A Physically Constrained Neural Network-Based Joint Inversion Method for Soil Moisture in the Surface and Root Zones
By employing a physically constrained neural network method and combining multi-source data to construct a neural network model with a joint loss function, the problem of joint inversion of surface and root zone soil moisture was solved, achieving high-precision, physically consistent multi-layer soil moisture estimation with cross-regional application capability.
Patent Information
- Application Number
- CN202611233362.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to achieve high-precision, physically consistent joint inversion of surface and root zone soil moisture. Traditional methods suffer from issues such as difficulty in establishing model parameters, high data requirements, and insufficient interpretability.
A physical constraint neural network method was adopted, which combined C-band SAR satellite remote sensing data, optical satellite data, meteorological data and topographic data to construct a neural network model with a joint loss function. The surface soil moisture was estimated by the Oh model and water cloud model, and the root zone soil moisture was estimated by the exponential filtering model, thus constructing a multi-layer soil moisture joint inversion architecture.
It achieves high-precision soil moisture estimation across time and space, maintains the physical interpretability of the model, and has cross-regional reliability and stability.
Smart Images

Figure CN123070179A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing soil moisture inversion technology, specifically relating to a joint inversion method for surface-root zone soil moisture using a physically constrained neural network. Background Technology
[0002] Soil moisture is a key state variable in the terrestrial surface water cycle, playing a central regulatory role in hydrological processes such as precipitation distribution, evapotranspiration, infiltration, and runoff. Specifically, surface soil moisture directly controls regional evapotranspiration, influencing energy exchange between the land surface and the atmosphere; while rhizosphere soil moisture controls the distribution of rainfall infiltration and surface runoff, providing crucial water supply for natural vegetation and crop growth, and is also an important indicator for agricultural drought monitoring and crop water stress assessment. Accurate estimation of surface and rhizosphere soil moisture is of significant scientific and practical value for improving the accuracy of hydrological modeling, enhancing drought prediction capabilities, and optimizing agricultural irrigation strategies.
[0003] However, acquiring high-precision, high-spatiotemporal-resolution surface and root zone soil moisture data faces numerous challenges. While traditional in-situ monitoring can provide accurate surface and root zone soil moisture observation data, it suffers from inherent drawbacks such as uneven spatial distribution of sites, high maintenance costs, and difficulties in data acquisition, making it difficult to meet the continuous monitoring needs of large-scale areas. The development of satellite remote sensing technology has provided an effective means for large-scale soil moisture monitoring. For example, satellite products based on microwave remote sensing, such as the Soil Moisture and Ocean Salinity Project (SMOS) and the Soil Moisture Active-Passive Detection Project (SMAP), can provide global-scale spatiotemporal soil moisture data. However, these remote sensing products are mainly limited to monitoring soil moisture in the shallow surface layer (0-5 cm depth), and cannot directly acquire soil moisture information at different depths, making it difficult to meet the application needs of agricultural drought monitoring, hydrological modeling, and agricultural production management for soil moisture information.
[0004] In recent years, machine learning models have provided new technical approaches for soil moisture estimation due to their powerful data processing capabilities, excellent prediction accuracy, and ability to model complex nonlinear relationships. Commonly used machine learning models include extreme gradient boosting numbers, random forests, artificial neural networks, and long short-term memory networks. These machine learning methods mainly involve two implementation paths: first, directly constructing a mapping relationship between surface soil moisture and rhizosphere soil moisture; second, utilizing surface and rhizosphere soil moisture observation data, combined with environmental factors, and using machine learning models to simultaneously fit the nonlinear relationship between independent variables and surface or rhizosphere soil moisture, thereby estimating soil moisture profile information. However, machine learning models are essentially "black box" models; they often neglect the physical laws governing soil moisture transport during the modeling process, resulting in weak interpretability of unknown data and insufficient physical consistency.
[0005] Furthermore, physical models based on the physical mechanisms of soil moisture transport simulate the vertical distribution and movement of soil moisture to achieve multi-layered soil moisture estimation, offering advantages such as describable physical processes, strong model interpretability, and excellent extrapolation potential. However, physical models suffer from problems such as large uncertainties in model structure, significant errors in empirical estimation of key parameters, large input data requirements, and high consumption of parameter calibration and computational resources. In recent years, scholars have proposed a new paradigm of physics-guided machine learning or deep learning, aiming to enhance the interpretability and extrapolation capabilities of machine learning models through physical constraints. Existing frameworks for implementing physics-guided machine learning mainly include: coupled frameworks that use the output of the physical model as the input of machine learning; soft-constraint frameworks that introduce physical constraints into the machine learning loss function; and hard-constraint frameworks that embed physical laws into the structure of the machine learning model. While the aforementioned framework achieves a certain degree of integration between physical laws and data-driven methods, it typically focuses on single-level soil moisture estimation, such as estimation of only the surface layer or only the root zone. Although some studies have attempted multi-depth estimation—first estimating surface soil moisture and then using it as input to estimate root zone soil moisture, or directly using a black-box model to simultaneously output multiple estimation results for surface and root zone soil moisture—the former suffers from the problem of error propagation and amplification along the "surface-root zone" path, while the latter lacks physical mechanism constraints and has insufficient interpretability. Neither approach has achieved joint inversion of surface and root zone soil moisture. The ultimate goal of soil moisture estimation is to obtain complete soil moisture profile information. Estimating surface or root zone soil moisture alone is insufficient to meet the needs of practical applications such as agricultural irrigation, drought monitoring, and hydrological simulation for multi-depth soil moisture data. When facing the task of joint estimation of surface and root zone soil moisture, surface soil moisture estimation models and root zone soil moisture estimation models differ significantly in the form of parameters and input data. For example, root zone soil moisture models typically require time series data of surface soil moisture as input data and update root zone soil moisture recursively, and their parameters have dynamic memory characteristics. The fundamental differences in the input data format and parameter attributes lead to problems such as data interface mismatch, inconsistent time scales, and inconsistent physical meanings of parameters when the two models are combined. This makes it difficult to form a closed-loop physical constraint and data transfer mechanism for surface and root zone soil moisture estimation, thus limiting the accuracy and reliability of multi-depth soil moisture co-inversion. In addition, the combined model still faces prominent problems such as difficulty in accurately determining parameters, insufficient model generalization ability under limited training data, and limited interpretability of intermediate physical parameters.
[0006] In summary, existing technologies, purely machine learning methods lack physical interpretability, while traditional physical models are limited by difficulties in model integration, parameter calibration, and high data requirements. Therefore, there is an urgent need for a joint inversion method for surface-rhizosphere soil moisture that can integrate physical models and machine learning models. This method should maintain the high accuracy of data-driven approaches while ensuring that the model output conforms to the physical laws of soil moisture transport, and achieve accurate, physically consistent, and interpretable joint estimation of surface and rhizosphere soil moisture. Summary of the Invention
[0007] (a) Technical problems to be solved
[0008] The purpose of this invention is to provide a method for joint inversion of surface-root zone soil moisture using a physically constrained neural network, primarily aimed at solving the following technical problems existing in the prior art:
[0009] (1) Due to the sparse number of ground soil moisture monitoring stations and the high maintenance costs, it is difficult to implement large-scale, multi-time-series surface and deep soil moisture monitoring.
[0010] (2) Existing satellite remote sensing data are only sensitive to changes in soil moisture in the shallow layer of 0-5cm. Although machine learning or deep learning models can be used to estimate the moisture in the deep soil, they are essentially black box models, lacking physical mechanism guidance and physical interpretability.
