Evapotranspiration inversion surface dry-wet monitoring system based on energy balance model

Abstract

The application discloses an evapotranspiration inversion land surface dry-wet monitoring system based on an energy balance model and relates to the technical field of land surface dry-wet monitoring.The system comprises a multi-source heterogeneous data acquisition module, a cloud intelligent processing module and a land surface dry-wet state evaluation module.The multi-source heterogeneous data acquisition module acquires thermal infrared, multispectral and active microwave remote sensing data and ground auxiliary data.The cloud intelligent processing module is internally provided with a data assimilation fusion unit, a physical data double-driving evapotranspiration inversion unit and a dynamic coupling self-correction unit, adopts a double-source energy balance physical model and a neural network model to calculate evapotranspiration, dynamically adjusts model weights and completes online training and self-correction, and obtains a target evapotranspiration value.The land surface dry-wet state evaluation module calculates a dry-wet index, constructs a dry-wet characteristic space and generates a monitoring atlas.The application effectively improves the spatiotemporal continuity and reliability of land surface dry-wet monitoring under complex underlying surfaces and cloudy conditions and is suitable for large-scale regional land surface dry-wet dynamic monitoring and drought early warning.

Description

Technical Field

[0001] This invention relates to the field of surface dryness and wetness monitoring technology, specifically to a surface dryness and wetness monitoring system based on evapotranspiration inversion using an energy balance model. Background Technology

[0002] Surface moisture status is a core indicator characterizing regional hydrothermal balance, eco-hydrological health, and drought risk, widely serving key areas such as agricultural irrigation, ecological protection, weather forecasting, and water resource management. Surface evapotranspiration, as the core process connecting the soil-vegetation-atmosphere continuum for water and energy exchange, directly determines surface moisture levels and is a key parameter for constructing a high-precision moisture monitoring system. Against the backdrop of global climate change and frequent extreme hydrological events, rapid, continuous, and accurate monitoring of regional and even large-scale surface moisture status has become a frontier and essential requirement for hydrological remote sensing, agro-meteorology, and ecological environment research, playing an irreplaceable supporting role in improving disaster early warning capabilities and optimizing water resource allocation.

[0003] Traditional surface wetness and dryness monitoring relies heavily on ground-based station observations. While offering high accuracy at single points, it suffers from limitations such as poor spatial representativeness, high deployment costs, and difficulty in covering complex terrains and large areas, failing to meet the dynamic monitoring needs at the watershed and national scales. Remote sensing technology, with its advantages of large-scale, periodic, and multi-band observation, has become the mainstream method for regional surface wetness and dryness inversion. It commonly uses thermal infrared data to invert surface temperature, multispectral data to extract vegetation information, and combines energy balance models to estimate evapotranspiration and assess wetness and dryness. However, problems such as frequent missing thermal infrared data due to cloud cover, poor spatiotemporal matching of multi-source remote sensing data, difficulty of adapting single physical models to nonlinear hydrothermal processes on heterogeneous underlying surfaces, and lack of physical constraints in data-driven models have long constrained monitoring accuracy and spatiotemporal continuity, particularly in cloudy and rainy areas and complex terrain regions, making it difficult to support all-weather, highly robust operational monitoring.

[0004] Current academic research continues to deepen around remote sensing evapotranspiration inversion and wet / dry monitoring, forming two mainstream technical paths: physical models and data-driven approaches. Meng Ying et al., in their article "Research Progress on Surface Evapotranspiration Based on Remote Sensing," published in *Remote Sensing Technology and Applications*, Vol. 37, No. 4, 2022, systematically reviewed the progress of single-source and dual-source energy balance models and machine learning inversion methods. They pointed out that traditional physical models are significantly affected by impedance parameterization and cloud interference, resulting in large errors due to heterogeneous underlying surfaces. While data-driven methods can fit nonlinear relationships, their generalization ability and physical interpretability are insufficient, and their long-term series stability is poor. Yao Yunjun et al., in their review of satellite remote sensing of land evapotranspiration published in *Ecological Processes*, comprehensively reviewed the theoretical foundation, algorithm system, and satellite products, clarifying that existing methods have bottlenecks in spatiotemporal continuity, component separation, and applicability to multi-cloud conditions. They emphasized that multi-source data fusion and physical-data dual-driven coupling are the breakthrough directions. Existing research focuses on optimizing a single model, failing to achieve deep synergy between physical mechanisms and data advantages. It lacks spatiotemporal interpolation, dynamic weighting, and online self-correction mechanisms, making it difficult to meet the needs of high-precision monitoring around the clock.

[0005] Chinese patents address these issues through engineering improvements. CN112014323B discloses a remote sensing inversion method and system for surface evapotranspiration, which improves the accuracy of daily evapotranspiration inversion by defining the dry-wet boundary temperature of soil and vegetation and calculating multiple humidity limiting factors, thus mitigating estimation bias under soil moisture stress to some extent. CN102176002B proposes a drought monitoring method and system based on remote sensing inversion of surface water and heat flux, relying on the energy balance equation to achieve pixel-by-pixel calculation of net radiation, flux, and drought index, thereby enhancing the spatial consistency of regional monitoring. However, these patents still have significant shortcomings: the former does not address data gaps caused by cloud cover and does not integrate active microwave data to achieve all-weather observation; the latter uses a single physical model, does not introduce deep learning to mine nonlinear relationships, lacks dynamic coupling and online self-correction, and suffers from large accuracy fluctuations under complex underlying surfaces and meteorological conditions, failing to achieve adaptive model optimization and long-term stable monitoring. Summary of the Invention

