Method for retrieving farmland soil moisture content based on multi-satellite dual-frequency GNSS signal intensity

Abstract

The present application relates to the global navigation satellite system remote sensing technical field, disclose a kind of farmland soil moisture content inversion method based on multi-satellite dual-frequency GNSS signal intensity, including obtaining the frequency point observation data satisfying preset condition, extract original signal-to-noise ratio parameter and convert into signal intensity sequence, separate out multipath interference signal sequence;Least square method is used to extract single-frequency delay phase, calculate error index and calculate adaptive weight accordingly, obtain dual-frequency fusion delay phase by weighted summation;Calculate normalized microwave reflectance index, with dual-frequency fusion delay phase to form multidimensional input feature vector;Establish nonlinear inversion prediction model, adopt non-stationary random bayesian optimization algorithm to optimize hyperparameter, input multidimensional input feature vector and output soil moisture content prediction value.By dynamically allocating multi-frequency weight and compensating vegetation attenuation, combined with optimization algorithm, the generalization ability and inversion precision of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology for global navigation satellite systems, specifically to a method for inverting farmland soil moisture content based on the intensity of multi-satellite dual-frequency GNSS signals. Background Technology

[0002] Soil moisture content is fundamental data for precision irrigation and drought monitoring in farmland. Global Navigation Satellite System (GNSS) interferometric reflectance measurement technology analyzes the multipath reflectance components of the Earth's surface in the signal-to-noise ratio data recorded by the receiver, enabling non-contact detection of surrounding surface environmental parameters. This technology is currently being applied to soil moisture content retrieval studies.

[0003] Existing inversion methods typically extract the delayed phase of the interferometric signal and establish a mapping model between it and the measured soil moisture content. However, conventional methods often rely on a single-frequency GNSS signal for inversion calculations. Since single-frequency signals are weakly resistant to environmental noise, they are prone to data fluctuations in complex observation environments, leading to insufficient stability of the inversion results. Furthermore, farmland surfaces are often accompanied by crop growth, and the vegetation canopy can obstruct and attenuate microwave signals reflected from the surface. Existing inversion methods lack quantitative compensation features for vegetation interference, making it impossible for the input single-phase parameter to accurately reflect changes in the dielectric constant of the underlying soil, resulting in large inversion errors.

[0004] Furthermore, when constructing inversion prediction models, traditional linear empirical models have limited fitting effects due to the complex nonlinear relationship between multi-source feature variables and actual soil moisture content. In addition, conventional nonlinear machine learning methods rely heavily on human experience in setting model hyperparameters and lack a global optimization mechanism that combines the model's internal structural information, making the algorithm prone to getting trapped in local optima and unable to maintain stable generalization ability and prediction accuracy on different observation datasets. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for inverting farmland soil moisture content based on the intensity of multi-satellite dual-frequency GNSS signals. This method solves the problems in existing GNSS interferometric reflection measurement techniques, such as the susceptibility of single-frequency signals to environmental noise interference leading to large fluctuations in inversion accuracy, the difficulty in effectively compensating for signal attenuation caused by surface vegetation cover, and the difficulty of conventional prediction models accurately fitting the complex nonlinear mapping relationship between multidimensional features and soil moisture content.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength, comprising: Obtain observation and navigation files from the Global Navigation Satellite System receiver, filter frequency observation data that meet the preset satellite elevation and azimuth range conditions, extract the raw signal-to-noise ratio parameters and convert them into signal strength time series; Low-pass filtering is applied to the signal intensity time series at different frequencies, the independent variable is converted into the sine value of the satellite elevation angle, and the slow variation trend of the direct signal is eliminated by polynomial fitting, thus separating the multipath interference signal sequence. Spectral analysis is performed on multipath interference signal sequences to identify the dominant frequency of the interference signal, and the single-frequency delay phase is extracted by fitting using the least squares method; Calculate the error index of the single-frequency delay phase in the preliminary estimation, and calculate the adaptive weight allocation value of each frequency data source accordingly. Then, sum the adaptive weight allocation value with the single-frequency delay phase to obtain the dual-frequency fused delay phase. Calculate the normalized microwave reflection index, and construct a multi-dimensional input feature vector by combining the normalized microwave reflection index with the dual-frequency fused delay phase; A nonlinear inversion prediction model incorporating a decision tree output accumulation mechanism is established. The hyperparameters are optimized using a nonstationary stochastic Bayesian optimization algorithm. The multidimensional input feature vector is input into the nonlinear inversion prediction model to output the predicted value of farmland soil moisture content.

[0007] Furthermore, the original signal-to-noise ratio (SNR) parameter is converted into a signal strength time series, including: dividing the original SNR parameter by decibels and performing a floor operation by dividing the step size; performing a maximum value selection operation on the floor result and the minimum allowable level threshold of the signal quality defined by the system; performing a minimum value selection operation on the maximum value selection result and the maximum record threshold level of the signal quality defined by the system to obtain a standardized signal strength parameter; and arranging the discretely distributed standardized signal strength parameters at each epoch in time sequence to generate a continuous signal strength time series.

