Soil moisture inversion method for deep transfer learning of scattering model

Abstract

The invention discloses a soil moisture inversion method for deep transfer learning of a scattering model, and belongs to the field of soil moisture inversion. The method comprises the following steps: generating a simulation data set based on an AIEM-Oh model; constructing a DET model; pre-training the DET model by using the simulation data set; obtaining a sample data set, wherein each sample comprises an on-site SSM target value and a preset predictive factor value; after feature migration is carried out on the predictive factor of each sample, the predictive factor is combined with an on-site SSM target value under the corresponding sample to be used for carrying out migration learning on the DET model which completes pre-training; in a transfer learning process, performing parameter fine tuning on the DET model after pre-training; and finally obtaining a DET model after transfer learning is completed for soil moisture inversion. The method is suitable for a vegetation coverage area, integrates a physical mechanism, and does not need to depend on actual measurement data of surface roughness and vegetation parameters.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of soil moisture inversion, and particularly relates to a soil moisture inversion method based on deep transfer learning of a scattering model. BACKGROUND

[0002] Surface soil moisture (SSM) is a key parameter affecting agricultural productivity and ecosystem stability. With its all-weather and high-resolution advantages, synthetic aperture radar (SAR) has become an increasingly important technical means for SSM inversion.

[0003] Machine learning (ML) provides an effective way for estimating SSM using SAR data through nonlinear modeling and efficient computing capabilities. However, most existing ML-based methods only use public auxiliary data and reference SSM for nonlinear modeling, lacking the integration of physical mechanisms. Coupling theoretical models with ML can improve model interpretability, but still faces two major challenges:

[0004] (1) Theoretical models have inherent uncertainties;

[0005] (2) Key input parameters (such as surface roughness) are difficult to obtain in practical applications.

[0006] In the prior art, there is a method of using AIEM-Oh model microwave scattering simulation data to pre-train DNN and transfer to data-scarce bare soil areas for SSM inversion. However, this method highly depends on measured data of surface roughness and is only applicable to bare soil areas, making it difficult to be extended to vegetation-covered areas.

[0007] Therefore, the present application proposes a soil moisture inversion method based on deep transfer learning of a scattering model. SUMMARY

[0008] The present application aims to overcome the shortcomings of the prior art and proposes a soil moisture inversion method based on deep transfer learning of a scattering model to achieve the following purposes: to propose a soil moisture inversion method applicable to vegetation-covered areas, integrating physical mechanisms, and not requiring measured data of surface roughness and vegetation parameters.

[0009] To achieve the above purposes, the technical solution adopted by the present application is as follows: a soil moisture inversion method based on deep transfer learning of a scattering model, comprising the following steps:

[0010] Step S1: generating a simulated data set based on an AIEM-Oh model;

[0011] Step S2, constructing a DET model; pre-training the DET model using the simulation data set;

[0012] Step S3, obtaining a sample data set, wherein each sample includes an in-situ SSM target value and a preset prediction factor value;

[0013] Step S4, after feature migration of the prediction factor of each sample, the in-situ SSM target value under the corresponding sample is used to perform transfer learning on the pre-trained DET model;

[0014] Step S5, during the transfer learning process, the pre-trained DET model is parameter fine-tuned; and finally, a transfer learning completed DET model is obtained for soil moisture inversion.

[0015] Preferably, in the step S1, the input of the AIEM-Oh model includes an incident angle θ of a soil surface, a root mean square height s, a correlation length l, a power exponent p and an SSM, and the generated simulation data includes VV polarization data of the soil surface VH polarization data incident angle θ, root mean square height s, correlation length l, power exponent p and SSM.

[0016] Preferably, in the step S2, the DET model includes a Transformer encoder and a DNN model, input data passes through the Transformer encoder and the DNN model in sequence to output a predicted SSM, and the input data includes VV polarization data of a soil surface VH polarization data incident angle θ, root mean square height s, correlation length l, power exponent p and SSM.

[0017] Preferably, in the step S3, the preset prediction factor includes a backscatter observation variable, a spectral reflectance variable, a soil texture variable, a terrain index variable and a vegetation coverage variable, wherein:

[0018] The backscatter observation variable includes VV polarization data at the top of the canopy VH polarization data at the top of the canopy incident angle θ at the top of the canopy can ;

[0019] The spectral reflectance variable includes visible light bands 2, 3 and 4, near-infrared band 8, and short-wave infrared bands 11 and 12;

[0020] The soil texture variable includes percentages of sand, silt and clay content;

[0021] The terrain index variable includes DEM data, slope, terrain surface convexity, terrain surface texture and terrain ruggedness index;

[0022] The vegetation cover variable includes vegetation types.

[0023] Preferably, in the step S4, the predictors of each sample are transferred by a fully connected layer, and the feature transfer method includes whole transfer, independent transfer and cross transfer.

[0024] Preferably, the whole transfer inputs the backscattering observation variable, the spectral reflectance variable, the soil texture variable, the terrain index variable and the vegetation cover variable into a single fully connected layer to obtain the VV polarization data of the soil surface VH polarization data The incident angle θ, the root mean square height s, the correlation length l and the power exponent p are input into the pre-trained DET model together with the SSM under the corresponding sample to perform transfer learning.