[0006] To address the aforementioned technical issues, this application discloses a surface dryness and wetness monitoring system based on an energy balance model for evapotranspiration inversion. Its features include a multi-source heterogeneous data acquisition module, a cloud-based intelligent processing module, and a surface dryness and wetness status assessment module, specifically: The multi-source heterogeneous data acquisition module acquires multi-source remote sensing data and ground-aided data of the target area. The multi-source remote sensing data includes at least: thermal infrared remote sensing data for retrieving surface temperature, multispectral remote sensing data for characterizing vegetation conditions, and active microwave remote sensing data for penetrating cloud layers to acquire surface information. The cloud-based intelligent processing module is communicatively connected to the multi-source heterogeneous data acquisition module, and integrates the following internal components: The data assimilation and fusion unit performs spatiotemporal interpolation and correction on thermal infrared remote sensing data obscured by clouds, and fuses multispectral remote sensing data to generate a dataset of land surface state parameters. The physical data dual-driven evapotranspiration inversion unit includes a physical model based on the principle of surface energy balance and a neural network model based on deep learning. The physical model takes the surface state parameter dataset and ground auxiliary data as input to calculate a first evapotranspiration estimate. The neural network model takes the same data as input and calculates a second evapotranspiration estimate by mining the nonlinear relationships in historical data. The dynamic coupling self-correction unit receives the first and second evapotranspiration estimates and dynamically adjusts the output weights of the physical model and the neural network model based on real-time ground verification data. It obtains the target evapotranspiration value through weighted fusion and feeds back the error of the fusion result to the neural network model for online training, thereby realizing the model's self-learning and self-correction. The surface dryness and wetness assessment module calculates the surface dryness and wetness index based on the target evapotranspiration value, combined with the surface net radiation and soil heat flux, and generates a surface dryness and wetness monitoring map.

[0007] Preferably, the active microwave remote sensing data acquired by the multi-source heterogeneous data acquisition module is synthetic aperture radar data, which is used to invert soil moisture content through the following empirical model. As part of the aforementioned surface state parameter dataset: ,in, and These are the vertical-vertical and vertical-horizontal backscattering coefficients of synthetic aperture radar data, respectively. Normalized Difference Vegetation Index; These are empirical coefficients.

[0008] Preferably, the data assimilation and fusion unit adopts a spatiotemporal fusion model based on variational assimilation, which minimizes the following cost function. The formula for generating the surface state parameter dataset x is as follows: ,in, The background field refers to the surface parameters initially retrieved from active microwave remote sensing data. The observation vector is the thermal infrared and multispectral remote sensing data. and These are the background field error covariance matrix and the observation error covariance matrix, respectively. It is a nonlinear observation operator used to map surface state parameters to the observation space.

[0009] Preferably, the physical model is a dual-source energy balance model, which treats the land surface as two independent sources: the vegetation canopy and the bare soil, and calculates the estimated value of the first evapotranspiration by solving the following set of coupled equations. The formula is: ; in, and These are the net radiation from the vegetation canopy and the soil surface, respectively. , , and , , These are the latent heat flux, sensible heat flux, and temperature of the canopy and soil, respectively. Soil heat flux; The temperature of the air inside the canopy; air density; The specific heat capacity of air; and The aerodynamic impedances of the canopy and soil are respectively. and These are the surface impedances of the canopy and soil, respectively. The slope of the saturated water vapor pressure-temperature curve; This is the constant of the wet / dry meter; This represents the saturated water vapor pressure difference.

[0010] Preferably, the neural network model employs a spatiotemporal attention-enhanced convolutional long short-term memory network architecture, comprising: The spatial feature extraction module, consisting of a multi-layer convolutional neural network, is used to extract spatial feature maps from the input surface state parameter dataset. The temporal dependency capture module, composed of a bidirectional long short-term memory network layer, is used to receive the spatial feature map and capture its forward and backward dependencies in time series, and output the hidden state sequence. The attention weighting module is used to calculate the attention weight at each time step in the hidden state sequence and generate a context feature vector accordingly. The output regression module, consisting of fully connected layers, is used to map the context feature vector to the second evapotranspiration estimate. The calculation formula for the attention weighting module is as follows: , ,in, For a moment The hidden state vector, For bidirectional long short-term memory networks at any point in the historical sequence The hidden state vector; For a moment Attention weight coefficient, , The weight matrix and weight vector are learnable. For bias vectors, This is the context feature vector.

[0011] Preferably, the dynamic coupling self-calibration unit receives ground verification data in real time via a data interface. The ground verification data includes eddy covariance system observation data from the flux tower, micrometeorological data from the automatic weather station, and soil moisture sensor data. The dynamic coupling self-calibration unit is configured as follows: The observed data from the eddy covariance system are subjected to quality control and flux calculation to obtain the measured evapotranspiration flux value; The measured evapotranspiration flux value is used as the model performance index. The calculation benchmark, and the loss function. Observations in .

[0012] Preferably, the dynamic coupling self-calibration unit calculates the target evapotranspiration value using the following adaptive weighted fusion model. The formula is: ,in The first estimated value of evapotranspiration. The second evapotranspiration estimate, time-varying weighting coefficient The following logistic function is dynamically determined: ,in, These are model performance metrics calculated based on real-time ground validation data. The preset performance threshold; This is the gain coefficient that controls the smoothness of the function.

