Ground surface soil humidity GNSS-R downscaling inversion method based on multi-model fusion

Abstract

The invention belongs to the technical field of microwave remote sensing, and provides an earth surface soil humidity GNSS-R downscaling inversion method based on multi-model fusion. Comprising the following steps: firstly, based on CYGNSS observation data, SMAP SM and auxiliary element data, constructing a training data set of a 36km grid and a test data set of a 3km grid through quality control; a limit gradient enhancement algorithm, a lightweight gradient elevator, a back propagation neural network and a random forest are adopted as base models, the training data set is divided randomly according to the proportion of 7: 3, and all the base models are constructed for training and verification; and finally, deeply learning a test result of each base model through a meta-learner (linear regression model), and automatically adjusting the weight of the test result to obtain a multi-model fused soil humidity downscaling inversion result. According to the method, the advantages of a plurality of base models are fully fused, and the soil humidity downscaling inversion precision can be effectively improved.

Description

Technical Field

[0001] The invention relates to a GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion, belonging to the technical field of microwave remote sensing. Background Art

[0002] In the fields of modern agriculture, environmental monitoring and geological survey, the monitoring of surface soil moisture is crucial. Soil moisture information plays an important role in understanding climate change, guiding agricultural irrigation, and predicting natural disasters. Traditional methods of monitoring surface soil moisture mainly rely on ground stations. Although these methods can provide high-precision measurement results, they are limited by the sparse distribution of stations and the maintenance cost of ground equipment. With the development of remote sensing technology, soil moisture monitoring using satellite remote sensing data has become an effective alternative method that can provide wider and more continuous monitoring coverage.

[0003] However, since remote sensing signals are affected by many factors such as atmospheric conditions, ground cover type and ground roughness, it is still a major challenge to directly extract accurate soil moisture information from remote sensing data. In addition, although Global Navigation Satellite System Reflectometry (GNSS-R) technology has been proven to be an effective means of remote sensing monitoring of soil moisture, its accuracy is low and it is difficult to meet the refined needs of monitoring in specific local areas. Therefore, improving the accuracy of inverting soil moisture information from GNSS-R data has become a research focus in this field.

[0004] In recent years, machine learning, as an important technology in the field of data processing and pattern recognition, has demonstrated its powerful capabilities in many scientific and engineering fields. By combining machine learning technology with remote sensing data, new possibilities are provided for improving the accuracy of soil moisture monitoring. In this context, the present invention provides an innovative GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion. This method can effectively improve the inversion accuracy of soil moisture information by comprehensively utilizing the advantages of different machine learning models, and provides an effective technical means for high-precision monitoring of soil moisture and related applications. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion, which includes the following steps: firstly, the Cyclone Global Navigation Satellite System (CYGNSS) observation data, active-passive remote sensing soil moisture (Soil Moisture Active and Passive, SMAP) data and auxiliary element data required for the experiment are obtained, the experimental data are quality controlled, and divided into a training data set corresponding to a 36km grid and a test data set corresponding to a 3km grid. Then, based on the training set, four machine learning methods, namely, eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Light Gradient Boosting Machine (LGBM) and Genetic Algorithm - Back Propagation (GA-BP) are used to construct training models respectively, and the parameters of the models are optimized. Finally, based on the trained base models, the soil moisture downscaling inversion results of each base model are obtained through the test set. At the same time, the meta-learner (linear regression model) is used to further learn the inversion results of the four base models, and their weights are automatically adjusted to obtain the final soil moisture downscaling inversion results. This method effectively improves the downscaling inversion accuracy of soil moisture by fully integrating the advantages of each base model.

