Satellite-borne microwave scatterometer snow depth inversion method based on machine learning and feature screening strategy

Abstract

The invention discloses a satellite-borne microwave scatterometer snow depth inversion method based on machine learning and a feature screening strategy, and belongs to the technical field of satellite remote sensing image processing and application. Active microwave scatterometer data are utilized, data such as accumulated snow density, surface temperature, elevation, longitude, latitude and land coverage types are combined, and by analyzing and counting correlation between input characteristics and snow depth, a strategy of combining correlation coefficient characteristic screening and ablation experiment characteristic screening is provided to obtain optimal input characteristics; and through evaluation of the machine learning model, an optimal snow depth inversion input feature and machine learning model combination of the research region can be obtained, and finally, a high-precision snow depth inversion result of the research region is obtained by using the combination.

Description

Technical Field

[0001] The invention belongs to the technical field of satellite remote sensing image processing and application. Background Art

[0002] Microwave refers to electromagnetic waves with a frequency of 0.3 to 300 GHz (wavelength 1 mm to 1 m). The working wavelength of microwave sensors is in the microwave spectrum region. By receiving microwave signals reflected by different objects, relevant information of the target object is extracted to analyze the relevant parameter characteristics. Compared with traditional visible light or infrared remote sensing, microwave remote sensing can operate both during the day and at night, and different weather conditions will not affect its normal operation. It can penetrate clouds and fog, reducing the impact of environmental factors on the detection results. This penetration has great advantages in detecting the physical shape and surface roughness of the target objects. Microwave remote sensing can be divided into active microwave remote sensing and passive microwave remote sensing according to the working mode. Active microwave remote sensing itself emits microwave signals. After the target object receives the microwave signal, it reflects or scatters it to form an echo, that is, backscattering. Active microwave remote sensing receives the echo signal. It mainly includes side-looking radar, synthetic aperture radar (SAR), radar altimeter, microwave scatterometer, etc. Passive microwave remote sensing does not emit microwave signals itself, but only receives echo signals reflected or scattered by natural microwaves such as sunlight from the target object for remote sensing monitoring. It mainly includes microwave radiometers. At present, large-scale snow depth parameter monitoring mainly uses passive microwave radiometers and active microwave scatterometers with low spatial resolution but high temporal resolution (see Tanniru S. Passive microwave remote sensing of SD: Techniques, challenges and future directions. Remote Sensing 15(4):1052. and Feng T. Estimating SD based on dual polarimetric radar index from Sentinel-1GRD data:A case study in the Scandinavian Mountains. International Journal of Applied Earth Observation and Geoinformation 130:103873.).

[0003] Microwave scatterometer is an active microwave remote sensor that reflects the scattering characteristics of the measured target object through the backscattering coefficient. The scatterometer transmits microwave signals to the target object and receives its backscattered echo signal to detect the information of the target object. The surface roughness of the target object and its own dielectric properties determine the strength of the echo signal. Both can be used as the basis for describing the characteristics of covering objects such as snow, sea ice and vegetation. Microwave scatterometer has low spatial resolution but high temporal resolution. It is generally used to study the backscattering characteristics of large-scale surfaces. It was originally used to measure the wind vector field of the global sea surface. Later, with the emergence of various different needs, it was also used for the study of snow, ice sheet, land, etc. In recent years, with the rapid development of machine learning technology, algorithms such as naive Bayes, logistic regression (LR), decision tree (DT), random forest (RF) and support vector machine (SVM) have shown great potential in the field of snow depth inversion. Different from the traditional physical model snow inversion method, machine learning methods can significantly improve the accuracy and reliability of snow depth inversion by simulating complex nonlinear relationships, especially in dealing with complex terrain and heterogeneous snow characteristics. In 2015, David Buckingham et al. proposed a low-cost, lightweight method to estimate the average snow water equivalent of the entire basin in real time using existing infrastructure, wireless sensor networks and machine learning algorithms (see Buckingham D. Inductive ML for improved estimation of catchment-scale snow water equivalent. Journal of Hydrology 524: 311-325.). In 2022, Nasim S et al. proposed a wearable IoT platform that uses machine learning models to estimate and classify snow depth categories, where the results of the random forest classifier showed an accuracy of 94% (see Nasim SMML model for SD estimation using a multisensory ubiquitous platform. Journal of Mountain Science 19 (9): 2506-2527.). In 2019, Yang Jianwei et al. studied the potential ability of random forest (RF) models in machine learning (ML) models to estimate snow depth (see Yang J. SD estimation and historical data reconstruction over China based on a RF ML approach. The Cryosphere, 14 (6): 1763-1778.).In 2021, Yang Jianwei et al. used the coupling microwave snow emission pattern and random forest (RF) technology to invert the snow depth in China (see Yang J. Comparison of ML-based SD estimation and development of a new operational retrieval algorithm over China. Remote Sensing 14 (12): 2800.). In 2022, Xiong Chuan et al. used optical and passive microwave remote sensing combined with the XGBoost machine model algorithm to invert the snow depth in mountainous areas (see Xiong C. Mountain SD retrieval from optical and passive microwave remote sensing using ML. IEEE Geoscience and Remote Sensing Letters 19: 1-5.). In 2023, Wei Yanlin et al. used the RF model to generate a snow cover fraction (NCFSC) product with a spatial resolution of 500 meters from 1980 to 2020. The comparison of NCFSC and MODIS FSC data (MDFSC) from 2015 to 2018 shows that NCFSC products are reliable and highly comparable (see Wei Y. Significant decreasing trends in snow cover and duration in Northeast China during the past 40 years from 1980 to 2020. Journal of Hydrology 626: 130318). At present, the inversion of snow cover parameters at home and abroad mainly uses passive microwave remote sensing data, and the accuracy of snow depth inversion results needs to be further improved.

