A remote sensing inversion method for soil moisture in China based on CYGNSS satellite data

Abstract

The present invention discloses a method for remote sensing inversion of soil moisture in China based on CYGNSS satellite data, and the steps are as follows: S1, pre-processing the collected measured soil moisture, delay-Doppler map, normalized vegetation index and land cover type; obtaining modeling samples matching all data through time-space matching; S2, taking soil moisture as the target variable, selecting CYGNSS trailing edge slope, incident angle, ground elevation, surface roughness and vegetation optical thickness as characteristic variables, and constructing a soil moisture remote sensing model; dividing the modeling samples into a training set and a test set, and training the soil moisture remote sensing model, and saving the best soil moisture remote sensing model; S3, according to the best soil moisture remote sensing model, inverting daily soil moisture data from the CYGNSS trailing edge slope and its related auxiliary data. The present invention improves the accuracy of remote sensing monitoring of soil moisture by spaceborne GNSS-R technology.

Description

Technical Field

[0001] The present invention relates to the technical field of soil moisture prediction, and particularly relates to a method for remotely sensing and retrieving soil moisture in the Chinese region based on CYGNSS satellite data. Background Technique

[0002] Soil moisture, as a key parameter of the water cycle, characterizes the evaporation loss of surface water, photosynthesis of plants, and activities of soil microorganisms. Traditional soil moisture monitoring methods are time-consuming, laborious, inefficient, and have few site distributions, and can no longer meet the current needs. In recent years, with the continuous development of the Global Navigation Satellite System Reflectometry (GNSS-R) technology, the L-band signal emitted by GNSS satellites is sensitive to soil moisture. After the GNSS satellite signal is reflected by the Earth's surface, it will interfere with the direct GNSS satellite signal at the GNSS receiver, and the interference signal will change with different surface states. Therefore, the GNSS reflectometry technology has become a new method for soil moisture monitoring. At present, the methods for measuring soil moisture by GNSS-R technology are mainly divided into ground-based and space-based. Ground-based GNSS-R measurements can obtain high-precision, high-observation-frequency, and long-time-series soil moisture at stations. For example, the Chinese invention patent "Method for Estimating Soil Moisture from Ground-Based GNSS-R Data Based on Modified Phase" with the publication number CN 115980317 B. However, due to the limited distribution of stations, this method cannot obtain soil moisture over a large area; spaceborne GNSS-R can quickly, periodically, and widely measure surface soil moisture, and has great advantages in soil moisture monitoring. Since the reflection of GNSS signals from the surface consists of two parts: coherent and incoherent components, surface roughness affects the surface reflection characteristics of GNSS signals. In areas with relatively small surface roughness, the reflection is mainly coherent; as the surface roughness increases, the effective reflection of the coherent component rapidly decreases, while the effective reflection of the incoherent component increases. However, the surface reflectivity used in current spaceborne GNSS-R remote sensing for monitoring soil moisture is often calculated under the assumption of smooth surface reflection and coherent components, ignoring the contribution of incoherent components to surface reflection. In addition, during the process of the GNSS signal from the navigation satellite incident on the surface and then reflected to the spaceborne receiver, it is attenuated twice by the vegetation covering the surface. Therefore, accurately determining surface reflection characteristics, surface roughness, and vegetation optical thickness are the key parameters for high-precision remote sensing monitoring of soil moisture by spaceborne GNSS-R.

[0003] Currently, when retrieving soil moisture by spaceborne GNSS-R remote sensing, the surface roughness and vegetation optical depth mainly come from the SMAP satellite soil moisture product. Even the soil moisture used as the true value in the modeling process comes from the SMAP satellite soil moisture product. For example, in the Chinese invention patent "A High-Precision GNSS-R Soil Moisture Estimation Method Based on CYGNSS Data" with the publication number CN 114371182 B, and the Chinese invention patent "Method for Co-Inverting Surface Soil Moisture and Vegetation Water Content by Fusing Spaceborne GNSS-R and Multi-Source RS Data" with the publication number CN 114371182 B, as well as the paper "Maria Paola Clarizia et al. Analysis of CYGNSS Data for Moisture Retrieval, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019, 12(7): 2227-2235". Due to the large-scale data gaps in the SMAP satellite soil moisture product and the limited accuracy of the satellite product data, error accumulation will occur during the soil moisture remote sensing modeling based on GNSS-R spaceborne data, making it difficult to achieve high-precision remote sensing retrieval of spaceborne GNSS-R soil moisture. Currently, the CYGNSS soil moisture product publicly released by NASA is retrieved using a simple linear regression model with the SMAP satellite soil moisture product as the true value and the CYGNSS surface reflectivity as the remote sensing variable, without correction for surface roughness and vegetation optical depth, and the accuracy of its soil moisture product needs to be further improved. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a method for remotely retrieving soil moisture in the Chinese region based on CYGNSS satellite data, so as to improve the accuracy of modeling in the process of remotely monitoring soil moisture by spaceborne GNSS-R.