[0008] Furthermore, the single-frequency delay phase is extracted using the least squares method, including: constructing a physical model of surface reflection interference based on the complex superposition of the direct and reflected signals, deriving a receiver output power expression that includes the direct signal, reflected signal, signal carrier wavelength, and delay phase; converting this receiver output power expression into a parameterized mathematical form that includes amplitude, bias term, interference signal main frequency, and delay phase, and substituting the identified interference signal main frequency into the parameterized mathematical form; applying the least squares method to perform nonlinear numerical fitting on the multipath interference signal sequence, and solving for the optimal amplitude, optimal bias term, and optimal delay phase with the minimum sum of squared residuals through iterative calculation, and using the optimal delay phase as the single-frequency delay phase.

[0009] Furthermore, the error index of single-frequency delay phase in the preliminary estimation is calculated, including: constructing a linear regression reference model independently for each frequency point based on historical measured soil moisture content samples and historical single-frequency delay phase; using test set data to input the single-frequency delay phase into the linear regression reference model to obtain a preliminary estimate; and calculating the mean square error between the preliminary estimate and the measured value as the error index.

[0010] Based on this, the adaptive weight allocation value of each frequency point data source is calculated, including: using an exponential mapping mechanism to convert the error index into a positive weight coefficient; using the negative constant multiple of the natural constant to the power of the error index as the numerator, the sum of the negative constant multiples of the error index corresponding to all frequency point data sources as the denominator, and the ratio of the numerator to the denominator as the adaptive weight allocation value.

[0011] Further, the normalized microwave reflection index is calculated, including: extracting the maximum and minimum amplitude values ​​of the multipath interference signal sequence within a set time observation window; calculating the difference between the maximum amplitude value and the optimal amplitude as the first difference value; calculating the difference between the maximum amplitude value and the minimum amplitude value as the second difference value; and using the ratio of the first difference value to the second difference value as the normalized microwave reflection index.

[0012] The dual-frequency fusion delayed phase is combined with the two-frequency fusion delayed phase to form a multi-dimensional input feature vector, including: setting the dual-frequency fusion delayed phase, which reflects the fluctuation of soil dielectric constant, as the dominant feature component, and setting the normalized microwave reflectance index, which reflects the vegetation shading effect, as the modified feature component; combining and splicing the dominant feature component and the modified feature component to form a two-dimensional input feature vector, and using the two-dimensional input feature vector as the multi-dimensional input feature vector.

[0013] Furthermore, before establishing a nonlinear inversion prediction model that includes a decision tree output accumulation mechanism, the process also includes: extracting historical multidimensional input feature vectors from observation files recorded within the historical observation period, obtaining synchronously measured in-situ soil moisture content values; establishing the historical multidimensional input feature vectors as the underlying input data, and using the in-situ soil moisture content values ​​as the physical state label data at the corresponding time to construct a basic dataset; dividing the basic dataset into training and testing sets; and constructing a nonlinear inversion prediction model that supports an extreme gradient boosting architecture.

[0014] Furthermore, the multidimensional input feature vector is input into the nonlinear inversion prediction model to output the farmland soil moisture content inversion prediction value, including: performing node splitting calculation on the multidimensional input feature vector by each regression decision tree according to the internal feature space partitioning rules to obtain the forward prediction output; accumulating the forward prediction outputs of all regression decision trees to obtain the overall estimate value, and using the overall estimate value as the farmland soil moisture content inversion prediction value.

[0015] Furthermore, a non-stationary stochastic Bayesian optimization algorithm is used to optimize hyperparameters, including: using the second derivative information of the nonlinear inversion prediction model to perform global optimization in the three-dimensional parameter space; using the mean square error between the predicted values ​​calculated from the training set data and the physical state label data as the loss function; driving the non-stationary stochastic Bayesian optimization algorithm to search for the optimal combination of hyperparameters that minimizes the loss function; and iteratively training and updating the nonlinear inversion prediction model based on the optimal hyperparameter combination.

[0016] This invention provides a method for inverting farmland soil moisture content based on the intensity of multi-satellite dual-frequency GNSS signals. It has the following beneficial effects: 1. This invention calculates the preliminary estimation error index of the single-frequency delay phase and uses an exponential mapping mechanism to calculate the adaptive weight allocation value of each frequency point. Then, it performs a weighted summation of data from different frequency points to obtain the dual-frequency fused delay phase. This technical feature dynamically allocates weights according to the actual quality of signal observation data at different frequencies, reducing the interference of frequency points with large errors on the overall inversion results. It avoids the problem of inversion accuracy fluctuations caused by environmental noise affecting a single frequency point, thus improving the stability and accuracy of the delay phase parameter.

[0017] 2. This invention calculates the normalized microwave reflectance index by extracting the extreme values ​​of a multipath interference signal sequence and subtracting them from the fitted optimal amplitude. This normalized reflectance index is then used as a corrected feature component and combined with the dual-frequency fusion delay phase to form a multi-dimensional input feature vector. This technique, while utilizing the delay phase to reflect the soil dielectric constant, introduces the normalized microwave reflectance index to compensate for the attenuation effect of surface vegetation cover on microwave signals. This allows the feature data input to the prediction model to more objectively reflect the complex actual surface physical state and reduces model inversion errors caused by vegetation interference.