[0013] Preferably, the dynamically coupled self-calibrating unit further minimizes the following loss function. The neural network model is trained online: in, For the first One ground verification data point; To verify the total number of data points; The regularization coefficient is used. Let KL divergence be denoted as KL divergence.

[0014] Preferably, the surface wetness and dryness assessment module calculates the evaporation rate. The formula for the surface wetness index is: ,in, Net surface radiation For soil heat flux, The target evaporation value.

[0015] Preferably, the surface wetness / dryness assessment module is further configured to assess the evaporation rate. With vegetation coverage Construct a dry and wet feature space, and calculate the relative distance from any cell to the dry and wet edges. To generate the aforementioned surface wet / dry monitoring map: ,in, and These are the equations for the dry side and the wet side, respectively.

[0016] Compared with the prior art, the technical solution of this application has the following technical effects: This invention, through multi-source heterogeneous data acquisition and cloud-based intelligent processing, can stably acquire thermal infrared, multispectral, and active microwave remote sensing data as well as ground-aided data, achieving complete acquisition of surface information of the target area. It effectively ensures the availability of data under conditions of cloud cover, fog, and complex underlying surfaces, providing a comprehensive and reliable data foundation for evapotranspiration inversion and enhancing the data support capability and operational stability of the entire monitoring process.

[0017] This invention employs a dual-drive architecture of physical model and neural network model, which can fully leverage the mechanistic constraints of physical model and the nonlinear fitting capability of deep learning to accurately characterize the surface water heat exchange process and reliably output stable evapotranspiration estimation results. Combined with a dynamic coupling self-correction unit, it realizes adaptive weight adjustment and online model optimization, continuously improving the accuracy of inversion results and the consistency of long-term series.

[0018] This invention performs spatiotemporal interpolation and correction on remote sensing data based on a data assimilation and fusion unit, which can effectively compensate for the data loss caused by cloud cover, generate a continuous and complete dataset of land surface state parameters, ensure that the monitoring process is not affected by weather conditions, realize uninterrupted land surface parameter analysis around the clock, and provide stable data and computing support for large-scale long-term land surface dryness and wetness monitoring.

[0019] This invention calculates the surface dryness and wetness index and generates monitoring maps based on target evapotranspiration values. It can intuitively present the spatial distribution and temporal variation characteristics of surface dryness and wetness in a region, providing standardized quantitative basis and visualization results for the assessment of surface dryness and wetness status. It supports high-precision, high-spatial-resolution surface dryness and wetness monitoring and meets the practical application needs of scenarios such as eco-hydrology and agricultural drought monitoring.

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0023] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows: Figure 1 A schematic diagram of the surface dryness and wetness monitoring system based on the energy balance model for evapotranspiration inversion; Figure 2 A schematic diagram of the internal unit composition and data flow structure of the cloud-based intelligent processing module; Figure 3 A schematic diagram of the composition and calculation process of the dual-drive evapotranspiration inversion unit model based on physical data; Figure 4 A comparison of the spatial distribution of surface temperature before and after data assimilation in the Huang-Huai-Hai Plain region; Figure 5 A time-series comparison of measured and retrieved surface evapotranspiration values ​​during the monitoring period in the Huang-Huai-Hai Plain; Figure 6 Scatter density plot of correlation between inverted and measured evapotranspiration values ​​in the Huang-Huai-Hai Plain; Figure 7 A three-dimensional schematic diagram of the spatial distribution of surface evapotranspiration in the arid Northwest region. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0025] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0026] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0027] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0028] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0029] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0030] Example 1 This embodiment mainly describes a surface dryness and wetness monitoring system based on an energy balance model for evapotranspiration inversion. Figure 1 As shown, it specifically includes a multi-source heterogeneous data acquisition module, a cloud-based intelligent processing module, and a surface dryness and wetness assessment module, specifically: The multi-source heterogeneous data acquisition module acquires multi-source remote sensing data and ground-aided data of the target area. The multi-source remote sensing data includes at least: thermal infrared remote sensing data for retrieving surface temperature, multispectral remote sensing data for characterizing vegetation conditions, and active microwave remote sensing data for penetrating cloud layers to acquire surface information. The cloud-based intelligent processing module is communicatively connected to the multi-source heterogeneous data acquisition module, and integrates the following internal components: The data assimilation and fusion unit performs spatiotemporal interpolation and correction on thermal infrared remote sensing data obscured by clouds, and fuses multispectral remote sensing data to generate a dataset of land surface state parameters. The physical data dual-driven evapotranspiration inversion unit includes a physical model based on the principle of surface energy balance and a neural network model based on deep learning. The physical model takes the surface state parameter dataset and ground auxiliary data as input to calculate a first evapotranspiration estimate. The neural network model takes the same data as input and calculates a second evapotranspiration estimate by mining the nonlinear relationships in historical data. The dynamic coupling self-correction unit receives the first and second evapotranspiration estimates and dynamically adjusts the output weights of the physical model and the neural network model based on real-time ground verification data. It obtains the target evapotranspiration value through weighted fusion and feeds back the error of the fusion result to the neural network model for online training, thereby realizing the model's self-learning and self-correction. The surface dryness and wetness assessment module calculates the surface dryness and wetness index based on the target evapotranspiration value, combined with the surface net radiation and soil heat flux, and generates a surface dryness and wetness monitoring map.