[0005] Technical Solution: A method for remotely retrieving soil moisture in the Chinese region based on CYGNSS satellite data includes the following steps:

[0006] S1, preprocess the collected measured soil moisture, delay-Doppler map, normalized difference vegetation index, and land cover type; obtain the modeling samples with all data matched through spatio-temporal matching; after processing the delay-Doppler map, obtain the CYGNSS trailing edge slope, the normalized difference vegetation index is used for calculating the vegetation optical depth, and the land cover type is used for calculating the surface roughness;

[0007] S2, taking soil moisture as the target variable, by calculating the Pearson correlation coefficient between the target variable and the characteristic variable, selecting CYGNSS trailing slope, incident angle, ground elevation, surface roughness and vegetation optical thickness as characteristic variables, and constructing a soil moisture remote sensing model; dividing the modeling samples into training sets and test sets, and training the soil moisture remote sensing model, and saving the best soil moisture remote sensing model;

[0008] S3, based on the best soil moisture remote sensing model, invert daily soil moisture data from the CYGNSS trailing slope and its related auxiliary data.

[0009] Furthermore, the measured soil moisture data of each meteorological station was screened, first eliminating missing values ​​and outliers, then calculating the mean and standard deviation of the overall data, and eliminating data that exceeded the range of 2 times the standard deviation of the mean.

[0010] Furthermore, the steps to obtain the CYGNSS trailing edge slope after delay-Doppler map processing are as follows:

[0011] SD1, subtract the noise floor value from the power value corresponding to each delay-Doppler graph, the noise calculation formula is:

[0012] , where 、 are the lower and upper boundaries of the delay chips selected for calculating the noise region, indicates the delay chip; 、 are the lower and upper boundaries of the frequency selected for calculating the noise area; M is the number of pixels in the selected noise area;

[0013] The calculation formula of the delay-Doppler map after denoising can be expressed as:

[0014] , where is the delay-Doppler diagram after denoising; DDM is the delay-Doppler diagram; Noise represents noise;

[0015] SD2, yes Integrate the reflected power in the Doppler dimension to obtain the integrated delay waveform. The calculation formula is:

[0016] , where is the integrated delay waveform; N is the number of selected Doppler frequencies; is the kth Doppler frequency;

[0017] SD3, normalize the one-dimensional integrated delay waveform, the calculation formula is:

[0018] , where is the normalized integrated delay waveform, is the maximum value in the integrated delay waveform, and i takes values from 1 to 17; represents the integrated delay waveform of the i-th delay chip;

[0019] SD4, using a linear function of the first order, performs a least squares fit on the trailing edge of the normalized integrated delay waveform, and the absolute value of the obtained slope is used as the trailing edge slope.

[0020] Furthermore, the implementation steps for obtaining the modeling samples that match all data are as follows:

[0021] S11, by calculating the distances between each ground reflection point of the CYGNSS satellite and all soil moisture automatic monitoring stations, the point pairs with the distance between the two less than 7 Km are taken as the spatially matching sites;

[0022] S12, performs a time matching on the CYGNSS data and the soil moisture data of the corresponding sites with the same observation date and time;

[0023] S13, matches the matched CYGNSS-measured soil moisture samples with the auxiliary data, and the auxiliary data includes vegetation optical depth and surface roughness;

[0024] S14, according to the observation time of the CYGNSS satellite data, selects the corresponding period of auxiliary data for matching; then, according to the longitude and latitude coordinates of each reflection point of the CYGNSS satellite, it corresponds to the pixels where the areal auxiliary data is located, so that the data of each reflection point of the CYGNSS satellite is matched with the corresponding pixel values of the auxiliary data, and modeling samples that match all data are generated.

[0025] Furthermore, in step S2, a multiple linear regression remote sensing model for soil moisture, a support vector remote sensing model for soil moisture, and a random forest remote sensing model for soil moisture are respectively constructed, and the three models are respectively trained and tested, and the best remote sensing models are respectively saved; for the three best remote sensing models, soil moisture prediction is respectively performed, and the remote sensing model with the highest accuracy is selected as the soil moisture remote sensing model.