[0018] 3. This invention establishes a nonlinear inversion prediction model incorporating a decision tree output accumulation mechanism, and utilizes a non-stationary stochastic Bayesian optimization algorithm combined with the model's second derivative information for global hyperparameter optimization. This technique can fit the complex nonlinear mapping relationship between multidimensional GNSS feature parameters and actual soil moisture content. Furthermore, relying on a loss function-driven optimal hyperparameter combination iterative update mechanism, it improves the model's convergence process under multivariate input conditions, enhancing the model's generalization ability to unknown data and the accuracy of the final output prediction. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to an embodiment of the present invention; Figure 2 This is a block diagram of the farmland soil moisture content inversion system based on multi-satellite dual-frequency GNSS signal strength according to an embodiment of the present invention. Figure 3 This is a graph showing the linear relationship between the delayed phase and soil moisture content in an embodiment of the present invention. Figure 4 This is a linear regression residual diagram of soil moisture content according to an embodiment of the present invention; Figure 5 This is a graph showing the distribution and correlation analysis of soil moisture content and related variables in an embodiment of the present invention. Figure 6 This is a linear fitting diagram of soil moisture content inversion according to an embodiment of the present invention; Figure 7 This is a comparison chart of the predicted values ​​of each model and the measured values ​​of soil moisture content in the embodiments of the present invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1 This invention provides a method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength, comprising the following steps: S100: Acquire observation files and navigation files from the global navigation satellite system receiver, filter frequency observation data that meet the preset satellite elevation angle range and azimuth angle range conditions, extract signal strength parameters from the observation data, and use a quantization function to convert the carrier noise ratio into a signal strength time series. S200 performs low-pass filtering on the signal strength time series at different frequency points to remove high-frequency noise, converts the independent variable of the time series into the sine value of the satellite elevation angle, removes the slow variation trend of the direct signal through polynomial fitting to separate the multipath interference signal sequence, constructs a surface reflection interference physical model containing direct and reflected signals, performs spectral analysis on the multipath interference signal sequence to identify the signal dominance frequency, uses the least squares method to fit the parameters of the output power expression of the interference physical model, and extracts the delay phase corresponding to each frequency point; S300 calculates the error index of the delay phase data in the preliminary estimation, dynamically calculates the adaptive weight allocation value of different frequency data sources based on the error index, and performs weighted summation calculation with the corresponding single-frequency delay phase to obtain the dual-frequency fused delay phase. S400 calculates the normalized microwave reflectance index, which characterizes the vegetation moisture content parameter. The normalized microwave reflectance index is used as a supplementary variable and together with the dual-frequency fusion delay phase, it forms a multi-dimensional input feature vector. S500 establishes a nonlinear inversion prediction model that includes a decision tree output accumulation mechanism. It uses a non-stationary stochastic Bayesian optimization algorithm to optimize the high-dimensional hyperparameter configuration and inputs the multidimensional input feature vector into the nonlinear inversion prediction model to output the farmland soil moisture content inversion prediction value.

[0022] See attached document Figure 2 This invention provides a system for retrieving farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength, comprising: The data acquisition module is used to acquire observation files and navigation files from the global navigation satellite system receiver, filter frequency observation data that meet the preset satellite elevation angle range and azimuth angle range conditions, extract signal strength parameters from the observation data, and use a quantization function to convert the carrier noise ratio into a signal strength time series. The phase extraction module is used to perform low-pass filtering on the signal strength time series at different frequency points to filter out high-frequency noise, convert the independent variable of the time series into the sine value of the satellite elevation angle, remove the slow variation trend of the direct signal through polynomial fitting to separate the multipath interference signal sequence, construct a surface reflection interference physical model containing direct and reflected signals, perform spectrum analysis on the multipath interference signal sequence to identify the signal dominance frequency, use the least squares method to fit the parameters of the output power expression of the interference physical model, and extract the delay phase corresponding to each frequency point. The data fusion module is used to calculate the error index of the delay phase data in the preliminary estimation, dynamically calculate the adaptive weight allocation value of different frequency data sources based on the error index, and perform weighted summation calculation with the corresponding single-frequency delay phase to obtain the dual-frequency fused delay phase. The feature building module is used to calculate the normalized microwave reflectance index, which characterizes the vegetation moisture content parameter. The normalized microwave reflectance index is used as a supplementary variable and together with the dual-frequency fusion delay phase, it is used to build a multi-dimensional input feature vector. The model inversion module is used to establish a nonlinear inversion prediction model that includes a decision tree output accumulation mechanism. It uses a nonstationary stochastic Bayesian optimization algorithm to optimize the high-dimensional hyperparameter configuration and inputs the multidimensional input feature vector into the nonlinear inversion prediction model to output the farmland soil moisture content inversion prediction value.

[0023] To further clarify the execution logic of the technical solution of this invention, the following describes in detail the process of obtaining observation files and navigation files from the global navigation satellite system receiver and constructing a time series, in conjunction with each data processing step.

[0024] S101, Obtain observation and navigation files from the Global Navigation Satellite System receiver.

[0025] In this embodiment, the Global Navigation Satellite System (GNSS) receiver generates basic record files during the data acquisition phase. These files primarily include observation files recording measurements such as carrier phase, pseudorange, and signal strength, as well as navigation files recording core operational statuses such as satellite ephemeris and satellite clock bias. By reading the observation and navigation files, satellite elevation angle parameters, satellite azimuth angle parameters, and raw signal-to-noise ratio parameters reflecting signal quality are extracted for each observation epoch. For the parsing and information extraction of GNSS standard data files, those skilled in the art can employ receiver-independent exchange format data parsing methods. The underlying file parsing rules are well-known in the field and will not be elaborated upon here.