[0004] Since the snow depth inversion results are mainly affected by the environmental characteristics of the observation area and snow-related characteristics, where environmental characteristics refer to geographical environment information that does not change with the changes in snow characteristics, such as elevation, longitude, latitude and land cover type, and snow-related characteristics refer to information that is greatly affected by changes in snow depth, such as surface temperature, snow density, backscattering coefficient, etc., and the correlation between the above data as input features and snow depth is not the same, reasonable input feature screening can effectively improve the performance and generalization ability of machine learning models, thereby improving the accuracy of snow depth inversion. Summary of the invention

[0005] In order to solve the above problems and improve the accuracy of snow depth inversion, the present invention proposes a satellite microwave scatterometer snow depth inversion method based on machine learning and feature screening strategy. It uses active microwave scatterometer data and combines data such as snow density, surface temperature, elevation, longitude, latitude and land cover type. By analyzing and counting the correlation between input features and snow depth, a strategy of joint correlation coefficient feature screening and ablation experiment feature screening is proposed to obtain the optimal input features. Through the evaluation of the machine learning model, the optimal snow depth inversion input features and machine learning model combination in the study area can be obtained. Finally, the combination is used to obtain high-precision snow depth inversion results for the study area.

[0006] The specific steps of the technical solution adopted by the present invention are as follows:

[0007] Step 1: Data acquisition and preprocessing:

[0008] (1) Obtain the following data:

[0009] a. Obtain the satellite data of the L1B level of the microwave scatterometer of the Ocean No. 2B satellite (HY-2B), split each L1B level data according to the different incident angles, and generate the daily HH polarization scattering coefficient and HH polarization scattering coefficient; then, use the HH polarization scattering coefficient and VV polarization scattering coefficient to calculate the dual polarization scattering coefficient polarization ratio (Ratio of Polarization, PR), polarization sum (Sum of Polarization, PS), polarization difference (Difference of Polarization, PD) and polarization product (Multiply of Polarization, PM);

[0010] Sigma PR =Sigma VV / Sigma HH (1)

[0011] Sigma PM =Sigma VV ×Sigma HH (2)

[0012] Sigma PS =Sigma VV +Sigma HH (3)

[0013] Sigma PD =Sigma VV -Sigma HH (4)

[0014] Sigma in formula (1) to formula (4) VVrepresents the VV polarization scattering coefficient, Sigma HH represents the HH polarization scattering coefficient, Sigma PR Represents the polarization ratio of the dual-polarization scattering coefficient, Sigma PM Represents the polarization product of the dual-polarization scattering coefficient, Sigma PS represents the polarization sum of the dual-polarization scattering coefficients, Sigma PD represents the polarization difference of dual-polarization scattering coefficient;

[0015] b. Obtain the snow density data from the "Daily Snow Density Dataset in China (2013-2020)" provided by the National Cryosphere and Desert Data Center;

[0016] c. Obtain the surface temperature data from the East Asia 0.02° Hourly Seamless Surface Temperature Dataset (2016-2021);

[0017] d. Obtain the coverage and distribution of land cover types in the study area from the Globeland30 data;

[0018] e. Obtain elevation data from the SRTM digital elevation model (DEM) 90M dataset.

