radiation, sensible heat flux, latent heat flux, and soil heat flux. The expression for the energy balance model is: Under constraints, this manifests as an upper limit constraint on available energy: .in, This represents the net radiation flux at the Earth's surface; Indicates latent heat flux; Indicates sensible heat flux; Indicates soil heat flux; This represents the initial fusion evaporation as latent heat. For example, an energy balance model may include upper bound calculation logic based on evaporation ratio correction derived from the feature space.

[0122] In some embodiments, the processor can obtain energy balance model data by receiving radiation data from a weather station and inputting it into the energy conservation equation.

[0123] A water balance model is a physical model that describes the conservation of mass between water input, output, and storage changes within a specific region. The expression for a water balance model is: .in, This represents the water balance reference evapotranspiration for the kth functional zone; Indicates precipitation; Indicates the amount of irrigation water diverted; Indicates runoff; This represents the change in soil water storage. For example, a water balance model can include calculation rules for algebraic addition and subtraction of hydrological elements on a monthly scale.

[0124] In some embodiments, the processor can obtain water balance model data by aggregating hydrological station and satellite gravity data and substituting them into the mass conservation equation.

[0125] In some embodiments, the formula for the water balance correction factor is as follows: ; in, This represents the water balance correction factor for the kth functional zone; This represents the water balance reference evapotranspiration for the kth functional zone; This represents the estimated total evapotranspiration in the region after initial integration.

[0126] In some embodiments, the formula for linear scaling correction of monthly evapotranspiration is as follows: ; in, This represents the final evapotranspiration at time t of the spatial pixel coordinate x after water balance correction. Indicates the water balance correction factor; This represents the evaporation rate after correction for energy balance upper limit constraints.

[0127] The final evapotranspiration field is the final high-precision water consumption distribution data after the initial evapotranspiration field has undergone dual physical self-consistency correction based on energy balance and water balance. For example, the final evapotranspiration field may include a pixel-level evapotranspiration numerical layer that meets the upper limit of energy supply and conforms to the monthly water income and expenditure.

[0128] In some embodiments, the processor can obtain the final evapotranspiration field by sequentially performing energy upper limit truncation and water quantity linear scaling correction on the preliminary results.

[0129] In some embodiments, the processor can, based on the preliminary fused evapotranspiration field, independently invert the evaporation ratio from the surface temperature and albedo characteristic space using an energy balance model to determine the upper limit of available energy, and perform energy balance upper limit constraint correction on the preliminary fused evapotranspiration field to obtain the corrected evapotranspiration field; based on the changes in precipitation, irrigation water diversion, runoff, and soil water storage, calculate using a water balance model to obtain the water balance reference evapotranspiration; and use the water balance reference evapotranspiration to perform water balance verification and linear scale correction on the corrected evapotranspiration field on a monthly scale to obtain the final evapotranspiration field of each functional area.

[0130] The upper limit of available energy is the maximum theoretical energy threshold that can be converted into latent heat of evapotranspiration after subtracting soil heat flux from net surface radiation under specific surface temperature and albedo conditions. For example, the upper limit of available energy may include the product of the evaporation ratio and the effective radiation difference derived from the characteristic space.

[0131] In some embodiments, the processor can obtain the upper limit of available energy by inverting the calculation using the surface energy balance index method.

[0132] The corrected evapotranspiration field is intermediate evapotranspiration spatial data obtained by forcibly removing abnormally high values ​​exceeding the physical upper limit of available energy from the initial fusion result. For example, the corrected evapotranspiration field may include a data matrix that compresses out-of-limit pixels to the upper energy limit threshold.

[0133] In some embodiments, the processor can obtain the corrected evaporation field by comparing the initial fusion latent heat with the upper limit of available energy and executing a minimum function.

[0134] In some embodiments, the expressions for the correction of the fusion uncertainty field and the initial fusion evaporation field by the energy balance upper limit constraint are as follows: ; ; in, The covariance matrix of the fusion uncertainty field represents the spatial pixel coordinate x at time t; This represents spatiotemporally adaptive dynamic weights; The S-th data source represents the error covariance matrix; N represents the total number of data sources. This represents the corrected latent heat flux of evapotranspiration; This represents the latent heat flux of evapotranspiration during the initial fusion process; This represents the evaporation ratio, which is spatially inverted from the characteristics of surface temperature and albedo. G represents the net surface radiation flux; G represents the soil heat flux.

[0135] The water balance reference evapotranspiration is a regional macroscopic evapotranspiration expectation calculated from independent hydrological observation data based on the law of conservation of mass, and is used as a benchmark for scale correction. For example, the water balance reference evapotranspiration can include scalar values ​​of various hydrological elements summed up or subtracted over a monthly scale.

[0136] In some embodiments, the processor can calculate the water balance reference evapotranspiration by performing a water balance model formula on the hydrological component data of a specific functional area.