[0026] S102, based on satellite elevation angle parameters and satellite azimuth angle parameters, selects specific frequency observation data that meet the preset satellite elevation angle range and azimuth angle range conditions.

[0027] A preset satellite elevation angle range is used to constrain the effective extraction space where multipath effects are strong. When the receiving antenna is at a low elevation angle, the satellite signal travels a longer path through the atmosphere, increasing its susceptibility to ground multipath reflection interference and allowing it to carry a certain degree of surface reflection physical characteristics. As a preferred approach, the preset satellite elevation angle range is set to 5° to 30°. A preset azimuth angle range is used to spatially define the effective reflective surface of the target farmland monitoring area, aiming to eliminate reflection interference from the surrounding non-target environment. In this embodiment, the preset azimuth angle range is set to 180° to 230°. While simultaneously satisfying the above-mentioned elevation angle spatial constraints and azimuth angle spatial constraints, observation data at specific frequencies in the corresponding epoch are retained. These specific frequencies are specifically configured as the L1 and L2 frequencies of the Global Navigation Satellite System.

[0028] S103 extracts the raw signal-to-noise ratio parameter from the observation data at a specific frequency point and uses a quantization function to convert it into a standardized signal strength time series.

[0029] Phase separation of multipath signals depends on the signal quality values ​​output by the equipment. In the output protocols of some stations, the observation files do not provide continuous floating-point signal-to-noise ratio (SNR) values, but only record them as discrete signal quality indicators. To adapt to the data structure of this type of station equipment, a quantization mapping function is constructed by setting specific boundary constraints and magnitude division rules. Based on the receiver independent switching format standard definition, continuous quantities with physical dimensions are converted into dimensionless scale values ​​conforming to the system transmission protocol. The calculation formula for the quantization function is as follows: ; In the formula, This represents the normalized signal strength parameter output after quantization mapping; This represents the raw signal-to-noise ratio parameter extracted from the data records of the observation file, and its unit of measurement is dBHz; This indicates the function operation of rounding down; This represents the function that retrieves the maximum value in a set. This represents the function that takes the minimum value in a set.

[0030] Mathematical operations are performed based on the quantization function, with the constant term 6 serving as the decibel step size. The original signal-to-noise ratio parameter is discretized and mapped to an integer range of 1 to 9. A numerical result of 1 represents the minimum permissible signal quality threshold defined by the system, while a numerical result of 9 represents the highest recorded signal quality threshold defined by the system. After numerical quantization, the standardized signal strength parameters discretely distributed at each epoch are arranged in time sequence to generate a continuous signal strength time series corresponding to each frequency point, which serves as the basic sequence data input for subsequent interferometric signal feature extraction.

[0031] Based on this, the execution details of the steps for multipath interference physical modeling and delayed phase extraction are explained.

[0032] S201 performs low-pass filtering on the signal strength time series at different frequency points to remove high-frequency noise and converts the independent variable of the time series into the sine value of the satellite elevation angle.

[0033] In this embodiment, high-frequency interference such as equipment thermal noise is often mixed in during the recording of signal strength by the global navigation satellite system receiver. In order to accurately extract the surface reflection characteristics, a low-pass filter is used to smooth the constructed signal strength time series to retain the low-frequency oscillation characteristics caused by multipath effects. After the filtering operation is completed, in order to establish the physical mapping relationship between signal fluctuations and satellite spatial geometric position, the original time independent variable corresponding to the time series is converted into the sine value of the satellite elevation angle parameter, forming a new series with the elevation angle sine value as the horizontal axis.

[0034] S202 uses polynomial fitting to eliminate the slow variation trend of direct signals and separates the multipath interference signal sequence caused by surface reflection.

[0035] The total signal energy acquired by the receiving antenna is the superposition of the direct signal energy from the satellite and the reflected signal energy after reflection from the Earth's surface. The intensity of the direct signal typically exhibits a smooth and slow variation with the satellite elevation angle parameter, forming a slow-varying trend background in the sequence. As a preferred approach, a low-order polynomial function is used to numerically fit the signal intensity sequence after the independent variable is transformed, calculating the slow-varying trend term reflecting the characteristics of the direct signal. Subtracting this slow-varying trend term from the original sequence strips away the energy distribution of the direct signal, and the remaining residual sequence is the multipath interference signal sequence caused by Earth's surface reflection.

[0036] S203. Construct a physical model of surface reflection interference that includes direct and reflected signals, and derive the expression for receiver output power.

[0037] During the propagation of signals in a global navigation satellite system, the antenna simultaneously receives both direct signals and signals reflected from environmental reflective surfaces. The interferometric signal can be represented as a complex superposition of these two physical signals, and its physical model formula for surface reflection interferometry is as follows: ; In the formula, This represents the superimposed interference signal; Indicates a direct signal; Indicates the reflected signal; Indicates the surface reflectance coefficient; Represents the imaginary unit; This represents the signal delay phase difference caused by different geometric paths.