[0011] (3) Although traditional physical models for estimating soil moisture in the surface or root zone are physically interpretable, they have many parameters that are difficult to obtain, require high-quality input data, and have low accuracy in soil moisture estimation results.
[0012] (II) Technical Solution
[0013] To address the above problems, the present invention provides the following technical process:
[0014] 1. A method for joint inversion of surface-root zone soil moisture using a physically constrained neural network, characterized by comprising the following steps:
[0015] S1. Data Collection and Preprocessing: Collect C-band SAR satellite remote sensing data, optical satellite remote sensing data, meteorological data, topographic data, and soil property data. Unify the spatiotemporal reference of all data and establish a feature dataset based on this. Collect measured datasets of soil volumetric water content at different depths.
[0016] S2. Construction of a Physical-Neural Network Model for Joint Estimation of Surface-Rhizosphere Soil Moisture: Shallow soil moisture (0-5cm) is used as a transitional indicator for calculating surface and rhizosphere soil moisture. The physical models for its estimation employ the Oh model and the water cloud model. The physical models for estimating surface (5-10cm) and rhizosphere (20-50cm) soil moisture use an exponential filtering model. Based on the shallow soil moisture output from the Oh model and the water cloud model, and combined with time-series cumulative precipitation and evapotranspiration, a neural network model is established to estimate the time-series weighted average of shallow soil moisture. This average is then used as the input data for the exponential filtering model, thus achieving a joint estimation of shallow and rhizosphere soil moisture.
[0017] S3. Construction of a Neural Network Model for Physical Parameter Estimation: The physical parameters required in step S2 can be divided into four categories: vegetation parameters, surface roughness parameters, surface soil moisture boundary parameters, and root zone soil moisture boundary parameters. Four feedforward neural network modules are constructed to establish the mapping relationship between the feature dataset and each type of parameter.
[0018] S4. Model Training and Validation: Based on the feature dataset extracted in step S1 and the measured soil moisture data, a sample dataset is constructed and divided into training samples and validation samples; a joint loss function containing surface soil moisture loss and root zone soil moisture loss is constructed; the weights of each neural network in steps S2 and S3 are updated synchronously through the backpropagation algorithm; each network is configured with an independent optimizer and learning rate scheduling strategy, and is optimized independently during training; at the same time, boundary constraints are applied to the physical parameters output by each network to keep them within a preset physical range; iteration continues until the joint loss converges to obtain the optimal parameters and weights of each neural network; subsequently, the model's temporal and spatial generalization capabilities are validated.
[0019] S5. Model Application: Using the regional feature dataset as input data, the model trained in S4 is used to predict the surface soil moisture and root zone soil moisture in the region.
[0020] 2. Preferably, in step S1, the data collection and preprocessing includes the following steps:
[0021] S1.1 C-band SAR Satellite Image Feature Extraction: Sentinel-1 or other C-band SAR satellite imagery data was collected, and preprocessed including terrain correction, speckle filtering, and noise removal. The VV polarization backscattering coefficients were converted to decibels, the incident angles were converted from degrees to radians, the image acquisition date was converted to day of year (DOY), and this DOY value was converted into constant raster data consistent with the image's spatial extent. The preprocessed VV polarization backscattering coefficients... Radar incident angle Image acquisition time (DOY) is used as a feature of SAR images.
[0022] S1.2 Optical Remote Sensing Feature Extraction: Select optical remote sensing images (such as Sentinel-2, Landsat, or MODIS) that are at the same or similar time to SAR satellite images, and perform preprocessing such as atmospheric correction, geometric correction, and cloud / snow / shadow masking; extract red light. Green light Near-infrared and shortwave infrared The surface reflectance of the band was measured, and vegetation and water indices were calculated, including the normalized difference vegetation index (NDVI). Green vegetation index Soil-regulated vegetation index Improve the normalized water index and normalized humidity index This serves as a feature of optical remote sensing images. The specific calculation formula is as follows:
[0023] (1);
[0024] (2);
[0025] (3);
[0026] (4);
[0027] (5);
[0028] S1.3 Meteorological Data Collection and Preprocessing: Collect meteorological datasets (such as the global high-resolution land surface reanalysis dataset ERA5-Land, etc.) from actual measurements at the collection sites or officially released datasets. Use bilinear interpolation to spatially interpolate and resample the raw data to ensure that its spatial resolution is consistent with that of the SAR image. Extract the average temperature of the day from the SAR image, and extract the precipitation and evapotranspiration of the day and the previous seven days.
[0029] S1.4, Topographic and Soil Attribute Feature Extraction: Collect digital elevation model datasets (such as SRTM DEM) and soil attribute datasets (such as the SoilDB dataset in OpenLandMap). Calculate topographic features such as slope and aspect using the DEM; select clay content, sand content, and soil bulk density at different depth levels as soil attribute features; spatially resample the above topographic and soil attribute feature data to ensure their spatial resolution is consistent with SAR or optical remote sensing imagery.
[0030] S1.5. Collect soil moisture data for the topsoil (5-10cm) and root zone (20-50cm or 50-100cm) on the day the SAR satellite passes overhead. This data can be collected from continuous soil moisture monitoring stations in the field, or from stratified soil samples collected before and after the SAR satellite's passage, measured using the drying and weighing method to determine the soil volumetric water content. Use the measured topsoil and root zone soil moisture values from the monitoring stations as the true label data for model training.
[0031] 3. Preferably, in step S2, the shallow soil moisture estimation model adopts a coupled model of the water cloud model and the Oh model, and the surface and root zone soil moisture estimation model adopts an exponential filtering model. Furthermore, the weighted average input data of the shallow soil moisture of the exponential filtering model is determined by transforming the output of the coupled model of the water cloud model and the Oh model. The specific steps are as follows:
[0032] S2.1 Shallow Soil Moisture Estimation Model: The water cloud model is used to remove the interference of vegetation on the VV backscattering coefficient, converting the backscattering coefficient of SAR imagery into the backscattering coefficient of bare soil. The specific formula is as follows:
[0033] (6);
[0034] In the formula, The VV polarization backscattering coefficient (dB) of the SAR image. The backscattering coefficient of bare soil after correction using the water cloud model. The radar incident angle is in radians, VWC is the vegetation water content (unit: kg / m²), and parameters A and B are the vegetation scattering coefficient and vegetation attenuation coefficient, respectively.
[0035] The Oh model describes the relationship between shallow soil volumetric water content and topographic roughness. Therefore, coupling the water cloud model with the Oh model can be used to calculate the surface soil volumetric water content. The specific formula is as follows:
[0036] (7);
[0037] (8);
[0038] In the formula, Let VV be the VV polarization backscattering coefficient of bare soil. For surface roughness parameters, For cross-polarization ratio, This represents the volumetric water content of the shallow soil layer (0-5cm) (unit: m³ / m³). The water cloud model, combined with the Oh model, can estimate the volumetric water content (i.e., soil moisture) of the surface soil using SAR backscattering coefficients, vegetation parameters, and surface roughness.
