A method for realizing high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data

Abstract

The application provides a method for realizing high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data, and belongs to the technical field of vegetation water content change monitoring. The application comprises the following steps: calculating an amplitude characteristic parameter in a GNSS-IR observation; processing hyperspectral data to obtain surface reflectivity of a red light and near-infrared band; using multiple linear and random forest regression algorithms to fuse GNSS-IR and hyperspectral data to establish a normalized difference vegetation index inversion model; using medium-resolution imaging spectrometer synchronous observation data as comparison and calculating inversion precision. Through the GNSS-IR and hyperspectral joint observation experiment, the data are fused to obtain vegetation water content change, and through comparison with MODIS vegetation water content index data, the method realizes high-precision monitoring of the vegetation index. Compared with a single data source, the method improves the inversion precision of the vegetation water content.

Description

Technical Field

[0001] This invention relates to the field of vegetation water content change monitoring technology, specifically to a method for high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data. Background Technology

[0002] Vegetation water content (VWC) is a key indicator for measuring plant water status, directly affecting plant growth, photosynthesis, and stress resistance. In ecosystems, changes in VWC reflect the degree of water stress in vegetation, providing important data for drought early warning, forest fire risk assessment, and agricultural irrigation optimization. Traditional VWC monitoring methods mainly include ground sampling, optical remote sensing, and microwave remote sensing, but these technologies all have certain limitations. For example, while ground sampling has high accuracy, it is highly destructive and inefficient, making it difficult to achieve large-scale dynamic monitoring; optical remote sensing is easily affected by cloud cover, and the inability to obtain high-quality observations when cloud cover is present limits its temporal resolution; microwave remote sensing, although capable of all-weather observation, has low spatial resolution, making it difficult to meet the needs of refined management. The emergence of Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) provides a new solution for VWC monitoring. This technology estimates VWC by analyzing the reflection characteristics of GNSS signals on the vegetation surface, inverting the dielectric constant of the vegetation. Compared to traditional methods, GNSS-IR offers advantages such as continuous observation, high spatiotemporal resolution, and low cost and ease of deployment. GNSS-IR is expected to become an important tool for global vegetation moisture monitoring, providing more efficient data support for precision agriculture, ecological protection, and disaster prevention.

[0003] In the fusion research of GNSS-IR and optical or microwave remote sensing data (such as MODIS, AMSR-E, etc.), Pan Yalong proposed a GNSS-IR and MODIS data fusion method based on GA-BP neural network, generating NMRI products with a resolution of 16 days / 500 meters. Liang Yueji used MODIS multi-band synthetic vegetation index to generate NMRI products with a resolution of 8 days / 500 meters, with an R-value and RMSE of 0.820 and 0.031 respectively during model training. Li Shuwen et al. integrated GNSS-IR NMRI datasets with AMSR-E and AMSR2 VOD datasets to generate spatiotemporally continuous NMRI products, achieving an R-value of 0.76 and a spatiotemporal resolution of 1 day / 25 kilometers in the generalized regression neural network model fitting.

[0004] GNSS-IR technology overcomes the limitations of continuous observation and low spatial resolution of optical and microwave remote sensing in VWC inversion. However, when using GNSS-IR alone to invert VWC, the inversion accuracy is low due to the influence of various factors on the signal. Therefore, many studies have attempted to fuse GNSS-IR with multi-source data to improve inversion accuracy. However, existing fusion methods are limited by the spatiotemporal resolution of auxiliary data, and the inversion results are still unsatisfactory. This study innovatively fuses GNSS-IR with ground-based hyperspectral data and proposes a high-precision VWC inversion method based on multiple linear and random forest regression models. This method enables continuous daily monitoring, significantly improves the inversion accuracy and spatiotemporal resolution of vegetation water content, and provides important technical support for agriculture, ecological protection, and natural disaster prediction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for high-precision inversion of vegetation water content (VWC) indicators by fusing GNSS-IR and hyperspectral data. This inversion method overcomes the limitations of traditional optical and microwave remote sensing in continuous observation and low accuracy for VWC monitoring, offering advantages such as high spatiotemporal resolution and high inversion accuracy. This invention provides a robust method for vegetation monitoring through multi-source data fusion, achieving higher accuracy compared to inversion results from a single data source.