[0025] The independent transfer sets the backscattering observation variable, the spectral reflectance variable and the vegetation cover variable as a first group, and sets the soil texture variable and the terrain index variable as a second group. After the two groups are respectively input into two independent fully connected layers, the first group obtains the VV polarization data of the soil surface VH polarization data The incident angle θ, the root mean square height s, the correlation length l and the power exponent p are input into the pre-trained DET model together with the SSM under the corresponding sample to perform transfer learning.

[0026] The cross transfer adds two cross connections on the basis of the independent transfer, that is, the output of the first group after the fully connected layer is additionally used as the input of the second group fully connected layer, and similarly, the output of the second group after the fully connected layer is additionally used as the input of the first group fully connected layer.

[0027] Preferably, the fully connected layer is configured with 128 neurons and adopts a ReLU activation function.

[0028] Preferably, in the step S5, the parameter fine-tuning method includes full unfreezing, encoder freezing, DNN freezing and full freezing.

[0029] Preferably, the full unfreezing simultaneously updates the parameters of the Transformer encoder and the DNN in the transfer learning process; the encoder freezing only updates the parameters of the DNN in the transfer learning process; the DNN freezing only updates the parameters of the Transformer encoder in the transfer learning process; and the full freezing keeps the parameters of the pre-trained Transformer encoder and the DNN in the transfer learning process.

[0030] Preferably, the method employs a combination of cross-migration and encoder freezing as the optimal transfer learning scheme.

[0031] The technical effects of the present application are:

[0032] (1) The present application constructs a DET model and pre-trains it using simulated knowledge generated based on the AIEM-Oh model, thereby migrating the simulated soil environment to the actual situation, and designs three feature migration strategies: overall migration, independent migration and cross-migration, thereby achieving accurate and high-resolution SSM estimation in areas where there is vegetation scattering interference and a lack of available surface roughness data.

[0033] (2) The present application designs different parameter fine-tuning strategies: full freezing, encoder freezing, DNN freezing and full unfreezing during the transfer learning process, thereby preserving the physical knowledge of the AIEM-Oh model while improving accuracy and generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 A flowchart of a scattering model deep transfer learning soil moisture inversion method is provided for the embodiments of the present application;

[0035] Figure 2 A DET model structure diagram is provided for the embodiments of the present application;

[0036] Figure 3 A transfer learning framework diagram is provided for the embodiments of the present application;

[0037] Figure 4 Scatter plots for SSM inversion using 12 different transfer learning methods: the three columns from left to right represent the results obtained by the three different migration strategies, namely overall migration, independent migration and cross-migration; while the four rows from top to bottom show the results using four fine-tuning methods, which are full unfreezing, encoder freezing, DNN freezing and full freezing; the color of the points represents the probability density, the red dotted line represents the 1:1 straight line, and the gray line represents the fitted linear regression line;

[0038] Figure 5 Scatter plots for SSM inversion using different inversion models in a dependent scenario (based on sample cross-validation scenario): (a) to (f) are the results of the optimal transfer model, random forest, XGBoost, DET, DNN and SVR in the dependent scenario;

[0039] Figure 6Scatter plots of SSM inversion under independent scenarios (based on independent site test scenarios) for different inversion models, and time series comparison schematic diagram of SSM estimated values based on different inversion models at different sites; wherein (a) to (f) respectively show the performance of the optimal transfer model, RF, XGBoost, DET, DNN and SVR based on independent scenarios; (g) to (j) are time series comparison schematic diagrams of SSM estimated values based on different inversion models (including the optimal transfer model, RF, DET and in-situ SSM) at four independent sites (i.e. C2, MS3523, MS3545 and MS3633), wherein the blue columnar chart represents the daily precipitation;

[0040] Figure 7 Comparison chart between SSM estimated by the optimal transfer model and other SSM products: (a) to (d) respectively show the time series of four independent scenario sites (i.e. C2, MS3523, MS3545 and MS3633); (e) and (f) are box plots of SSM estimated by four independent scenario sites in the test area and in-situ SSM;

[0041] Figure 8 SSM spatial distribution map of the Nagqu River Basin; wherein: (a) to (c) are respectively the median SSM inference maps of the Nagqu River Basin with a resolution of 10 meters in April, July and October 2019 generated by using the optimal transfer strategy, (d) to (l) are comparison charts of median SSM products in the same period, i.e. (d) to (f) are SSM_China, (g) to (i) are BTCH, and (j) to (l) are TP_SM. DETAILED DESCRIPTION

[0042] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings, and the purpose is to help the technical personnel in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present application, and to help its implementation. It should be noted that the words "first", "second" and the like described in this application are only convenient for the description of the technical solution and used as a distinction between components, and the corresponding component configuration may be the same or different, and this application is not limited to this. In order to make the technical solution of the present application more clear, the present application is explained and described by the following examples.

[0043] The embodiment provides a soil moisture inversion method based on scattering model deep transfer learning, which is suitable for vegetation coverage area, fuses physical mechanism, and can realize soil moisture inversion without relying on measured data of surface roughness and vegetation parameters, and specifically comprises the following steps:

[0044] Step S1, generating a simulation data set based on the AIEM-Oh model;

[0045] Step S2, constructing a DET model; pre-training the DET model using the simulation dataset;

[0046] Step S3, obtaining a sample dataset, wherein each sample includes an in-situ SSM target value and preset predictor values;

[0047] Step S4, performing feature migration on the predictor of each sample, and combining the in-situ SSM target value under the corresponding sample to perform transfer learning on the pre-trained DET model;

[0048] Step S5, during the transfer learning process, performing parameter fine-tuning on the pre-trained DET model; and finally obtaining a transfer-learned DET model for soil moisture inversion.