[0039] S2.2, Surface and Root Zone Soil Moisture Estimation Model: The exponential filtering model estimates deeper soil moisture based on temporal information of shallow surface soil moisture combined with soil physical property parameters. The specific formula is as follows:
[0040] (9);
[0041] (10);
[0042] Formula (9) is the formula for calculating the exponentially weighted average of shallow soil moisture, where t is the current time (day). i Let SWI(t) be the historical time (day i), T be the time feature length (days), and SWI(t) be the soil moisture index at time t. The weight of SWI(t) decreases exponentially with the increase of the time difference between each historical time and time t. i ) and SVWC(t) are the shallow soil volumetric water content and the surface or root zone soil volumetric water content above the shallow soil, respectively. Formula (10) is the soil moisture index scaling, where W min and W max These are the permanent wilting point and field water holding capacity of the soil at the corresponding depths, respectively, and are collectively referred to as soil moisture boundary parameters.
[0043] S2.3, Physics-Neural Network Model Combination: Equations (9) and (10) require the first t i To simplify the calculation of shallow soil water at time t, this method reduces the soil moisture index SWI(t) to a nonlinear function of the current shallow soil water at time t and the cumulative precipitation and evapotranspiration of the previous few days. Therefore, formula (10) can be expressed as:
[0044] (11);
[0045] In the formula, SVWC(t) is the soil volumetric water content in the surface or root zone (unit: m³ / m³). These are the shallow soil moisture outputs from the Water Cloud Model and the Oh Model, respectively, as well as the cumulative precipitation and evapotranspiration of the previous few days. Representative to A nonlinear transformation is performed to replace the process of calculating SWI(t) by time series weighted average in formula (9).
[0046] S2.4, Parameter and Input Data Simulation: In formula (6), and The parameters A, B, and ks, as well as the input data VWC and W, are directly given by the VV polarization backscattering coefficients and the radar incident angle, respectively, from the sample dataset. min Wmax The features extracted in step S1 are processed by a neural network. Since the water cloud model and the Oh model are only suitable for estimating the volumetric water content of the 0-5 cm shallow soil layer, in order to avoid the inconsistency error caused by directly comparing the model output with the measured values of the surface layer with a depth of more than 5 cm, the estimation process of surface soil moisture with a depth of more than 5 cm also needs to be converted to the corresponding measured depth by formula (11); among which, the difference between the joint inversion of surface soil moisture and root zone soil moisture is reflected in W min With W max The difference lies in the fact that the two values are taken as the soil permanent wilting point and soil field water holding capacity corresponding to their respective depths. In formula (11), the values for... The nonlinear transformation is also replaced by a neural network model to achieve the desired result. Optimization.
[0047] 4. Preferably, in step S3, five feedforward neural network modules are constructed to respectively handle the vegetation parameters A, B, VWC, ks, and SWI(t) in the simultaneous formula, as well as W at different depths. min W max Perform independent prediction and optimization; each module configures input and output dimensions based on differentiated feature-parameter mapping relationships.
[0048] S3.1 Feedforward Neural Network Structure Design: Each feedforward neural network module adopts a unified basic network architecture, consisting of an input layer, three hidden layers, and an output layer connected sequentially. The first hidden layer has 256 neurons, the second hidden layer has 128 neurons, and the third hidden layer has 64 neurons. Each hidden layer uses the GELU activation function for nonlinear mapping. Batch normalization layers and Dropout regularization layers are introduced after the first and second hidden layers, respectively, with a dropout rate of 0.5 to suppress overfitting and accelerate network convergence. The third hidden layer outputs through a fully connected linear layer.
[0049] S3.2, Feedforward Neural Network for Vegetation Parameters or Input Data: The first feedforward neural network takes the vegetation / water index feature group and annual accumulated day feature as input, and outputs the vegetation scattering coefficient A, vegetation attenuation coefficient B and vegetation water content VWC.
[0050] S3.3, Surface Roughness Neural Network: The second feedforward neural network takes the terrain feature group and the surface soil attribute feature group as inputs and outputs the surface roughness parameter ks.
[0051] S3.4 Soil Moisture Index Optimization Neural Network: The third neural network takes the volumetric water content of the 0-5 cm surface soil, the cumulative precipitation and evapotranspiration of the previous seven days estimated by the water cloud model and the Oh model as input, and outputs SWI(t) in formula (9) and formula (10).
[0052] S3.5, Feedforward Neural Network for Surface Soil Moisture Boundary Parameters: The fourth feedforward neural network takes topographic features, soil property features at a depth of 10 cm, and meteorological features as inputs, and outputs the boundary parameter W of the 10 cm surface soil moisture layer. min With W max .
[0053] S3.6, Feedforward Neural Network for Root Zone Soil Moisture Boundary Parameters: The fifth feedforward neural network takes topographic features, soil property features near 50 cm depth, and meteorological features as inputs, and outputs the root zone depth soil moisture boundary parameter W. min With W max .
[0054] S3.7 Constructing the loss function: Substitute the parameters output by each feedforward neural network module and the input data into the simultaneous formula described in S2, and perform forward modeling to calculate the predicted values of surface and root zone soil moisture; construct a joint loss function that includes surface soil moisture loss terms and root zone soil moisture loss terms, wherein the surface loss term and the root zone loss term are the mean square errors between the predicted and measured values of soil moisture at the corresponding depths; sum the surface loss term and the root zone loss term to obtain the total loss function (15).
[0055] (12);
[0056] (13);
[0057] (14);
[0058] In the formula, and These represent the predicted surface soil moisture and the measured surface soil moisture, respectively. For predicted root zone soil moisture and measured root zone soil moisture. and These are the surface soil moisture loss function and the root zone soil moisture loss function, respectively. This is the total loss function.
[0059] 5. Preferably, the surface-root zone soil moisture joint inversion method of the physical constraint neural network according to claim 1 is characterized in that, in step S4, the specific steps of model training are as follows:
[0060] S4.1 Constructing a sample dataset: Stack the multi-source features extracted in S1.1-S1.4 by channel dimension to construct a multi-dimensional feature dataset; based on the spatial coordinates of the measured soil moisture points, extract the multi-source feature values of the corresponding pixel positions of each point to form a training sample set; all feature variables are used as independent variables for model input, and the measured values of surface soil moisture and root zone soil moisture corresponding to each point are used as label data for model output.
[0061] S4.2 Optimizer Configuration: An independent parameter optimizer is configured for each feedforward neural network module, and an initial learning rate is set for each. A learning rate decay strategy is configured so that the learning rate automatically decreases with each training iteration according to a preset step size and decay coefficient to stabilize the convergence process. Each feedforward neural network module independently performs forward and backward propagation during training. After calculating the prediction error, the optimizer performs error backpropagation, synchronously updating the weights of the five neural networks to complete the overall model training. Each optimizer only updates the internal weights of its corresponding module, and the modules do not affect each other.
[0062] S4.3, Range of Constraint Parameter Prediction Results: Physical boundary constraints are applied to the physical parameters output by each feedforward neural network module. A nonlinear activation function maps vegetation parameters, surface roughness parameters, and soil moisture boundary parameters to a preset reasonable physical range. For the soil moisture boundary parameter W... min With W max Sequential constraints are applied to ensure that the upper limit of soil moisture in the root zone is always greater than the lower limit, thus ensuring that all physical parameters meet the requirements of physical interpretability and numerical stability.
[0063] S4.4 Result Accuracy Assessment: The accuracy assessment process includes calculating different error indices and using measured soil moisture data to evaluate the final results. Specific error indices include the coefficient of determination (R²). 2 Pearson correlation coefficient (R), root mean square error (RMSE), and absolute bias (Bias).