[0006] To solve the above problems, the design technical solution of the present invention is as follows:

[0007] The above-mentioned GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion includes the following steps:

[0008] Step S1, collecting data sets, including spaceborne GNSS-R observation data, SMAP SM data and auxiliary element data; preprocessing the collected data sets to obtain sample data sets, and dividing them into training data sets corresponding to 36km grids and test data sets corresponding to 3km grids;

[0009] Step S2, randomly shuffle the training data set, use 70% as the modeling training set, and the remaining 30% as the modeling test set; construct four base models, XGBoost, RF, LGBM, and GA-BP, respectively, reasonably set the hyperparameters of XGBoost, RF, and LGBM, and use ten-fold cross-validation to verify the reliability of model training;

[0010] Step S3, downscaling and inverting soil moisture using the four trained base models respectively, and calculating the accuracy indexes of the four base models on the test set;

[0011] Step S4, using the linear regression model as a meta-learner, automatically adjusting its weight by learning the inversion results of the four base models, and training the meta-learner to obtain the final soil moisture downscaling inversion result;

[0012] Step S5, using the data from the in-situ soil moisture measuring station to test and evaluate the soil moisture downscaling inversion results mentioned above;

[0013] In the above-mentioned GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion, step S1 includes the following sub-steps:

[0014] Step S1.1, the characteristic parameters of the L1V2.0 version of the satellite-borne GNSS-R data: leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), surface reflectivity (SR), and auxiliary elements including terrain elements, vegetation elements, and meteorological elements are used as input features of the training set, and the SPL3SMP 8th version of the global daily 36km spatial resolution soil moisture product SMAP data is used as the output feature of the training set;

[0015] In step S1.2, the quality of the satellite-borne GNSS-R data is controlled and aggregated into the SMAP SM 36km grid together with the auxiliary element data by averaging method as the training sample set; at the same time, the characteristic parameters of the satellite-borne GNSS-R data and the auxiliary elements are resampled to 3km spatial resolution as the test set samples.

[0016] Based on the above-mentioned surface soil moisture GNSS-R downscaling inversion method based on multi-model fusion, the optimization of the hyperparameters in step S2 uses a grid search method to carefully adjust the above-mentioned four base models.

[0017] In the above-mentioned GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion, in step S3, the accuracy indicators include root mean square error, mean absolute error and correlation coefficient.

[0018] In the above-mentioned GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion, the adjustment of the weight in step S4 includes the following steps:

[0019] Step S4.1, training of the base learner: first, four base models are trained respectively using the training set, so as to obtain inversion models of the four base models;

[0020] Step S4.2, constructing a training data set for the meta-learner; using the inversion results of the base learner as input features of the meta-learner to construct a new training data set;

[0021] Step S4.3, training of the meta-learner: Use the new training data set to train the meta-learner, and use the meta-learner (linear regression model) to learn the inversion results of multiple base learners to determine the weight of each base learner, and then automatically adjust the weight of each base learner; the weight expression of the meta-learner is as follows:

[0022]

[0023] Where: is the weighted fusion value, is the number of models, Indicates The inversion results of the model are Indicates The weight of the model.

[0024] In the above-mentioned surface soil moisture GNSS-R downscaling inversion method based on multi-model fusion, the weights obtained by the meta-learner are used to weightedly fuse the inversion results of the base learner to obtain the final soil moisture downscaling inversion result.

[0025] Compared with the prior art, the GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion proposed in the present invention has the advantage of automatically adjusting its weight by learning the inversion results of the base model, thereby highlighting the advantages of each model while suppressing its weaknesses; at the same time, the inversion results of all base models are weighted and fused through the meta-model to form a soil moisture downscaling inversion fusion model, which can significantly improve the accuracy and generalization ability of the soil moisture downscaling inversion model. Compared with the traditional single machine learning model, the fusion inversion model provided by the present invention can more effectively handle complex nonlinear relationships, thereby maximizing the overall inversion performance while maintaining model diversity, which is of great significance for improving the accuracy of soil moisture monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Flow chart of the present invention

[0027] Figure 2 Scatter plot of the fusion model on the training set and test set.

