Ground surface evapotranspiration calculation method, device and equipment and storage medium

Abstract

The invention relates to the field of artificial intelligence, and discloses a surface evapotranspiration calculation method, device and equipment and a storage medium. Performing layer-by-layer decomposition on the remote sensing time sequence data through an earth surface evapotranspiration model, and extracting features of different time scales to obtain time sequence features; carrying out residual feature fusion on the time sequence features through a residual feature fusion module; inputting the fused feature tensor into a time sequence feature coding module, establishing a dynamic dependency relationship between time steps through a self-attention mechanism, and outputting a tail end time step feature; and taking the end time step characteristics as intermediate variables of evapotranspiration, taking output of an intermediate variable simulation layer as input of a corresponding physical model in a forward propagation process, and calculating and outputting the surface evapotranspiration amount of the target area through a physical constraint module. And a more accurate and reliable result is provided for surface evapotranspiration prediction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and storage medium for calculating surface evapotranspiration. Background Art

[0002] As the core link in the water cycle, evapotranspiration (ET) has a profound impact on regional and global climate, water resource management, and the stable operation of ecosystems. Accurate modeling of ET is of great practical significance for rationally planning water resources, improving agricultural irrigation efficiency, and assessing the impact of climate change on the ecological environment. Precise estimation of ET not only improves the accuracy of crop yield estimates but also plays a vital role in the management of water and agricultural resources. Broadly speaking, existing ET models can be categorized as either physically driven or data-driven.

[0003] In terms of physical process-driven models (hereafter referred to as physical models), traditional physical models such as the Penman-Monteith model and the SEBAL model, based on a physical understanding of evapotranspiration, use a series of mathematical equations to describe the dynamic process of water vapor exchange. The main advantage of physical models is their strong interpretability because they are based on known physical mechanisms. However, such models also have significant limitations. Many physical models require a large amount of field observation data for calibration, which is sometimes difficult to obtain. At the same time, the operation of such models requires a large number of input parameters and computing resources, and the empirical or semi-empirical methods for determining key parameters are mostly rigid and difficult to take into account most of the interacting processes in the atmosphere-ecosphere system, resulting in poor adaptability to complex environments.

[0004] In contrast to physical models, data-driven models are available. Some studies are based on traditional regression models, such as random forests and Gaussian process regression, and model evapotranspiration by fitting relationships between evapotranspiration and meteorological factors at a single time point. This approach has the advantages of a simple model structure, ease of implementation, and, under certain conditions, good simulation results. However, traditional regression models also have limitations, such as their inability to capture dynamic changes and complex nonlinear relationships in the data. To overcome these limitations, time series models based on time windows, such as spARIMA and Prophet, have emerged. These models divide time series data into multiple time windows and use the data within these windows to establish fitted relationships, thereby capturing dynamic changes and complex nonlinear characteristics in the data. Compared to traditional regression modeling, which only uses a single time point to fit data relationships, time series modeling methods based on time windows can better adapt to dynamic data changes and improve simulation accuracy. Both traditional regression and time series models, as purely data-driven models, rely on simple statistical methods, shallow machine learning, or deep learning methods, trained on large amounts of historical data to simulate evapotranspiration. Such models are often regarded as black box models because they only care about the accuracy of the simulation results and ignore the physical mechanisms behind them, resulting in poor fitting results when there is insufficient training data and extreme meteorological conditions.

[0005] Therefore, how to accurately calculate surface evapotranspiration is a technical problem that needs to be solved urgently.

[0006] The above content is only used to assist in understanding the technical solution of the present invention and does not constitute an admission that the above content is prior art. Summary of the Invention

[0007] The main purpose of the present invention is to provide a method, device, equipment and storage medium for calculating surface evapotranspiration, aiming to solve the technical problem of poor accuracy of surface evapotranspiration calculation in the prior art.

[0008] To achieve the above object, the present invention provides a method for calculating surface evapotranspiration, which comprises:

[0009] Obtain remote sensing time series data of the target area;

[0010] Decomposing the remote sensing time series data layer by layer through a surface evapotranspiration model, extracting features of different time scales, and obtaining time series features;

[0011] Performing residual feature fusion on the time series features through a residual feature fusion module;

[0012] The fused feature tensor is input into the temporal feature encoding module, and the dynamic dependency between time steps is established through the self-attention mechanism, and the terminal time step features are output;

[0013] The terminal time step feature is used as an intermediate variable of evapotranspiration. In the forward propagation process, the output of the intermediate variable simulation layer is used as the input of the corresponding physical model. The surface evapotranspiration of the target area is output through calculation through the physical constraint module.