[0049] The steps S1-S5 of the embodiment will be described below. Figure 1 The steps S1-S5 of the embodiment will be described below.

[0050] Referring to step S1, a simulation dataset is generated based on an AIEM-Oh model. The AIEM-Oh model is a combination of AIEM and Oh models. The AIEM model has high simulation accuracy in co-polarization, and the Oh model also has high simulation accuracy in cross-polarization. The simulation data obtained thereby has high reliability. In the embodiment, the input of the AIEM-Oh model includes a soil surface incident angle θ, a root mean square height s, a correlation length l, a power exponent p, and an SSM. The generated simulation data includes VV-polarization data

[0051] VH-polarization data The incident angle θ, the root mean square height s, the correlation length l, the power exponent p, and the SSM, wherein s, l, and p are surface roughness variables. The parameter ranges used in the AIEM-Oh model simulation are shown in Table 1.

[0052]

[0053] Table 1

[0054] The simulation dataset generated based on the AIEM-Oh model is a dataset based on physical principles, covering the microwave scattering mechanism for subsequent model pre-training. The dataset is randomly divided into training (70%) and testing (30%) subsets to support the pre-training and verification of the subsequent model.

[0055] Referring to step S2, a DET model is constructed for soil moisture inversion in the embodiment. The DET model is as shown in Figure 2As shown, the input data, including SSM and 6 input variables, i.e., VV polarization data of soil surface VH polarization data Incident angle θ, root mean square height s, correlation length l, power exponent p.

[0056] Specifically, the core architecture of Transformer consists of self-attention mechanism and parallelized encoder-decoder structure. The encoder utilizes self-attention mechanism and feed-forward neural network to capture long-distance dependencies in input sequences by stacking multiple encoder and decoder layers. The decoder further adopts self-attention mechanism and encoder-decoder attention mechanism to realize sequence-to-sequence mapping. Initially, the input data will be projected to a high-dimensional space through a linear embedding layer, while introducing position encoding to preserve the position information of the input sequence to enable the self-attention mechanism to capture long-distance dependencies in the sequence. The self-attention mechanism simulates the dependencies between inputs by calculating the dot product attention between query (Q), key (K) and value (V) representations. The corresponding mathematical formula is as follows:

[0057]

[0058] where d k represents the dimension of the key (K); T represents matrix transpose. In addition, attention can be calculated in parallel by dividing the input data into multiple "heads". For the sake of simplicity, the single-head attention mechanism is used to calculate the attention value in this embodiment. The expression is as follows:

[0059] MultiHead(Q,K,V)=Concat(head)W O ;

[0060] head=Attention(QW Q ,KW K ,VW V );

[0061] where head represents the output of the attention head; W O represents the output weight matrix; W Q , W K , W V represent the weight matrix of query (Q), key (K) and value (V), respectively. The output of the attention mechanism and the input are added through a residual connection and use layer normalization to stabilize the training process, while the feed-forward neural network is used to further process and transform the features processed by the self-attention mechanism to enhance the expression ability of the model.

[0062] In a DNN architecture, network layers are divided into three types: input, hidden, and output, with full connection established through weight matrices. This means that each neuron in the i+1 layer is connected to every neuron in the i layer, and each neuron applies a nonlinear activation function before passing the input from the previous layer to the next. Through this hierarchical process, the network progressively extracts complex high-dimensional features, enhancing its representational capacity. The training process involves forward and backward propagation, enabling the network to learn data representations for prediction and performance optimization. The mathematical formulation of this process is as follows:

[0063] a m+1 =f m+1 (W m+1 a m +b m+1 ),m=0,1…,M-1;a 0 =p;a=a M ;

[0064] where a 0 represents the initial input matrix of the network, a m+1 represents the input matrix of the m+1 layer, a represents the output matrix of the network, W m+1 and b m+1 are the weight matrix and bias term of the m+1 layer, respectively, f m+1 represents the activation function of the corresponding layer, and M is the total number of layers in the network. When the network is initialized for training, the weights and bias terms are randomly generated to set the initial parameters.

[0065] The DET model of this embodiment fully utilizes the ability of the Transformer encoder to capture global feature interactions and the powerful nonlinear mapping ability of the DNN, and realizes the joint optimization of feature extraction and regression. The Transformer encoder automatically reveals global feature correlations, ranks feature importance, and captures dependencies between multiple source inputs, ultimately generating high-dimensional representations that can be accurately retrieved by the DNN. DET achieves higher computational efficiency by directly combining hierarchical feature learning and SSM estimation through an end-to-end architecture, while still maintaining high regression accuracy, making it particularly suitable for handling the inherent spatiotemporal complexity of SSM retrieval tasks.

[0066] After the DET model is constructed, it is pre-trained using the simulated data generated by the AIEM-Oh model, where the SSM is designated as the target variable of the DET model, and six input features are used, namely the soil surface VV polarization data VH polarization data incident angle θ, root mean square height s, correlation length l, and power exponent p, resulting in a pre-trained DET model.