[0026] Compared with the prior art, the remarkable effects of the present invention are as follows:

[0027] 1. The present invention comprehensively considers the effects of coherent and incoherent reflections of surface GNSS signals. Based on the shape characteristics of the time-delay Doppler map (DDM) of spaceborne GNSS-R observables, the trailing edge slope (TES) is proposed as the main variable for remotely sensing soil moisture. This addresses the drawback in previous soil moisture remote sensing where surface reflectivity only considered the coherent component and ignored the contribution of the incoherent component to surface reflection, improving the accuracy of remotely sensing soil moisture using spaceborne GNSS-R technology.

[0028] 2. Since surface roughness affects the reflection characteristics of surface GNSS signals and vegetation cover attenuates GNSS signals, surface roughness and vegetation optical depth are two important factors affecting the remotely sensed soil moisture using spaceborne GNSS-R. The present invention proposes a method for accurately calculating these two parameters using the EOS satellite MODIS surface products, providing accurate auxiliary variables for the high-precision remote sensing inversion of soil moisture using spaceborne GNSS-R.

[0029] 3. Aiming at the problem of inaccurate modeling parameters in the current process of remotely sensing soil moisture using spaceborne GNSS-R, with the measured soil moisture data in the Chinese region as the target variable and the trailing edge slope data of CYGNSS satellites as the remote sensing variable, combined with vegetation optical depth and surface roughness data, multiple regression and machine learning methods are used to construct an optimal CYGNSS soil moisture remote sensing inversion model suitable for the Chinese region, solving the problem of inaccurate modeling parameters in previous spaceborne GNSS-R soil moisture remote sensing methods, and obtaining daily and monthly soil moisture remote sensing products for the Chinese region retrieved from CYGNSS data, providing an important data source for agricultural climate drought and hydro-meteorological research in the Chinese region. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the flow chart of the present invention;

[0031] Figure 2 in (a) is the time-delay Doppler map of CYGNSS satellites, and (b) is the normalized integrated time-delay waveform map;

[0032] Figure 3 is the scatter plot of measured values and predicted values during the training and testing of the multiple linear remote sensing model of soil moisture in the embodiment of the present invention;

[0033] Figure 4 is the scatter plot of measured values and predicted values during the training and testing of the support vector remote sensing model of soil moisture in the embodiment of the present invention;

[0034] Figure 5 is the scatter plot of measured values and predicted values during the training and testing of the random forest remote sensing model of soil moisture in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0035] The present invention will be further described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0036] The present invention provides a method for remotely sensing and retrieving soil moisture in the Chinese region based on CYGNSS satellite data. In order to verify the method proposed by the present invention, data for four typical months of January, April, July, and October 2018 in the Chinese region, representing different seasons respectively, are used to construct an optimal CYGNSS soil moisture remote sensing inversion model suitable for the Chinese region, and remotely sense and retrieve soil moisture data for four typical months of February, May, August, and November 2019, generating daily and monthly soil moisture remote sensing products for the Chinese region. The implementation process of the technical solution is as follows Figure 1 shown, including the following steps:

[0037] Step 1, input of data;

[0038] The data that needs to be input in the present invention mainly includes:

[0039] Soil moisture data measured by meteorological stations in the Chinese region, including information such as site longitude, latitude, elevation, and observation date, is used for the construction of the remote sensing inversion model;

[0040] CYGNSS satellite data includes data sets such as delay-Doppler maps (DDM), specular reflection point latitude (sp_lat), specular reflection point longitude (sp_lon), specular reflection point incident angle (sp_inc_angle), etc. The DDM is used for the calculation of ground reflectivity and is the main variable for soil moisture remote sensing;

[0041] TERRA or AQUA satellite MODIS vegetation index products, and the included normalized vegetation index is used for the calculation of vegetation optical thickness;

[0042] MODIS land cover type data classified by the International Geosphere Biosphere Programme (IGBP) is used for the calculation of surface roughness.

[0043] Delay-Doppler maps (DDM) are signal powers around mirror reflection points in the ground blaze zone obtained by the satellite-borne GNSS-R receiver. They are two-dimensional distribution diagrams of the related power of delay and Doppler frequency shift. Their shape and amplitude can be used to infer the surface reflection characteristics. When the surface is relatively smooth, the signal reflected by the coherent component will show a typical "horseshoe" shape on the DDM; with the influence of surface roughness and soil moisture, the signal reflected by the incoherent component will show various delay wave shapes. Therefore, the leading edge slope (LES) and trailing edge slope (TES) calculated from the DDM normalized delay waveform can well reflect the degree of coherent / incoherent reflection on the surface.