[0014] Preferably, the temporal feature encoding module adopts a dual-channel convolution structure to achieve multi-scale decomposition of the signal;

[0015] The temporal feature encoding module includes two one-dimensional convolution operators P and U, both of which adopt depth-wise separable convolution design;

[0016] Given input signal Where B is the batch size, C is the number of channels, T is the time series length, and the decomposition process is achieved by constructing a residual map;

[0017] First generate the approximate component residual p res =xP(x), and then generate the adaptive adjustment term through U convolution to calculate the approximate component a=x+U(p res ), and finally output detail component d=xP(a).

[0018] Preferably, the residual feature fusion module is a combination structure of linear transformation and nonlinear activation;

[0019] The residual feature fusion module receives the dimension B×T×(C×2 levels +C) and maps it to C×2 through a fully connected layer levels dimensional space, and apply the GELU activation function to achieve nonlinear transformation.

[0020] Preferably, the temporal feature encoding module adopts a standard Transformer encoder architecture;

[0021] The temporal feature encoding module consists of two encoder layers, each of which contains a multi-head self-attention mechanism and a feedforward neural network, and the temporal features adopt a sinusoidal encoding strategy.

[0022] Preferably, the physical constraint module is a PM model or a PT-JPL model.

[0023] Preferably, the intermediate variable simulation layer before the physical constraint module is composed of a fully connected network;

[0024] The fully connected network adopts a two-layer linear structure. The intermediate variables selected are the surface impedance input variable in the PM model and the soil moisture constraint factor input variable in the PT-JPL model. The specific coupling method is that the output of the intermediate variable simulation layer of the previous physical model is used as one of the inputs of the two physical models, and the physical model is used as part of the loss function.

[0025] In addition, the present invention also provides a surface evapotranspiration calculation device, which includes:

[0026] Acquisition module, used to obtain remote sensing time series data of the target area;

[0027] An extraction module is used to decompose the remote sensing time series data layer by layer through a surface evapotranspiration model, extract features of different time scales, and obtain time series features;

[0028] A feature fusion module, configured to perform residual feature fusion on the temporal features through a residual feature fusion module;

[0029] The feature encoding module is used to input the fused feature tensor into the time series feature encoding module, establish dynamic dependencies between time steps through the self-attention mechanism, and output the terminal time step features;

[0030] The physical constraint module is used to use the terminal time step characteristics as the intermediate variable of evapotranspiration, use the output of the intermediate variable simulation layer as the corresponding physical model input during the forward propagation process, calculate through the physical constraint module, and output the surface evapotranspiration of the target area.

[0031] In addition, to achieve the above-mentioned objectives, the present invention further provides a surface evapotranspiration calculation device, wherein a surface evapotranspiration calculation program is stored on the surface evapotranspiration calculation device, and when the surface evapotranspiration calculation program is executed by a processor, the steps of the surface evapotranspiration calculation method described above are implemented.

[0032] In addition, to achieve the above-mentioned object, the present invention further proposes a storage medium, on which a surface evapotranspiration calculation program is stored. When the surface evapotranspiration calculation program is executed by a processor, the steps of the surface evapotranspiration calculation method described above are implemented.

[0033] The present invention has the following advantages and positive effects:

[0034] (1) Effectively improve the stability and accuracy of evapotranspiration simulation under extreme meteorological conditions. The present invention effectively solves the problem of insufficient adaptability of a single physical model under extreme meteorological conditions through the deep coupling of physical constraints and time series decomposition. Traditional physical models use fixed parameterization schemes, which are difficult to accurately simulate nonlinear processes such as extreme high temperatures and droughts; and although pure data-driven models can fit complex relationships, the lack of physical constraints can easily produce outputs that violate the laws of nature. Physical constraints such as the energy balance equation ensure that the output conforms to the laws of thermodynamics and avoids non-physical solutions; on the other hand, multi-level time series decomposition is used to separate the characteristics of extreme events, so that the model can adjust the response strategy in a targeted manner.