[0006] The present invention adopts the following solution:

[0007] A method for high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data includes:

[0008] S1: Obtain the amplitude (Amp) characteristic index, which is correlated with the Normalized Differential Vegetation Index (NDVI), through GNSS-IR observations;

[0009] S2: Processing hyperspectral data to obtain surface reflectance in the red and near-infrared (NIR) bands;

[0010] S3: A normalized differential vegetation index (NDVI) inversion model was established by fusing GNSS-IR and hyperspectral data using a multiple linear regression model, and the calculation accuracy was compared with that of MODIS data from a medium resolution imaging spectrometer.

[0011] S4: A normalized differential vegetation index (NDVI) inversion model was established by fusing GNSS-IR and hyperspectral data using a random forest regression model, and the calculation accuracy was compared with that of MODIS data from a medium resolution imaging spectrometer.

[0012] As a further improvement of the present invention, the specific implementation process of step S1 is as follows:

[0013] S11: Set up GNSS-IR observation equipment, including an upward-looking GNSS right-hand circularly polarized antenna and a GNSS receiver. Set up the antenna in the observation area, collect low-elevation-angle interferometric signal observation SNR data, and store it in the receiver.

[0014] S12: Calculate the satellite elevation and azimuth angles for each observation epoch based on the precise ephemeris and the location of the receiving antenna in the stored files, and select the signal-to-noise ratios that meet the vegetation monitoring constraints. ;

[0015] S13: Signal-to-noise ratio expressed in decibels in logarithmic form. The signal is converted into a linear SNR value in Volts / Volts units, and after fitting an approximation of the direct signal, the multipath reflection component is obtained. ;

[0016] S14: Through the Lomb-Scargle spectral analysis was performed to obtain the oscillation frequency. Vertical height of the receiver antenna relative to the reflector surface ;

[0017] S15: Use the IGGⅢ weighted least squares method to fit the SNR observations and solve for the optimal amplitude A for the fitting period of a single GNSS satellite. m ;

[0018] S16: Outliers were removed by the median absolute difference (MAD) method, the amplitude was normalized, and effective satellites with a Pearson correlation coefficient R>0.5 and a number of fitted points>4 were selected. Their normalized amplitudes were summarized and the daily average value was calculated as the daily observed amplitude Amp of GNSS-IR.

[0019] As a further improvement of the present invention, in step S15:

[0020] The IGGⅢ weighted least squares method is a weighted iterative method that employs robust robust estimation and calculates equivalent weights using the IGGⅢ model. The formula is as follows:

[0021]

[0022] In the formula, the subscript for Number, For observation epochs; for Weights of each observation; The residuals of the observed values; The standard error of the observations; Standardized error; For quantile parameters; This is the elimination point.

[0023] As a further improvement of the present invention, the specific implementation process of step S2 is as follows:

[0024] S21: Construct a hyperspectral observation device at the same location as the GNSS-IR observation site, including a portable hyperspectral camera;

[0025] S22: Perform atmospheric correction and minimum noise separation processing on the raw hyperspectral data to extract the red light reflectance (Red) and near-infrared reflectance (NIR) at wavelengths of 668 nm and 845 nm, respectively.

[0026] As a further improvement of the present invention, the specific implementation process of step S3 is as follows:

[0027] S31: Download the MOD09GQ and MYD09GQ data from MODIS to calculate the daily NDVI values ​​for the study area. The formula is as follows:

[0028]

[0029] In the formula, band1 and band2 are the surface reflectance of the red light and near-infrared bands, respectively;

[0030] S32: Combine the red reflectance (Red) and near-infrared reflectance (NIR) into VIR_HS according to the calculation formula in step S31. Use the amplitude (Amp), VIR_HS, and both obtained from daily GNSS-IR to construct an inversion model with MODIS NDVI using the multiple linear regression method, and calculate the constant coefficients of the model.

[0031] S33: Using the daily NDVI values ​​obtained in step S31 as the baseline true values, the model accuracy is verified by the correlation coefficient R, root mean square error RMSE, and mean absolute error MAE.

[0032] As a further improvement of the present invention, the calculation formulas for the constant coefficients of each model in step S32 are as follows:

[0033]

[0034] in, This represents the inversion value obtained from the multiple linear regression model after data fusion; , , , , , , This represents the constant coefficients that need to be calculated when using different data models.