[0064] This invention constructs a deep learning architecture for soil moisture estimation that embeds physical laws. This architecture uses differentiable physical constraints as its core, coupling and simplifying the physical models of surface soil moisture and root zone soil moisture to form a unified physical model of soil moisture. Then, through differentiable processing of the physical model, this unified physical model of soil moisture is embedded into a deep learning framework.
[0065] This invention combines the high interpretability of physical models with the high accuracy of deep learning models. On the one hand, it ensures that the model output conforms to the physical laws of soil moisture transport through a differentiable physical constraint module, preserving the interpretability of the physical process; on the other hand, it utilizes the powerful nonlinear fitting capability of feedforward neural networks to achieve high-precision estimation of surface and root zone soil moisture within the physical constraint framework.
[0066] This invention possesses cross-time and cross-spatial generalization capabilities: when trained using measured data from a subset of years and independently validated using data from other years, the model maintains high estimation accuracy; similarly, when trained using data from a subset of stations and independently validated using data from other stations not involved in the training, the model also demonstrates stable performance. The differentiable physical constraints are based on the universal physical laws of soil moisture transport, enabling the model to operate independently of the statistical characteristics of specific years or the local empirical relationships of specific stations, effectively achieving reliable soil moisture estimation across years and regions. Attached Figure Description
[0067] Figure 1 This is a flowchart of the method of the present invention.
[0068] Figure 2 This is a comparison chart of the verification accuracy of this invention with other methods.
[0069] Figure 3 This is a comparison chart of the time generalization accuracy verification of the present invention with other methods.
[0070] Figure 4 This is a comparison chart of the spatial generalization accuracy verification of this invention and other methods.
[0071] Figure 5 This is a comparison diagram of the soil moisture mapping results of this invention and other methods. Detailed Implementation
[0072] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention may be implemented in other embodiments without these specific details.
[0073] Example 1
[0074] Reference Figure 1 A method for joint inversion of surface-root zone soil moisture using a physically constrained neural network includes the following steps:
[0075] S1. Data Collection and Preprocessing: Collect C-band SAR satellite remote sensing data, optical satellite remote sensing data, meteorological data, topographic data, and soil property data. Unify the spatiotemporal reference of all data and establish a feature dataset based on this. Collect measured datasets of soil volumetric water content at different depths.
[0076] S1.1 C-band SAR Satellite Image Acquisition and Preprocessing: Collect Sentinel-1 or other C-band SAR satellite imagery data, and perform preprocessing such as terrain correction, speckle filtering, and noise removal. Convert the VV polarization backscattering coefficient to decibels, convert the incident angle from degrees to radians, convert the image acquisition date to day of year (DOY), and convert this DOY value into constant raster data consistent with the image's spatial extent. Preprocessed VV polarization backscattering coefficients. Radar incident angle Image acquisition time (DOY) is used as a feature of SAR images.
[0077] S1.2 Optical Remote Sensing Image Acquisition and Preprocessing: Select optical remote sensing images acquired at the same or similar time as SAR satellite images (such as Sentinel-2, Landsat, or MODIS), and perform preprocessing such as atmospheric correction, geometric correction, and cloud / snow / shadow masking; extract red light. Green light Near-infrared and shortwave infrared The surface reflectance of the band was measured, and vegetation and water indices, including the normalized difference in vegetation index (NDVI), were calculated. Green vegetation index Soil-regulated vegetation index Improve the normalized water index and normalized humidity index This serves as a feature of optical remote sensing images. The specific calculation formula is as follows:
[0078] (1);
[0079] (2);
[0080] (3);
[0081] (4);
[0082] (5);
[0083] S1.3 Meteorological Data Collection and Preprocessing: Collect meteorological datasets (such as the global high-resolution land surface reanalysis dataset ERA5-Land, etc.) from actual measurements at the collection sites or officially released datasets. Use bilinear interpolation to spatially interpolate and resample the raw data to ensure that its spatial resolution is consistent with that of the SAR image. Extract the average temperature of the day from the SAR image, and extract the precipitation and evapotranspiration of the day and the previous seven days.
[0084] S1.4. Topographic and Soil Attribute Data Collection and Preprocessing: Collect digital elevation model datasets (such as SRTM DEM) and soil attribute datasets (such as the SoilDB dataset in OpenLandMap). Calculate topographic features such as slope and aspect using the DEM; select clay content, sand content, and soil bulk density at different depth levels as soil attribute features; spatially resample the above topographic and soil attribute data to ensure their spatial resolution is consistent with SAR or optical remote sensing imagery.
[0085] S1.5, Measured Soil Moisture Data: Collect surface (5-10cm) and root zone (20-50cm or 50-100cm) soil moisture data on the day the SAR satellite passes overhead. This data can be collected from continuous soil moisture monitoring stations in the field, or from stratified soil samples collected before and after the SAR satellite passes, measured using the drying and weighing method to determine the volumetric water content. The measured surface and root zone soil moisture values from the monitoring stations will be used as the real label data for model training.
[0086] S2. Construction of a Physical-Neural Network Model for Joint Estimation of Surface-Rhizosphere Soil Moisture: Shallow soil moisture (0-5cm) is used as a transitional indicator for calculating surface and rhizosphere soil moisture. The physical models for its estimation employ the Oh model and the water cloud model. The physical models for estimating surface (5-10cm) and rhizosphere (20-50cm) soil moisture use an exponential filtering model. Based on the shallow soil moisture output from the Oh model and the water cloud model, and combined with time-series cumulative precipitation and evapotranspiration, a neural network model is established to estimate the time-series weighted average of shallow soil moisture. This average is then used as the input data for the exponential filtering model, thus achieving a joint estimation of shallow and rhizosphere soil moisture.
[0087] S2.1 Shallow Soil Moisture Estimation Model: The water cloud model is used to remove the interference of vegetation on the VV backscattering coefficient, converting the backscattering coefficient of SAR imagery into the backscattering coefficient of bare soil. The specific formula is as follows:
[0088] (6);
[0089] In the formula, The VV polarization backscattering coefficient (dB) of the SAR image. The backscattering coefficient of bare soil after correction using the water cloud model. The radar incident angle is in radians, VWC is the vegetation water content (unit: kg / m²), and parameters A and B are the vegetation scattering coefficient and vegetation attenuation coefficient, respectively.
[0090] The Oh model describes the relationship between shallow soil volumetric water content and topographic roughness, with the following formula:
[0091] (7);
[0092] (8);
[0093] In the formula, Let VV be the VV polarization backscattering coefficient of bare soil. For surface roughness parameters, For cross-polarization ratio, This represents the volumetric water content of the shallow soil layer (0-5cm) (unit: m³ / m³). The water cloud model, combined with the Oh model, can estimate the volumetric water content (i.e., soil moisture) of the surface soil using SAR backscattering coefficients, vegetation parameters, and surface roughness.
[0094] S2.2, Surface and Root Zone Soil Moisture Estimation Model: The exponential filtering model estimates deeper soil moisture based on temporal information of shallow surface soil moisture combined with soil physical property parameters. The specific formula is as follows:
[0095] (9);
[0096] (10);
[0097] Formula (9) is the formula for calculating the exponentially weighted average of shallow soil moisture, where t is the current time (day). i Let SWI(t) be the historical time (day i), T be the time feature length (days), and SWI(t) be the soil moisture index at time t. The weight of SWI(t) decreases exponentially with the increase of the time difference between each historical time and time t. i ) and SVWC(t) are the shallow soil volumetric water content and the surface or root zone soil volumetric water content above the shallow soil, respectively. Formula (10) is the soil moisture index scaling, where W min and W max These are the permanent wilting point and field water holding capacity of the soil at the corresponding depths, respectively, and are collectively referred to as soil moisture boundary parameters.