[0031] Furthermore, the multi-source heterogeneous data acquisition module acquires multi-source remote sensing data and ground-aided data for the target area; the multi-source remote sensing data includes: thermal infrared remote sensing data for retrieving surface temperature, which is obtained through thermal infrared sensors (such as Landsat). Land surface radiance and temperature information is acquired using the Landsat TIRS sensor (or the MODIS thermal infrared band). The data format is raster image, with spatial resolution ranging from 30 meters (Landsat) to 1000 meters (MODIS). Temporal resolution depends on the satellite transit cycle (e.g., 16 days for Landsat, 1-2 days for MODIS). The radiance and temperature values ​​after radiometric calibration are used to retrieve the true surface temperature via Planck function inversion. During the inversion process, parameters such as atmospheric transmittance, upward and downward atmospheric radiation must be considered, and atmospheric correction is performed using a single-window or split-window algorithm. Multispectral remote sensing data characterizing vegetation status is acquired using multispectral sensors (such as Landsat's OLI sensor and Sentinel-2's MSI sensor) to obtain visible light (blue, green, and red bands, wavelength range 400-700 nm) and near-infrared band (wavelength range 700-1300 nm) reflectance. This data is used to calculate the Normalized Difference Vegetation Index (NDVI). The formula for calculating NDVI is... ,in For near-infrared reflectivity, The reflectance data is red band reflectance, and undergoes radiometric calibration, atmospheric correction (e.g., using the FLAASH model) and geometric correction to ensure accuracy and consistency. Active microwave remote sensing data, which penetrates clouds to acquire surface information, uses synthetic aperture radar (e.g., Sentinel-1's C-band SAR, ALOS-2's L-band SAR) to obtain backscattering coefficients. The backscattering coefficient is closely related to surface roughness and soil moisture. The data format is a complex image. After multi-view processing, filtering (such as Lee filtering) and radiometric calibration, an intensity image is obtained, and then the backscattering coefficient is calculated. The penetrability of microwaves (approximately 5 cm for C-band and approximately 10 cm for L-band) is used to obtain surface roughness and soil moisture information under the cloud layer. Soil moisture inversion uses a water cloud model or an integral equation model. The formula for the water cloud model is... ,in The backscattering coefficient of the vegetation layer. The backscattering coefficient of the soil layer, The transmittance of the vegetation layer; surface auxiliary data includes air temperature observed by meteorological stations ( ), relative humidity ( ), wind speed ( ), air pressure ( ) and radiation observation data (net surface radiation) Solar shortwave radiation Meteorological data is collected in real time through automatic weather stations (AWS), with a time resolution of up to the minute level. Radiation data is obtained through sensors such as total radiation meters and net radiation meters. All data is transmitted in real time to the cloud intelligent processing module through communication interfaces (such as 4G / 5G, satellite communication).

[0032] Furthermore, such as Figure 2 As shown, the cloud-based intelligent processing module is responsible for data fusion, model calculation, and dynamic correction, and includes a data assimilation and fusion unit, a physical data dual-drive evapotranspiration inversion unit, and a dynamic coupling self-correction unit.

[0033] The data assimilation and fusion unit preprocesses and fuses the received multi-source data. To address the issue of missing thermal infrared data due to cloud cover, the system employs a variational assimilation-based spatiotemporal fusion model for reconstruction. This model constructs a cost function. The aim is to find the optimal surface state parameters. This ensures that the estimated value satisfies both the constraints of the background field (preliminary surface parameters retrieved from active microwave data, such as soil moisture and surface temperature) and minimizes the difference from the observed data (thermal infrared and multispectral data). The cost function is formulated as follows: ,in, The background field is the surface parameters (such as soil moisture and surface temperature) that are initially retrieved from active microwave remote sensing data. The observation vectors are the brightness and temperature of the thermal infrared remote sensing data and the reflectance of the multispectral remote sensing data. It is a nonlinear observation operator used to convert surface state parameters Mapping to the observation space, for example, simulating the radiance of the thermal infrared band using radiative transfer models (such as MODTRAN); The background field error covariance matrix reflects the uncertainty of the background field. The observation error covariance matrix reflects the error characteristics of the observed data. This is achieved by minimizing the cost function. The optimal surface state parameters are solved using variational methods (such as three-dimensional variational 3D-Var or four-dimensional variational 4D-Var). This generates a dataset of land surface state parameters, including parameters such as surface temperature, soil moisture, and vegetation cover. At the same time, by integrating NDVI calculated from multispectral data, the spatial distribution of vegetation parameters is further refined, improving the accuracy of land surface state parameters.

[0034] like Figure 3 As shown, the physical data dual-driven evapotranspiration inversion unit includes a physical model and a neural network model, which calculate evapotranspiration from the perspectives of physical mechanisms and data-driven approaches, respectively.

[0035] The physical model employs a two-source energy balance model (TSEB), treating the surface vegetation canopy and bare soil as two independent sources, and calculating their energy balance separately. The energy balance equation for the vegetation canopy is as follows: ,in Net radiation from the vegetation canopy, This refers to the latent heat flux (i.e., evapotranspiration) of the vegetation canopy. The sensible heat flux of the vegetation canopy is given by: The energy balance equation for the soil layer is: ,in Net radiation of the soil layer, The latent heat flux of the soil layer, For the sensible heat flux of the soil layer, This represents soil heat flux. Sensible heat flux was calculated using aerodynamic methods, specifically the sensible heat flux from the vegetation canopy. ,in air density, The specific heat capacity of air at constant pressure. For vegetation canopy temperature, The temperature of the air inside the canopy. Aerodynamic impedance of the canopy; sensible heat flux of the soil layer. ,in This is the constant of the wet and dry meter. For soil surface temperature, The aerodynamic impedance of the soil surface is given; the latent heat flux is calculated using a simplified form of the Penman-Monteith formula, and the latent heat flux of the vegetation canopy is given. ,in The slope of the saturated water vapor pressure-temperature curve. For saturated water vapor pressure difference, Surface impedance of the vegetation canopy; latent heat flux of the soil layer. ,in The surface impedance of the soil surface. Total estimated first evapotranspiration. .