[0044] The data used in the soil moisture remote sensing modeling of the embodiment of the present invention are: the measured soil moisture data of four typical months in my country, namely January, April, July and October 2018, the CYGNSS Delay-Doppler Map (DDM) data and its auxiliary parameters including the incident angle, the MODIS (Moderate-resolution Imaging Spectroradiometer) Normalized Difference Vegetation Index product and MODIS IGBP global land cover type data of the same period.

[0045] Step 2, data preprocessing;

[0046] (A1) Quality control of measured soil moisture data

[0047] In order to ensure the validity of data in soil moisture remote sensing modeling, the measured soil moisture data of each meteorological station was screened. First, missing values ​​and outliers were removed, and then the mean and standard deviation of the overall data were calculated. Data that exceeded the range of 2 times the standard deviation of the mean were removed to ensure the accuracy of the measured soil moisture data and improve the accuracy of the soil moisture remote sensing model.

[0048] (A2) Processing of CYGNSS satellite data

[0049] Quality control is performed on CYGNSS satellite data. For data in the CYGNSS data set with a DDM signal-to-noise ratio less than 0db, a receiver antenna gain less than 0dB, and an incident angle greater than 65°, it means that the received signal quality is relatively poor and should be eliminated.

[0050] The CYGNSS satellite GNSS-R receiver has 4 reflection measurement channels. Each channel observes DDM data of nearly 100,000 reflection points globally. The DDM data of each reflection point is a two-dimensional correlation power distribution map composed of 17 delay cells (Bins) and 11 Doppler shift cells (Bins), as shown in Figure 2 as shown in (a) of it. Calculating the CYGNSS trailing edge slope (TES) from the DDM data is divided into 4 steps:

[0051] (D1) Denoising the DDM data

[0052] Subtract the noise floor value from the power value corresponding to each DDM. The noise calculation formula is:

[0053] (1)

[0054] In the formula, , are respectively the lower boundary and the upper boundary of the delay chips selected for calculating the noise area; , are respectively the lower boundary and the upper boundary of the frequency selected for calculating the noise area; M is the number of pixels in the selected noise area. The noise area selected in the present invention is the pixels of all frequencies in the first 4 delays, and M is 44 pixels. The formula for the denoised DDM can be expressed as:

[0055] (2)

[0056] In the formula, is the denoised DDM, and Noise represents the noise.

[0057] (D2) Calculating the integrated delay waveform

[0058] Integrate the reflected power of the denoised DDM in the Doppler dimension to obtain the integrated delay waveform (Integrated Delay Waveform, IDW). The calculation formula is:

[0059] (3)

[0060] In the formula, is the normalized integrated delay waveform; N is the number of selected Doppler frequencies. In the present invention, all frequencies are taken as 11; is the kth Doppler frequency.

[0061] (D3) Normalization of the integrated delay waveform

[0062] Perform normalization processing on the one-dimensional integrated delay waveform. Its calculation formula can be expressed as:

[0063] (4)

[0064] Wherein, is the normalized integrated delay waveform, is the maximum value in the integrated delay waveform, and i takes values from 1 to 17; represents the integrated delay waveform of the i-th delay chip. Figure 2 The (b) in is a normalized integrated delay waveform diagram in an embodiment of the present invention.

[0065] Calculation of the trailing edge slope of (D4)

[0066] The trailing edge of the normalized integrated delay waveform is composed of N sampling points after the waveform peak and decreases with the delay. A linear function can be used to perform least squares fitting on the trailing edge to obtain the trailing edge slope. In the present invention, 3 waveform values after the waveform peak delay ( ) are used for linear regression, and the absolute value of the obtained slope is the CYGNSS trailing edge slope TES.

[0067] Calculation of vegetation optical depth of (A3)

[0068] When the electromagnetic wave signal transmitted by the GNSS satellite encounters vegetation, it will propagate from the top of the vegetation canopy to the ground and then from the ground to the top of the vegetation canopy, and will experience double attenuation of the vegetation during the GNSS signal transmission process. The transmittance of the GNSS information penetrating the vegetation depends on the vegetation optical depth and the incident angle, and the calculation formula is as follows:

[0069] (5)

[0070] (6)

[0071] Wherein, represents the incident angle, represents the vegetation optical depth, VWC represents the vegetation water content, and b represents the proportionality factor related to the land cover type.