[0098] S2.3, Physics-Neural Network Model Combination: Equations (9) and (10) require the first t i To simplify the calculation of shallow soil water at time t, this method reduces the soil moisture index SWI(t) to a nonlinear function of the current shallow soil water at time t and the cumulative precipitation and evapotranspiration of the previous few days. Therefore, formula (10) can be expressed as:
[0099] (11);
[0100] In the formula, SVWC(t) is the soil volumetric water content in the surface or root zone (unit: m³ / m³). These are the shallow soil moisture outputs from the Water Cloud Model and the Oh Model, respectively, as well as the cumulative precipitation and evapotranspiration of the previous few days. Representative to A nonlinear transformation is performed to replace the process of calculating SWI(t) by time series weighted average in formula (9).
[0101] S2.4, Parameter and Input Data Simulation: In formula (6), and The parameters A, B, and ks, as well as the input data VWC and W, are directly given by the VV polarization backscattering coefficients and the radar incident angle, respectively, from the sample dataset. min W max The features extracted in step S1 are processed by a neural network. Since the water cloud model and the Oh model are only suitable for estimating the volumetric water content of the 0-5 cm shallow soil layer, in order to avoid the inconsistency error caused by directly comparing the model output with the measured values of the surface layer with a depth of more than 5 cm, the estimation process of surface soil moisture also needs to be converted to the corresponding measured depth by formula (11). Among them, the difference between the joint inversion of surface soil moisture and root zone soil moisture is reflected in W min With W max The difference lies in the fact that the two values are taken as the soil permanent wilting point and soil field water holding capacity corresponding to their respective depths. In formula (11), the values for... The nonlinear transformation is also replaced by a neural network model to achieve the desired result. Optimization.
[0102] S3. Construction of a Neural Network Model for Physical Parameter Estimation: The physical parameters required in step S2 can be divided into four categories: vegetation parameters, surface roughness parameters, surface soil moisture boundary parameters, and root zone soil moisture boundary parameters. Four feedforward neural network modules are constructed to establish the mapping relationship between the feature dataset and each type of parameter.
[0103] S3.1 Feedforward Neural Network Structure Design: Each feedforward neural network module adopts a unified basic network architecture, consisting of an input layer, three hidden layers, and an output layer connected sequentially. The first hidden layer has 256 neurons, the second hidden layer has 128 neurons, and the third hidden layer has 64 neurons. Each hidden layer uses the GELU activation function for nonlinear mapping. Batch normalization layers and Dropout regularization layers are introduced after the first and second hidden layers, respectively, with a dropout rate of 0.5 to suppress overfitting and accelerate network convergence. The third hidden layer outputs through a fully connected linear layer.
[0104] S3.2, Feedforward Neural Network for Vegetation Parameters or Input Data: The first feedforward neural network takes the vegetation / water index feature group and annual accumulated day feature as input, and outputs the vegetation scattering coefficient A, vegetation attenuation coefficient B and vegetation water content VWC.
[0105] S3.3, Surface Roughness Neural Network: The second feedforward neural network takes the terrain feature group and the surface soil attribute feature group as inputs and outputs the surface roughness parameter ks.
[0106] S3.4 Soil Moisture Index Optimization Neural Network: The third neural network takes the surface soil volumetric water content, cumulative precipitation and evapotranspiration of the previous seven days estimated by the water cloud model and the Oh model as input, and outputs SWI(t) in formula (9) and formula (10).
[0107] S3.5, Feedforward Neural Network for Surface Soil Moisture Boundary Parameters: The fourth feedforward neural network takes topographic features, surface soil depth properties, and meteorological features as inputs, and outputs the surface soil moisture boundary parameter W. min With W max .
[0108] S3.6, Feedforward Neural Network for Root Zone Soil Moisture Boundary Parameters: The fifth feedforward neural network takes topographic features, soil property features near the root zone depth, and meteorological features as inputs, and outputs the root zone depth soil moisture boundary parameter W. min With W max .
[0109] S3.7 Constructing the loss function: Substitute the parameters output by each feedforward neural network module and the input data into the simultaneous formula described in S2, and perform forward modeling to calculate the predicted values of surface and root zone soil moisture; construct a joint loss function that includes surface soil moisture loss terms and root zone soil moisture loss terms, wherein the surface loss term and the root zone loss term are the mean square errors between the predicted and measured values of soil moisture at the corresponding depths; sum the surface loss term and the root zone loss term to obtain the total loss function (15).
[0110] (12);
[0111] (13);
[0112] (14);
[0113] In the formula, and These represent the predicted surface soil moisture and the measured surface soil moisture, respectively. For predicted root zone soil moisture and measured root zone soil moisture. and These are the surface soil moisture loss function and the root zone soil moisture loss function, respectively. This is the total loss function.
[0114] S4. Model Training and Validation: Based on the feature dataset extracted in step S1 and the measured soil moisture data, a sample dataset is constructed and divided into training samples and validation samples; a joint loss function containing surface soil moisture loss and root zone soil moisture loss is constructed; the weights of each neural network in steps S2 and S3 are updated synchronously through the backpropagation algorithm; each network is configured with an independent optimizer and learning rate scheduling strategy, and is optimized independently during training; at the same time, boundary constraints are applied to the physical parameters output by each network to keep them within a preset physical range; iteration continues until the joint loss converges to obtain the optimal parameters and weights of each neural network; subsequently, the model's temporal and spatial generalization capabilities are validated.
[0115] S4.1 Constructing a sample dataset: Stack the multi-source features extracted in S1.1-S1.4 by channel dimension to construct a multi-dimensional feature dataset; based on the spatial coordinates of the measured soil moisture points, extract the multi-source feature values of the corresponding pixel positions of each point to form a training sample set; all feature variables are used as independent variables for model input, and the measured values of surface soil moisture and root zone soil moisture corresponding to each point are used as label data for model output.
[0116] S4.2 Optimizer Configuration: An independent parameter optimizer is configured for each feedforward neural network module, and an initial learning rate is set for each. A learning rate decay strategy is configured so that the learning rate automatically decreases with each training iteration according to a preset step size and decay coefficient to stabilize the convergence process. Each feedforward neural network module independently performs forward and backward propagation during training. After calculating the prediction error, the optimizer performs error backpropagation, synchronously updating the weights of the five neural networks to complete the overall model training. Each optimizer only updates the internal weights of its corresponding module, and the modules do not affect each other.
[0117] S4.3, Range of Constraint Parameter Prediction Results: Physical boundary constraints are applied to the physical parameters output by each feedforward neural network module. A nonlinear activation function maps vegetation parameters, surface roughness parameters, and soil moisture boundary parameters to a preset reasonable physical range. For the soil moisture boundary parameter W... min With W max Sequential constraints are applied to ensure that the upper limit of soil moisture in the root zone is always greater than the lower limit, thus ensuring that all physical parameters meet the requirements of physical interpretability and numerical stability.
[0118] S4.4 Result Accuracy Assessment: The accuracy assessment process includes calculating different error indices and using measured soil moisture data to evaluate the final results. Specific error indices include the coefficient of determination (R²). 2Pearson correlation coefficient (R), root mean square error (RMSE), absolute value of deviation ( ).
[0119] S5. Model Application: Using the regional feature dataset as input data, the model trained in S4 is used to predict the surface soil moisture and root zone soil moisture in the region.