[0036] The neural network model employs a spatiotemporally attention-enhanced convolutional long short-term memory network (ST-AttentionConvLSTM). This network architecture includes a spatial feature extraction module, a temporal dependency capture module, an attention weighting module, and an output regression module. The spatial feature extraction module consists of multiple convolutional neural network (CNN) layers, with input being a dataset of surface state parameters (such as surface temperature, soil moisture, NDVI, etc., with a size of [missing information - likely a unit of measurement]). ,in For height, For width, (Number of channels), spatial features are extracted using convolutional kernels (e.g., 3×3, 5×5), and the output feature map size is [size missing]. The temporal dependency capture module consists of a bidirectional long short-term memory (BiLSTM) network layer, and its input is the feature sequence (of length 1) output by the spatial feature extraction module. (i.e., time step) captures the dependencies between time series using forward and backward LSTM units, and outputs the hidden state vector. (dimension is) The attention weighting module is used to calculate the attention weight at each time step in the hidden state vector sequence. The weight calculation formula is as follows: ,in For a moment The hidden state vector, For bidirectional long short-term memory networks at any point in the historical sequence The hidden state vector; For a moment Attention weight coefficient, , The weight matrix and weight vector are learnable. For bias vectors, For the context feature vector, the formula is: The output regression module consists of fully connected layers that map the context feature vectors to the second evapotranspiration estimate. .

[0037] The computational process is as follows: the spatial feature extraction module extracts the spatial feature map of the input data through a multi-layer CNN; the temporal dependency capture module captures the forward and backward dependencies of the time series through a bidirectional LSTM and outputs the hidden state vector. The attention weighting module calculates attention weights. And generate context feature vectors The output regression module will Mapped to the second evapotranspiration estimate .

[0038] The dynamic coupling self-calibration unit receives the first evapotranspiration estimate from the physical model. The second evapotranspiration estimate from the neural network model And based on real-time ground verification data (such as evapotranspiration observed by eddy covariance systems) Dynamic correction is performed using micro-meteorological data from automatic weather stations and soil moisture sensor data. This is achieved through data quality control (such as removing outliers and filling missing values) and flux calculation (such as using eddy covariance method). The measured evapotranspiration flux was obtained and used as a performance index of the model. Computational benchmark; model performance indicators The Nash efficiency coefficient is used, and the calculation formula is as follows: ,in These are the simulated values ​​from the model. The average of the measured values. For the first Measured evapotranspiration flux values ​​corresponding to each ground validation data point; variable weighting coefficients The value is dynamically determined by the logistic function, and the formula is as follows: ,in This is the gain coefficient that controls the smoothness of the function. This is a preset performance threshold (e.g., 0.7). Target evaporation value. The formula is obtained through weighted fusion. At the same time, the error of the fusion result ( The feedback is fed back to the neural network model, minimizing the loss function. For online training, the loss function formula is: ,in To verify the total number of data points, The regularization coefficient is . KL divergence is used to constrain the consistency between the output distribution of the neural network model and the output distribution of the physical model. , The mean and variance output by the physical model. , This represents the mean and variance of the neural network model's output. The weight parameters of the neural network are updated using the backpropagation algorithm, enabling the model to self-learn and self-correct.

[0039] As a supplement, the dynamic coupling self-calibration unit receives ground validation data in real time via a data interface. This ground validation data includes eddy covariance system observation data from the flux tower, micrometeorological data from automatic weather stations, and soil moisture sensor data. The dynamic coupling self-calibration unit is configured to: perform quality control and flux calculation on the eddy covariance system observation data to obtain measured evapotranspiration flux values; and use these measured evapotranspiration flux values ​​as model performance indicators. The calculation benchmark, and as a loss function Measured evapotranspiration flux values .

[0040] The surface wet / dry condition assessment module is based on the target evapotranspiration value. Combined with net surface radiation and soil heat flux The surface wetness index (EF) is calculated using the following formula: The surface wetness index EF reflects the energy available for evapotranspiration on the surface relative to the surface's effective energy (EF). The ratio of EF (Earth Expansion Rate) to the surrounding land surface is calculated. A value closer to 1 indicates a wetter surface, while a value closer to 0 indicates a drier surface. To further refine the classification of wet and dry conditions, the system calculates the relative evapotranspiration of any pixel relative to its surrounding land. To generate a dry-wet monitoring map. The calculation formula is: ,in This is the dry-side equation (i.e., the evapotranspiration index under completely dry conditions). ), This is the wet-edge equation (i.e., the evapotranspiration index under fully wet conditions). ).according to The value is used to classify the surface dryness and wetness state into extremely dry ( ),drought( ), slightly dry ( ),suitable( ), moist ( Five levels are used to generate surface dryness and wetness monitoring maps. The spatial resolution of the maps is consistent with the input remote sensing data, while the temporal resolution is determined according to the data update frequency.