[0072] The proportionality factor b is related to the vegetation type, canopy structure, growth status, and the frequency of the signal transmitted by the GNSS satellite. In this embodiment, the proportionality factor b is determined by the look-up table method using the land cover type data classified by MODIS IGBP in 2018 (see Table 1).

[0073] Table 1 Vegetation parameters of different land cover types

[0074]

[0075] The vegetation water content VWC can be calculated according to the following empirical formula (3) through the normalized difference vegetation index NDVI in the corresponding period of 2018, the maximum value of the normalized difference vegetation index in the past 10 years and the minimum value , as well as the vegetation stem factor:

[0076] (7)

[0077] In the formula, the vegetation stem factor is determined by the look-up table method according to the MODIS IGBP land cover type in 2018; the maximum value and the minimum value of the normalized difference vegetation index in the past 10 years can be obtained by programming calculation on the Google Earth Engine (GEE) platform.

[0078] (A4) Determination of surface roughness

[0079] When the signal emitted by the GNSS satellite is incident on the smooth ground surface, it will exhibit the characteristics of specular reflection, and the signal received by the GNSS receiver is mainly coherent scattering. As the surface roughness increases, the amount of incoherent scattering on the ground gradually increases, and the incoherent scattering is insensitive to polarization characteristics, and the difference between its horizontal and vertical polarizations is small. For surface roughness, the root mean square height h is usually used to represent it in the vertical direction of the ground, which is related to the surface vegetation cover type, terrain undulation, polarization mode and frequency of electromagnetic waves. In this embodiment, the surface roughness is determined by the look-up table method using the MODIS IGBP land cover type data in 2018 (Table 1).

[0080] (A5) Temporal and spatial matching of data

[0081] Since the distance between the ground reflection points in the same scan band of the CYGNSS satellite is about 7 Km, the position of the ground reflection point signal received by the CYGNSS satellite changes every day, while the position of the soil moisture automatic monitoring station is fixed. When performing spatial matching of the two types of data, calculate the distance between each ground reflection point of the CYGNSS satellite and all soil moisture automatic monitoring stations every day, and take the point pairs with a distance less than 7 Km between the two as the spatial matching sites.

[0082] The CYGNSS satellite, as a polar orbiting satellite, passes over at 0-23 hours every day, while the soil moisture is observed hourly from 0-23 o'clock every day. After completing the spatial matching of the two types of data, perform temporal matching of the CYGNSS data with the soil moisture data at the corresponding sites on the same date and time, that is, complete the temporal matching of the two types of data.

[0083] Since the CYGNSS satellite data are point data, while the two auxiliary data (vegetation optical depth and surface roughness) are area data. First, according to the observation time of the CYGNSS satellite data, the corresponding auxiliary data for the corresponding period are selected; second, according to the longitude and latitude coordinates of each reflection point of the CYGNSS satellite, they are mapped to the pixels where the area auxiliary data are located, so that the data of each reflection point of the CYGNSS satellite are matched with the corresponding pixel values of the auxiliary data, generating samples in which the two types of data are matched.

[0084] To obtain the modeling samples in which all data are matched, the steps are as follows:

[0085] Step 11, by calculating the distances between each ground reflection point of the CYGNSS satellite and all soil moisture automatic monitoring stations, the point pairs with the distance between the two less than 7 Km are taken as the spatially matched sites.

[0086] Step 12, perform time matching on the CYGNSS data and the soil moisture data of the corresponding sites with the same observation date and time, that is, complete the spatio-temporal matching of the two types of data.

[0087] Step 13, match the matched CYGNSS-measured soil moisture sample points with the two auxiliary data (vegetation optical depth and surface roughness).

[0088] Step 14, according to the observation time of the CYGNSS satellite data, select the corresponding auxiliary data for the corresponding period; then, according to the longitude and latitude coordinates of each reflection point of the CYGNSS satellite, map them to the pixels where the area auxiliary data are located, so that the data of each reflection point of the CYGNSS satellite are matched with the corresponding pixel values of the auxiliary data, generating the modeling samples in which all data are matched.