[0120] Example 2
[0121] This embodiment uses Inner Mongolia Autonomous Region as the study area, selecting soil moisture monitoring data from 119 meteorological stations in the region between 2017 and 2022 as labels: the soil volumetric water content at a depth of 10 cm represents the surface soil moisture, and the average soil volumetric water content at a depth of 20-50 cm represents the root zone soil moisture. Based on the above labeled data, combined with multi-source feature datasets within the study area, a joint inversion method for surface-root zone soil moisture using a physically constrained neural network is employed to achieve accurate prediction of soil moisture content.
[0122] Reference Figures 2-5 A method for joint inversion of surface-root zone soil moisture using a physically constrained neural network includes the following steps:
[0123] S1. Sample Dataset Construction: The latitude and longitude coordinates of 119 soil moisture monitoring stations were uploaded to the remote sensing cloud computing platform in SHP format. The platform's built-in multi-source data, including remote sensing images, meteorological data, topographic data, and soil properties, were used for preprocessing. The specific process is as follows:
[0124] S1.1 SAR Image Feature Extraction: The Sentinel-1 SAR GRD dataset built into the cloud computing platform was selected and filtered according to the spatial range of the study area and the target time period. The filtered images were batch-processed with preprocessing operations such as terrain correction, speckle filtering and noise removal, and the VV band backscattering coefficient was uniformly converted to decibel (dB) unit. At the same time, the incident angle band was converted to radians, and the annual day band was calculated and added according to the imaging time of each image to complete the SAR feature construction.
[0125] S1.2, Feature extraction of optical remote sensing images: Select the MCD43A4 surface reflectance dataset that is consistent with the time phase of the SAR image, crop it according to the spatial range of the SAR image, and perform cloud masking; then extract key bands, calculate vegetation indices (NDVI, GNDVI, SAVI) and water indices (MNDWI, NDMI) to construct a multi-source optical remote sensing feature set.
[0126] S1.3 Meteorological Feature Extraction: Select the ERA5-LAND Daily climate dataset, which has a daily resolution. After selecting climate data with the same date as the SAR, crop the data according to the SAR image range and extract the daily precipitation and temperature bands from the dataset.
[0127] S1.4 Topographic and Soil Attribute Feature Extraction: Topographic features, including elevation, slope and aspect, were extracted using SRTM data. Three subsets of soil attributes, namely clay content, sand content and soil bulk density, from the OpenLandMap soil database were used as static features of soil attributes. Soil attribute values at four depth levels (0 cm, 10 cm, 30 cm and 60 cm) were extracted for each subset to construct a joint static feature set of topographic and soil attributes.
[0128] S1.5 Matching Feature Data with Measured Data: Combine SAR image features, optical remote sensing image features, meteorological features, and static features from the same date into a multi-source joint feature image; based on the uploaded coordinates of 119 stations, extract the multi-source feature information corresponding to each station on the same day, output it in CSV format and download it to the local machine; match and merge the extracted feature data with the measured soil moisture data of the corresponding date to construct a sample dataset.
[0129] S2. Combining the semi-empirical model of surface soil moisture with the physical model of surface-root zone soil moisture: The formula obtained by combining the water cloud model, the Oh model, and the simplified exponential filter model is written as a function in the Python 3.9 environment, with the following input parameters: A, B, VWC, ks W min W max The returned value is the volumetric water content of the topsoil or root zone at a depth greater than 5 cm.
[0130] S3. Neural Network Model Construction: Using the PyTorch deep learning framework in a Python 3.9 environment, four basic feedforward neural network training modules are built. The specific process is as follows:
[0131] S3.1 Basic Feedforward Neural Network Module Structure Design: A fully connected feedforward neural network is constructed, consisting of an input layer, three hidden layers, and an output layer. The first hidden layer has 256 neurons, which are sequentially connected to batch normalization, activation functions, and randomized regularization. The second hidden layer has 128 neurons, configured similarly. The third hidden layer has 64 neurons, which are connected to the output layer after being connected to the activation function. The dimensions of each layer are adaptively configured according to the number of input features and the number of output target parameters, and forward propagation is achieved through sequential containers.
[0132] S3.2 Mapping relationship between different feature combinations and target parameters: The first feedforward neural network takes vegetation / water index and annual accumulated days as input and outputs parameters A, B, and VWC; the second neural network takes topographic parameters and surface soil property data as input and outputs parameter ks; the third neural network takes topographic parameters and soil property data at a depth of 10cm as input and outputs W of the soil at a depth of 10cm. min and W max Parameters; the fourth neural network takes topographic parameters and soil property data near the 50cm depth as input, and outputs W of the soil at a depth of 50cm. min and W max parameter.
[0133] S4. Loss Function Establishment: The parameters output by the four feedforward neural networks are substituted into the physical model of soil moisture in the surface and root zones to calculate the volumetric water content of the surface soil at a depth of 10 cm and the volumetric water content of the root zone at a depth of 20–50 cm. The mean square error is used to calculate the deviation between the calculated and measured values of soil moisture in the surface and root zones, and the two are added together to obtain the total error. The four neural networks are backpropagated synchronously based on the total error to realize the joint update of the internal weights.
[0134] S5. Model Training: An adaptive moment estimation optimizer is configured for each feedforward neural network, with an initial learning rate of 0.007. A learning rate scheduler is introduced to dynamically adjust the learning rate based on the training progress, suppressing gradient vanishing or gradient exploding phenomena. The predicted parameters output by each network are constrained within their numerical range using activation functions to ensure that the parameter values conform to physical meaning. The maximum number of iterations is set to 3000, and early stopping is not activated.
[0135] S6. Accuracy Evaluation: Three training strategies were employed to systematically verify the model's performance: First, random cross-validation, where samples were randomly shuffled, and 10-fold cross-validation was used to optimize the number of deep learning iterations and evaluate generalization performance. Second, temporal generalization validation, using samples from 2017–2020 for training and samples from 2021–2022 for validation, to test the model's temporal extrapolation ability. Third, spatial generalization validation, using data from a subset of locations for training and data from the remaining locations for validation, to test the model's spatial transfer ability. Validation metrics included the coefficient of determination (R²), correlation coefficient (R), root mean square error (RMSE), and absolute bias. Random forest (RF), unconstrained feedforward neural networks (FNN), and convolutional neural networks (CNN) were also introduced as benchmarks to comprehensively evaluate the accuracy advantages and generalization robustness of the proposed method.
[0136] S6.1 10-fold cross-validation results: After 10-fold cross-validation, the differentiable physical constraint model showed the best accuracy. The coefficient of determination (R²) for predicting the top 10 cm soil moisture content was 0.7544, the correlation coefficient (R) was 0.8685, the root mean square error (RMSE) was 0.0401 m³ / m³, and the bias was 0.0011. The coefficient of determination (R²) for predicting the volumetric water content of the root zone (20–50 cm) was 0.7962, the correlation coefficient (R) was 0.8923, the root mean square error (RMSE) was 0.0366 m³ / m³, and the bias was 0.0016.
[0137] S6.2 Time Generalization Test Results: After time generalization verification, the differentiable physical constraint model showed the best accuracy. The coefficient of determination (R²) for predicting the top 10 cm soil moisture was 0.5955, the correlation coefficient (R) was 0.7717, the root mean square error (RMSE) was 0.0522 m³ / m³, and the bias was 0.0101. The coefficient of determination (R²) for predicting the volumetric water content of the root zone (20–50 cm) was 0.6215, the correlation coefficient (R) was 0.7884, the root mean square error (RMSE) was 0.0516 m³ / m³, and the bias was 0.0152. The physical constraint neural network-based joint inversion method for surface-root zone soil moisture performed best.