[0019] (2) The above data are cropped, spliced ​​and resampled according to the selected research area to obtain a multi-feature dataset with the same spatial resolution.

[0020] Step 2: Create a snow depth inversion dataset:

[0021] The snow depth data measured by ground observation stations are used as label data. This dataset includes the station number, station level information, and longitude, latitude, altitude, year, month, day, and snow depth attributes of each ground observation station. A portion of observation stations in the study area are randomly selected as verification stations, and the remaining stations are used as training and testing stations. The verification stations are divided into verification stations of different land cover types according to the area ratio of the land cover type.

[0022] The snow density, longitude, latitude, average scattering coefficient (SigmaAve), VV polarization scattering coefficient (SigmaVV), HH polarization scattering coefficient (SigmaHH), dual polarization scattering coefficient polarization ratio (SigmaPR), dual polarization scattering coefficient polarization sum (SigmaPM), dual polarization scattering coefficient polarization sum (SigmaPS), dual polarization scattering coefficient polarization difference (SigmaPD), surface temperature and elevation (DEM) corresponding to the geographical location of the ground observation station are extracted as the initial input, and the measured snow depth value is used as the output to obtain the snow depth inversion data set. The calculation formula of the average scattering coefficient (SigmaAve) is as follows:

[0023]

[0024] (5) where SigmaAve represents the average scattering coefficient, n is the number of orbits that scan the study area on that day, and Sigma i Represents the backscatter coefficient for each swept track data.

[0025] Step 3: Feature screening:

[0026] (a) Correlation coefficient feature screening:

[0027] (1) Calculate the three correlation coefficients of snow depth and other 12 characteristics: Pearson correlation coefficient PCC, Spearman correlation coefficient SCC, and Kendall correlation coefficient KCC;

[0028] The calculation methods of Pearson correlation coefficient (PCC), Spearman correlation coefficient (SCC) and Kendall correlation coefficient (KCC) are as follows:

[0029]

[0030] (6) Where PCC represents the Pearson correlation coefficient, and denote the average values ​​of Y and X, respectively, and n is the number of features used for evaluation; (7) where SCC denotes the Spearman correlation coefficient, di denotes the difference in rank values ​​of the i-th data pair, and n is the number of features used for evaluation; (8) where KCC denotes the Kendall correlation coefficient, C denotes the number of pairs of elements in XY that are consistent (two elements form a pair), and D denotes the number of pairs of elements in XY that are inconsistent.

[0031] (2) Calculate the average correlation coefficient ACC (Average correlation coefficient) between snow depth and each feature

[0032] ACC=(PCC+SCC+KCC) / 3.

[0033] (3) Sort the average correlation coefficient ACC from high to low, retain the N features with the highest average correlation coefficient ACC, and finally discard the features with ACC less than M among the N features, and retain the remaining features, where M and N are obtained through experiments and the significance of the correlation coefficient.

[0034] (b) Ablation experiment feature screening: In this step, the random forest algorithm is used to invert the snow depth data, and the ratio of the training set data to the test set data is 8:2;

[0035] (1) First, the N features obtained by screening in step (a) are used as the initial input features to obtain the snow depth inversion result A0. The N features are eliminated in sequence to obtain N groups of N-1 input features, each of which has different input features. The inversion results A1 to A1 are obtained by training them respectively. N ;

[0036] (2) A1~A N Compared with A0, if the evaluation index of the snow depth inversion result of a certain group compared with A0 is lower, it means that the corresponding feature quantity removed has a positive effect on the snow depth inversion and is retained; if the evaluation index of the snow depth inversion result of a certain group compared with A0 remains unchanged or increases, it means that the corresponding feature quantity removed has no effect on the snow depth inversion and may even reduce the snow depth inversion accuracy. The number of these features is recorded as Na, and they are all discarded, and N=N-Na;

[0037] (3) Repeat the above steps until N no longer changes. The features finally retained are the optimal input features and are used as the input features for subsequent snow depth inversion experiments.