[0028] Figure 3 Schematic diagram of the comparison between the soil moisture inverted by the base model and the fusion model and the number of in-situ soil moisture stations. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the present invention clearer, the embodiments disclosed in the present invention will be further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0030] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0031] The specific steps are as follows:

[0032] Step S1, collecting data sets, including spaceborne GNSS-R observation data, SMAP SM data and auxiliary element data; preprocessing the collected data sets to obtain sample data sets, and dividing them into training data sets corresponding to 36km grids and test data sets corresponding to 3km grids;

[0033] Step S2, randomly shuffle the training data set, use 70% as the modeling training set, and the remaining 30% as the modeling validation set; construct four base models, XGBoost, RF, LGBM, and GA-BP, respectively, reasonably set the hyperparameters of XGBoost, RF, LGBM, and GA-BP, and use ten-fold cross-validation to verify the reliability of model training;

[0034] Step S3, downscaling and inverting soil moisture using the four trained base models respectively, and calculating the accuracy indexes of the four base models on the test set;

[0035] Step S4, using the linear regression model as a meta-learner, automatically adjusting its weight by learning the inversion results of the four base models, and training the meta-learner to obtain the final soil moisture downscaling inversion result;

[0036] Step S5, using the data from the in-situ soil moisture measuring station to test and evaluate the soil moisture downscaling inversion results mentioned above;

[0037] As an implementation of this embodiment, step S1 includes the following sub-steps:

[0038] Step S1.1. The dataset refers to the GNSS data from January 1, 2019 to December 31, 2020 in the southern continental United States (24°N-38°N, -123°W-75°W), including SMAPSM data obtained from the National Snow and Ice Data Center (https: / / nsidc.org / ) and auxiliary element data obtained from the GEE platform; the collected datasets are preprocessed: the CYGNSS data in the study area are strictly screened to exclude data that do not meet the quality standards; and then secondary refinement processing is performed to ensure the overall reliability of the data.

[0039] Step S1.2: After quality control of the onboard GNSS-R data, the CYGNSS observation data and auxiliary element data are aggregated into the SMAP 36 km spatial resolution grid. At the same time, the CYGNSS observation data and auxiliary element data are resampled to a high spatial resolution of 3 km using the nearest neighbor method and subjected to spatiotemporal matching processing, thereby obtaining a training data set corresponding to the 36 km grid and an inversion sample set corresponding to the 3 km grid.

[0040] Step S1.3, the characteristic parameters of the satellite-borne GNSS-R data: leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), surface reflectivity (SR) and auxiliary factors: land cover type, elevation, vegetation index (NDVI) and rainfall are used as input features of the training set, and the SMAP soil moisture data are used as output features of the training set;

[0041] In the present invention, the parameter setting method in step S2 includes but is not limited to the grid search method. Before the grid search method is performed, the value ranges of the four machine learning model feature screening model parameters are preliminarily set according to the empirical values, and the parameter settings are as follows:

[0042] For the XGBoost model, the maximum depth of the tree was set to 8, the minimum child node weight was set to 1, the learning rate was set to 0.25, the number of trees was set to 100, the subsampling ratio was set to 0.9, the tree column sampling ratio was set to 0.6, and the minimum loss reduction of node splitting was set to 0; for the RF model, the number of trees was set to 100, the maximum depth was set to 6, the maximum number of leaf nodes was set to None, the minimum number of leaf samples was set to 1, and the minimum split sample tree was set to 2; for the LGBM model, the learning rate was set to 0.09, the number of trees was set to 100, the minimum gain split was set to 0.1, the maximum depth was set to 6, the maximum leaf tree was set to 50, the subsampling was set to 0.8, and the column sampling ratio was set to 0.8; for the GA-BP model, the population size was set to 50, the learning rate was set to 0.001, the number of iterations was set to 100, and the number of hidden layers was set to 10; in the preset parameter grid, every possible parameter combination was systematically enumerated; among all the parameter combinations tested, the parameter combination with the best performance in cross validation was selected as the optimal parameter. Using the determined optimal parameter combination, the model is retrained on all training data to obtain the final optimized model.