[0138] S6.3 Spatial Generalization Test Results: After time generalization verification, the differentiable physical constraint model showed the best accuracy. The coefficient of determination (R²) for predicting soil moisture in the top 10 cm layer reached 0.6283, the correlation coefficient (R) was 0.7927, the root mean square error (RMSE) was 0.0512 m³ / m³, and the bias was 0.0018. The coefficient of determination (R²) for predicting soil volumetric water content in the root zone (20–50 cm layer) reached 0.5469, the correlation coefficient (R) was 0.7395, the root mean square error (RMSE) was 0.0589 m³ / m³, and the bias was 0.0016. The physical constraint neural network-based joint inversion method for surface-root zone soil moisture performed best.
Claims
1. A method for joint inversion of surface-root zone soil moisture using a physically constrained neural network, characterized in that, Includes the following steps: S1. Data Collection and Preprocessing: Collect C-band SAR satellite remote sensing data, optical satellite remote sensing data, meteorological data, topographic data, and soil property data. Unify the spatiotemporal reference of all data and establish a feature dataset based on this. Collect measured datasets of soil volumetric water content at different depths. S2. Construction of a physical-neural network model for joint estimation of surface-root zone soil moisture: Shallow soil moisture (0-5 cm) is used as an intermediate indicator for calculating surface and root zone soil moisture. The physical models for its estimation employ the Oh model and the water cloud model. The physical models for estimating surface (5-10 cm) and root zone (20-50 cm) soil moisture use an exponential filtering model. Based on the shallow soil moisture output from the Oh model and the water cloud model, and combined with time-series cumulative precipitation and evapotranspiration, a neural network model is established to estimate the time-series weighted average of shallow soil moisture. This average is then used as input data for the exponential filtering model, achieving a joint estimation of shallow and root zone soil moisture. S3. Construction of a Neural Network Model for Physical Parameter Estimation: The physical parameters required in step S2 can be divided into four categories: vegetation parameters, surface roughness parameters, surface soil moisture boundary parameters, and root zone soil moisture boundary parameters. Four feedforward neural network modules are constructed to establish the mapping relationship between the feature dataset and each type of parameter. S4. Model Training and Validation: Based on the feature dataset extracted in step S1 and the measured soil moisture data, a sample dataset is constructed and divided into training samples and validation samples; a joint loss function containing surface soil moisture loss and root zone soil moisture loss is constructed; the weights of each neural network in steps S2 and S3 are updated synchronously through the backpropagation algorithm; each network is configured with an independent optimizer and learning rate scheduling strategy, and is optimized independently during training; at the same time, boundary constraints are applied to the physical parameters output by each network to keep them within a preset physical range; iteration continues until the joint loss converges to obtain the optimal parameters and weights of each neural network; subsequently, the model's temporal and spatial generalization capabilities are validated. S5. Model Application: Using the regional feature dataset as input data, the model trained in S4 is used to predict the surface soil moisture and root zone soil moisture in the region.
2. The method for joint inversion of surface-root zone soil moisture using a physically constrained neural network according to claim 1, characterized in that, In step S1, the data collection and preprocessing includes the following steps: S1.1 C-band SAR Satellite Image Acquisition and Preprocessing: Collect Sentinel-1 or other C-band SAR satellite imagery data, and perform preprocessing such as terrain correction, speckle filtering, and noise removal. Convert the VV polarization backscattering coefficient to decibels, convert the incident angle from degrees to radians, convert the image acquisition date to day of year (DOY), and convert this DOY value into constant raster data consistent with the image's spatial extent. Preprocessed VV polarization backscattering coefficients. Radar incident angle Image acquisition time (DOY) is used as a feature of SAR images. S1.2 Optical Remote Sensing Image Acquisition and Preprocessing: Select optical remote sensing images acquired at the same or similar time as SAR satellite images (such as Sentinel-2, Landsat, or MODIS), and perform preprocessing such as atmospheric correction, geometric correction, and cloud / snow / shadow masking; extract red light. Green light Near-infrared and shortwave infrared The surface reflectance of the band was measured, and vegetation and water indices, including the normalized difference in vegetation index (NDVI), were calculated. Green vegetation index Soil-regulated vegetation index Improve the normalized water index and normalized humidity index This serves as a feature of optical remote sensing images. The specific calculation formula is as follows: (1); (2); (3); (4); (5); S1.3 Meteorological Data Collection and Preprocessing: Collect meteorological datasets (such as the global high-resolution land surface reanalysis dataset ERA5-Land, etc.) from actual measurements at the collection sites or officially released datasets. Use bilinear interpolation to spatially interpolate and resample the raw data to ensure that its spatial resolution is consistent with that of the SAR image. Extract the average temperature of the day from the SAR image, and extract the precipitation and evapotranspiration of the day and the previous seven days. S1.
4. Topographic and Soil Attribute Data Collection and Preprocessing: Collect digital elevation model datasets (such as SRTM DEM) and soil attribute datasets (such as the SoilDB dataset in OpenLandMap). Calculate topographic features such as slope and aspect using the DEM; select clay content, sand content, and soil bulk density at different depth levels as soil attribute features; spatially resample the above topographic and soil attribute data to ensure their spatial resolution is consistent with SAR or optical remote sensing imagery. S1.5, Measured Soil Moisture Data: Collect surface (5-10cm) and root zone (20-50cm or 50-100cm) soil moisture data on the day the SAR satellite passes overhead. This data can be collected from continuous soil moisture monitoring stations in the field, or from stratified soil samples collected before and after the SAR satellite passes, measured using the drying and weighing method to determine the volumetric water content. The measured surface and root zone soil moisture values from the monitoring stations will be used as the real label data for model training.
