A vegetation water consumption estimation method based on satellite remote sensing

Abstract

This application discloses a method for estimating vegetation water consumption based on satellite remote sensing. The method may include: constructing an input dataset of multi-source remote sensing, hydrological, and meteorological data; constructing a dual physical-spatial constraint objective from both bottom-up and top-down dimensions; constructing a U-Net deep learning model and training it based on a hybrid physical loss function; and inputting the input dataset into the trained U-Net deep learning model to estimate vegetation water consumption for the target time period. This invention achieves high-precision, high spatiotemporal consistency, and physical interpretability quantitative monitoring and assessment of vegetation water consumption in complex underlying surface areas, providing reliable technical support for precise regional water resource management, ecological protection decision-making, and water-saving policy formulation.

Description

Technical Field

[0001] This invention relates to the field of evapotranspiration estimation, and more specifically, to a method for estimating vegetation water consumption based on satellite remote sensing. Background Technology

[0002] Vegetation water consumption is typically characterized by actual evapotranspiration (ET), a key hydrological element in the terrestrial water and energy cycles, directly reflecting the comprehensive characteristics of water and energy exchange within the vegetation-soil-atmosphere continuum (SPAC). As a crucial component of watershed water balance, evapotranspiration is a core parameter in water resource assessment, eco-hydrological process simulation, and regional water resource regulation and management. Particularly in arid and semi-arid regions, evapotranspiration often constitutes a major proportion of terrestrial precipitation expenditure, and the accuracy of its estimation directly impacts the rational allocation of regional water resources, the effectiveness evaluation of ecological restoration projects, and the scientific validity of soil and water conservation and related engineering decisions.

[0003] In existing technologies, representative traditional methods for evapotranspiration observation mainly include lysimeter methods, Bowen ratio-energy balance methods, and eddy covariance methods. Although these methods have clear physical mechanisms and high single-point observation accuracy, they are limited by the spatial distribution density of observation stations and surface heterogeneity. They are essentially "point"-scale observations, resulting in limited spatial representation of the results. When dealing with complex terrain or large-scale watershed studies, it is difficult to effectively extrapolate the spatiotemporal distribution characteristics of regional-scale evapotranspiration using limited station data.

[0004] With the development of remote sensing technology, satellite remote sensing, due to its advantages of macroscopic, continuous, and dynamic observation, has provided favorable conditions for large-scale evapotranspiration estimation. Currently, various evapotranspiration estimation models based on remote sensing data have been developed both domestically and internationally. Representative models include the Penman-Monteith-Leuning (PML) model, the Global Land Evapotranspiration Amsterdam (GLEAM) model, and the Shuttleworth-Wallace two-layer model. These methods are mainly based on the surface energy balance remainder, evapotranspiration complementarity theory, or empirical statistical relationships, and retrieve evapotranspiration information through parameters such as surface temperature, vegetation index, and meteorological elements. Although such methods are widely used in long-term, global, or large-scale studies, they still have shortcomings in small- to medium-scale applications and complex underlying surfaces: existing physical models are highly sensitive to the accuracy of meteorological driving data and underlying surface parameters (such as impedance and conductivity), and the regionalization of parameters is difficult, which limits the universality of models under different climatic zones and underlying surface conditions; on the other hand, most remote sensing inversion methods focus on instantaneous or daily scale "top-down" energy balance estimation, lacking effective coupling with watershed macro-water balance and surface hydrological processes, resulting in deviations in the estimation results in terms of temporal continuity and physical consistency (such as water conservation).

[0005] In recent years, with the advancement of artificial intelligence technology, deep learning methods have been gradually introduced into the field of evapotranspiration estimation. By exploring the nonlinear relationship between multi-source remote sensing data and evapotranspiration, the accuracy of inversion has been improved. Deep learning models have, to some extent, improved inversion accuracy by uncovering the complex nonlinear mapping relationship between multi-source remote sensing data and evapotranspiration. However, existing data-driven deep learning models are often considered "black box" models, focusing primarily on the fitting performance of statistical indicators while neglecting the inherent physical constraints of hydrological processes. This lack of physical mechanisms results in limited generalization ability of the models when faced with scenarios outside the training data, and the output results often fail to meet the water balance closure requirements at the watershed scale.