[0137] S7: Based on the final evapotranspiration field and the fusion uncertainty field, spatial integration and error propagation calculations are performed to obtain the total water consumption of the oasis and its confidence interval, thus completing the oasis water consumption estimation.

[0138] Total water consumption of an oasis refers to the total amount of water resources lost to the atmosphere through evaporation and transpiration from all surface functional zones within the entire oasis study area over a specified time period. For example, total water consumption of an oasis can include the cumulative value of the final evaporation field of each functional zone in both time and spatial domains.

[0139] In some embodiments, the processor can obtain the total water consumption of the oasis by performing spatial area integration and time accumulation on the final evaporation field.

[0140] A confidence interval is a range of values ​​that contains the true total water consumption of the oasis with a certain probability, used to quantify the overall uncertainty of the estimation results. For example, a confidence interval may include the total water consumption estimate plus or minus the error boundary amount calculated based on the propagation of the fused uncertainty field.

[0141] In some embodiments, the processor can obtain the confidence interval by calculating the area integral superposition propagation term on the covariance matrix of the fused uncertainty field.

[0142] In some embodiments, the expressions for the total water consumption of the oasis and its confidence interval are as follows: ; ; in, This indicates the total water consumption of the oasis; This represents the area of ​​the k-th functional region; This represents the final evapotranspiration field data of the k-th functional area at time t, representing the spatial pixel coordinate x. This represents a confidence interval at a 95% confidence level. Represents the trace of the covariance matrix of the fused uncertainty field; It represents the error covariance between the coordinates x and x' of different spatial pixels.

Claims

1. A method for estimating oasis water consumption using multi-source evapotranspiration fusion and dynamic weight optimization, characterized in that, include: S1: Based on the acquired multi-source evapotranspiration data, an evaluation mechanism is used to obtain the data quality index of each data source. S2: Based on the acquired multi-temporal high-resolution remote sensing images, the oasis study area is divided into multiple functional zones using a surface classification strategy to obtain a partition mask. S3: Based on multi-source evapotranspiration data and partitioning masks, the Riemannian manifold theory of information geometry is used to quantify the multi-source uncertainty of data from each data source in each functional area, and the error covariance matrix of each data source is obtained. S4: Based on the data quality index and multi-source evapotranspiration data, the spatiotemporal adaptive dynamic weights of each data source in each functional area are calculated using the physically enhanced variational Bayesian adaptive weight update mechanism. S5: Based on spatiotemporal adaptive dynamic weights, multi-source evapotranspiration data and error covariance matrix, combined with partitioning mask, hierarchical weighted fusion and covariance propagation are performed to obtain the preliminary fused evapotranspiration field and fused uncertainty field. S6: Based on the preliminary evapotranspiration field, the final evapotranspiration field of each functional area is obtained by using the energy balance model and the water balance model for dual physical constraint correction. S7: Based on the final evapotranspiration field and the fusion uncertainty field, spatial integration and error propagation calculations are performed to obtain the total water consumption of the oasis and its confidence interval, thus completing the oasis water consumption estimation.

2. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 1, characterized in that, S1 includes: Based on the acquired multi-source evapotranspiration data and corresponding cloud cover data, temporal continuity indicators, irrigation pulse flags and crop coefficients, the cloud pollution self-enhancing penalty term, the cloud pollution modulation temporal continuity coupling term and the land surface type-specific nonlinear deviation penalty term are calculated respectively. By combining the cloud pollution self-enhancing penalty term, the cloud pollution modulation temporal continuity coupling term, and the surface type-specific nonlinear deviation penalty term, the data quality index of each data source is obtained.

3. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 2, characterized in that, The expression for the cloud pollution self-enhancing penalty term is: ； The expression for the time continuity coupling term of the cloud pollution modulation is: ； The expression for the surface type-specific nonlinear deviation penalty term is: ； ； in, This indicates a self-enhancing penalty for cloud pollution. This represents the cloud pollution index of the S-th data source. This represents the empirical value coefficient of optical remote sensing in arid areas. This represents the time-continuity coupling term of cloud contamination modulation. This represents the time continuity index of the S-th data source. This represents the cloud pollution index of the S-th data source. Represents the attenuation factor constant. Indicates the adjustment parameter. Indicates irrigation status. This indicates the irrigation pulse flag, where t represents the time. This represents a nonlinear deviation penalty term specific to land surface type. This represents the decay rate of the k-th functional region. Indicates the penalty index. This indicates the deviation under the penalty index. Indicates the deviation amount. This represents the crop coefficient of the S-th data source. This represents the upper limit of the optimal crop coefficient range for the k-th functional zone. denoted by , represents the lower limit of the optimal crop coefficient interval for the k-th functional zone, and x represents the spatial pixel coordinates.

4. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 1, characterized in that, S2 includes: Based on the acquired multi-temporal high-resolution remote sensing images, combined with vegetation phenology information, impermeable surface extraction results and auxiliary geographic data, the oasis study area is divided into six functional zones: irrigated farmland, natural vegetation, water and wetland, oasis-to-desert transition zone, bare soil and saline-alkali area, and urban built-up area. A zoning mask is constructed and obtained.

5. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 1, characterized in that, S3 includes: Based on the multi-source evapotranspiration data and the partition mask, mutually independent data sources are selected to construct the error covariance estimation matrix within the time window; The error covariance estimation matrix is ​​mapped to a symmetric positive definite matrix space, and the geodesic distance between the covariance matrices of each data source is obtained by using the affine invariant Riemannian metric. Based on the geodesic distance, the error covariance matrix of each data source is obtained by solving the Riemann trigonometric calibration equations and combining them with the physical reliability modulation factor determined by the spatial variance of the cloud pollution index and vegetation index for triple weighted aggregation.

6. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 5, characterized in that, The expression for the physical reliability modulation factor is: ； in, Indicates the physical reliability modulation factor. Indicates the cloud pollution suppression coefficient. Represents a triplet The average cloud pollution index at the current moment for the three data sources. Indicates the heterogeneity adjustment coefficient. Represents the x-neighborhood window of spatial cell coordinates Spatial variance of normalized vegetation index within the region denoted as the mean of the spatial variance of the normalized vegetation index for the entire region, t represents time, and x represents the spatial pixel coordinates.

7. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 5, characterized in that, When performing weighted aggregation of triples, if singularities occur in the triple estimation, an adaptive regularization perturbation is used to correct the aggregation weights. The expression for the corrected aggregation weights is as follows: ； ； in, This represents the corrected aggregate weights. Represents the original aggregate weights. This represents the adaptive regularization perturbation. Represents a triplet The determinant of the error covariance estimation matrix, This represents the basic disturbance quantity. The trace of the covariance matrix. Indicates the heterogeneity sensitivity parameter. This represents the spatial variance of the normalized vegetation index within the x-neighborhood window of the spatial cell coordinates. This represents the mean spatial variance of the normalized vegetation index for the entire region. Let represent the basic error covariance matrix of the triplet, and x represent the spatial pixel coordinates.

8. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 1, characterized in that, S4 includes: Based on the data quality index and the multi-source evapotranspiration data, an energy balance penalty term representing the upper limit of available energy and a water balance penalty term representing the monthly water mismatch are embedded in the lower bound objective function of variational inference to construct a physical enhancement evidence lower bound under the physical enhancement variational Bayesian adaptive weight update mechanism. By maximizing the lower bound of the physical augmentation evidence, the weight parameters are iteratively updated to obtain the spatiotemporally adaptive dynamic weights of each data source within each functional area.

9. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 8, characterized in that, The expressions for the lower bound of the physical enhancement evidence and the included energy balance penalty term and water balance penalty term are as follows: ； ； ； The expression for the spatiotemporal adaptive dynamic weight is: ； ； ； in, Let q denote the lower bound of the physical enhancement evidence, q denote the variational approximation distribution, and p denote the prior distribution. Let q represent the expectation under distribution q. Represents the relative entropy divergence. This represents observed evapotranspiration data. Indicates the surface type adaptability coefficient. This represents the adaptive penalty coefficient for energy balance at the current time t. This represents the adaptive penalty coefficient for water balance in the k-th functional zone at time t. Represents the energy balance residual term. This represents the water balance residual term. Indicates the amount of fusion evaporation. This indicates the evaporation ratio at the current moment. G represents the net surface radiation flux, and G represents the soil heat flux. This represents the joint error variance of available energy observations. This represents the reference evapotranspiration for water balance. This represents the joint error variance of the observations of each component of the water balance. An indicator function representing the availability of water balance data. and These represent the posterior accuracy parameters before and after the iterative update, respectively. and Let represent the posterior mean parameters before and after the iterative update, respectively. Indicates the lower bound of physically enhanced evidence. This represents a truncation operator that projects a value onto the interval [0.5, 2.0]. This represents spatiotemporal adaptive dynamic weights. This represents the base weight of each data source, and N represents the total number of data sources.

10. The oasis water consumption estimation method based on multi-source evapotranspiration fusion and dynamic weight optimization according to claim 1, characterized in that, S6 includes: Based on the preliminary evapotranspiration field, the evaporation ratio is independently retrieved from the surface temperature and albedo characteristic space using an energy balance model to determine the upper limit of available energy. The energy balance upper limit constraint correction is then applied to the preliminary evapotranspiration field to obtain the corrected evapotranspiration field. Based on the changes in precipitation, irrigation water diversion, runoff and soil water storage, the water balance model is used to calculate the reference evapotranspiration for water balance. The water balance reference evapotranspiration is used to perform water balance verification and linear scale correction on the corrected evapotranspiration field on a monthly scale to obtain the final evapotranspiration field of each functional area.

Images