[0067] Referring to step S3, the sample dataset includes in-situ SSM values. In this embodiment, the Nagqu River Basin (31-32°N, 91.5-92.5°E) in the Tibet Autonomous Region of China was used as the experimental area. In-situ SSM values ​​were obtained through the Tibetan Plateau Mesoscale Soil Temperature and Humidity Observation Network, which contains 57 stations recording soil moisture data at depths of 5cm, 10cm, 20cm, and 40cm at 30-minute intervals (publicly available from the National Tibetan Plateau Scientific Data Center). Given that the C-band penetration depth is usually less than 5cm, records at a depth of 5cm from 2019 to 2021 were extracted. Considering the impact of freeze-thaw cycles in the basin, only SSM data from the non-freezing period were used and processed by daily averaging. The station data were matched with a 10-meter grid of Sentinel-1 imagery, and the arithmetic mean within the multi-station grid was used to obtain representative in-situ SSM values.

[0068] Meanwhile, the sample dataset also includes preset predictor values. This is because, in addition to radar observations, SSM is also affected by various factors through surface water and energy balance, which interact to shape the spatiotemporal pattern of soil surface moisture. In this embodiment, the preset predictor values ​​include backscattering observation variables, spectral reflectance variables, soil texture variables, topographic index variables, and vegetation cover variables, wherein:

[0069] Backscattering observation variables include VV polarization data at the top of the canopy. VH polarization data at the top of the canopy angle of incidence θ at the top of the canopy can This embodiment uses backscattering observation data from Sentinel-1 in the C-band (5.405 GHz), including both VV and VH polarizations. Furthermore, due to its significant impact on backscattering, the incident angle was also considered. The Sentinel-1GRD dataset used in this study was obtained from the Google Earth Engine (GEE) platform. This data underwent standard processing, including thermal noise removal, radiometer calibration, and terrain correction. The spatial resolution of this data is 10 meters, and the temporal resolution is 5 to 7 days over the study area, achieved by combining ascent and descent data. This embodiment extracts Sentinel-1GRD pixels covering all monitoring stations from 2019 to 2021. To reduce the difference in incident angles between ascent and descent orbits, the SAR incident angle was corrected to the local incident angle adjusted for slope.

[0070]

[0071] Where, θ can θ represents the incident angle at the top of the canopy after slope correction. sardenotes the SAR incidence angle to be corrected. The slope angle (a) and the azimuth angle (b) are derived from the global 30-meter Digital Elevation Model (DEM), the SAR azimuth angle is extracted from the platform heading attribute embedded in the Sentinel-1 GRD image metadata.

[0072] The spectral reflectance variables include the visible bands 2, 3, 4, the near-infrared band 8, and the shortwave infrared bands 11, 12. The spectral reflectance variables are derived from the Sentinel-2 MSI Level-2A dataset, which is also available through the GEE platform. Since these bands all reflect vegetation information, this embodiment selected three 10-meter visible bands (bands 2, 3, and 4), one 10-meter near-infrared band (band 8), and two 20-meter shortwave infrared bands (bands 11 and 12) as the spectral reflectance variables for estimating the soil surface temperature (SSM), which has a high temporal resolution of 5 days. This embodiment extracted the Sentinel-2 pixels covering all sites during 2019-2021, while to ensure the temporal alignment with the Synthetic Aperture Radar (SAR) data, only those MSI pixels that have been cleared of cloud and snow cover and acquired within a range of 3 days from the corresponding SAR acquisition date were retained.

[0073] The soil texture variables include the percentage of sand, silt, and clay content. The soil texture variables come from the Chinese High-Resolution National Soil Information Grid dataset, which provides soil physical and chemical properties including six depth intervals (0-5 cm, 5-15 cm, 15-30 cm, 30-60 cm, 60-100 cm, and 100-200 cm) with a spatial resolution of 90 meters. This embodiment extracted the percentage of sand, silt, and clay content (by weight) at the 0-5 cm depth in the Nyangqu River Basin.

[0074] The terrain index variables include DEM data, slope, terrain surface convexity, terrain surface texture, and terrain ruggedness index. The terrain index variables are derived based on the DEM GLO-30 data product provided on the GEE platform. In this embodiment, four terrain features are extracted to represent the surface variation, including slope, terrain surface convexity (TSC), terrain surface texture (TST), and terrain ruggedness index (TRI). Specifically, the slope is calculated by the differential gradient algorithm to determine the maximum rate of elevation change in the plane coordinate system to describe the inclination of the ground element; the TSC is derived by the second derivative surface fitting method to quantify the concave-convex degree of the terrain surface relative to the locally optimized fitting plane; the TST reflects the frequency and intensity of micro-topographic changes by calculating the coefficient of variation of the gradient field; and the TRI is defined as the standard deviation of the elevation difference between the central pixel and its adjacent pixels, representing the complexity of the regional terrain. These parameters collectively depict the surface roughness features in multiple dimensions from terrain curvature to macro-relief undulations, thereby forming a set of features with practical significance.

[0075] The vegetation cover variables include vegetation types. The data of the vegetation cover variables is derived from the SinoLC-1 land cover product. SinoLC-1 is the first 1-meter resolution land cover thematic product in China, which is mainly generated by a cross-resolution deep transfer learning framework. The framework effectively transfers low-resolution labeled data to sub-meter image interpretation using a weakly supervised training mode, achieving large-scale and high-precision land cover classification. The vegetation types are mainly numerically encoded by OneHotEncoder.