[0035] (2) Effectively improve the generalization performance of evapotranspiration simulation under different ecosystem conditions. The present invention successfully overcomes the limitations of traditional methods in cross-ecosystem applications through the innovative design of using physical model parameters as intermediate variables for deep learning. A single physical model requires parameter adjustment for different ecosystems (such as forests, farmlands, deserts, etc.), while pure data-driven models often lack the generalization ability in new areas. The coupling framework proposed in the present invention can automatically learn the physical response laws of different underlying surfaces and can adapt to various ecosystems without manual parameter adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the structure of the surface evapotranspiration calculation method in the hardware operating environment involved in the embodiment of the present invention;

[0037] Figure 2 This is a schematic flow chart of a first embodiment of a method for calculating surface evapotranspiration according to the present invention;

[0038] Figure 3 This is a flow chart of the model construction in the embodiment of the surface evapotranspiration calculation method of the present invention;

[0039] Figure 4 This is an overall flow chart of data preprocessing in an embodiment of the surface evapotranspiration calculation method of the present invention;

[0040] Figure 5 The Pearson correlation heat map of each feature in the embodiment of the surface evapotranspiration calculation method of the present invention;

[0041] Figure 6 This is a diagram of the Tranformer model structure in an embodiment of the surface evapotranspiration calculation method of the present invention;

[0042] Figure 7 This is a fitting performance diagram of the physical model and the coupling model in the embodiment of the surface evapotranspiration calculation method of the present invention;

[0043] Figure 8A scatter plot of the overall generalization performance of the coupled model and the data-driven model in the embodiment of the surface evapotranspiration calculation method of the present invention;

[0044] Figure 9 This is a structural block diagram of the first embodiment of the surface evapotranspiration calculation device of the present invention.

[0045] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0046] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a surface evapotranspiration calculation device in the hardware operating environment involved in an embodiment of the present invention.

[0048] like Figure 1 As shown, the surface evapotranspiration calculation device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen. Optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. In the present invention, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a wireless fidelity (WI-FI) interface). The memory 1005 may be a high-speed random access memory (RAM) or a non-volatile memory (NVM), such as a disk drive. The memory 1005 may also be a storage device independent of the processor 1001.

[0049] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the surface evapotranspiration calculation device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0050] like Figure 1 As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a surface evapotranspiration calculation program.

[0051] exist Figure 1 In the illustrated evapotranspiration calculation device, the network interface 1004 is primarily used to connect to a backend server and communicate data with the backend server; the user interface 1003 is primarily used to connect to a user device; the evapotranspiration calculation device invokes a evapotranspiration calculation program stored in the memory 1005 via the processor 1001 and executes the evapotranspiration calculation method provided in an embodiment of the present invention.

[0052] Based on the above hardware structure, an embodiment of a method for calculating surface evapotranspiration of the present invention is proposed.

[0053] Reference Figure 2 , Figure 2 FIG1 is a flow chart of the first embodiment of the method for calculating surface evapotranspiration according to the present invention, and provides the first embodiment of the method for calculating surface evapotranspiration according to the present invention.

[0054] By decomposing and reconstructing the evapotranspiration time series data through the time series model, the characteristics of different scales and periods in the data are extracted to better capture the dynamic changes and nonlinear relationships of the data; while physical constraints can provide physical rationality and prior knowledge for the time series model, thereby improving the model's predictive robustness. Coupling the deep learning time series model with physical constraints can give full play to the advantages of both, effectively overcome the limitations of both, and provide more accurate and reliable results for surface evapotranspiration prediction. The model construction process is as follows: Figure 3 The main steps include:

[0055] Step 1: Obtain remote sensing time series data of the target area and perform preprocessing work including three main processes: data cleaning, spatial and temporal resolution alignment, and feature engineering;

[0056] Step 2: Use the cascaded time series decomposition module to extract the time series features of the input sequence layer by layer;

[0057] Step 3: Perform residual feature fusion on the time series features obtained in step 2 to integrate the original input signal and the decomposed features to prevent excessive attenuation of low-frequency information during deep feature extraction.

[0058] Step 4: Input the fused feature tensor into the temporal feature encoding module and establish the dynamic dependency between time steps through the self-attention mechanism;

[0059] In step 5, the physical constraint module receives the terminal time step features output by the encoder layer in step 4 and uses them as intermediate variables for evapotranspiration. During the forward propagation process, the model uses the output of the intermediate variable simulation layer as the input of the corresponding physical model and calculates the final output through the physical model.

[0060] Specifically, in the first embodiment, the surface evapotranspiration calculation method includes the following steps:

[0061] Step S10: Acquire remote sensing time series data of the target area.