[0089] Step 3, construction of the soil moisture remote sensing model;

[0090] (B1) Selection of model characteristic parameters

[0091] In the present invention, the soil moisture data are used as the target variable, and the trailing edge slope TES of CYGNSS, the incident angle , the ground elevation , the surface roughness and the vegetation optical depth As the characteristic variables, the characteristic parameters of the soil moisture remote sensing model are screened. Since the selection of characteristic variables has a great influence on the model accuracy, when the correlation between the characteristic variables and the dependent variable is low, it is difficult to obtain a good fitting effect for the established model; when the correlation is too high, the model will fall into flatness, resulting in low model efficiency. In order to obtain a model with high accuracy and high efficiency, it is necessary to screen the characteristic variables of the model, calculate the Pearson correlation coefficients between the characteristic variables and between the characteristic variables and the dependent variable, and calculate the variance inflation factor (VIF) between the dependent variable and each characteristic variable. Finally, based on the Pearson correlation coefficient being less than 0.8 and the variance inflation factor being less than 5, the optimal characteristic parameters for constructing the soil moisture remote sensing model are determined. Table 2 lists the Pearson correlation coefficients and variance inflation factors between the soil moisture and each characteristic variable.

[0092] Table 2 Statistics of Soil Moisture and Each Characteristic Variable

[0093]

[0094] As can be seen from Table 2, all the correlation coefficients are less than 0.8 and the variance inflation factors are lower than 5, indicating that there is no problem of multicollinearity between the soil moisture and each characteristic variable, and the 5 characteristic variables can be used for soil moisture remote sensing modeling.

[0095] (B2) Construct a soil moisture remote sensing model

[0096] After the selection of the model characteristic parameters, 50% of all the modeling samples are randomly selected as the training data set for constructing the soil moisture remote sensing model, and the remaining 50% are used as the test set of the model. Three methods, namely multiple linear regression, support vector regression, and random forest, are respectively selected to construct the soil moisture remote sensing model.

[0097] In the embodiments of the present invention, the total number of pairs of samples in which the soil moisture of each measured site matches the data of 5 characteristic variables is 128,518. 50% of the randomly selected modeling samples are used as the training set for establishing the model, with a total of 64,259 pairs of samples, and the remaining 50% of the modeling samples are used as the test set for model prediction, with a total of 64,259 pairs of samples.

[0098] (i) Multiple linear regression model

[0099] Assume that the soil moisture (SM) and the 5 characteristic variables, the trailing slope TES of CYGNSS and the incident angle , ground elevation , surface roughness , and vegetation optical depth There is a multiple linear correlation relationship. The multiple linear regression analysis method is adopted, and a soil moisture remote sensing model is established using the training samples. The general form of the model is as follows:

[0100] (8)

[0101] In the formula: is the constant term of the model, is the regression coefficient of the model. The multiple linear regression is performed on the training set of the soil moisture samples by the least squares method to construct the soil moisture remote sensing model.

[0102] For the training samples, the multiple linear regression analysis method is adopted, and the constructed soil moisture remote sensing model is:

[0103]

[0104] In the formula: SM is the soil volume water content, with the unit of cm 3 / cm 3 ; the incident angle V z is in degrees ( ); the ground elevation is in kilometers (Km); the surface roughness and the vegetation optical thickness are dimensionless.

[0105] After the t-test, the p-value of the regression coefficient of each characteristic variable is less than 0.01, indicating that the linear influence of each characteristic variable on the corresponding variable reaches an extremely significant level. The scatter plots of the measured values and the predicted values during the training and testing of the soil moisture multiple linear regression remote sensing model are shown in Figure 3 .

[0106] (ii) Support vector regression model

[0107] In the support vector regression (SVR) model, a function and two parameters have a greater impact on the results of the model, namely the kernel function, the penalty coefficient C, and the kernel coefficient gamma. The kernel function determines the accuracy and universality of the SVR model. The commonly used kernel functions include: linear kernel function, polynomial kernel function, Gaussian kernel function, and Sigmoid kernel function. In the present invention, through experiments on 4 kinds of kernel functions, the Gaussian kernel function with better smoothing performance is selected for the construction of the soil moisture support vector remote sensing model.

[0108] The default value of the penalty coefficient C is 1.0. The value of C controls the complexity of the model and the tolerance for errors. The larger the value of C, the higher the degree of fitting of the model to the training data, that is, the model is less likely to tolerate errors on the training set, which may lead to overfitting of the model; on the contrary, a smaller C will cause the model to easily tolerate more errors, which helps to improve the generalization ability of the model. The present invention adopts a method of trial numbers, taking the C values as [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0] respectively, testing all training sample sets, and selecting the optimal C value by calculating the correlation coefficient R and the root mean square error RMSE between the measured value of soil moisture and the predicted value of the SVR model.