[0076] All these predictor datasets from 2019 to 2021 are extracted based on the coordinates of 57 field observation points, and are resampled to a spatial resolution of 10 meters using bilinear interpolation to be consistent with the Sentinel-1 data. Among the 57 sites, 29 sites cannot generate valid prediction attributes due to cloud coverage or temporal and spatial inconsistency of the product, so the remaining 28 sites are successfully matched to form a sample dataset containing 3382 samples. Each sample includes the field SSM target value and the preset predictor values of this embodiment. 3025 samples selected from the 24 sites are used as the dependent scene, which are repeatedly randomly divided into a training set (70%) and a validation set (30%) for model training, hyperparameter adjustment, overfitting monitoring, and cross-validation. The remaining 4 sites (357 samples) are used as an independent scene test set to evaluate the generalization ability of the model.

[0077] The pre-trained DET model requires 6 input variables, three of which are roughness variables (s, l, p) and cannot be used directly for SSM estimation. AIEM-Oh is used for soil surface scattering modeling, and due to the complex interaction between the real observation values of VV and VH of Sentinel-1 and the potential vegetation canopy, combined with the isotropic assumption of soil surface roughness in the AIEM-Oh model, it may lead to poor soil moisture retrieval accuracy by only using the pre-trained DET model. Therefore, the application introduces a method of transfer learning, which overcomes the influence of the lack of vegetation scattering model and measured ground roughness input through feature transfer, and obtains an optimized model that combines physical mechanism and strong generalization performance through parameter fine-tuning

[0078] Referring to step S4, the prediction factors of each sample are subjected to feature transfer through a fully connected layer, as shown in Figure 3 The method of feature transfer includes overall transfer, independent transfer and cross transfer.

[0079] Overall transfer inputs the backscatter observation variable, spectral reflectance variable, soil texture variable, terrain index variable and vegetation coverage variable into a single fully connected layer to obtain the VV polarization data VH polarization data of the soil surface The incidence angle θ, root mean square height s, correlation length l and power exponent p are input into the pre-trained DET model together with the SSM under the corresponding sample to perform transfer learning. In this way, the vegetation scattering can be reduced to obtain the soil surface scattering component, and the influence of the soil surface roughness can be eliminated through the deep learning back propagation mechanism, thereby avoiding the challenges of obtaining soil roughness parameters and calibrating the vegetation scattering model.

[0080] Independent transfer sets the backscatter observation variable, spectral reflectance variable and vegetation coverage variable as a first group, aiming to reduce the influence of vegetation scattering to obtain the soil surface scattering component. The soil texture variable and terrain index variable are set as a second group, which is used to estimate the soil surface roughness parameter. After being input into two independent fully connected layers respectively, the first group obtains the VV polarization data VH polarization data of the soil surface The incidence angle θ, root mean square height s, correlation length l and power exponent p of the second group are obtained, and the data of the two groups are input into the pre-trained DET model together with the SSM under the corresponding sample to perform transfer learning. Eliminating the influence of vegetation scattering and soil surface roughness is a different process with different mechanisms. Therefore, setting two independent fully connected layers can learn these mechanisms respectively, thereby eliminating the influence of vegetation scattering and soil surface roughness respectively.

[0081] Cross-attention adds two cross-connections on the basis of the independent migration, that is, the output of the first group of fully connected layers is additionally taken as the input of the second group of fully connected layers, and the output of the second group of fully connected layers is additionally taken as the input of the first group of fully connected layers. The principle of cross-connection is that there are multiple interaction scattering effects between the soil interface and the vegetation, and the use of the cross-migration strategy helps to eliminate these interactions, so as to obtain a more reasonable soil surface scattering component and eliminate the influence of the soil surface roughness.

[0082] In the present embodiment, all the above-mentioned fully connected layers are configured with 128 neurons and adopt the ReLU activation function to facilitate gradient flow and improve convergence speed. During the training process of transfer learning, the Adam optimizer (learning rate = 0.001, ε = 1e-7) is used for optimization, the batch size is 32, the training is completed by minimizing the mean absolute error in 100 cycles, and the early stopping mechanism is implemented to prevent overfitting.

[0083] The present embodiment uses the AIEM-Oh model to generate simulation data for pre-training the DET model, but the AIEM-Oh model can reflect the scattering law between the microwave signal and the ground surface, but it itself has errors. Therefore, in order to retain the physical scattering knowledge described by the AIEM-Oh model and eliminate the errors of the AIEM-Oh model, the parameters of the pre-trained DET model need to be fine-tuned. Referring to step S5, the parameter fine-tuning method includes full unfreezing, encoder freezing, DNN freezing and full freezing. Among them:

[0084] Full unfreezing updates the parameters of the Transformer encoder and the DNN at the same time in the transfer learning process, which provides the greatest flexibility in SSM retrieval, but increases the risk of overfitting due to limited data; encoder freezing only updates the parameters of the DNN in the transfer learning process, so as to retain the pre-trained feature extraction in the encoder and enable efficient adaptation with low risk of overfitting; DNN freezing only updates the parameters of the Transformer encoder in the transfer learning process, at which time the fixed pre-training parameters of the DNN preferentially extract novel features from constantly changing data patterns; full freezing maintains the pre-trained parameters of the Transformer encoder and the DNN in the transfer learning process, which makes it suitable for simulated soil environments that are very similar to real soil conditions.

[0085] To verify the effectiveness of the method of the embodiment, four traditional machine learning methods are used in the embodiment for comparison, which are random forest (RF), extreme gradient boosting (XGBoost), deep neural network (DNN) and support vector regression (SVR), and the DET model without transfer learning is also used as a benchmark. All machine learning models are optimized by using different random seeds to perform repeated random partitioning of the training set and the validation set. For the machine learning model, the hyperparameters include RF (maximum depth = 20, estimator number = 100), SVR (kernel function = 'rbf') and XGBoost (estimator number = 100, learning rate = 0.001, maximum depth = 20). The DNN architecture consists of three fully connected layers containing 64 neurons, and early stopping (stop time = 10) is used to prevent overfitting. The non-migrated DET retains the same hyperparameter configuration and architecture strategy as the migrated version to ensure comparability.