[0062] In a specific implementation, the execution subject of this embodiment is the surface evapotranspiration calculation device, wherein the surface evapotranspiration calculation device can be an electronic device such as a personal computer or a server, and this embodiment does not limit this. This embodiment provides a surface evapotranspiration simulation method that couples physical constraints with deep learning time series models. A basic data system for evapotranspiration modeling is constructed using multi-source heterogeneous data sets, and a spatiotemporal continuous feature input and verification benchmark is formed through the fusion processing of ground observations, satellite remote sensing and reanalysis data. The data set covers three major data sources: flux observation data based on the eddy covariance method, multispectral remote sensing inversion products, and reanalysis meteorological driven data sets. Each type of data forms complementary support for each other. The following is the detailed information of each data set:

[0063] (1) Eddy covariance station observation data

[0064] Eddy covariance flux observation data is used as the ground truth benchmark, and the observation process relies on the theoretical system of turbulent flux measurement. Based on the eddy covariance principle, the instantaneous sensible heat flux (H) and latent heat flux (λET) can be calculated as the covariance of the vertical wind speed fluctuation w′ and the water vapor concentration fluctuation q′: Where λ is the latent heat of vaporization and ρ is the air density. This study integrates the locations of 25 representative ecological sites of the China Flux Observation Alliance (ChinaFlux), covering a geographical range of 21.92°–51.78°N and 75.05°–128.98°E from 2003 to 2020, encompassing six ecosystems: farmland, forest, grassland, shrubland, wetland, and wasteland.

[0065] (2) Remote sensing inversion data

[0066] Remote sensing inversion data provides regional-scale surface dynamic information, including three key parameters: vegetation index, surface albedo, and soil moisture. The vegetation index uses the MODIS Normalized Difference Vegetation Index (NDVI MOD13A1). This data product is based on Terra / Aqua binary satellite observations, with a temporal and spatial resolution of 16 days and a resolution of 1 km. The surface albedo data is derived from the MCD43A3 product, with a temporal and spatial resolution of days and 500 resolution. The black sky albedo of the first band (red light) was selected in this study. The soil moisture data uses the ESACCI soil moisture composite product of the European Space Agency's Climate Change Initiative project, with a spatial resolution of 0.25° and a temporal resolution of days.

[0067] (3) Meteorological driving data

[0068] The meteorological driving data used is the China Regional Surface Meteorological Element Driven Dataset v2.0 (CMFD v2.0). This dataset is compiled using the internationally available Princeton reanalysis data, GLDAS data, GEWEX-SRB radiation data, and TRMM precipitation data as background fields, integrated with conventional meteorological observations from the China Meteorological Administration. The overall accuracy of the dataset is between that of the Meteorological Administration's observational data and satellite remote sensing data, and is superior to the accuracy of existing international reanalysis data. The dataset has a temporal resolution of daily levels and a spatial resolution of 0.1°, spanning the period from 2003 to 2020, and covers seven key variables, including temperature, precipitation, and air pressure.

[0069] Furthermore, the data preprocessing of the acquired multivariate heterogeneous data can be divided into three processes: data cleaning, spatiotemporal resolution alignment, and feature engineering. The sub-steps of each process are as follows: Figure 4 shown.

[0070] (1) Data cleaning

[0071] The first step in data cleaning involves physically constraining feature boundaries and detecting anomalies. This process incorporates principles of meteorology, surface energy balance, and hydrological processes to define the appropriate range for each feature variable. It is important to note that detected outliers are further validated; if they are due to extreme meteorological events, they are retained.

[0072] (2) Temporal and spatial resolution alignment

[0073] Through the steps of projection unification, preliminary cropping, time series interpolation and spatial grid alignment, the problem of differences in spatiotemporal resolution of multi-source data is resolved, providing a high-quality and consistent data foundation for subsequent data analysis and model building.

[0074] (3) Feature Engineering

[0075] In the feature engineering phase, the focus is on feature construction and correlation verification of the original observation data. The feature construction part defines and calculates four key features: net radiation flux, water vapor pressure difference, time series characteristics, and site ecosystem:

[0076] The net radiation flux (Rn) is calculated by the difference between the net shortwave radiation and the net longwave radiation. The specific formula is:

[0077] Rn=(1-albedo)×Rs down +(Rl down -ε×σ×Ts 4 )#(1)

[0078] Among them, albedo is the surface albedo, Rs down For downward shortwave radiation, Rldowm is the downward longwave radiation, ε is the surface emissivity, σ is the Boltzmann constant, and Ts is the surface temperature.