[0109] The kernel coefficient gamma defines the influence range of the kernel function, which affects the fitting of the model to the training data. A larger gamma value will cause the model to only focus on the sample points closer to the support vectors, which may lead to overfitting of the model to the training data; a smaller gamma value will make the range of support vectors larger, which may lead to underfitting of the model. The present invention adopts a method of trial numbers, taking the gamma values as [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0] respectively, testing all training sample sets, and selecting the optimal gamma value by calculating the R and RMSE between the measured value of soil moisture and the predicted value of the SVR model.

[0110] Thus, a soil moisture support vector remote sensing model with optimal parameters is constructed.

[0111] For the training samples, the support vector regression method is adopted, and the Gaussian kernel function with better smoothing performance is selected as the kernel function. The penalty coefficient C values are taken as [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0] respectively, and the kernel coefficient gamma values are taken as [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0] respectively. All training sample sets are tested to determine the optimal parameters in the support vector regression model, and a soil moisture support vector regression remote sensing model is established. The scatter plot of the measured value and the predicted value during model training and testing is shown in Figure 4 。

[0112] (iii) Random forest model

[0113] In this embodiment, two key parameters (the number of decision trees and the feature quantity of each decision tree) in the random forest model are optimized and selected.

[0114] For the training samples, the random forest method is adopted. By setting the set of the number of decision trees as {50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000} and the set of the feature quantity of each decision tree as {1, 2, 3, 4, 5}, the 5-fold cross-validation method is used for the training sample data set. The number of decision trees and the feature quantity of each decision tree are continuously optimized from different combinations to improve the model accuracy, and the optimal number of decision trees and the feature quantity of each decision tree are determined to construct a random forest remote sensing model for soil moisture. The scatter plots of the measured values and the predicted values during the model training and testing are shown in Figure 5 .

[0115] Table 3 lists the statistical parameters of the three soil moisture remote sensing models during training and testing. From the correlation coefficient R, root mean square error RMSE, and mean absolute error MAE, the training accuracy and prediction accuracy of the soil moisture multiple linear remote sensing model are comparable, while the training accuracy of the soil moisture support vector remote sensing model and random forest remote sensing model is better than the prediction accuracy. Especially, the training accuracy of the random forest model is very high. Among the three soil moisture remote sensing models, the accuracy of the multiple linear model is lower, the accuracy of the support vector model is the second, and the accuracy of the random forest model is the highest. Therefore, the random forest remote sensing model is the best model for soil moisture remote sensing inversion in the embodiments of the present invention.

[0116] Table 3 Statistical parameters of different soil moisture remote sensing models during training and testing

[0117]

[0118] Step 4, remote sensing inversion of soil moisture;

[0119] (C1) Preparation of inversion data

[0120] First, calculate the trailing edge slope data of the 8 small satellites of the CYGNSS constellation every day and extract the corresponding incident angles. For the convenience of data processing, the data of the 8 satellites of the CYGNSS every day can be merged.

[0121] In the embodiments of the present invention, the established optimal random forest remote sensing model for soil moisture is used to retrieve the soil moisture data for four typical months, namely February, May, August, and November 2019. First, the trailing edge slope data of the 8 small satellites of the CYGNSS constellation for each day in the 4 months of 2019 are calculated, the corresponding incident angles are extracted, and the data of the 8 satellites of CYGNSS for each day are merged. Then, according to the longitude and latitude coordinates of the CYGNSS ground reflection points, the elevation values corresponding to the CYGNSS ground reflection points are extracted from the Digital Elevation Model (DEM) of the Chinese region; from the vegetation optical depth data calculated from the MODIS global vegetation index product, the vegetation optical depth values corresponding to the CYGNSS ground reflection points are extracted; from the surface roughness data calculated from the MODIS IGBP land cover type product, the surface roughness values corresponding to the CYGNSS ground reflection points are extracted. Finally, the daily CYGNSS trailing edge slope and incident angle data, as well as the corresponding ground elevation, surface roughness, and vegetation optical depth, are merged as independent variables for the remote sensing retrieval of soil moisture, preparing for the remote sensing retrieval of soil moisture.

[0122] (C2)Generation of remote sensing products for soil moisture

[0123] According to the established optimal remote sensing model for soil moisture, the daily soil moisture data are retrieved from the CYGNSS trailing edge slope and its related auxiliary data. Based on the longitude and latitude coordinates of the CYGNSS ground reflection points, a daily soil moisture point vector file is generated, and then the inverse distance weighting method (IDW) is used to perform spatial interpolation on the soil moisture data of each reflection point to generate the daily remote sensing product data for soil moisture. The daily remote sensing product data for soil moisture for each month are averaged to generate the monthly remote sensing product data for soil moisture.