[0035] As a further improvement of the present invention, the specific implementation process of step S4 is as follows:

[0036] S41: Determine the core hyperparameters of the random forest regression model and use 10-fold cross-validation to screen the optimal hyperparameter combination;

[0037] S42: For different data combination schemes, the random forest model is independently trained using the optimal hyperparameter combination; when the input contains hyperspectral data, the model directly uses the original red light and near-infrared band surface reflectance.

[0038] S43: Using the daily NDVI obtained in step S31 as the baseline true value, verify the model accuracy through the correlation coefficient R, root mean square error RMSE, and mean absolute error MAE.

[0039] As a further improvement of the present invention, the core hyperparameters in step S41 include the number of decision trees, the maximum tree depth, and the minimum number of samples in the leaf nodes.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] (1) By fusing GNSS-IR and hyperspectral data, compared with the existing inversion methods that only use GNSS-IR or hyperspectral data, the method of the present invention significantly improves the inversion accuracy of NDVI, avoids the low inversion accuracy caused by the influence of various factors on GNSS-IR signals, and has a more obvious monitoring effect. The results show that the method of the present invention has smaller errors and higher accuracy.

[0042] (2) The method of the present invention solves the problem of observation failure caused by weather factors such as cloud cover in hyperspectral observation, and ensures the continuity of vegetation water content inversion. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0044] Figure 2 This is a schematic diagram of the device composition of the present invention;

[0045] Figure 3 The graph shows the correlation analysis of linear regression modeling of Amp, VIR_HS individually and their fusion with NDVI;

[0046] Figure 4 This is a correlation analysis diagram of random forest regression modeling using Amp, Red, and NIR individually, as well as their fusion and NDVI;

[0047] Figure 5 This is a comparative evaluation chart showing the degree of improvement of the multi-source fusion data model compared to the single-source data model;

[0048] Figure 6 The graphs show a comparison between the NDVI model values ​​retrieved from fused data based on multiple linear regression and random forest models, and the MODIS NDVI. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0050] The present invention will now be described in further detail with reference to the accompanying drawings:

[0051] A method for high-precision vegetation water content inversion by fusing GNSS-IR and hyperspectral data, according to Figure 2 A high-precision vegetation water content change monitoring device integrating GNSS-IR and hyperspectral data was constructed, as shown. A monitoring experiment was conducted in an experimental field in Zetou Town, Wendeng District, Weihai City, from September 10th to October 8th, 2024. Multiple systems of GNSS-IR and hyperspectral data were collected. Specific steps are as follows (refer to...). Figure 1 (as shown)

[0052] S1. Obtain amplitude characteristic indicators correlated with the Normalized Differential Vegetation Index (NDVI) through GNSS-IR observations. The specific implementation process is as follows:

[0053] S11: Set up GNSS-IR observation equipment to obtain low elevation angles. SNR observation data

[0054] The monitoring device includes: an upward-looking GNSS right-hand circularly polarized antenna and a geodetic GNSS receiver. The receiving antenna is erected within the observation plot at a height of 2.039m. Figure 1 The image shown in the upper left corner is of a hyperspectral camera;

[0055] The receiver stores the SNR (Signal Noise Ratio) of the interferometric signal synthesized from the direct and reflected signals, recorded in the receiver in RINEX 3.04 format;

[0056] S12: Calculate the satellite elevation and azimuth angles for each observation epoch based on the positions of the receiving antennas in the precise ephemeris and RINEX files, determine whether they meet the vegetation azimuth and elevation angle constraints, and filter out the signal-to-noise ratios that meet the conditions from the observation files. ;

[0057] S13: Raw signal-to-noise ratio expressed in decibels in logarithmic form. Convert to a linear value SNR in Volts / Volts:

[0058] ;

[0059] An approximate value for the direct signal is obtained by fitting the SNR using a quadratic polynomial. The difference between the SNR and this approximate value is then calculated; this difference signal represents the multipath reflection component related to the characteristics of the reflecting surface. , can be represented as:

[0060] ;

[0061] in, The amplitude of the reflected signal. The height of the receiver antenna relative to the reflector surface. For satellite signal wavelength, For GNSS satellite elevation angle, This represents the phase difference between the direct signal and the reflected signal.