[0079] The vapor pressure difference (VPD) is calculated using a common meteorological calculation method. Based on relative humidity data, the difference between the saturated vapor pressure and the actual vapor pressure is calculated to characterize the evaporation capacity of the air. The formula is as follows:

[0080]

[0081] Among them, Ta is the air temperature and RH is the relative humidity.

[0082] The time series feature (time_embed) uses a sinusoidal encoding strategy with a year as the cycle, converting the date into the dth day of the year, and constructing it as

[0083]

[0084] Where DOY is the number of days in the year.

[0085] The site ecosystem type (site_type) maps the six ecosystem types (farmland, forest, grassland, shrubland, wetland, and wasteland) to corresponding numbers based on the surface cover classification information in the site metadata to generate learnable features. In the feature correlation test stage, for each input feature, the Pearson correlation coefficient matrix between it and the target latent heat flux LE (which can be converted to actual evapotranspiration) is calculated, and the feature components with an absolute value of the correlation coefficient greater than 0.3 and less than 0.9 are screened. The Pearson correlation heat map of each feature is shown in the figure below. Figure 5 The figure shows the features after deletion. Among them, air pressure, soil moisture, wind speed, and ecosystem type do not reach a correlation of 0.3 with the target variable. However, since these variables are input parameters of the physical model, they need to be retained. It should be noted that these variables do not participate in model training.

[0086] Step S20: Decomposing the remote sensing time series data layer by layer through the surface evapotranspiration model, extracting features of different time scales, and obtaining time series features.

[0087] It is understandable that a time series input matrix is constructed with dimensions of B×T×C, where B represents the batch size, T is the time step, and C is the number of channels. The input sequence is decomposed into a multi-level time series. The multi-level time series decomposition module is the basic unit for feature extraction of the model. Inspired by wavelet lifting decomposition, the module uses a dual-channel convolution structure to achieve multi-scale decomposition of the signal. The module contains two one-dimensional convolution operators P and U, both of which use a depth-separable convolution design. Given an input signal Where B is the batch size, C is the number of channels, and T is the time series length. The decomposition process is achieved by constructing a residual map. First, the approximate component residual p is generated.res =xP(x), and then generate the adaptive adjustment term through U convolution to calculate the approximate component a=x+U(p res ), and finally outputs detail component d = xP(a). The mathematical expression of this process can be summarized as:

[0088]

[0089] Where l is the decomposition level. This design achieves step-by-step separation of time series signals through alternating residual calculations. The approximate component a preserves low-frequency trend information, while the detail component d captures high-frequency fluctuations. The introduction of depthwise separable convolution enables independent feature extraction for each channel, reducing the number of parameters while maintaining decoupling between channels.

[0090] The multi-level decomposition architecture achieves hierarchical feature abstraction by stacking adaptive layers. In the initialization phase, the channel dimension is gradually expanded according to the preset decomposition layer parameter levels, and the number of channels is doubled at each level of decomposition. In the forward propagation of the model, the above-mentioned adaptive time series decomposition modules are added to the module list, and the module list is iterated to obtain the module instances. The input signal is decomposed in turn, and the output approximate components and detail components are spliced and input into the next level module instance. For the levels layer decomposition, the final channel dimension is expanded to C×2 levels , forming a tensor containing multi-scale features This pyramid structure allows the model to capture signal characteristics at different temporal resolutions, providing sufficient feature representation for subsequent processing.

[0091] Furthermore, in this embodiment, the temporal feature encoding module adopts a dual-channel convolution structure to achieve multi-scale decomposition of the signal;

[0092] The temporal feature encoding module includes two one-dimensional convolution operators P and U, both of which adopt depth-wise separable convolution design;

[0093] Given input signal Where B is the batch size, C is the number of channels, T is the time series length, and the decomposition process is achieved by constructing a residual map;

[0094] First generate the approximate component residual p res =xP(x), and then generate the adaptive adjustment term through U convolution to calculate the approximate component a=x+U(p res ), and finally output detail component d=xP(a).

[0095] It should be noted that the multi-level time series decomposition module, as the basic unit for feature extraction of the model, uses a dual-channel convolution structure to achieve multi-scale decomposition of the signal. The module contains two one-dimensional convolution operators P and U, both of which use a depth-separable convolution design. Given an input signal Where B is the batch size, C is the number of channels, and T is the time series length. The decomposition process is achieved by constructing a residual map. First, the approximate component residual p is generated. res =xP(x), and then generate the adaptive adjustment term through U convolution to calculate the approximate component a=x+U(p res ), and finally output detail component d=xP(a).