[0043] As an implementation method of this embodiment, the ten-fold cross-validation process of step S2 includes: in the ten-fold cross-validation, the training set is randomly divided into 10 subsets of equal size. For each model training and evaluation process, one of the subsets is retained as verification data (validation set), and the remaining 9 subsets are used for model training (training subsets). This process will be repeated 10 times (corresponding to 10 subsets), and a different subset is selected as the validation set each time to ensure that each subset has the opportunity to be used as validation data. Through ten-fold cross-validation, the evaluation results of the model performance can be obtained after each training. The average value of these 10 evaluation results is calculated as the accuracy index of the current model. For the four models of XGBoost, RF, LGBM, and GA-BP, the above ten-fold cross-validation process is performed respectively to obtain the average accuracy index of the four base models.

[0044] In the present invention, the accuracy indicators in step S3 include but are not limited to root mean square error, mean absolute error and correlation coefficient. In the training and testing process of RF, XGBoost, LGBM, and GA-BP soil moisture downscaling models, mean absolute error (MAE), correlation coefficient (R) and root mean square error (RMSE) are used as accuracy evaluation indicators of the inversion model, where the calculation formula of the mean absolute error MAE is as follows:

[0045]

[0046] Among them, O i is the actual observation value of the ith i is the i-th inversion value, and n is the total number of observations.

[0047] The calculation formula of root mean square error RMSE is as follows:

[0048]

[0049] Among them, O i is the actual observed value of the ith i is the i-th inversion value, and n is the total number of observations.

[0050] The calculation formula of the correlation coefficient R is as follows:

[0051]

[0052] in, and represent the average values ​​of actual observations and inverted values, respectively.

[0053] As an implementation of this embodiment, step S4 includes the following sub-steps:

[0054] Step S4.1, during the training and weight assignment process of the meta-learner, the weights of each base learner are determined by minimizing the variance; at the same time, the minimization loss function of the linear regression model is:

[0055] L(w)=

[0056] Among them, m is the number of samples, is the true value of each j-th sample, n is the number of base learners, For the The weights of the base learners, .

[0057] By solving the minimum value of the loss function, the weight vector of the meta-learner can be obtained:

[0058]

[0059] Among them, H is the inversion result matrix of the base learner, is a real value vector.

[0060] The prediction results of the base learner and the weights of the meta learner are weighted fused to obtain the final inversion result of the meta learner:

[0061]

[0062] in, is the weighted fusion value, is the number of models, Indicates The inversion results of the model are Indicates The weight of the model.

[0063] Step S4.2, after weighted fusion, the linear regression model is used as a meta-learner, the weights are automatically adjusted by learning the inversion results of the base learner, and the meta-learner is trained to obtain a soil moisture inversion fusion model;

[0064] The present invention uses four in-situ soil moisture station monitoring networks, namely SCAN, USCRN, SNOTEL and ARM, to compare and evaluate the inversion accuracy of the soil moisture inversion model; pre-processes the collected in-situ station data, including data cleaning, outlier processing and interpolation, to ensure the quality and availability of the data, and performs spatiotemporal matching between the results of the fusion model and the in-situ station data. The accuracy of the fusion model is comprehensively evaluated using three indicators, namely, mean absolute error (MAE), correlation coefficient (R) and root mean square error (RMSE). Among them: MAE reflects the average absolute deviation between the inversion value and the true value, and a smaller MAE indicates a higher inversion accuracy. R measures the correlation between the inversion value and the true value. The closer the R value is to 1, the stronger the correlation is and the better the inversion performance is. RMSE reflects the square root of the average deviation between the inversion value and the true value. A smaller RMSE indicates a smaller error and a higher model inversion accuracy.