[0086] In addition, the embodiment also selects three 1km resolution SSM datasets for cross comparison to further evaluate the method proposed in the embodiment. These datasets include the 1km resolution SSM product of China (SSM_China dataset), the 1km resolution daily SSM dataset of the Qinghai-Tibet Plateau (BTCH dataset) and the SSM dataset of the Qinghai-Tibet Plateau (TP_SM dataset). These datasets provide daily SSM prediction values at 1km resolution, covering a longer period of 2003-2023, 2001-2020 and 2003-2022 respectively. In addition, the embodiment also obtains daily precipitation data (in the form of daily total) from the ERA5-Land database from 2019 to 2021, and the daily total precipitation values are also given in the form of daily total. These data are mainly used to analyze the impact of precipitation events on SSM and evaluate whether the inversion model can capture the SSM changes caused by precipitation.

[0087] Correspondingly, the embodiment uses two methods for evaluation and verification: 1) model cross-validation in a dependent scenario; 2) model testing in an independent scenario to evaluate its generalization ability in unfamiliar geographical areas. Four commonly used statistical indicators are used to evaluate accuracy, including Pearson correlation coefficient (R), root mean square error (RMSE), bias and unbiased root mean square error (ubRMSE).

[0088] The specific experimental results are as follows:

[0089] 1. Effectiveness of different transfer strategies:

[0090] The optimal transfer strategy is determined by combining three feature transfer methods (overall transfer, independent transfer, and cross transfer) with four parameter fine-tuning methods (full thaw, encoder freezing, DNN freezing, and full freezing) in pairs, i.e., the combination of cross transfer and encoder freezing.

[0091] Figure 4 (a), (b), and (c) show the results of the three transfer methods using the full thaw method. The overall transfer strategy achieves an acceptable R value (0.854) and ubRMSE value (0.054 m 3 / m 3 ), which is better than the ensemble target value of radar inversion 0.06 m 3 / m 3 . For the range of 0.1 to 0.35 m 3 / m 3 , the SSM is well captured. However, a large underestimation and overestimation phenomenon is observed at high and low values, respectively. The inversion accuracy is improved using the independent transfer strategy, with R and RMSE values of 0.861 and 0.053 m 3 / m 3 , respectively. However, a large overestimation phenomenon is still observed at low values (<0.1 m 3 / m 3 ). This problem is partially solved by using the cross transfer strategy, with R and RMSE values of 0.887 and 0.049 m 3 / m 3 , respectively. Similar results are achieved using other fine-tuning methods, with the cross transfer strategy being the most effective, followed by the independent transfer and overall transfer strategies. Therefore, the cross transfer strategy is subsequently selected as the optimal feature transfer method.

[0092] Similarly, Figure 4 (a), (d), (g), and (j) show the results of the four parameter fine-tuning methods under the overall transfer strategy. The encoder freezing achieves the best estimation accuracy, with an R value of 0.858 and an ubRMSE of 0.054 m 3 / m 3 ; the full thaw model shows a slight decrease in performance compared to the encoder freezing, with an R value decrease of 0.01 and an ubRMSE that remains essentially unchanged; the DNN freezing method shows a quantifiable difference compared to the encoder freezing, with an R value of 0.03 and an ubRMSE of 0.004 m 3 / m 3 ; the full freezing method performs significantly worse, with an R decrease of 0.113 and an ubRMSE increase of 0.015 m 3 / m 3Under the independent strategy, this ranking remained unchanged. The fully unfrozen and DNN-frozen methods showed only slight differences in metrics compared to encoder freezing, with R values ​​0.01 to 0.02 lower and ubRMSE values ​​0.001 to 0.005m higher, respectively. 3 / m 3 The performance degradation of the fully frozen method was more significant, with R decreasing by 0.075 and ubRMSE increasing by 0.011m. 3 / m 3 In cross-transfer strategies, encoder freezing still performs better. Both the full-unfreeze method and the DNN-freeze method show a slight performance decrease compared to the baseline, specifically a decrease in R-value of 0.01 to 0.04, while ubRMSE increases of 0.001 to 0.007m. 3 / m 3 The fully frozen method showed a significant difference, with its R-value decreasing by 0.05 and its ubRMSE increasing by 0.01m. 3 / m 3 .

[0093] Overall, the cross-transfer method with encoder freezing achieved the highest accuracy, while the total transfer method with full freezing had the lowest accuracy.