[0096] Step S30: performing residual feature fusion on the temporal features through a residual feature fusion module.

[0097] It should be noted that the residual feature fusion is performed on the time series decomposition features. The residual feature fusion is designed as a combination of linear transformation and nonlinear activation. Its main function is to integrate the original input signal and the decomposition features. The module receives the dimension B×T×(C×2 levels +C) and maps it to C×2 through a fully connected layer levels dimensional space and applies the GELU activation function to achieve nonlinear transformation. The introduction of retaining the original input signal here aims to prevent excessive attenuation of low-frequency information during deep feature extraction and provide the time series feature encoding module with a feature representation that combines local details and global trends.

[0098] Furthermore, in this embodiment, the residual feature fusion module is a combination structure of linear transformation and nonlinear activation;

[0099] The residual feature fusion module receives the dimension B×T×(C×2 levels +C) and maps it to C×2 through a fully connected layer leveks dimensional space, and apply the GELU activation function to achieve nonlinear transformation.

[0100] In the specific implementation, the residual feature fusion design is a combination of linear transformation and nonlinear activation. The module receives the dimension B×T×(C×2 kevels +C) and maps it to C×2 through a fully connected layer levels dimensional space, and apply the GELU activation function to achieve nonlinear transformation.

[0101] Step S40: Input the fused feature tensor into the temporal feature encoding module, establish a dynamic dependency relationship between time steps through the self-attention mechanism, and output the terminal time step feature.

[0102] It should be understood that the features obtained in the above steps are encoded using time series features. The time series feature encoding module uses the following method: Figure 6 The standard Transformer encoder architecture shown here consists of two encoder layers. Each encoder layer incorporates a multi-head self-attention mechanism and a feedforward neural network. This module receives the fused feature tensor and uses the self-attention mechanism to establish dynamic dependencies between time steps. In particular, this module uses a temporal encoding explicitly constructed during feature engineering rather than a positional encoding layer. This design preserves the Transformer's strengths in modeling long-range dependencies while also meeting the input requirements of time series tasks.

[0103] Furthermore, in this embodiment, the temporal feature encoding module adopts a standard Transformer encoder architecture;

[0104] The temporal feature encoding module consists of two encoder layers, each of which contains a multi-head self-attention mechanism and a feedforward neural network, and the temporal features adopt a sinusoidal encoding strategy.

[0105] It should be understood that the time series feature encoding module uses a standard Transformer encoder architecture, consisting of two encoder layers. Each encoder layer incorporates a multi-head self-attention mechanism and a feedforward neural network. The time series feature (time_embed) uses a sinusoidal encoding strategy, with a year as the period, converting dates to the dth day of the year.

[0106] Step S50: Using the terminal time step feature as an intermediate variable of evapotranspiration, using the output of the intermediate variable simulation layer as the corresponding physical model input during the forward propagation process, calculating through the physical constraint module, and outputting the surface evapotranspiration of the target area.

[0107] In the specific implementation, the results obtained in the above steps are input into the intermediate variable simulation layer. The intermediate variable simulation layer before the physical constraint module is composed of a fully connected network. The fully connected part uses a two-layer linear structure to achieve feature dimensionality reduction: receiving the terminal time step feature output by the previous layer encoder And further mapped to scalar output. The physical constraint module has two optional physical models, namely the PM model and the PT-JPL model mentioned above. By coupling these two physical models based on different evapotranspiration explanation theories, the physical interpretability of the model is more comprehensive. The output of the intermediate variable simulation layer serves as one of the inputs of the two physical models, and the physical model serves as part of the loss function. The intermediate variables selected here are the inputs that are difficult to parameterize and difficult to measure in the two models: the PM model based on the energy balance equation, the surface impedance rs in its input reflects the dynamic regulation of stomatal conductance on transpiration at the plant physiological level, and thereby reflects the transpiration component in evapotranspiration; the PT-JPL model based on radiation transfer theory, the soil moisture constraint factor f in its input sm , which reflects the limitation of soil moisture on evapotranspiration, and mainly reflects the soil evaporation component in evapotranspiration. In the forward propagation process, the model simulates the output of the intermediate variable layer. As the corresponding physical model input, the final output is calculated through the physical model.