[0006] Furthermore, a common challenge faced by existing technologies lies in the singularity and external dependence of data sources. Whether it's traditional physical models or emerging machine learning methods, their core inputs heavily rely on a few international satellite data sources (such as MODIS data carried by NASA's Terra / Aqua satellites) and international reanalysis data (such as ERA5 and GLDAS). With the Terra / Aqua satellites exceeding their service life, sensor aging, and uncertainties surrounding subsequent replacement plans, potential risks have been introduced to the long-term continuous and stable acquisition of ET data. This high dependence on a single foreign data source has become a bottleneck in building an independent and controllable regional water resources monitoring system.

[0007] Therefore, it is necessary to develop a method for estimating vegetation water consumption based on satellite remote sensing.

[0008] The information disclosed in the background section of this invention is intended only to enhance the understanding of the general background of this invention, and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0009] This invention proposes a method for estimating vegetation water consumption based on satellite remote sensing. It enables quantitative monitoring and assessment of vegetation water consumption (evapotranspiration) in complex underlying surface areas with high accuracy, high spatiotemporal consistency, and physical interpretability, providing reliable technical support for precise regional water resource management, ecological protection decision-making, and water conservation policy formulation.

[0010] This disclosure provides a method for estimating vegetation water consumption based on satellite remote sensing, including: Construct an input dataset of multi-source remote sensing, hydrological, and meteorological data; Construct physical-spatial dual-constraint objectives from both bottom-up and top-down dimensions; Construct a U-Net deep learning model and train it based on a hybrid physical loss function; The input dataset is fed into the trained U-Net deep learning model to estimate the vegetation water consumption during the target time period.

[0011] Preferably, the input dataset for constructing multi-source remote sensing, hydrological, and meteorological data includes: Acquire vegetation and surface remote sensing data, as well as meteorological driving data; The acquired data is subjected to projection transformation, spatial resampling, temporal aggregation, and normalization to obtain the input dataset.

[0012] Preferably, the physical constraints for water balance are constructed from a bottom-up perspective as follows:

[0013] In the formula, For the physical constraints of water balance, P is the monthly precipitation of the basin, Q is the monthly runoff of the basin, and ΔS is the monthly change in the total water storage of the basin.

[0014] Preferably, spatial distribution constraints are constructed from a top-down perspective, including: Spatial distribution constrained data are generated based on a water-carbon coupled remote sensing evapotranspiration model driven by leaf area index.

[0015] Preferably, the U-Net deep learning model is constructed by using a land use type embedding layer and a channel attention mechanism module.

[0016] Preferably, the land use feature embedding layer maps discrete categories of land use types into high-dimensional dense vectors.

[0017] Preferably, the channel attention mechanism module is a compression-excitation attention module. Compression involves performing global average pooling on the feature map to obtain the global receptive field for each channel; The incentive is to learn the dependencies between channels through two fully connected layers, generate channel weight coefficients, and multiply the weight coefficients back into the original feature map.

[0018] Preferably, model training based on a hybrid physical loss function includes: Based on the model's output, the water balance loss and spatial distribution loss are calculated separately, and then the loss function is calculated as follows:

[0019] in, For loss function, For water balance loss, For spatial distribution loss, These are the weighting coefficients; The Adam optimizer is used for parameter updates, and a dynamic learning rate adjustment strategy is introduced. When the loss function does not decrease significantly in multiple consecutive training cycles, the learning rate is automatically reduced in order to find the global optimum. During training, the model weights with the smallest loss value are saved as the trained U-Net deep learning model.

[0020] Preferably, the water balance loss is:

[0021] in, Let n be the output of the model corresponding to the i-th sample in the input dataset, and n be the number of samples in the input dataset. The physical constraint for water balance corresponding to the i-th sample in the input dataset.