[0062] S14: Through the Lomb-Scargle spectral analysis can be used to obtain the oscillation frequency of multipath signals. and the height of the receiver antenna relative to the reflector ,have:

[0063] ;

[0064] S15: Using IGGⅢ weighting function weighting / least squares method By fitting the SNR observations, the optimal amplitude for the fitting period of a single star is determined. .

[0065] Specifically, the IGGⅢ weighted least squares method employs a robust and robust estimation-based weight selection iteration method, using the IGGⅢ model to calculate equivalent weights. :

[0066] ;

[0067] In the formula, the subscript for Number, For observation epochs; for Weights of each observation; The residuals of the observed values; The standard error of the observations; Standardized error; For quantile parameters; This is the point of elimination. In this technology... It is 2. The weight is 4. This weighting function divides the observed data into valid information, usable information, and harmful information according to quality, thereby enabling the handling of excessively large errors. The points are re-weighted.

[0068] S16: Normalize the amplitude.

[0069] Before amplitude normalization, outliers were removed from the raw amplitude observations for each satellite using the median absolute difference (MAD) method. Normalization was then performed for each satellite.

[0070]

[0071]

[0072] in It is a constant (set to 2 in this technology). This represents the median of the amplitude observed by the satellite during the observation period.

[0073] Calculate the normalized amplitude of single-star observations Correlation analysis was performed using the Pearson correlation coefficient R between the sequence and the alignment value NDVI;

[0074] in ;

[0075] Satellites with R > 0.5 and more than 4 fitting points were selected as valid satellites. During the multi-satellite fusion process, the normalized amplitudes of all eligible satellites were summarized and the daily average amplitude was calculated as the final daily observation amplitude of GNSS-IR.

[0076] S2. Process hyperspectral data to obtain surface reflectance in the red and near-infrared (NIR) bands; the specific implementation process is as follows:

[0077] S21: To establish a hyperspectral observation facility in the same location as the GNSS-IR equipment, including a GaiaField Pro-V10 portable hyperspectral camera;

[0078] S22: After atmospheric correction and minimum noise separation of the original hyperspectral data, the surface reflectance data (Red and NIR) at 668nm and 845nm in the red and near-infrared bands are obtained by single-band calculation.

[0079] S3. A normalized differential vegetation index (NDVI) inversion model was established by fusing GNSS-IR and hyperspectral data using a multiple linear regression model, and the calculation accuracy was compared with that of the medium resolution imaging spectrometer (MODIS) data.

[0080] S31: Download MOD09GQ and MYD09GQ data from MODIS and calculate daily NDVI values ​​for the study area:

[0081] ;

[0082] Among them, band1 and band2 are the surface reflectance of red light and near-infrared bands, respectively.

[0083] S32: Combine the red reflectance (Red) and near-infrared reflectance (NIR) according to the calculation formula in step S31 to form VIR_HS. Using the amplitude (Amp), VIR_HS, and both obtained from daily GNSS-IR, construct an inversion model with MODIS NDVI using a multiple linear regression method, and calculate the model's constant coefficients:

[0084] ;

[0085] in, This represents the inversion value obtained from the multiple linear regression model after data fusion; , , , , , , , This represents the constant coefficients that need to be calculated when using different data fusion methods.

[0086] S33: To evaluate the method's performance, the daily NDVI obtained in step S31 was used as the baseline truth value to verify the inversion results of the three data models in step S32. Specifically, the following indicators were calculated for each of the three cases: correlation coefficient R, root mean square error RMSE, and mean absolute error MAE, thereby comprehensively verifying the effectiveness and accuracy of the method. The results are shown in Table 1.