[0038] The strengths of the three characteristic parameters mentioned above change with variations in the satellite elevation angle parameter. Based on the physical model of surface reflection interferometry, since the signal strength parameter used by the receiver is closely related to the received signal power, the expression for the receiver output power is derived as follows: ; In the formula, Indicates the receiver output power; This represents the superimposed interference signal that varies with the satellite elevation angle parameter; Indicates the satellite elevation angle parameter; A direct signal representing the variation of satellite elevation angle parameters; This represents the reflected signal that enters the antenna from the ground. Indicates the wavelength of the signal carrier; This indicates the known antenna height above the ground; This represents the delayed phase. The delayed phase is independent of the satellite elevation angle parameter, but it is modulated by changes in the surface dielectric constant within the target monitoring area, and thus has a close physical correlation with fluctuations in soil moisture content.

[0039] S204 performs spectral analysis on the multipath interference signal sequence to identify the signal's dominant frequency, uses the least squares method to fit the parameters of the output power expression, and extracts the delay phase of the multipath reflected signal corresponding to each frequency point.

[0040] To facilitate parameter calculation by the computer, the above receiver output power expression is converted into the following parameterized mathematical form: ; In the formula, Indicates amplitude, and ; This represents the sine value of the elevation angle, and ; Indicates the dominant frequency of the interference signal, and ; Indicates the bias term, and .

[0041] For the acquired multipath interferometric signal sequence, the dominant frequency with the most concentrated energy in the sequence is extracted using spectral analysis as the dominant frequency of the interferometric signal. For the identification and extraction of the dominant frequency of the interferometric signal, those skilled in the art can use the Lomb-Scargle periodogram method, whose spectral peak location rules are well-known in the field and will not be elaborated here. After determining the dominant frequency of the interferometric signal, the calculated dominant frequency value is substituted into the parameterized mathematical form, and the least squares method is applied to perform nonlinear numerical fitting on the multipath interferometric signal sequence. Through iterative calculation, the optimal amplitude, optimal bias term, and optimal delay phase that minimize the sum of squared residuals are obtained. The delay phase calculated at each specific frequency point is recorded as the core basic data for subsequent assessment of soil parameter changes; the delay phase extracted at each specific frequency point is referred to as the single-frequency delay phase.

[0042] See attached document Figure 3 See attached Figure 4 This paper elaborates on the specific execution details of dynamically calculating the adaptive weight allocation value of different frequency data sources based on the error index of the delay phase data in the preliminary estimation, and then performing weighted summation calculation of the adaptive weight allocation value and the corresponding single-frequency delay phase to obtain the dual-frequency fused delay phase.

[0043] S301, calculate the error index of the delay phase data at each frequency point in the preliminary estimation.

[0044] In this embodiment, observation data at a single frequency point is susceptible to interference from local environmental noise when reflecting changes in the surface dielectric constant. To effectively utilize the complementary information of multi-source reflection signals to improve the stability of inversion features, it is necessary to establish a reliability measurement mechanism for the single-frequency delay phase extracted for each specific frequency point. Specifically, based on historical measured soil moisture samples and historical single-frequency delay phases, a mechanism is independently constructed for each specific frequency point as shown in the attached figure. Figure 3 The linear regression reference model is shown. Using the test set data, the single-frequency delay phase at each specific frequency point is input into the corresponding linear regression reference model to obtain preliminary estimates. (Combined with the appendix...) Figure 4The regression residual distribution characteristics shown (where the horizontal axis Doy represents the accumulated days of the year) are used to calculate the deviation between the preliminary estimate and the measured value. The mean square error is used as the error index characterizing the reliability of the data source at this frequency point. The construction of the linear regression reference model and the statistical solution process for the mean square error can be achieved by those skilled in the art using conventional statistical analysis operations. The basic error measurement methods are well-known techniques in this field and will not be elaborated upon here.

[0045] S302 dynamically calculates adaptive weight allocation values ​​for data sources at different frequency points based on error indicators.

[0046] Given the error metrics for each specific frequency point, the weight of each data source in the fusion parameters is dynamically adjusted based on its performance. As a preferred approach, an exponential mapping mechanism is used to convert negatively correlated error metrics into positive weight coefficients, thereby calculating the adaptive weight allocation values ​​for data sources at different frequencies. The formula for calculating the adaptive weight allocation values ​​is as follows: ; In the formula, Indicates the first Adaptive weight allocation values ​​corresponding to a specific frequency point data source; This represents the total number of data sources at a specific frequency point, which is specifically configured as 2 in this embodiment, corresponding to the L1 and L2 frequency points of the Global Navigation Satellite System. Indicates the first The error index corresponding to a specific frequency point is substituted into the mean square error value obtained by the above steps. Represents the natural constant; This represents the adjustment parameter. The adjustment parameter is set as a constant greater than zero, primarily used to control the degree of influence of error on the adaptive weight allocation value. Those skilled in the art can set the specific value of this constant according to the noise distribution level of the actual observation station. Increasing this value will exponentially amplify small error differences between different frequency points, causing the adaptive weight allocation value to significantly tilt towards the frequency points with smaller errors.

[0047] S303 calculates the weighted sum of the adaptive weight allocation value and the corresponding single-frequency delay phase to obtain the dual-frequency fused delay phase.

[0048] After weight allocation, a hierarchical merging operation of multi-frequency data is performed. The single-frequency delay phase extracted independently for each specific frequency point is numerically multiplied with its corresponding adaptive weight allocation value, and the product results for all specific frequency points are accumulated to obtain a dual-frequency fused delay phase with higher observation stability. The weighted summation formula for the dual-frequency fused delay phase is as follows: ; In the formula, Indicates the dual-frequency fusion delay phase; Indicates the first The single-frequency delay phase extracted from a specific frequency point.