[0022] Preferably, the spatial distribution loss is:

[0023] in, Let n be the output of the model corresponding to the i-th sample in the input dataset, and n be the number of samples in the input dataset. The spatial distribution constraint is defined for the i-th sample in the input dataset.

[0024] Its beneficial effects are as follows: 1) Combining deep learning models with watershed hydrological processes effectively avoids the physical inconsistency problem that may occur in the extrapolation application of pure data-driven models, and significantly improves the physical rationality and hydrological consistency of evapotranspiration estimation results. 2) By comprehensively utilizing multi-source information such as domestic satellite remote sensing data, GRACE water storage change data, CLDAS meteorological data, and measured runoff data from hydrological stations, the vegetation water consumption estimation process is constrained by multiple data sources, which improves the model's ability to characterize regional water cycle characteristics and thus enhances the overall accuracy and stability of vegetation water consumption estimation. 3) This invention innovatively introduces an embedding layer and channel attention mechanism guided by land use type data (CLCD) into the U-Net model. This feature enables the model to "perceive" the properties of the underlying surface, and can adaptively adjust the weight distribution of each feature channel according to the physical characteristics of different land types such as farmland, forest, grassland, and desert. This greatly enhances the model's feature extraction capability under non-uniform underlying surface conditions, and significantly improves the simulation accuracy and robustness in areas with large topographic relief and high vegetation fragmentation.

[0025] The method of the present invention has other features and advantages that will be apparent from or will be set forth in detail in the accompanying drawings and following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0026] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same parts.

[0027] Figure 1 A flowchart illustrating the steps of a satellite remote sensing-based vegetation water consumption estimation method according to an embodiment of the present invention is shown.

[0028] Figure 2 A schematic diagram showing the monthly verification results of vegetation water consumption in the Yellow River Basin according to an embodiment of the present invention is provided.

[0029] Figure 3 A schematic diagram showing the spatial distribution of vegetation water consumption in the Yellow River Basin according to an embodiment of the present invention is provided. Detailed Implementation

[0030] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0031] Figure 1 A flowchart illustrating the steps of a satellite remote sensing-based vegetation water consumption estimation method according to an embodiment of the present invention is shown.

[0032] like Figure 1 As shown, this satellite remote sensing-based vegetation water consumption estimation method includes: Step 101: Construct the input dataset of multi-source remote sensing, hydrological and meteorological data; Step 102: Construct a dual physical-spatial constraint objective from both bottom-up and top-down dimensions; Step 103: Construct the U-Net deep learning model and train the model based on the hybrid physical loss function; Step 104: Input the input dataset into the trained U-Net deep learning model to estimate the vegetation water consumption during the target time period.

[0033] In one example, the input dataset for constructing multi-source remote sensing, hydrological, and meteorological data includes: Acquire vegetation and surface remote sensing data, as well as meteorological driving data; The acquired data is subjected to projection transformation, spatial resampling, temporal aggregation, and normalization to obtain the input dataset.

[0034] In one example, the water balance physical constraints are constructed from a bottom-up perspective as follows:

[0035] In the formula, For the physical constraints of water balance, P is the monthly precipitation of the basin, Q is the monthly runoff of the basin, and ΔS is the monthly change in the total water storage of the basin.

[0036] In one example, spatial distribution constraints are constructed from a top-down perspective, including: Spatial distribution constrained data are generated based on a water-carbon coupled remote sensing evapotranspiration model driven by leaf area index.

[0037] In one example, a U-Net deep learning model is built using a land use type embedding layer and a channel attention mechanism module.

[0038] In one example, the land use feature embedding layer maps discrete categories of land use types to high-dimensional dense vectors.

[0039] In one example, the channel attention mechanism module is a compression-stimulation attention module. Compression involves performing global average pooling on the feature map to obtain the global receptive field for each channel; The incentive is to learn the dependencies between channels through two fully connected layers, generate channel weight coefficients, and multiply the weight coefficients back into the original feature map.