[0087] The effectiveness of this method was verified by analyzing the results of a linear model established through an experiment that integrates GNSS-IR and hyperspectral data to achieve high-precision monitoring of vegetation water content changes. Figure 3The results of the correlation analysis of a linear regression model constructed using single-source Amp, single-source VIR_HS, multi-source fused GNSS-IR, and hyperspectral data as independent variables and NDVI as the target variable are presented. Different symbols are used in the figure to distinguish the input data types; the dashed line in the middle represents the ideal fit line, i.e., the theoretical fit when the NDVI model value is completely consistent with the MODIS NDVI reference value. The results show that the Pearson correlation coefficient between the single data source and NDVI is below 0.8, but the R-squared value increases to 0.8497 after fusion. The statistical results of the linear regression evaluation indicators R, RMSE, and MAE are shown in Table 1. The NDVI values ​​retrieved from amplitude inversion (dots) show a clear linear growth trend. This relationship is strongly supported by the positive correlation of the MODIS NDVI product (R = 0.7357) and the low overall error (RMSE = 0.0553, MAE = 0.0466). The VIR_HS model is a further optimization compared to the Amp model, achieving an R-value of 0.7883, with RMSE and MAE reduced to 0.0502 and 0.0408, respectively. The results based solely on Amp (dots) show significant deviations and dispersion from the ideal fitted line (y = x), indicating limited stability and accuracy. In contrast, the VIR_HS model (squares) shows a more tightly packed point distribution around the reference line, reflecting superior inversion performance. Notably, the fused Amp + VIR_HS model (triangles) demonstrates the best fit, with the most concentrated distribution of data points along the ideal line and minimal dispersion. The fused model achieves a high correlation coefficient (R = 0.8497) through multi-source collaboration, with RMSE and MAE optimized to 0.0430 and 0.0352, respectively, confirming the significant potential of multi-source data fusion in leveraging complementary advantages and improving the accuracy of vegetation parameter inversion. This intuitive assessment is strongly supported by quantitative indicators, confirming that multi-source data fusion effectively improves the accuracy and robustness of NDVI inversion.

[0088] S4. An NDVI inversion model is established by fusing GNSS-IR and hyperspectral data using a random forest regression model, and the accuracy is compared with that of the Moderate Resolution Imaging Spectroradiometer (MODIS) data; the specific implementation method is as follows:

[0089] S41: Model Construction and Hyperparameter Optimization. The core hyperparameters of the random forest regression model were determined, including the number of decision trees, maximum tree depth, and minimum number of samples per leaf node. Ten-fold cross-validation was used to optimize different combinations of hyperparameters. The mean squared errors of the ten-fold validations were summed and their average value was calculated as an indicator to evaluate the performance of the hyperparameter combinations, ultimately determining the optimal configuration. Table 2 shows the optimal hyperparameter combinations of the random forest model under different input variables and their evaluation metrics.

[0090] S42: Multi-source data input and model training. For three different data combination schemes, three random forest models are independently trained using the optimal hyperparameter combinations obtained in step S41. Unlike linear regression models, in this scheme, when the input includes hyperspectral data, the model directly uses the raw red and near-infrared band surface reflectance. This design stems from the random forest algorithm's inherent ability to handle nonlinear relationships and feature interactions, without relying on manually constructed linear combination features.

[0091] S41: Model Validation. Using the daily NDVI obtained in step S31 as the baseline truth value, the inversion results of the three data models in step S42 are validated. Specifically, the following indicators are calculated for each of the three cases: correlation coefficient R, root mean square error RMSE, and mean absolute error MAE, thereby comprehensively validating the effectiveness and accuracy of the method.

[0092] The analysis of the results of the random forest regression model established by the experiment of high-precision monitoring of vegetation water content changes by fusing GNSS-IR and hyperspectral data verifies the superiority of the machine learning random forest regression algorithm and re-emphasizes the effectiveness of multi-source data fusion. Figure 4The correlation analysis plots for random forest regression models, with single-source Amp, single-source Red and NIR, multi-source fused GNSS-IR, and hyperspectral data as independent variables and NDVI as the target variable, are shown in the figures. Different symbols are used to distinguish the input data types, and the dashed line in the middle represents the ideal fit line, i.e., the theoretical fit when the NDVI model value is completely consistent with the MODIS NDVI reference value. The model inversion results under different data inputs show relative performance consistent with linear regression: the inversion results of Amp (dots) are relatively scattered, while the fusion of Amp, NIR, and Red (triangles) has the best fit, with its data points tightly clustered along a 1:1 straight line. Specific relevant evaluation indicators are shown in Table 3: the combination of Amp + NIR + Red is the most accurate, followed by Red + NIR, while Amp alone performs the worst. Specifically, compared to the single Amp model (R=0.9424, RMSE=0.0280, MAE=0.0230), the Red + NIR combination improved the correlation coefficient (R=0.9697, RMSE=0.0237, MAE=0.0202) by 2.90%, while reducing the root mean square error (RMSE) and mean absolute error (MAE) by 15.36% and 12.17%, respectively. The Amp + NIR + Red fusion model achieved the best performance, with an R value of 0.9752, an RMSE of 0.0207, and an MAE of 0.0169, confirming its superiority among all evaluated combinations. These results demonstrate that the random forest regression method can significantly improve the accuracy of NDVI inversion and further enhance performance by capturing complex nonlinear relationships that linear models cannot fully reflect.