[0049] By employing the aforementioned physical process of evaluation and weighted merging, the synergistic effect can be dynamically intervened during system operation based on the verification characteristics of the data source. Frequency data with smaller error indices are assigned a higher numerical weight, while the weight of frequency data with local abnormal fluctuations is correspondingly reduced. This mechanism effectively overcomes the limitations of single-frequency observations, enabling the output dual-frequency fused delayed phase to more accurately characterize the true physical state of the target area.

[0050] See attached document Figure 5 The paper elaborates on the specific execution details of calculating the normalized microwave reflectance index, which characterizes vegetation moisture content, as a supplementary variable and together with the dual-frequency fusion delay phase to form the input feature vector.

[0051] S401, calculates the normalized microwave reflectance index, which characterizes the vegetation moisture content parameter.

[0052] In this embodiment, Global Navigation Satellite System (GNSS) signals typically pass through crop vegetation layers on their way to the farmland surface. The moisture contained within the vegetation stems and leaves absorbs and scatters the microwave signals, causing additional energy attenuation and phase shift in the multipath reflection signals from the ground surface. To eliminate the interference of vegetation growth status on soil moisture retrieval, a normalized microwave reflectance index is introduced as an environmental parameter characterizing vegetation moisture content. Specifically, the moisture content of vegetation directly causes attenuation of the multipath reflection signal amplitude. Based on the amplitude characteristics extracted using the aforementioned least squares fitting method, the extreme values ​​within the historical observation window are used for normalization. The formula for calculating the normalized microwave reflectance index is as follows: ; In the formula, Represents the normalized microwave reflectance index; Indicates amplitude; This indicates the maximum amplitude of the interference signal at a specific frequency point within a set time observation window; This represents the minimum amplitude of the interference signal at a specific frequency point within a set time observation window.

[0053] By solving the above calculation formula, the amplitude attenuation value with physical dimensions is converted into a dimensionless exponent in the interval [0,1], and the corresponding exponent value is obtained. Its basic calculation model can effectively reflect the current degree of vegetation shading.

[0054] S402, based on the correlation distribution characteristics, the normalized microwave reflectance index was established as a supplementary variable.

[0055] After obtaining the above environmental parameters, it is necessary to further clarify their functional positioning in the inversion prediction model. (See attached...) Figure 5 The morphological characteristics of the scatter plot distribution curve and confidence ellipse (where temp represents temperature and NMRI represents the normalized microwave reflectance index) are used to quantitatively evaluate the correlation between various physical quantities and soil moisture content. Based on the distribution pattern of the lower triangular region in the figure, both the single-frequency delayed phase and the dual-frequency fused delayed phase extracted at specific frequency points show a significant positive correlation with soil moisture content, while the normalized microwave reflectance index shows a weak correlation with soil moisture content. The physical essence of this weak correlation lies in the fact that the normalized microwave reflectance index mainly responds to the moisture fluctuations within the vegetation layer itself, rather than changes in moisture in the deeper soil layers. As a preferred approach, the normalized microwave reflectance index is established as a supplementary variable describing the surface vegetation cover state, used to help capture and reduce the bias in dielectric constant calculation caused by crop growth in subsequent algorithms.

[0056] S403 combines the supplementary variables with the dual-frequency fusion delay phase to form a multi-dimensional input feature vector.

[0057] After extracting the core interference features and quantifying the vegetation environment state, a dimension merging operation is performed at the data level. The dual-frequency fused delay phase, reflecting fluctuations in the soil dielectric constant, is set as the dominant feature component, and the normalized microwave reflectance index, reflecting the vegetation shading effect, is set as the modified feature component. These two physical parameters are combined and concatenated to construct a two-dimensional input feature vector suitable for machine learning algorithms. The mathematical expression of this input feature vector is as follows: ; In the formula, This represents the multidimensional input feature vector of the assembly; Indicates the dual-frequency fusion delay phase; This represents the normalized microwave reflectance index.

[0058] By combining the aforementioned physical parameters, a joint representation mechanism was established. This input feature vector, while transmitting the core dielectric characteristics of the land surface to the backend prediction algorithm, also includes the background vegetation environment state at the current observation epoch, thus providing a clear and complete basic data input structure for the stable operation of the nonlinear inversion model in complex farmland environments.

[0059] See attached document Figure 6 See attached Figure 7 This paper elaborates on the specific execution details of acquiring historical observation data to construct training samples, establishing a nonlinear inversion prediction model that includes multi-scale feature extraction and decision tree integration, performing hyperparameter optimization and training, and calculating the final soil moisture content.

[0060] S501: Acquire observation data and in-situ measurement benchmarks to construct a training sample set, and define the input feature dimensions and the physical state of the model output.

[0061] In this embodiment, supervised training of the inversion prediction model relies on a real surface physical state benchmark. A foundational dataset for model training and validation is constructed by combining GPS signal files recorded during historical observation periods and synchronously measured in-situ soil moisture content values. The multi-dimensional input feature vector extracted and concatenated in the preceding steps is established as the model's underlying input data. The in-situ measured soil moisture content is used as the physical state label data for the corresponding time point. By appropriately dividing the foundational dataset into training and testing sets, the model can subsequently learn the mapping relationship between the input multipath interference features and real soil moisture fluctuations.