[0040] In one example, model training based on a hybrid physics loss function includes: Based on the model's output, the water balance loss and spatial distribution loss are calculated separately, and then the loss function is calculated as follows:

[0041] in, For loss function, For water balance loss, For spatial distribution loss, These are the weighting coefficients; The Adam optimizer is used for parameter updates, and a dynamic learning rate adjustment strategy is introduced. When the loss function does not decrease significantly in multiple consecutive training cycles, the learning rate is automatically reduced in order to find the global optimum. During training, the model weights with the smallest loss value are saved as the trained U-Net deep learning model.

[0042] In one example, the water balance loss is:

[0043] in, Let n be the output of the model corresponding to the i-th sample in the input dataset, and n be the number of samples in the input dataset. The physical constraint for water balance corresponding to the i-th sample in the input dataset.

[0044] In one example, the spatial distribution loss is:

[0045] in, Let n be the output of the model corresponding to the i-th sample in the input dataset, and n be the number of samples in the input dataset. The spatial distribution constraint is defined for the i-th sample in the input dataset.

[0046] Specifically, the core of this invention lies in constructing a deep learning-based vegetation water consumption estimation method that combines data-driven and physical mechanism collaboration. This method overcomes the dependence of traditional remote sensing inversion on foreign satellite surface parameter data, relying entirely on domestic Fengyun satellite data, China's first-generation global land surface reanalysis daily product (CRA-Land) reanalysis data, and meteorological-driven data from the China Meteorological Administration's land surface data assimilation system (CLDAS-V2.0). It utilizes a U-Net network with an attention mechanism to extract nonlinear features from multi-source data. Simultaneously, this invention innovatively designs a hybrid physical loss function, constructing a dual constraint system: a "bottom-up" physical mechanism hard constraint based on the true value of watershed water balance calculated using the GRACE gravity satellite, and a "top-down" spatial distribution soft constraint based on the spatial distribution information of existing remote sensing products. This achieves high-precision, physically consistent estimation of vegetation water consumption in complex areas such as the Yellow River's "U-shaped bend." The specific steps of this invention are as follows: Step 1: Construct the input dataset of multi-source remote sensing, hydrological and meteorological data This step is used to acquire and preprocess multi-source remote sensing data and meteorological drivers for vegetation water consumption estimation, and to establish a model input dataset with a consistent spatial reference system, time scale, and resolution.

[0047] (1) Data acquisition: Vegetation and surface remote sensing data: Leaf area index (LAI) products retrieved from the medium resolution spectral imager (MERSI-II) of the domestically produced Fengyun-3D satellite (FY-3D), used to characterize vegetation growth status and its temporal variation characteristics; Albedo from China's first-generation global land surface reanalysis daily product (CMD-Land), used to characterize surface energy characteristics; China's annual land cover dataset (CLCD), used to distinguish different land use types and underlying surface characteristics.

[0048] Meteorological driving data: Near-surface temperature, specific humidity, and precipitation data provided by the high-resolution near-real-time product dataset of the China Meteorological Administration's Land Surface Data Assimilation System (CLDAS-V2.0) are used to describe the land-atmosphere interaction conditions.

[0049] (2) Data preprocessing and spatiotemporal unification: Projection transformation: Reproject all raster data to the Albers equal-area conic projection coordinate system. Spatial resampling: All raster data were resampled to a standard grid with a 1km resolution. Bilinear interpolation was used for continuous variables (LAI, Albedo, meteorological data); mode interpolation was used for discrete variables (CLCD land use).

[0050] Temporal aggregation: Data with varying temporal resolutions (daily, ten-day) are aggregated into a monthly scale. Specifically, precipitation data uses monthly cumulative values, while other meteorological and vegetation parameters use monthly average values.

[0051] Normalization: The dynamic input variables (LAI, Albedo, meteorological data) are standardized and their values ​​are mapped to the [0, 1] interval to eliminate dimensional differences and accelerate model convergence.

[0052] Step 2: Construct a dual "physical-spatial" constraint objective for the model simulation results.