[0041] This implementation details how a multi-source heterogeneous data acquisition module acquires thermal infrared, multispectral, and active microwave remote sensing data, combines this with ground-aided data, and utilizes a cloud-based intelligent processing module's data assimilation and fusion unit to eliminate cloud cover interference and generate a land surface state parameter dataset. A dual-drive evapotranspiration inversion unit combines an energy balance physical model with a deep learning neural network model to calculate the first and second evapotranspiration estimates, respectively. A dynamic coupling self-correction unit then dynamically adjusts the weights and merges them based on ground validation data, enabling model self-learning. A land surface dryness and wetness assessment module calculates the dryness and wetness index and generates monitoring maps, improving the accuracy of evapotranspiration inversion and the reliability of land surface dryness and wetness monitoring, providing precise data support for water resource management and drought early warning.

[0042] Based on Example 1, Example 2 details the specific implementation and verification of this application, selecting a typical agricultural area in the Huang-Huai-Hai Plain (36°N, 116°E) as the research object. This area is mainly composed of winter wheat-summer maize rotation farmland, with bare soil and water bodies, and relatively diverse surface cover types. To verify the system's data repair capability and the accuracy of the dual-drive model under cloud cover conditions, the monitoring window was selected from June 15, 2025 (summer maize seedling stage) to September 15, 2025 (grain-filling stage).

[0043] During the multi-source data acquisition phase, the system acquired thermal infrared band (TIRS) data from the Landsat 8 satellite, with a raw spatial resolution of 100 meters (resampled to 30 meters), and C-band dual-polarization (VV / VH) data from the Sentinel-1 SAR satellite. Ground-aided data was provided by five flux towers (numbered FLUX-01 to FLUX-05) distributed throughout the region, with a sampling frequency of 30 minutes per sampling.

[0044] In response to the cloud cover issue caused by severe convective weather in the region on June 28, the data assimilation and fusion unit initiated a spatiotemporal interpolation procedure. The system utilized the backscattering coefficients obtained by Sentinel-1 penetrating the cloud layer. and The missing land surface temperature (LST) was reconstructed by combining a lookup table with historical optical data from the same period.

[0045] like Figure 4 As shown, the spatial distribution of surface temperature before and after data assimilation for that day is compared. Figure 4 In the original image on the left, the central area of ​​the study region is covered by thick clouds, appearing as a large white area of ​​invalid data, making it impossible to directly obtain the surface thermal state; while Figure 4In the assimilation result image on the right, the areas originally obscured by clouds were filled with continuous color-level data, clearly showing that the temperature in the farmland area was approximately 305K-308K, while the water temperature in the northwest corner was lower, at approximately 298K. This restoration was not a simple pixel copy, but rather generated through physical constraints based on prior knowledge of soil moisture derived from microwave data inversion. The restored data maintained a high degree of spatial continuity with the ground temperature gradient in the clear-sky area on the same day, with the root mean square error (RMSE) controlled within 1.2K.

[0046] Based on the restored surface state parameters, the physical data dual-drive evapotranspiration inversion unit began operation. The physical model (dual-source energy balance model) first calculated the first estimated value of evapotranspiration. Meanwhile, the neural network model was trained using a five-year historical meteorological and remote sensing database of the region, outputting a second estimate of evapotranspiration. The dynamic coupling self-calibration unit calculates the weights of the two data points based on the daily measured data from the ground flux tower.

[0047] The table below details the comparison between the model inversion values ​​and measured values ​​for each of the five flux tower sites at 12:00 on June 28th: Table 1: Comparison of Evapotranspiration Inversion Values ​​and Measured Values ​​for Different Models in Example 2

[0048] Table 1 shows that a single physical model tends to underestimate evapotranspiration in bare soil areas (error -3.3%), while a single neural network model tends to overestimate evapotranspiration in irrigated farmland areas due to the lag in training data (error +4.0%). After weighted fusion using dynamically coupled self-calibrating units... The relative error at all sites was controlled within ±2%, which is significantly better than a single model.

[0049] The surface dryness and wetness assessment module calculates the surface dryness and wetness index based on the merged evapotranspiration value, combined with net radiation and soil heat flux.

[0050] The spatial distribution of the surface moisture index (SDDI) in the study area on June 28th showed that in the southeastern river and large irrigation areas, the SDDI values ​​ranged from -0.8 to -0.5, indicating sufficient moisture. In the central summer maize planting area, the SDDI values ​​ranged from 0.1 to 0.3, indicating a suitable moisture level, consistent with the water requirements of this crop during its growth period. In the northwest, the unirrigated bare soil areas appeared orange-red, with SDDI values ​​exceeding 0.6, and were classified as experiencing mild drought. These results are highly consistent with soil moisture content distribution recorded by manual ground inspections, validating the accuracy of the system's monitoring of agricultural drought.

[0051] To further verify the long-term stability of the system, the daily average evapotranspiration during the monitoring period from June to September 2025 was continuously tracked, such as... Figure 5 The comparison curves of the system inversion values ​​and the measured values ​​of the flux tower over time are shown. Figure 5 The solid black line represents the measured daily average evapotranspiration from the ground flux tower, and the dashed red line represents the daily average evapotranspiration retrieved by this system. Figure 5 The curves show a high degree of synchronicity between the two curves. During the seedling stage in late June, evapotranspiration was at a low level (approximately 2.5 mm / d). Entering the peak summer heat period from mid-July to early August, both curves synchronously climbed to their peak values, with the highest measured value reaching 6.8 mm / d, followed closely by the system-retrieved value at 6.75 mm / d, accurately capturing the peak water demand of the crop. In September, as the crop matured and temperatures decreased, both curves synchronously declined. Throughout the entire monitoring period, the red dashed line (retrieved value) consistently fluctuated closely around the black solid line (measured value), without significant phase lag or amplitude distortion, directly demonstrating the system's excellent stability and tracking ability in long-term monitoring.