[0094] 2. Performance of the optimal migration strategy:

[0095] This embodiment compares the performance of the optimal migration strategy with other methods in both dependent and independent scenarios. Figure 5 As shown, in dependency scenarios, the optimal migration strategy outperforms other retrieval methods, achieving an R-value of 0.891 and an ubRMSE of 0.048m. 3 / m 3 The integrated methods (RF and XGBoost) performed well, with R-values ​​ranging from 0.832 to 0.841 and ubRMSE ranging from 0.056 to 0.058m. 3 / m 3 The results of directly using unadjusted deep learning methods (DET and DNN) are slightly inferior to ensemble methods, with R values ​​ranging from 0.805 to 0.822 and ubRMSE ranging from 0.059 to 0.063m. 3 / m 3 Between these two, DET's R-value is 0.017 higher than DNN's, while its ubRMSE is 0.004m lower. 3 / m 3 In contrast, SVR performed the worst, with an R-value of 0.726 and an ubRMSE of 0.072m. 3 / m 3 The main range of SSM in this area is between 0.1 and 0.35 m. 3 / m3 Between 0.1 and 0.35 m 3 / m 3 , the optimal transfer strategy performed better in terms of sample density and retrieved SSM more effectively. In the low SSM region (<0.1 m 3 / m 3 ), the sample density of all models was low, and SSM was slightly overestimated. In the high SSM region (>0.35 m 3 / m 3 ), the sample density of all models was low, and SSM was slightly underestimated.

[0096] Figure 6 (a)-(f) The optimal transfer model had higher sample density and better estimation accuracy than other methods, with an R value of 0.865 and an ubRMSE of 0.050 m 3 / m 3 . As expected, the results in the independent scenario performed slightly worse, and the performance of all methods decreased similarly. Consistent with the results in the dependent scenario, the R values of the two ensemble methods were 0.832-0.841, and the ubRMSE was 0.056-0.058 m 3 / m 3 , although they were better than the two deep learning methods (R from 0.805 to 0.822, and ubRMSE from 0.059 to 0.063 m 3 / m 3 ), but still not as good as the optimal transfer model. SVR performed the worst (R = 0.674, and ubRMSE = 0.071 m 3 / m 3 ). Sensitivity evaluation in different scenarios showed that the optimal transfer strategy had minimal performance differences between different scenarios: the R value decreased by 0.026, and the ubRMSE increased by 0.002 m 3 / m 3 in the independent scenario. The ensemble methods, deep learning, and SVR performed similarly, with a decrease in R value of 0.005-0.02 and an increase in ubRMSE of about 0.002 m 3 / m 3 .

[0097] The SSM time series estimated by the optimal model were compared with those obtained using other methods and field measurements (g-j). To make the results clearer, this study only selected the better-performing RF and DET (h-i) for comparison, and included daily precipitation data from ERA5-Land to explore the correlation between SSM estimates and precipitation. The results showed that the optimal transfer strategy tracked the temporal variation of SSM more accurately than RF and DET in response to field values and precipitation events. Figure 6 Figure 5 Figure 6

[0098] 3. Comparative tests between various SSM products:

[0099] Figure 7 (a)-(d) show that the optimal migration model outperforms SSM_China in capturing the field SSM changes at the four independent scenario sites. It also shows comparable accuracy compared to TP_SM and BTCH, all of which have a bias of less than 0.05 m 3 / m 3 to the field SSM. The optimal strategy has similar bias to the three products: at MS3523 and MS3545 sites, the SSM curve overestimates the field SSM by about 0.05 m 3 / m 3 , while at C2 site, the SSM curve underestimates the field SSM by about 0.02 m 3 / m 3 . At MS3633 site, the optimal migration model agrees very well with the field SSM, but the other three products show more pronounced overestimation or underestimation.

[0100] Figure 8 (e)-(f) show the evaluation results at the four sites under independent scenarios: the results show that the optimal migration model significantly outperforms the other products, with an average R value of 0.81 and an average ubRMSE of 0.039 m 3 / m 3 . In comparison, the performance of the reference products is in the following order: BTCH > TP_SM > SSM_China, with average R values of 0.790, 0.760 and 0.730, and average ubRMSE of 0.041 m 3 / m 3 , 0.043 m 3 / m 3 and 0.062 m 3 / m 3 , respectively. In addition, the error bars in the box plot of the BTCH dataset are shorter than those of the other two products, which indicates that the BTCH dataset provides more stable SSM estimates under independent scenarios. Overall, under independent scenarios at the four sites, the optimal migration model effectively captures the temporal changes of the field SSM, with better accuracy than TP_SM, BTCH and SSM_China.

[0101] Meanwhile, in order to explore the application prospect of the optimal migration model, the embodiment generates a SSM spatial distribution map of the Nagqu River Basin. Figure 8(a)-(c) show the 10 m median SSM maps for April (spring), July (summer), and October (autumn) 2019, generated by the optimal migration strategy and compared to the SSM_China, BTCH, and TP_SM products (SSM_China is Figure 8 d-f, BTCH is Figure 8 g-i, and TP_SM is Figure 8 j-l).

[0102] It can be seen that similar to the in-situ SSM trends, the optimal migration strategy captures the seasonal variations: the spring SSM increases from 0.1 m 3 / m 3 to 0.25 m 3 / m 3 , the summer from 0.25 m 3 / m 3 to 0.45 m 3 / m 3 , and then the autumn decreases to 0.1 m 3 / m 3 to 0.25 m 3 / m 3 . SSM_China, BTCH, and TP_SM also capture this trend, but when comparing the similarity of SSM values in these products, the optimal migration model has a range of SSM values in all seasons that is closer to BTCH, while SSM_China and TP_SM significantly underestimate the spring and summer SSMs. Moreover, considering that the in-situ SSM values at the four sites are 0.15 to 0.2 m 3 / m 3 in April and October and 0.25 to 0.35 m 3 / m 3 ( Figure 8 a-d), the optimal migration model estimates agree well with the in-situ SSM at these sites in different seasons, while the other three products show seasonal discrepancies. Specifically, SSM_China underestimates by about 0.1 m 3 / m 3 in spring, by about 0.05 m 3 / m 3 in autumn, and similarly in summer; TP_SM underestimates by about 0.05 m 3 / m 3 in spring, and similarly in summer and autumn; and BTCH shows similar patterns in all three seasons.