[0053] This step aims to generate "truth" labels for training deep learning models, including physical-mechanism-based watershed totals labels and spatial-prior-based distribution labels.

[0054] (1) Constructing bottom-up physical constraints for water balance ( Based on the assumption of a closed-loop water balance in a watershed, the water balance method is an effective method for verifying the accuracy of evapotranspiration data and evaluating the accuracy of evapotranspiration products at the watershed scale. The formula is as follows:

[0055] In the formula, P represents the monthly precipitation of the watershed, in mm / mon. -1 Q represents the monthly runoff of the basin, in mm yr⁻¹; ΔS represents the monthly change in total water storage of the basin, in mm mon. -1 The quantification of ΔS used data from three “Gravity Recovery and Climate Experiment” satellite products, including the Center for Space Research (CSR), the Jet Propulsion Laboratory (JPL), and the Goddard Space Flight Center (GSFC).

[0056] The "bottom-up" dimension refers to the total quantity constraint derived and constructed at the overall watershed scale by using basic hydrological and meteorological observations and remote sensing inversion elements (changes in precipitation, runoff, and water storage) within the watershed. This process does not rely on complex empirical parameters but strictly follows the physical mechanism of water balance to calculate the regional evapotranspiration baseline. By introducing this bottom-up watershed water conservation constraint, it is possible to effectively prevent overfitting, numerical divergence, or violations of macroscopic physical laws in data-driven deep learning models during iterative training, ensuring the physical authenticity and lower limit guarantee of vegetation water consumption estimation results, thereby improving the physical rationality, consistency, and stability of vegetation water consumption estimation results.

[0057] This invention introduces a water balance constraint mechanism into a deep learning model to construct a vegetation water consumption estimation framework that combines physical constraints and data-driven approaches. This enables a fine depiction of the spatial distribution of remote sensing inversion results and a synergistic unification of water conservation constraints at the watershed scale, thereby improving the physical rationality, consistency, and stability of vegetation water consumption estimation results.

[0058] (2) Constructing top-down spatial distribution constraints ( ): The PML-V2 (Penman-Monteith-Leuning Version 2) model is a diagnostic remote sensing evapotranspiration model based on the water-carbon coupling theory. Specifically, this model uses the Penman-Monteith equation as its theoretical framework and couples it with an improved Ball-Berry-Leuning stomatal conductance model. Driven by meteorological data (including temperature, precipitation, air pressure, wind speed, net radiation, water vapor pressure difference, etc.) and remote sensing surface characteristic parameters (including LAI, albedo, etc.), it achieves regional-scale evapotranspiration simulation.

[0059] This invention uses LAI products from the Fengyun-3D satellite as the main remote sensing data and the CLDAS-V2.0 meteorological dataset as input to drive the PML-V2 model, generating ET data with a spatial resolution of 1km in the Yellow River Basin, which serves as a spatial distribution constraint. ).

[0060] Unlike traditional methods that extend from single-point observations to regional scales, the model generated based on the PML-V2 model... The data can more accurately capture and characterize the spatial heterogeneity of vegetation water consumption in the Yellow River Basin under complex terrain and diverse underlying surface conditions. Although the data may have biases in its performance on water balance, as a high-resolution (1km) "spatial prior" information, it can provide deep learning models with fine spatial distribution patterns and texture features "from top to bottom".

[0061] By constructing the above-mentioned "physical-spatial" dual constraint objective, the present invention enables the model to strictly satisfy the macroscopic water conservation at the watershed scale and reconstruct spatial distribution characteristics that conform to the mechanism during the iterative optimization process.

[0062] Step 3: Construct the U-Net deep learning model with an attention mechanism

[0063] This invention constructs an improved U-Net convolutional neural network, characterized by the introduction of land use type embedding and channel attention mechanism (SE-Attention).

[0064] (1) Input layer design: The model input tensor has dimensions (B, C, H, W), where B is the model batch size, which is adjusted according to the model training and running equipment configuration; H and W are the image height and width; C is the number of input channels, which is divided into two parts: the first 5 channels are dynamic variables (LAI, Albedo, Temperature, Specific Humidity, Precipitation), and the 6th channel is a static variable (CLCD land use type code).