[0108] Furthermore, in this embodiment, the physical constraint module is a PM model or a PT-JPL model. The intermediate variable simulation layer preceding the physical constraint module is composed of a fully connected network; this fully connected network adopts a two-layer linear structure, and the intermediate variables selected are the surface impedance input variable in the PM model and the soil moisture constraint factor input variable in the PT-JPL model. Specifically, the coupling method is that the output of the intermediate variable simulation layer of the previous physical model serves as one of the inputs of the two physical models, and the physical model serves as part of the loss function.

[0109] It should be noted that the physical constraint module has two optional physical models: the PM (Penman-Monteith) model and the PT-JPL (Priestley-Taylor Jet Propulsion Laboratory) model. The intermediate variable simulation layer before the physical constraint module is composed of a fully connected network. The fully connected part uses a two-layer linear structure to achieve feature dimensionality reduction. The intermediate variables selected here are the surface impedance input variable in the PM model and the soil moisture constraint factor input variable in the PT-JPL model. The specific coupling method is that the output of the intermediate variable simulation layer of the previous physical model is used as one of the inputs of the two physical models, and the physical model is used as part of the loss function.

[0110] Figure 7The scatter plot of fitting performance obtained through experiments is shown, in which the PhyTSNet model is the model corresponding to the present invention. The experiment includes two versions of the PhyTSNet model coupled with different physical models for comparison with the corresponding versions of the physical model. From the results, it can be seen that after the introduction of the coupled modeling method of physical constraints, the accuracy of the PhyTSNet-PM model is excellent, and the systematic deviation of the PM physical model is significantly improved. The corresponding PhyTSNet-PTJPL model also performs well, and various indicators are close to the PhyTSNet-PM model (the difference is less than 2%), which proves that the coupling framework has strong adaptability to the choice of physical models.

[0111] The overall accuracy of the model is important, but what is more important is the generalization performance of the model in untrained areas. Generalization performance is an important guarantee for the usability of the model. In the experiment, Dinghushan Station, Sanjiangyuan Station, Yuanjiang Station, Jinzhou Station, Panjin Station and Muztagh Ata Station were excluded from the model training data set as untrained areas. These six stations are different ecosystems, namely forests, grasslands, shrubs, farmlands, wetlands and wastelands. Among them, Muztagh Ata Station is the only station of wasteland ecosystem. There is no data of this ecosystem in the training set. The purpose of selecting this station is to test the generalization performance of the model in the case of extreme data loss. The scatter plot of the fitting effect is as follows Figure 8 As shown in Figure 1, TSNet is a purely data-driven model that removes the physical constraint module from the aforementioned PhyTSNet model, allowing the intermediate variable simulation layer of the original physical model to directly output the target value. The results show that the physical constraint-based coupled model, PhyTSNet-PM, achieves optimal performance across all four evaluation metrics. Constraining the model using physical mechanisms as prior knowledge effectively improves the model's generalization performance in unknown regions, and the theoretical accuracy of the physical framework is positively correlated with the actual performance of the coupled model.

[0112] This embodiment provides a surface evapotranspiration simulation method that couples physical constraints with a deep learning time series model. The time series model decomposes and reconstructs evapotranspiration time series data, extracting features at different scales and periods within the data to better capture the data's dynamic changes and nonlinear relationships. Physical constraints provide the time series model with physical rationality and prior knowledge, thereby improving its predictive robustness. Coupling the deep learning time series model with physical constraints leverages the strengths of both, effectively overcoming their limitations and providing more accurate and reliable results for surface evapotranspiration prediction.

[0113] In addition, an embodiment of the present invention further provides a storage medium storing a surface evapotranspiration calculation program. When the surface evapotranspiration calculation program is executed by a processor, the steps of the surface evapotranspiration calculation method described above are implemented.

[0114] In addition, refer to Figure 9 The embodiment of the present invention further provides a surface evapotranspiration calculation device, the surface evapotranspiration calculation device comprising:

[0115] An acquisition module 10 is used to acquire remote sensing time series data of a target area;

[0116] An extraction module 20 is configured to decompose the remote sensing time series data layer by layer using a surface evapotranspiration model, extract features at different time scales, and obtain time series features;

[0117] A feature fusion module 30 is configured to perform residual feature fusion on the temporal features through a residual feature fusion module;

[0118] The feature encoding module 40 is used to input the fused feature tensor into the time series feature encoding module, establish a dynamic dependency relationship between time steps through the self-attention mechanism, and output the terminal time step feature;

[0119] The physical constraint module 50 is used to use the terminal time step characteristics as the intermediate variable of evapotranspiration, use the output of the intermediate variable simulation layer as the corresponding physical model input during the forward propagation process, calculate through the physical constraint module, and output the surface evapotranspiration of the target area.