[0093] To evaluate the enhanced accuracy and robustness of multi-source data fusion in NDVI retrieval, Figure 5 compares the performance of the fusion model with that of a baseline model using only a single data source (GNSS-IR, hyperspectral). Dark and light colors represent the improvement in performance of each indicator based on different single-source data for the Random Forest Regression (RFR) and Multiple Linear Regression (MLR) models, respectively. The "AMP" column on the left indicates the improvement of the fusion model compared to using only GNSS-IR single-source data; the "HSI" column on the right indicates the improvement of the fusion model compared to using only hyperspectral single-source data. This figure demonstrates the relative improvements in R, RMSE, and MAE achieved by Random Forest Regression (RFR) and Multiple Linear Regression (MLR), clearly highlighting the value of data fusion. In the linear regression model, compared to single-source hyperspectral data, data fusion improved R, RMSE, and MAE by 7.79%, 14.34%, and 13.73%, respectively, indicating that data fusion brings comprehensive performance improvements. For the random forest regression model, the improvement in R² was relatively small (0.57%), but the reduction in error metrics was more significant, with RMSE and MAE decreasing by 12.66% and 16.34%, respectively. Compared to regression models using only single-source GNSS-IR Amp data, multi-source data fusion brought more significant improvements. In the linear regression model, the R² value increased significantly by 15.50%; for the random forest regression model, the R² value increased by 3.48%. In terms of inversion accuracy, the random forest regression model showed a more obvious advantage, with its root mean square error (RMSE) and mean absolute error (MAE) improving by 26.07% and 26.52%, respectively, exceeding the corresponding improvements of 22.24% and 24.46% achieved by linear regression. Furthermore, the above findings reveal the superior performance of the random forest algorithm in solving this inversion problem. Its advantages stem from the nonlinear nature of the algorithm, which can accurately capture complex patterns and effectively control prediction errors through an ensemble learning mechanism.

[0094] Figure 6 The charts show a comparison between the NDVI model values ​​retrieved from data fusion using multiple linear regression and random forest models, respectively, and the MODIS NDVI. The NDVI time series (dashed line) retrieved using the multiple linear regression model with fused data captures the overall trend, but shows significant deviations at certain time points (e.g., DOY 260, 267, and 273). In contrast, the random forest model (light solid line) using the same fused data shows a high degree of consistency with the measured NDVI, accurately retrieving its fine temporal variations. These observations intuitively confirm a key synergy: multi-source data fusion provides the necessary foundation for performance improvement, while the random forest algorithm is crucial for fully realizing this potential.

[0095] Table 1. Linear regression functions and evaluation indicators for Amp, VIR_HS, and Amp + VIR_HS with NDVI.

[0096]

[0097] Table 2. Optimal hyperparameter combinations and evaluation metrics of the random forest model under different input variables.

[0098] enter Number of decision trees Maximum depth of tree Minimum number of samples per leaf node Average MSE Amp 150 10 3 0.0017 Red+NIR 150 5 1 0.0031 Amp+Red+NIR 100 5 1 0.0018

[0099] Table 3 Evaluation metrics for establishing random forest regression models with NDVI using Amp, Red+NIR, and Amp+Red+NIR.

[0100] enter R RMSE MAE Amp 0.9424 0.0280 0.0230 Red+NIR 0.9697 0.0237 0.0202 Amp+NIR+Red 0.9752 0.0207 0.0169