[0065] (2) Land Use Feature Embedding Layer: Since land use types (1=farmland, 2=forest, 3=open shrubland, 4=grassland, 5=water body, 6=ice / snow, 7=desert, 8=impermeable surface / urban, 9=wetland) are discrete integers, directly inputting them into a convolutional network cannot reflect the physical relationships between underlying surfaces. This invention designs an embedding layer that maps the discrete categories of CLCD to a 5-dimensional vector composed of continuous numerical values. This embedding layer is then concatenated with dynamic variable features in the channel dimension, enabling the model to adaptively adjust the evapotranspiration estimation strategy according to different underlying surface types.

[0066] (3) Encoder: Employs a 3-stage downsampling structure. Each stage contains two consecutive 3-stage downsampling units. 3 convolutional layers (followed by batch size and ReLU activation function) and a 2 2. Max pooling layer. As the layer depth increases, the feature map size is gradually halved, and the number of channels is gradually doubled (64->128->256) to extract multi-scale features from local texture to global semantics.

[0067] (4) Channel Attention Mechanism Module (SE-Attention): A squeeze-and-excitation attention module is introduced in the deepest layer (bottleneck layer) of the encoder. Squeeze-and-excitation involves global average pooling of the feature map to obtain the global receptive field of each channel. Excitation learns the dependencies between channels through two fully connected layers to generate channel weight coefficients (range 0-1). Finally, the weight coefficients are multiplied back into the original feature map. This mechanism enables the model to automatically "focus" on the feature channels that contribute the most to the current evapotranspiration estimation (e.g., automatically assigning higher weights to precipitation in arid areas and higher weights to LAI during the growing season). This invention, by constructing a deep learning model that incorporates an attention mechanism, fully explores the nonlinear relationship between multi-source remote sensing data, meteorological elements, vegetation cover type, and topographic factors and vegetation water consumption, thereby improving the accuracy and spatiotemporal continuity of vegetation water consumption estimation results under complex underlying surface conditions.

[0068] (5) Decoder: A 3-level upsampling structure is adopted. Each level uses transposed convolution to restore the feature map size. The U-Net model uses skip connections to concatenate the decoder features with the corresponding level features of the encoder to compensate for the spatial details lost during downsampling and ensure high resolution of the output.

[0069] (6) Output layer: via 1 The first convolutional layer maps the feature map to a single-channel output and uses the Sigmoid activation function to constrain the result to a normalized range, outputting a predicted spatial distribution map of vegetation water consumption. ).

[0070] Step 4: Model Training Based on Hybrid Physical Loss Function

[0071] To ensure that the model output conforms to both physical conservation laws and has a reasonable spatial distribution, this invention designs a hybrid physical loss function. It is used to guide the iterative optimization of model parameters.

[0072] (1) Definition of loss function:

[0073] In the formula, This is a weighting coefficient used to balance the importance of physical and spatial constraints in water balance.

[0074] (2) Water balance loss ): The calculation model outputs a prediction map ( The ET value within the Yellow River basin is then compared with the value obtained in step two based on the water balance equation. The mean squared error (MSE) between the two sides, this constraint, through bottom-up physical constraints, ensures that the model follows the water balance equation at the macroscopic scale of the region:

[0075] in, Let n be the output of the model corresponding to the i-th sample in the input dataset, and n be the number of samples in the input dataset. The physical constraint for water balance corresponding to the i-th sample in the input dataset.

[0076] (3) Spatial distribution loss ): The computational model outputs prediction results and spatial constraints ( At the pixel-scale mean square error, this constraint, applied top-down, makes the model estimation results more spatially reasonable:

[0077] in, Let n be the output of the model corresponding to the i-th sample in the input dataset, and n be the number of samples in the input dataset. The spatial distribution constraint is defined for the i-th sample in the input dataset.