[0103] Overall, the outstanding performance of the optimal transfer strategy demonstrates the importance of high-resolution data for capturing SSM changes in areas with strong spatial heterogeneity. Due to the complex terrain and variable climate in the Zhongdu River Basin, low-resolution SSM products are difficult to effectively function. In contrast, the 10-meter resolution map generated using the optimal transfer learning method is superior to the 1-kilometer resolution product in terms of spatial resolution, time series analysis, and accuracy.

[0104] The present application is described above with reference to the drawings. Obviously, the specific implementation of the present application is not limited to the above-described manner. As long as various non-essential improvements are made using the inventive concept and technical solutions of the present application; or without improvement, the above-mentioned concept and technical solutions of the present application are directly applied to other occasions, which are within the protection scope of the present application.

Claims

1. A soil moisture inversion method based on scattering model deep transfer learning, characterized in that: The method comprises the following steps: Step S1, generating a simulation data set based on an AIEM-Oh model; Step S2, constructing a DET model; pre-training the DET model using the simulation data set; Step S3, obtaining a sample data set, wherein each sample comprises an in-situ SSM target value and preset predictor values; Step S4, performing feature migration on the predictor of each sample, and combining the in-situ SSM target value under the corresponding sample to perform transfer learning on the pre-trained DET model; Step S5, during the transfer learning process, performing parameter fine-tuning on the pre-trained DET model; and finally obtaining a DET model after transfer learning for soil moisture inversion.

2. The soil moisture inversion method of claim 1, wherein: In the step S1, the input of the AIEM-Oh model includes the incidence angle θ, the root mean square height s, the correlation length l, the power exponent p and the SSM of the soil surface, and the generated simulation data includes the VV polarization data of the soil surface VH polarization data The incidence angle θ, the root mean square height s, the correlation length l, the power exponent p and the SSM.

3. The soil moisture inversion method of claim 2, wherein: In the step S2, the DET model comprises a Transformer encoder and a DNN model, input data are sequentially subjected to the Transformer encoder and the DNN model, and predicted SSM is output, wherein the input data comprise soil surface VV polarization data soil surface VH polarization data soil surface incidence angle θ, root mean square height s, correlation length l, power exponent p, and SSM.

4. The soil moisture inversion method of claim 3, wherein: In the step S3, the preset predictors comprise a backscattering observation variable, a spectral reflectance variable, a soil texture variable, a terrain index variable, and a vegetation coverage variable, wherein: Backscatter observation variables include canopy top VV polarization data Canopy top VH polarization data Canopy top incident angle theta can ; The spectral reflectance variable comprises visible light bands 2, 3, and 4, a near-infrared band 8, and short-wave infrared bands 11 and 12; The soil texture variable comprises percentages of sand, silt, and clay content; The terrain index variable comprises DEM data, slope, terrain surface convexity, terrain surface texture, and terrain ruggedness index; The vegetation coverage variable comprises vegetation types.

5. The soil moisture inversion method of claim 4, wherein: In the step S4, the predictors of each sample are subjected to feature migration through a fully connected layer, and the feature migration method comprises overall migration, independent migration, and cross migration.

6. The soil moisture inversion method of claim 5, wherein: The backscatter observation variable, the spectral reflectance variable, the soil texture variable, the topographic index variable, and the vegetation cover variable are input into a single fully connected layer to obtain VV polarization data of the soil surface VH polarization data The incident angle θ, the root mean square height s, the correlation length l, and the power exponent p are input into the pre-trained DET model together with the SSM under the corresponding sample to perform transfer learning. The independent migration sets the backscattering observation variable, the spectral reflectance variable, and the vegetation coverage variable as a first group, sets the soil texture variable and the terrain index variable as a second group, and respectively inputs the two groups to two independent fully connected layers. The first group obtains the VV polarization data of the soil surface The VH polarization data The incident angle θ, the second group obtains the root mean square height s of the soil surface, the correlation length l, and the power exponent p. The two groups of data are input into the pre-trained DET model together with the SSM under the corresponding sample to perform the transfer learning. The cross migration adds two cross connections to the independent migration, that is, the output of the first group of fully connected layers is additionally used as the input of the second group of fully connected layers, and similarly, the output of the second group of fully connected layers is additionally used as the input of the first group of fully connected layers.

7. The soil moisture inversion method of claim 6, wherein: The fully connected layer is configured with 128 neurons and adopts a ReLU activation function.

8. The scattering model deep transfer learning method for soil moisture inversion according to any one of claims 5-7, characterized in that: In the step S5, the parameter fine-tuning method comprises full unfreezing, encoder freezing, DNN freezing, and full freezing.

9. The soil moisture inversion method of claim 8, wherein: The full unfreezing simultaneously updates the parameters of the Transformer encoder and the DNN during the transfer learning process; the encoder freezing only updates the parameters of the DNN during the transfer learning process; the DNN freezing only updates the parameters of the Transformer encoder during the transfer learning process; and the full freezing maintains the parameters of the pre-trained Transformer encoder and DNN during the transfer learning process.

10. The soil moisture inversion method of claim 9, wherein: The method adopts the combination of cross migration and encoder freezing as the optimal transfer learning scheme.

Images