[0065] In order to verify the effectiveness of the present invention, the soil moisture downscaling inversion results mentioned above were tested and evaluated using the data from the in-situ soil moisture station; a schematic diagram of the inversion accuracy of the machine learning model and the fusion model was obtained (see Figure 2). Figure 2 ), a schematic diagram comparing the number of soil moisture inverted by the base model and the fusion model with the number of in-situ soil moisture stations (as shown in Figure 3 As shown); The experimental results show that the fusion model can accurately reflect the changing trend of soil moisture at most in-situ stations; in addition, most in-situ stations show low absolute error and root mean square error. It can be seen that compared with a single model, the inversion method based on multi-model fusion can significantly improve the inversion accuracy of soil moisture, especially under complex surface conditions, its superiority is more obvious.

Claims

1. A GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion, comprising the following steps: Step S1, collecting data sets, including spaceborne GNSS-R observation data, SMAP SM data and auxiliary element data; preprocessing the collected data sets to obtain sample data sets, and dividing them into training data sets corresponding to 36km grids and test data sets corresponding to 3km grids; Step S2, randomly shuffle the training data set, use 70% as the modeling training set, and the remaining 30% as the modeling test set; construct four base models, XGBoost, RF, LGBM, and GA-BP, respectively, reasonably set the hyperparameters of XGBoost, RF, and LGBM, and use ten-fold cross-validation to verify the reliability of model training; Step S3, downscaling and inverting soil moisture using the four trained base models respectively, and calculating the accuracy indexes of the four base models on the test set; Step S4, using the linear regression model as a meta-learner, automatically adjusting its weight by learning the inversion results of the four base models, and training the meta-learner to obtain the final soil moisture downscaling inversion result; Step S5: Use the data from the in-situ soil moisture measuring station to verify and evaluate the soil moisture downscaling inversion results mentioned above.

2. According to the method of GNSS-R downscaling inversion of surface soil moisture based on multi-model fusion according to claim 1, it is characterized in that: Step S1 includes the following sub-steps: Step S1.1, the characteristic parameters of the L1V2.0 version of the satellite-borne GNSS-R data: leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), surface reflectivity (SR), and auxiliary elements including terrain elements, vegetation elements, and meteorological elements are used as input features of the training set, and the SPL3SMP 8th version of the global daily 36km spatial resolution soil moisture product SMAP data is used as the output feature of the training set; In step S1.2, the quality of the satellite-borne GNSS-R data is controlled and aggregated into the SMAP SM 36km grid together with the auxiliary element data by averaging method as the training sample set; at the same time, the characteristic parameters of the satellite-borne GNSS-R data and the auxiliary elements are resampled to 3km spatial resolution as the test set samples.

3. The method for surface soil wet GNSS-R downscaling inversion based on multi-model fusion according to claim 1 is characterized in that: The hyperparameters in step S2 are adjusted using a grid search method.

4. The GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion according to claim 1 is characterized in that: In step S3, the accuracy indicators include root mean square error, mean absolute error and correlation coefficient.

5. The method for GNSS-R downscaling inversion of surface soil moisture based on multi-model fusion according to claim 1 is characterized in that: The weight adjustment in step S4 includes the following steps: Step S4.1, training of the base learner: first, four base models are trained respectively using the training set, so as to obtain inversion models of the four base models; Step S4.2, constructing a training data set for the meta-learner; using the inversion results of the base learner as input features of the meta-learner to construct a new training data set; Step S4.3, training of the meta-learner: Use the new training data set to train the meta-learner, and use the meta-learner (linear regression model) to learn the inversion results of multiple base learners to determine the weight of each base learner, and then automatically adjust the weight of each base learner; the weight expression of the meta-learner is as follows: , where is the weighted fusion value, is the number of models, Indicates The prediction results of the model are Indicates The weight of each model; Step S4.4, use the weights obtained by the meta-learner to perform weighted fusion on the inversion results of the base learner, so as to obtain the final soil moisture downscaling inversion result.

6. The GNSS-R downscaling inversion method for surface soil moisture based on multi-model fusion according to claim 1 is characterized in that: The data from in-situ soil moisture stations are used to verify and evaluate the soil moisture downscaling inversion results mentioned above.

Images