[0052] To further verify the long-term stability of the system, the correlation between the daily average evapotranspiration retrieved by the system and the measured data of the flux tower was statistically analyzed from June to September 2025. Figure 6 The scatter density plot and regression analysis results for the entire monitoring period are shown below. Figure 6 Each scatter point represents an observation at a given moment, and the color intensity of the scatter points indicates the density of data points (the darker the color, the higher the density). The coefficient of determination reached 0.91, and the slope was very close to 1. This indicates that during the monitoring period of up to 3 months, the system maintained extremely high inversion accuracy under different weather conditions (sunny, cloudy) and different crop growth stages, without any obvious systematic bias.

[0053] Example 3: This example selects the arid inland region of Northwest China (such as the oasis on the edge of the Tarim Basin, 41°N, 86°E), which has a more severe environment and stronger surface heterogeneity, for verification. This region has a "oasis-desert" binary structure, huge differences in surface temperature, and is affected by dust aerosols all year round, which poses greater challenges to atmospheric correction and model robustness.

[0054] The monitoring date was set for July 20, 2025, during a period of high temperatures and intense evaporation in farmland on the edge of the oasis. The data acquisition module obtained MODIS thermal infrared data (spatial resolution 1000 meters) and Sentinel-1 SAR data. Due to the low resolution of MODIS and the fragmented surface of the area, the system first utilized the high-resolution texture information of the SAR data for spatial downscaling.

[0055] When processing data for this region, the physical model introduced an atmospheric transmittance correction parameter $\tau_{dust}$ for dust aerosols. The calculation results show that in the core oasis area, the evapotranspiration retrieved by the physical model is approximately 7.2 mm / d, while the neural network model, due to a lack of samples with both extreme high temperatures (>315K) and high humidity in its training data, initially outputs only 5.8 mm / d, indicating a significant underestimation.

[0056] After detecting this deviation, the dynamic coupling self-calibration unit automatically reduced the weight of the neural network model (from the default 0.5 to 0.2), increased the weight of the physical model (to 0.8), and incorporated the extreme high temperature parameters observed by the oasis edge weather station on that day for online fine-tuning.

[0057] Table 2 shows a comparison of the inversion data of three typical pixels in the oasis-desert transition zone on that day: Table 2: Comparison of evapotranspiration inversion data in the oasis-desert transition zone in Example 3

[0058] Note: The accuracy improvement refers to the percentage reduction in error of the fused value relative to that of a single neural network model.

[0059] The data in Table 2 shows that in extreme drought conditions, simple deep learning models are prone to failure. However, this system effectively utilizes the stability of physical models under extreme conditions by dynamically adjusting weights, reducing the average inversion error from 25% of the neural network model to less than 5%. like Figure 7 The area shown represents surface evapotranspiration, with vertical height representing the magnitude of evapotranspiration. Figure 7 A towering evapotranspiration plateau can be clearly seen, corresponding to oasis farmland areas, with evapotranspiration values ​​consistently high at 5.5-6.5 mm / d; while the surrounding area is a low-lying desert plain with values ​​close to 0. At the boundary between the two, Figure 7 The system accurately captured steep evapotranspiration cliffs and dramatic spatial gradient changes without exhibiting the smoothing blurring common in conventional algorithms. This demonstrates the system's superior performance in monitoring surface wetness and dryness in highly heterogeneous and extreme environments, fully meeting the application requirements in complex environments.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A surface dryness and wetness monitoring system based on an energy balance model for evapotranspiration inversion, characterized in that, It includes a multi-source heterogeneous data acquisition module, a cloud-based intelligent processing module, and a surface dryness and wetness assessment module, specifically: The multi-source heterogeneous data acquisition module acquires multi-source remote sensing data and ground-aided data of the target area. The multi-source remote sensing data includes at least: thermal infrared remote sensing data for retrieving surface temperature, multispectral remote sensing data for characterizing vegetation conditions, and active microwave remote sensing data for penetrating cloud layers to acquire surface information. The cloud-based intelligent processing module is communicatively connected to the multi-source heterogeneous data acquisition module, and integrates the following internal components: The data assimilation and fusion unit performs spatiotemporal interpolation and correction on thermal infrared remote sensing data obscured by clouds, and fuses multispectral remote sensing data to generate a dataset of land surface state parameters. The physical data dual-driven evapotranspiration inversion unit includes a physical model based on the principle of surface energy balance and a neural network model based on deep learning. The physical model takes the surface state parameter dataset and ground auxiliary data as input to calculate a first evapotranspiration estimate. The neural network model takes the same data as input and calculates a second evapotranspiration estimate by mining the nonlinear relationships in historical data. The dynamic coupling self-correction unit receives the first and second evapotranspiration estimates and dynamically adjusts the output weights of the physical model and the neural network model based on real-time ground verification data. It obtains the target evapotranspiration value through weighted fusion and feeds back the error of the fusion result to the neural network model for online training, thereby realizing the model's self-learning and self-correction. The surface dryness and wetness assessment module calculates the surface dryness and wetness index based on the target evapotranspiration value, combined with the surface net radiation and soil heat flux, and generates a surface dryness and wetness monitoring map.