[0078] (4) Training strategy: The Adam optimizer is used for parameter updates, and a dynamic learning rate adjustment strategy is introduced. When the loss function does not decrease significantly within a certain number of consecutive training epochs, the learning rate is automatically reduced to find the global optimum. During training, only the model weights with the smallest loss value on the validation set are saved as the optimal model.

[0079] Step 5: Estimation and Output of Regional Vegetation Water Consumption

[0080] The optimal U-Net deep learning model is trained to estimate the target time period, and finally high-resolution vegetation water consumption data of the Yellow River Basin is generated. Furthermore, the vegetation water consumption of each sub-basin and each administrative region can be statistically analyzed.

[0081] Overall, this invention is entirely driven by domestic Fengyun satellites and domestic meteorological assimilation data, overcoming the dependence on foreign satellite data such as MODIS. Furthermore, the U-Net deep learning model constructed in this invention embeds the water balance equation as a hard constraint into the loss function, constraining the simulation process from the bottom up and significantly improving physical reliability. Simultaneously, thanks to the spatial feature extraction capabilities of the U-Net architecture and the channel weighting capability of the attention mechanism, it can effectively capture the spatial heterogeneity of evapotranspiration in complex terrains and underlying surfaces (such as the Yellow River's "U-shaped bend" and the surrounding desert and farmland transition zone), solving the problem of traditional hydrological methods struggling to obtain high-resolution spatial distribution data.

[0082] To facilitate understanding of the solutions and effects of the embodiments of the present invention, a specific application example is given below. Those skilled in the art should understand that this example is merely for the purpose of understanding the present invention, and any specific details therein are not intended to limit the present invention in any way.

[0083] Example 1

[0084] To verify the effectiveness and superiority of the vegetation water consumption estimation method based on domestic satellites and physical constraints proposed in this invention, this embodiment selects the Yellow River Basin as a typical research area for application demonstration.

[0085] This embodiment selects 2020 to 2024 as the study period. In the model construction process, it first utilizes precipitation data from the China Meteorological Administration's CLDAS-V2.0 system, land water storage data retrieved from the GRACE gravity satellite, and measured runoff data from the Lijin hydrological station, the outlet control station of the Yellow River basin, based on the water balance equation (…). ) Calculate the true value of evapotranspiration in the Yellow River Basin ( This is used to construct a "bottom-up" hard constraint on the total physical quantity of the model; at the same time, evapotranspiration data generated by the PML-V2 model driven by the domestic Fengyun-3D satellite are used. As prior knowledge, the model is constructed using "top-down" spatial distribution soft constraints.

[0086] In terms of dataset partitioning, the multi-source spatiotemporal data from 2020 to 2023 were used as the training set to drive the U-Net deep learning model constructed in this invention to learn the nonlinear mapping relationship and physical characteristics between vegetation water consumption and multi-source factors; the data from 2024 were used as an independent test set to verify the model's prediction accuracy and generalization ability.

[0087] Figure 2 A schematic diagram showing the monthly verification results of vegetation water consumption in the Yellow River Basin according to an embodiment of the present invention is provided.

[0088] Figure 2This figure presents a time-series comparison of monthly ET values ​​in the Yellow River Basin from 2020 to 2024. The curves in the figure represent: the physical true values ​​calculated based on the water balance equation (…). ), the evapotranspiration estimated by the U-Net deep learning model of this invention ( Reference values ​​retrieved from Fengyun satellites ( (and, in contrast, mainstream international satellite remote sensing products) and As can be seen from the figure, the present invention estimates... With physical truth value The model maintains a high degree of consistency in seasonal fluctuation trends and peak characteristics, indicating that it has successfully learned the water balance laws at the watershed scale. Table 1 details the algorithm of this invention. , , and With watershed water balance calculation Error index between them.

[0089] Table 1. Algorithm Predictions and Error Statistics for Each ET Product

[0090] Figure 3 A schematic diagram showing the spatial distribution of vegetation water consumption in the Yellow River Basin according to an embodiment of the present invention is provided.