3. The method for joint inversion of surface-root zone soil moisture using a physically constrained neural network according to claim 1, characterized in that, In step S2, the physical model for estimating shallow soil moisture adopts the Oh model and the water cloud model; the physical model for estimating surface and root zone soil moisture adopts the exponential filtering model; based on the shallow soil moisture output by the Oh model and the water cloud model, combined with time-series cumulative precipitation and evapotranspiration, a neural network model is established to estimate the time-series weighted average of shallow soil moisture, and this is used as the input data for the exponential filtering model to realize the simultaneous establishment of shallow-surface-root zone soil moisture estimation models. S2.1 Shallow Soil Moisture Estimation Model: The water cloud model is used to remove the interference of vegetation on the VV backscattering coefficient, converting the backscattering coefficient of SAR imagery into the backscattering coefficient of bare soil. The specific formula is as follows: (6); In the formula, The VV polarization backscattering coefficient (dB) of the SAR image. The backscattering coefficient of bare soil after correction using the water cloud model. The radar incident angle is in radians, VWC is the vegetation water content (unit: kg / m²), and parameters A and B are the vegetation scattering coefficient and vegetation attenuation coefficient, respectively. The Oh model describes the relationship between shallow soil volumetric water content and topographic roughness, with the following formula: (7); (8); In the formula, Let VV be the VV polarization backscattering coefficient of bare soil. For surface roughness parameters, For cross-polarization ratio, This represents the volumetric water content of the shallow soil layer (0-5cm) (unit: m³ / m³). The water cloud model, combined with the Oh model, can estimate the volumetric water content (i.e., soil moisture) of the surface soil using SAR backscattering coefficients, vegetation parameters, and surface roughness. S2.2, Surface and Root Zone Soil Moisture Estimation Model: The exponential filtering model estimates deeper soil moisture based on temporal information of shallow surface soil moisture combined with soil physical property parameters. The specific formula is as follows: (9); (10); Formula (9) is the formula for calculating the exponentially weighted average of shallow soil moisture, where t is the current time (day). i Let SWI(t) be the historical time (day i), T be the time feature length (days), and SWI(t) be the soil moisture index at time t. The weight of SWI(t) decreases exponentially with the increase of the time difference between each historical time and time t. i ) and SVWC(t) are the shallow soil volumetric water content and the surface or root zone soil volumetric water content above the shallow soil, respectively. Formula (10) is the soil moisture index scaling, where W min and W max These are the permanent wilting point and field water holding capacity of the soil at the corresponding depths, respectively, and are collectively referred to as soil moisture boundary parameters. S2.3, Physics-Neural Network Model Combination: Equations (9) and (10) require the first t i To simplify the calculation of shallow soil water at time t, this method reduces the soil moisture index SWI(t) to a nonlinear function of the current shallow soil water at time t and the cumulative precipitation and evapotranspiration of the previous few days. Therefore, formula (10) can be expressed as: (11); In the formula, SVWC(t) is the soil volumetric water content in the surface or root zone (unit: m³ / m³). These are the shallow soil moisture outputs from the Water Cloud Model and the Oh Model, respectively, as well as the cumulative precipitation and evapotranspiration of the previous few days. Representative to A nonlinear transformation is performed to replace the process of calculating SWI(t) by time series weighted average in formula (9). S2.4, Parameter and Input Data Simulation: In formula (6), and The parameters A, B, and ks, as well as the input data VWC and W, are directly given by the VV polarization backscattering coefficients and the radar incident angle, respectively, from the sample dataset. min W max The features extracted in step S1 are processed by a neural network. Since the water cloud model and the Oh model are only suitable for estimating the volumetric water content of the 0-5 cm shallow soil layer, in order to avoid the inconsistency error caused by directly comparing the model output with the measured values of the surface layer with a depth of more than 5 cm, the estimation process of surface soil moisture with a depth of more than 5 cm also needs to be converted to the corresponding measured depth by formula (11); among which, the difference between the joint inversion of surface soil moisture and root zone soil moisture is reflected in W min With W max The difference lies in the fact that the two values are taken as the soil permanent wilting point and soil field water holding capacity corresponding to their respective depths. In formula (11), the values for... The nonlinear transformation is also replaced by a neural network model to achieve the desired result. Optimization.
4. The method for joint inversion of surface-root zone soil moisture using a physically constrained neural network according to claim 1, characterized in that, In step S3, five feedforward neural network modules are constructed to handle the vegetation parameters A, B, VWC, ks, and SWI(t) in the simultaneous formula, and different depths of W... min W max Perform independent prediction and optimization; each module configures input and output dimensions based on differentiated feature-parameter mapping relationships. S3.1 Feedforward Neural Network Structure Design: Each feedforward neural network module adopts a unified basic network architecture, consisting of an input layer, three hidden layers, and an output layer connected sequentially. The first hidden layer has 256 neurons, the second hidden layer has 128 neurons, and the third hidden layer has 64 neurons. Each hidden layer uses the GELU activation function for nonlinear mapping. Batch normalization layers and Dropout regularization layers are introduced after the first and second hidden layers, respectively, with a dropout rate of 0.5 to suppress overfitting and accelerate network convergence. The third hidden layer outputs through a fully connected linear layer. S3.2, Feedforward Neural Network for Vegetation Parameters or Input Data: The first feedforward neural network takes the vegetation / water index feature group and annual accumulated day feature as input, and outputs the vegetation scattering coefficient A, vegetation attenuation coefficient B and vegetation water content VWC. S3.3, Surface Roughness Neural Network: The second feedforward neural network takes the terrain feature group and the surface soil attribute feature group as inputs and outputs the surface roughness parameter ks. S3.4, Soil Moisture Index Optimization Neural Network: The third neural network takes the surface soil volumetric water content, cumulative precipitation and evapotranspiration of the previous seven days estimated by the Water Cloud Model and the Oh Model as input, and outputs the soil moisture index SWI(t) in formula (9) and formula (10). S3.5, Feedforward Neural Network for Surface Soil Moisture Boundary Parameters: The fourth feedforward neural network takes topographic features, surface soil depth properties, and meteorological features as inputs, and outputs the surface soil moisture boundary parameter W. min With W max . S3.6, Feedforward Neural Network for Root Zone Soil Moisture Boundary Parameters: The fifth feedforward neural network takes topographic features, soil property features near the root zone depth, and meteorological features as inputs, and outputs the root zone depth soil moisture boundary parameter W. min With W max . S3.7 Constructing the loss function: Substitute the parameters output by each feedforward neural network module and the input data into the simultaneous formula described in S2, and perform forward modeling to calculate the predicted values of surface and root zone soil moisture; construct a joint loss function that includes surface soil moisture loss terms and root zone soil moisture loss terms, wherein the surface loss term and the root zone loss term are the mean square errors between the predicted and measured values of soil moisture at the corresponding depths; sum the surface loss term and the root zone loss term to obtain the total loss function (15). (12); (13); (14); In the formula, and These represent the predicted surface soil moisture and the measured surface soil moisture, respectively. For predicted root zone soil moisture and measured root zone soil moisture. and These are the surface soil moisture loss function and the root zone soil moisture loss function, respectively. This is the total loss function.
5. The method for joint inversion of surface-root zone soil moisture using a physically constrained neural network according to claim 1, characterized in that, In step S4, the specific steps for model training are as follows: S4.1 Constructing a sample dataset: Stack the multi-source features extracted in S1.1-S1.4 by channel dimension to construct a multi-dimensional feature dataset; based on the spatial coordinates of the measured soil moisture points, extract the multi-source feature values of the corresponding pixel positions of each point to form a training sample set; all feature variables are used as independent variables for model input, and the measured values of surface soil moisture and root zone soil moisture corresponding to each point are used as label data for model output. S4.2 Optimizer Configuration: Configure an independent parameter optimizer for each feedforward neural network module and set the initial learning rate for each; configure a learning rate step decay strategy so that the learning rate automatically decreases with each training iteration according to a preset step size and decay coefficient to stabilize the convergence process; Each feedforward neural network module independently performs forward and backward propagation during training. After calculating the prediction error, the optimizer performs error backpropagation, synchronously updating the weights of the five neural networks to complete the overall model training. Each optimizer only updates the internal weights of its corresponding module, and the modules do not affect each other. S4.3, Range of prediction results for constraint parameters: Apply physical boundary constraints to the physical parameters output by each feedforward neural network module, and map vegetation parameters, surface roughness parameters and soil moisture boundary parameters to a preset reasonable physical range through a nonlinear activation function; For soil moisture boundary parameter W min With W max Sequential constraints are applied to ensure that the upper limit of soil moisture in the root zone is always greater than the lower limit, thus ensuring that all physical parameters meet the requirements of physical interpretability and numerical stability. S4.4 Result Accuracy Assessment: The accuracy assessment process includes calculating different error indices and using measured soil moisture data to evaluate the final results. Specific error indices include the coefficient of determination (R²). 2 Pearson correlation coefficient (R), root mean square error (RMSE), absolute value of deviation ( ).