Claims

1. A remote sensing inversion method for soil moisture in China based on CYGNSS satellite data, characterized in that: The steps include: S1, pre-processing the collected measured soil moisture, delay-Doppler map, normalized vegetation index and land cover type; obtaining modeling samples matching all data through time-space matching; obtaining the CYGNSS trailing edge slope after processing the delay-Doppler map, the normalized vegetation index is used for calculating the vegetation optical thickness, and the land cover type is used for calculating the surface roughness; S2, taking soil moisture as the target variable, by calculating the Pearson correlation coefficient between the target variable and the characteristic variable, selecting CYGNSS trailing slope, incident angle, ground elevation, surface roughness and vegetation optical thickness as characteristic variables, and constructing a soil moisture remote sensing model; dividing the modeling samples into a training set and a test set, and training the soil moisture remote sensing model, and saving the best soil moisture remote sensing model; S3, based on the best soil moisture remote sensing model, invert daily soil moisture data from the CYGNSS trailing edge slope and its related auxiliary data.

2. The method for remote sensing inversion of soil moisture in China based on CYGNSS satellite data according to claim 1 is characterized in that: The measured soil moisture data of each meteorological station were screened. First, missing values ​​and outliers were eliminated, and then the mean and standard deviation of the overall data were calculated. Data that exceeded 2 times the standard deviation of the mean were eliminated.

3. The method for remote sensing inversion of soil moisture in China based on CYGNSS satellite data according to claim 1 is characterized in that: The steps to obtain the CYGNSS trailing edge slope after processing the delay-Doppler map are as follows: SD1, subtract the noise floor value from the power value corresponding to each delay-Doppler graph, and the noise calculation formula is: , where , are the lower and upper boundaries of the delay chips selected for calculating the noise region, represents the delay chip; , are the lower and upper boundaries of the frequency selected for calculating the noise area; M is the number of pixels in the selected noise area; Then the calculation formula of the delay-Doppler map after denoising can be expressed as: , where is the delay-Doppler diagram after denoising; DDM is the delay-Doppler diagram; Noise represents noise; SD2, yes Integrate the reflected power in the Doppler dimension to obtain the integrated delay waveform. The calculation formula is: , where is the integrated time delay waveform; N is the number of selected Doppler frequencies; is the kth Doppler frequency; SD3, normalizes the one-dimensional integrated delay waveform, and the calculation formula is: , where is the normalized integrated delay waveform, is the maximum value in the integrated delay waveform, i takes values ​​from 1 to 17; represents the integrated delay waveform of the i-th delay chip; SD4, a linear function is used to perform least squares fitting on the trailing edge of the normalized integrated delay waveform, and the absolute value of the slope is taken as the trailing edge slope.

4. The method for remote sensing inversion of soil moisture in China based on CYGNSS satellite data according to claim 1 is characterized in that: The implementation steps for obtaining modeling samples that match all data are as follows: S11, by calculating the distance between each ground reflection point of the CYGNSS satellite and all soil moisture automatic monitoring stations, the point pairs with a distance less than 7 km between the two are taken as spatial matching sites; S12, time matching the CYGNSS data with the soil moisture data of the corresponding station with the same observation date and time; S13, matching the matched CYGNSS-measured soil moisture sample points with auxiliary data, wherein the auxiliary data includes vegetation optical thickness and surface roughness; S14, according to the observation time of the CYGNSS satellite data, selecting the auxiliary data of the corresponding period for matching; Then, according to the longitude and latitude coordinates of each reflection point of the CYGNSS satellite, it corresponds to the pixel where the planar auxiliary data is located, so that the data of each reflection point of the CYGNSS satellite is matched with the corresponding pixel value of the auxiliary data to generate a modeling sample with matching data.

5. The method for remote sensing inversion of soil moisture in China based on CYGNSS satellite data according to claim 1 is characterized in that: In step S2, a soil moisture multivariate linear regression remote sensing model, a soil moisture support vector remote sensing model and a soil moisture random forest remote sensing model are constructed respectively, and the three models are trained and tested respectively, and the best remote sensing models are saved respectively; soil moisture prediction is performed for the three best remote sensing models respectively, and the remote sensing model with the highest accuracy is selected as the soil moisture remote sensing model.

Images