[0120] Other embodiments or specific implementations of the surface evapotranspiration calculation device of the present invention can refer to the above-mentioned method embodiments and will not be described in detail here.

[0121] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0122] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same item of hardware. The use of the terms first, second, and third, etc., does not denote any order and should be construed as identifiers.

[0123] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as a magnetic disk or optical disk) and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0124] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for calculating surface evapotranspiration, characterized in that: The surface evapotranspiration calculation method includes: Obtain remote sensing time series data of the target area; Decomposing the remote sensing time series data layer by layer through a surface evapotranspiration model, extracting features of different time scales, and obtaining time series features; Performing residual feature fusion on the time series features through a residual feature fusion module; The fused feature tensor is input into the temporal feature encoding module, and the dynamic dependency between time steps is established through the self-attention mechanism, and the terminal time step features are output; The terminal time step feature is used as an intermediate variable of evapotranspiration. In the forward propagation process, the output of the intermediate variable simulation layer is used as the input of the corresponding physical model. The surface evapotranspiration of the target area is output through calculation through the physical constraint module.

2. The method for calculating surface evapotranspiration according to claim 1, wherein: The temporal feature encoding module adopts a dual-channel convolution structure to achieve multi-scale decomposition of the signal; The temporal feature encoding module includes two one-dimensional convolution operators P and U, both of which adopt depth-wise separable convolution design; Given input signal , where B is the batch size, C is the number of channels, T is the time series length, and the decomposition process is achieved by constructing a residual map; First generate the approximate component residuals , and then generate adaptive adjustment items through U convolution to calculate the approximate components , and finally output the detail component d=x - P(a).

3. The method for calculating surface evapotranspiration according to claim 1, wherein: The residual feature fusion module is a combination of linear transformation and nonlinear activation; The residual feature fusion module receives the dimension The concatenated features are mapped to dimensional space, and apply the GELU activation function to achieve nonlinear transformation.

4. The method for calculating surface evapotranspiration according to claim 1, wherein: The temporal feature encoding module adopts the standard Transformer encoder architecture; The temporal feature encoding module consists of two encoder layers, each of which contains a multi-head self-attention mechanism and a feedforward neural network, and the temporal features adopt a sinusoidal encoding strategy.

5. The method for calculating surface evapotranspiration according to any one of claims 1 to 4, wherein: The physical constraint module is a PM model or a PT-JPL model.

6. The method for calculating surface evapotranspiration according to claim 5, wherein: The intermediate variable simulation layer before the physical constraint module is composed of a fully connected network; The fully connected network adopts a two-layer linear structure. The intermediate variables selected are the surface impedance input variable in the PM model and the soil moisture constraint factor input variable in the PT-JPL model. The specific coupling method is that the output of the intermediate variable simulation layer of the previous physical model is used as one of the inputs of the two physical models, and the physical model is used as part of the loss function.

7. A surface evapotranspiration calculation device, characterized in that: The surface evapotranspiration calculation device comprises: Acquisition module, used to obtain remote sensing time series data of the target area; An extraction module is used to decompose the remote sensing time series data layer by layer through a surface evapotranspiration model, extract features of different time scales, and obtain time series features; A feature fusion module, configured to perform residual feature fusion on the temporal features through a residual feature fusion module; The feature encoding module is used to input the fused feature tensor into the time series feature encoding module, establish dynamic dependencies between time steps through the self-attention mechanism, and output the terminal time step features; The physical constraint module is used to use the terminal time step characteristics as the intermediate variable of evapotranspiration, use the output of the intermediate variable simulation layer as the corresponding physical model input during the forward propagation process, calculate through the physical constraint module, and output the surface evapotranspiration of the target area.

8. A surface evapotranspiration calculation device, characterized in that: The surface evapotranspiration calculation device stores a surface evapotranspiration calculation program, and when the surface evapotranspiration calculation program is executed by a processor, the steps of the surface evapotranspiration calculation method according to any one of claims 1 to 6 are implemented.

9. A storage medium, characterized in that: The storage medium stores a surface evapotranspiration calculation program, which, when executed by a processor, implements the steps of the surface evapotranspiration calculation method according to any one of claims 1 to 6.

Images