[0062] S502, builds nonlinear inversion prediction models that support multiple linear regression, convolutional neural networks, or extreme gradient boosting architectures.

[0063] Multipath interference signals in natural surface environments are affected by surface roughness and spatial heterogeneity of soil moisture during propagation, exhibiting complex nonlinear characteristics. To characterize these nonlinear characteristics and provide a multi-adaptive prediction scheme, an inversion prediction model is established. Assuming a mapping relationship exists between soil moisture content and multiple characteristic variables, the unified calculation logic of this model is expressed in matrix form as follows: ; In the formula, This represents the estimated soil moisture content output by the model. This represents the multidimensional input feature vector of the assembly; This represents the dual-frequency fused delayed phase component in the multidimensional input feature vector; This represents the normalized microwave reflection index component in the multidimensional input feature vector.

[0064] As a preferred approach, when the model is configured with a multiple linear regression architecture (as shown in the first term of the formula matrix), Represents the intercept term; This represents the partial regression coefficient corresponding to the delayed phase component of the dual-frequency fusion; This represents the partial regression coefficient corresponding to the normalized microwave reflectance index component. This represents the random error term that reflects unobserved factors.

[0065] When the model is configured as a convolutional neural network architecture (as shown in the second term of the formula matrix), its internal hierarchical structure includes an input layer, two convolutional layers, and a fully connected output layer. The first convolutional layer has 16 convolutional kernels, and the second convolutional layer has 32 convolutional kernels. This indicates the introduction of a linear rectified function (ReLU) to achieve nonlinear activation, with weight initialization using the Glorot strategy. to They represent the first Layer to the first Weight matrix in a layered network structure; to They represent the first Layer to the first Bias terms in layered network structures; This indicates the total number of layers in the network structure. In the data flow direction, after the multi-dimensional input feature vector enters the input layer, the hidden features of multi-source associations are extracted through the local receptive field of the convolutional layer, and finally the estimated soil moisture content is output through the fully connected layer.

[0066] When the model is configured with an extreme gradient boosting architecture (as shown in the third term of the formula matrix), its prediction logic is composed of an integrated combination of multiple regression decision trees. This represents the total number of decision trees. Indicates the first Each decision tree performs forward prediction on the multi-dimensional input feature vector. Each decision tree performs node splitting calculations on the input variables according to its internal feature space partitioning rules, and the outputs of all decision trees are summed to obtain the overall estimate.

[0067] S503 combines a non-stationary stochastic Bayesian optimization algorithm to dynamically optimize hyperparameters and iteratively train the prediction model.

[0068] The inversion accuracy of the model is directly affected by the network layer or ensemble parameter settings. For extreme gradient boosting architectures, a non-stationary stochastic Bayesian optimization (NRBO) algorithm is used to dynamically adjust the model's hyperparameter configuration. The model's second derivative information is utilized to accelerate the convergence process within the optimized three-dimensional parameter space (number of parameters). Global optimization is performed within the training set, with the maximum number of iterations limited to 20. The mean squared error between the predicted values ​​calculated from the training set data and the true label data is used as the model's loss function, driving the optimization algorithm to search for the optimal combination of hyperparameters that minimizes the loss function, aiming to avoid overfitting and underfitting during model training. For the convolutional neural network architecture, the momentum gradient descent algorithm (SGDM) is used to iteratively update the weight matrix on the training set.

[0069] S504 inputs the multi-dimensional input feature vectors acquired in real time into the trained inversion prediction model to calculate the soil moisture content inversion result of the target monitoring area.

[0070] During the online operation phase of the monitoring system, a multi-dimensional input feature vector containing multi-band interferometric parameters and vegetation environment parameters, generated in real time, is continuously fed into the nonlinear inversion prediction model that has undergone hyperparameter optimization and structural training. The model performs forward propagation numerical calculations according to a determined internal weight distribution or decision tree combination rule, outputting continuous soil moisture content values ​​for the current observation epoch. (Combined with attached...) Figure 6 The linear fit distribution state and attached Figure 7 As shown in the attached figure, the measured values ​​compared with the tracking trajectory indicate that (in the attached figure, MLR, CNN, and NRBO-XGBoost represent the multiple linear regression model, the convolutional neural network model, and the extreme gradient boosting model based on the non-stationary stochastic Bayesian optimization algorithm, respectively; rmse and mae represent the root mean square error and the mean absolute error, respectively). The above nonlinear inversion prediction model can stably track changes in surface humidity, thereby achieving dynamic and high-precision quantitative monitoring of soil moisture status in complex farmland environments.

Claims

1. A method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength, characterized in that, include: Obtain observation and navigation files from the Global Navigation Satellite System receiver, filter frequency observation data that meet the preset satellite elevation and azimuth range conditions, extract the raw signal-to-noise ratio parameters and convert them into signal strength time series; Low-pass filtering is performed on the signal intensity time series at different frequency points to convert the independent variable into the sine value of the satellite elevation angle. Multi-path interference signal sequences are separated by eliminating the slow variation trend of direct signals through polynomial fitting. Spectral analysis is used to identify the main frequency of the interference signal, and single-frequency delay phase is extracted by least squares fitting. Calculate the error index of the single-frequency delay phase in the preliminary estimation, and calculate the adaptive weight allocation value of each frequency data source accordingly. Then, sum the adaptive weight allocation value with the single-frequency delay phase to obtain the dual-frequency fused delay phase. The normalized microwave reflection index is calculated and combined with the dual-frequency fused delay phase to form a multi-dimensional input feature vector; A nonlinear inversion prediction model incorporating a decision tree output accumulation mechanism is established. The hyperparameters are optimized using a nonstationary stochastic Bayesian optimization algorithm. The multidimensional input feature vector is input into the nonlinear inversion prediction model to output the predicted value of farmland soil moisture content.

2. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 1, characterized in that, The step of extracting the original signal-to-noise ratio parameter and converting it into a signal intensity time series includes: The original signal-to-noise ratio parameter is divided by decibels and the step size is divided to perform a rounding down operation. The result of the rounding down operation is compared with the minimum allowable threshold level of signal quality defined by the system to perform a maximum value selection operation. The result of the maximum value selection operation is compared with the maximum record threshold level of signal quality defined by the system to perform a minimum value selection operation, thereby obtaining the standardized signal strength parameter. The standardized signal strength parameters, which are discretely distributed at each epoch, are arranged in time sequence to generate a continuous signal strength time series.

3. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 1, characterized in that, The extraction of single-frequency delay phase using least squares fitting includes: A physical model of surface reflection interference is constructed based on the complex superposition of the direct signal and the reflected signal. Based on the physical model of surface reflection interference, an expression for the receiver output power, which includes the direct signal, the reflected signal, the signal carrier wavelength, and the delayed phase, is derived. The receiver output power expression is converted into a parameterized mathematical form that includes amplitude, bias term, main frequency of interference signal and delay phase, and the identified main frequency of interference signal is substituted into the parameterized mathematical form; The least squares method is applied to perform nonlinear numerical fitting on the multipath interference signal sequence. The optimal amplitude, optimal bias term, and optimal delay phase with the minimum residual sum of squares are obtained through iterative calculation. The optimal delay phase is then used as the single-frequency delay phase.

4. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 1, characterized in that, The error indices for calculating the single-frequency delay phase in the preliminary estimation include: Based on historical measured soil moisture content samples and historical single-frequency delayed phase, a linear regression reference model is independently constructed for each frequency point; The single-frequency delay phase is input into the linear regression reference model using test set data to obtain a preliminary estimate. Calculate the mean square error between the preliminary estimate and the measured value, and use the mean square error as the error index.

5. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 4, characterized in that, The calculation of adaptive weight allocation values ​​for each frequency data source includes: An exponential mapping mechanism is used to convert the error index into positive weighting coefficients; The numerator is the negative constant multiple of the error index raised to the power of the natural constant, and the denominator is the sum of the negative constant multiples of the error index raised to the power of the natural constant corresponding to all frequency data sources. The ratio of the numerator to the denominator is used as the adaptive weight allocation value.

6. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 3, characterized in that, The calculation of the normalized microwave reflectance index includes: Extract the maximum and minimum amplitude values ​​of the multipath interference signal sequence within the set time observation window; The difference between the maximum amplitude and the optimal amplitude is calculated as the first difference, and the difference between the maximum amplitude and the minimum amplitude is calculated as the second difference. The ratio of the first difference to the second difference is used as the normalized microwave reflectance index.

7. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 1, characterized in that, The component is a multi-dimensional input feature vector, including: The dual-frequency fusion delay phase, which reflects the fluctuation of soil dielectric constant, is set as the dominant characteristic component, and the normalized microwave reflectance index, which reflects the vegetation shading effect, is set as the modified characteristic component. The dominant feature component and the modified feature component are combined and concatenated to form a two-dimensional input feature vector, which is then used as the multi-dimensional input feature vector.

8. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 1, characterized in that, Before establishing a nonlinear inversion prediction model that includes a decision tree output accumulation mechanism, the method further includes: By combining the observation files recorded within the historical observation period, historical multidimensional input feature vectors are extracted to obtain the in-situ soil moisture content values ​​measured synchronously. The historical multidimensional input feature vector is established as the bottom-level input data, and the in-situ soil moisture content value is used as the physical state label data at the corresponding time to build the basic dataset. The basic dataset is divided into a training set and a test set; Construct a nonlinear inversion prediction model that supports an extreme gradient boosting architecture.

9. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 8, characterized in that, The step of inputting multidimensional input feature vectors into a nonlinear inversion prediction model to output farmland soil moisture content inversion prediction values ​​includes: The forward prediction output is obtained by performing node splitting calculation on the multidimensional input feature vector according to the internal feature space partitioning rules of each regression decision tree; The overall estimated value is obtained by summing the forward prediction outputs of all regression decision trees, and the overall estimated value is used as the inversion prediction value of farmland soil moisture content.

10. The method for inverting farmland soil moisture content based on multi-satellite dual-frequency GNSS signal strength according to claim 8, characterized in that, The method of using a non-stationary stochastic Bayesian optimization algorithm to find hyperparameters includes: Global optimization is performed in the three-dimensional parameter space using the second derivative information of the nonlinear inversion prediction model. The mean square error between the predicted value calculated from the training set data and the physical state label data is used as the loss function; The nonstationary stochastic Bayesian optimization algorithm is driven to search for the optimal hyperparameter combination that minimizes the loss function, and the nonlinear inversion prediction model is iteratively trained and updated based on the optimal hyperparameter combination.

Images