[0101] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for high-precision inversion of vegetation water content indicators by integrating GNSS-IR and hyperspectral data, characterized in that, include: S1: Obtain the amplitude (Amp) characteristic index, which is correlated with the Normalized Differential Vegetation Index (NDVI), through GNSS-IR observations; S2: Processing hyperspectral data to obtain surface reflectance in the red and near-infrared (NIR) bands; S3: A normalized differential vegetation index (NDVI) inversion model was established by fusing GNSS-IR and hyperspectral data using a multiple linear regression model, and the calculation accuracy was compared with that of MODIS data from a medium resolution imaging spectrometer. S4: A normalized differential vegetation index (NDVI) inversion model was established by fusing GNSS-IR and hyperspectral data using a random forest regression model, and the calculation accuracy was compared with that of MODIS data from a medium resolution imaging spectrometer. The specific implementation process of step S1 is as follows: S11: Set up GNSS-IR observation equipment, including an upward-looking GNSS right-hand circularly polarized antenna and a GNSS receiver. Set up the antenna in the observation area, collect low-elevation-angle interferometric signal observation SNR data, and store it in the receiver. S12: Calculate the satellite elevation and azimuth angles for each observation epoch based on the precise ephemeris and the location of the receiving antenna in the stored files, and select the signal-to-noise ratios that meet the vegetation monitoring constraints. ; S13: Signal-to-noise ratio expressed in decibels in logarithmic form. The signal is converted into a linear SNR value in Volts / Volts units, and after fitting an approximation of the direct signal, the multipath reflection component is obtained. ; S14: Through the Lomb-Scargle spectral analysis was performed to obtain the oscillation frequency. Vertical height of the receiver antenna relative to the reflector surface ; S15: Use the IGGⅢ weighted least squares method to fit the SNR observations and solve for the optimal amplitude A for the fitting period of a single GNSS satellite. m ; S16: Outliers were removed by the absolute median difference method, the amplitude was normalized, and effective satellites with Pearson correlation coefficient R>0.5 and number of fitted points>4 were selected. Their normalized amplitudes were summarized and the daily average value was calculated as the daily observed amplitude Amp of GNSS-IR. In step S15: The IGGⅢ weighted least squares method is a weighted iterative method that employs robust robust estimation and calculates equivalent weights using the IGGⅢ model. The formula is as follows: In the formula, the subscript for Number, For observation epochs; for Weights of each observation; The residuals of the observed values; The standard error of the observations; Standardized error; For quantile parameters; This is the elimination point.

2. The method for high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data according to claim 1, characterized in that, The specific implementation process of step S2 is as follows: S21: Construct a hyperspectral observation device at the same location as the GNSS-IR observation site, including a portable hyperspectral camera; S22: Perform atmospheric correction and minimum noise separation processing on the raw hyperspectral data to extract the red light reflectance (Red) and near-infrared reflectance (NIR) at wavelengths of 668 nm and 845 nm, respectively.

3. The method for high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data according to claim 1, characterized in that, The specific implementation process of step S3 is as follows: S31: Download the MOD09GQ and MYD09GQ data from MODIS to calculate the daily NDVI values ​​for the study area. The formula is as follows: In the formula, band1 and band2 are the surface reflectance of the red light and near-infrared bands, respectively; S32: Combine the red reflectance (Red) and near-infrared reflectance (NIR) into VIR_HS according to the calculation formula in step S31. Use the amplitude (Amp), VIR_HS, and both obtained from daily GNSS-IR to construct an inversion model with MODIS NDVI using the multiple linear regression method, and calculate the constant coefficients of the model. S33: Using the daily NDVI values ​​obtained in step S31 as the baseline true values, the model accuracy is verified by the correlation coefficient R, root mean square error RMSE, and mean absolute error MAE.

4. The method for high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data according to claim 3, characterized in that, The formulas for calculating the constant coefficients of each model in step S32 are as follows: in, This represents the inversion value obtained from the multiple linear regression model after data fusion; , , , , , , This represents the constant coefficients that need to be calculated when using different data models.

5. The method for high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data according to claim 1, characterized in that, The specific implementation process of step S4 is as follows: S41: Determine the core hyperparameters of the random forest regression model and use 10-fold cross-validation to screen the optimal hyperparameter combination; S42: For different data combination schemes, the random forest model is independently trained using the optimal hyperparameter combination; when the input contains hyperspectral data, the model directly uses the original red light and near-infrared band surface reflectance. S43: Using the daily NDVI obtained in step S31 as the baseline true value, verify the model accuracy through the correlation coefficient R, root mean square error RMSE, and mean absolute error MAE.

6. The method for high-precision vegetation water content index inversion by fusing GNSS-IR and hyperspectral data according to claim 5, characterized in that, The core hyperparameters in step S41 include the number of decision trees, the maximum tree depth, and the minimum number of samples per leaf node.

Images