2. The surface dryness and wetness monitoring system based on energy balance model for evapotranspiration inversion according to claim 1, characterized in that, The active microwave remote sensing data acquired by the multi-source heterogeneous data acquisition module is synthetic aperture radar data, which inversely determines soil moisture content using the following empirical model. As part of the aforementioned surface state parameter dataset: ,in, and These are the vertical-vertical and vertical-horizontal backscattering coefficients of synthetic aperture radar data, respectively. Normalized Difference Vegetation Index; These are empirical coefficients.

3. The surface dryness and wetness monitoring system based on the energy balance model for evapotranspiration inversion according to claim 1, characterized in that, The data assimilation and fusion unit adopts a spatiotemporal fusion model based on variational assimilation, which minimizes the following cost function. The formula for generating the surface state parameter dataset x is as follows: ,in, The background field refers to the surface parameters initially retrieved from active microwave remote sensing data. The observation vector is the thermal infrared and multispectral remote sensing data. and These are the background field error covariance matrix and the observation error covariance matrix, respectively. It is a nonlinear observation operator used to map surface state parameters to the observation space.

4. The surface dryness and wetness monitoring system based on an energy balance model for evapotranspiration inversion according to claim 2 or 3, characterized in that, The physical model described is a dual-source energy balance model, which treats the Earth's surface as two independent sources: the vegetation canopy and the bare soil. The estimated value of the first evapotranspiration is calculated by solving the following set of coupled equations. The formula is: ； in, and These are the net radiation from the vegetation canopy and the soil surface, respectively. , , and , , These are the latent heat flux, sensible heat flux, and temperature of the canopy and soil, respectively. Soil heat flux; The temperature of the air inside the canopy; air density; The specific heat capacity of air; and The aerodynamic impedances of the canopy and soil are respectively. and These are the surface impedances of the canopy and soil, respectively. The slope of the saturated water vapor pressure-temperature curve; This is the constant of the wet / dry meter; This represents the saturated water vapor pressure difference.

5. The surface dryness and wetness monitoring system based on energy balance model for evapotranspiration inversion according to claim 1, characterized in that, The neural network model employs a spatiotemporal attention-enhanced convolutional long short-term memory network architecture, including: The spatial feature extraction module, consisting of a multi-layer convolutional neural network, is used to extract spatial feature maps from the input surface state parameter dataset. The temporal dependency capture module, composed of a bidirectional long short-term memory network layer, is used to receive the spatial feature map and capture its forward and backward dependencies in time series, and output the hidden state sequence. The attention weighting module is used to calculate the attention weight at each time step in the hidden state sequence and generate a context feature vector accordingly. The output regression module, consisting of fully connected layers, is used to map the context feature vector to the second evapotranspiration estimate. The calculation formula for the attention weighting module is as follows: , Where T is the time step of the input sequence, For a moment The hidden state vector, For bidirectional long short-term memory networks at any point in the historical sequence The hidden state vector; For a moment Attention weight coefficient, , The weight matrix and weight vector are learnable. For bias vectors, This is the context feature vector.

6. The surface dryness and wetness monitoring system based on energy balance model for evapotranspiration inversion according to claim 1, characterized in that, The dynamic coupling self-calibration unit receives ground verification data in real time via a data interface. This ground verification data includes eddy covariance system observation data from the flux tower, micrometeorological data from the automatic weather station, and soil moisture sensor data. The dynamic coupling self-calibration unit is configured as follows: The observed data from the eddy covariance system are subjected to quality control and flux calculation to obtain the measured evapotranspiration flux value; The measured evapotranspiration flux value is used as the model performance index. The calculation benchmark, and as a loss function Observations in .

7. The surface dryness and wetness monitoring system based on energy balance model for evapotranspiration inversion according to claim 6, characterized in that, The dynamic coupling self-calibration unit calculates the target evapotranspiration value using the following adaptive weighted fusion model. The formula is: ,in The first estimated value of evapotranspiration. The second evapotranspiration estimate, with time-varying weighting coefficients. The following logistic function is dynamically determined: ,in, These are model performance metrics calculated based on real-time ground validation data. The preset performance threshold; This is the gain coefficient that controls the smoothness of the function.

8. The surface dryness and wetness monitoring system based on energy balance model for evapotranspiration inversion according to claim 7, characterized in that, The dynamic coupling self-calibration unit also minimizes the following loss function. The neural network model is trained online: in, For the first One ground verification data point; To verify the total number of data points; The regularization coefficient is used. Let KL divergence be the KL divergence. , The mean and variance output by the physical model. , The mean and variance of the neural network model output.

9. The surface dryness and wetness monitoring system based on the energy balance model for evapotranspiration inversion according to claim 1, characterized in that, The surface wetness / dryness assessment module calculates the evaporation ratio. The formula for the surface wetness index is: ,in, Net surface radiation For soil heat flux, The target evaporation value.

10. The surface dryness and wetness monitoring system based on the energy balance model for evapotranspiration inversion according to claim 9, characterized in that, The surface wetness and dryness assessment module is also used to assess the evaporation rate. With vegetation coverage Construct a dry and wet feature space, and calculate the relative distance from any cell to the dry and wet edges. The formula for generating the surface wetness and dryness monitoring map is as follows: ,in, and These are the equations for the dry side and the wet side, respectively.

Images