[0091] Figure 3 The results demonstrate the spatial distribution pattern of annual cumulative vegetation water consumption during the test period (2024). Among them: Figure 3 (a) is a spatial distribution map of vegetation water consumption predicted by the model of this invention. ), Figure 3 (b) Distribution map of Fengyun satellites retrieved as a space reference ( ); Figure 3 (c) is a spatial distribution map of the differences between the model predictions and the reference values. The comparative results show that the method of the present invention can not only accurately depict the macroscopic spatial texture features of the Yellow River Basin, but also effectively correct the deviations in local areas, thus achieving high-precision spatial mapping.

[0092] Those skilled in the art should understand that the above description of the embodiments of the present invention is only intended to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any of the examples given.

[0093] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for estimating vegetation water consumption based on satellite remote sensing, characterized in that, include: Construct an input dataset of multi-source remote sensing, hydrological, and meteorological data; Construct physical-spatial dual-constraint objectives from both bottom-up and top-down dimensions; Construct a U-Net deep learning model and train it based on a hybrid physical loss function; The input dataset is fed into the trained U-Net deep learning model to estimate the vegetation water consumption during the target time period.

2. The satellite remote sensing based method of estimating vegetation water use of claim 1, wherein, The input dataset for constructing multi-source remote sensing, hydrological, and meteorological data includes: Acquire vegetation and surface remote sensing data, as well as meteorological driving data; The acquired data is subjected to projection transformation, spatial resampling, temporal aggregation, and normalization to obtain the input dataset.

3. The satellite remote sensing based method of estimating vegetation water use of claim 1, wherein, The physical constraints for water balance are constructed from a bottom-up perspective as follows: In the formula, For the physical constraints of water balance, P is the monthly precipitation of the basin, Q is the monthly runoff of the basin, and ΔS is the monthly change in total water storage of the basin.

4. The method for estimating vegetation water consumption based on satellite remote sensing according to claim 3, wherein, Constructing spatial distribution constraints from a top-down perspective includes: A spatial distribution constraint is generated based on a water-carbon coupled remote sensing evapotranspiration model driven by leaf area index.

5. The satellite remote sensing based method of estimating vegetation water use of claim 4, wherein, The U-Net deep learning model is constructed using a land use type embedding layer and a channel attention mechanism module.

6. The satellite remote sensing based method of estimating vegetation water use of claim 5, wherein, The land use feature embedding layer maps discrete categories of land use types into high-dimensional dense vectors.

7. The satellite remote sensing based method of estimating evapotranspiration from vegetation as claimed in claim 5, wherein, The channel attention mechanism module is a compression-excitation attention module. Compression involves performing global average pooling on the feature map to obtain the global receptive field for each channel; The incentive is to learn the dependencies between channels through two fully connected layers, generate channel weight coefficients, and multiply the weight coefficients back into the original feature map.

8. The satellite remote sensing based method of estimating vegetation water use of claim 7, wherein, Model training based on hybrid physical loss functions includes: Based on the model's output, the water balance loss and spatial distribution loss are calculated separately, and then the loss function is calculated as follows: wherein, is a loss function, is a water balance loss, is a spatial distribution loss, is a weight coefficient; The Adam optimizer is used for parameter updates, and a dynamic learning rate adjustment strategy is introduced. When the loss function does not decrease significantly in multiple consecutive training cycles, the learning rate is automatically reduced in order to find the global optimum. During training, the model weights with the smallest loss value are saved as the trained U-Net deep learning model.

9. The satellite remote sensing based method of estimating evapotranspiration from vegetation as claimed in claim 8, wherein, The water balance loss is: wherein, is the output of the model corresponding to the i-th sample in the input dataset, n is the number of samples in the input dataset, is the water balance physical constraint corresponding to the i-th sample in the input dataset.

10. The satellite remote sensing based method of estimating vegetation water use of claim 8, wherein, The spatial distribution loss is: in, Let n be the output of the model corresponding to the i-th sample in the input dataset, and n be the number of samples in the input dataset. The spatial distribution constraint is defined for the i-th sample in the input dataset.

Images