[0038] Step 4: Snow depth inversion based on optimal features and machine learning:

[0039] The features finally retained after feature screening in step 3 were used as input features. Four machine learning regression models, namely support vector regression model (SVR), K-nearest neighbor algorithm model (KNN), extreme gradient boosting model (XGBoost) and random forest model, were used to invert snow depth. The snow depth results of the study area were obtained. The root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ), and select the machine learning regression model with the best results.

[0040] Step 5: Take the features finally retained after feature screening in step 3 as input features, input them into the machine learning regression model with the best results screened out in step 4, and invert to obtain the final snow depth value.

[0041] Beneficial effects of the present invention:

[0042] The present invention uses satellite-borne microwave scatterometer data and proposes a strategy of combining correlation coefficient feature screening and ablation experiment feature screening to perform optimal feature screening based on the correlation between multi-parameter features and snow depth in the study area, and combines the machine learning model with the ablation experiment feature screening strategy to perform snow depth inversion. By making full use of the relationship between snow depth and features, and the relationship between features, the optimal input features are obtained by combining the machine learning model with the ablation experiment feature screening strategy, and high-precision snow depth inversion is achieved in the study area. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1It is a flow chart of the snow depth inversion method of spaceborne microwave scatterometer based on machine learning and feature screening strategy.

[0044] Figure 2 It is the flow chart of the correlation coefficient feature screening algorithm (MCCFS).

[0045] Figure 3 This is the flow chart of feature screening for ablation experiments.

[0046] Figure 4 It is a comparison chart of the M0 mode test set results in the N1=10, N2=7 and N3=6 experiments.

[0047] Figure 5 It is a comparison chart of snow depth inversion results after inputting original features, features after correlation coefficient screening, and features after ablation experiment screening.

[0048] Figure 6 These are the snow depth inversion results and fitting curves of the training and test sets of the four models.

[0049] Figure 7 This is a comparison chart of the snow depth inversion results of the training set and test set of the RF model and passive microwave snow product. DETAILED DESCRIPTION

[0050] The specific steps of the embodiment are as follows:

[0051] Step 1: Data collection and preprocessing

[0052] (a) The present invention uses the snow depth data measured by ground observation stations as label data. This data set includes the station number, station level information, and attributes such as longitude, latitude, altitude, year, month, day, and snow depth of each ground observation station.

[0053] (b) Satellite data uses the L1B-level data of the Ocean No. 2B satellite (HY-2B) microwave scatterometer. The HY-2B microwave scatterometer operates in the Ku band and has two inner and outer pencil beams. The incident angles of the inner and outer beams are 41°, corresponding to HH polarization; the incident angles of the outer beams are 48°, corresponding to VV polarization. The operating frequency is 13.256GHz, the spatial resolution is 25KM, and the data can be downloaded from https: / / osdds.nsoas.org.cn / OceanDynamics. The L1B-level data used in the present invention is a 13000×96 size HH polarization and VV polarization fusion data, which includes HH polarization and VV polarization scattering coefficients, so it is necessary to split each L1B-level data according to the different incident angles to generate daily HH polarization scattering coefficients and VV polarization scattering coefficients.

[0054] The present invention performs polarization calculations on the basis of the dual-polarization scattering coefficients to obtain new eigenvalues ​​as preliminary feature inputs for subsequent experiments. This section defines the calculations of the polarization ratio (PR), the sum of polarization (PS), the difference of polarization (PD), and the multiply of polarization (PM) as follows:

[0055] SigmaPR=SigmaVV / SigmaHH (1)

[0056] SigmaPM=SigmaVV×SigmaHH (2)

[0057] SigmaPS=SigmaVV+SigmaHH (3)

[0058] SigmaPD=SigmaVV-SigmaHH (4)

[0059] (c) The snow density dataset comes from the “Daily Snow Density Dataset for China (2013-2020)” provided by the National Cryosphere and Desert Data Center (http: / / www.ncdc.ac.cn). This dataset includes snow density data at 25 km per day over the land area of ​​China. It combines ground survey data and ground station observation data from 2017 to 2019, as well as meteorological observation data from Chinese meteorological stations from 2013 to 2020. In addition, it also combines satellite remote sensing snow products, terrain factors, and meteorological elements obtained from reanalysis data. The dataset is generated using a spatiotemporal simulation model of snow density by using a spatiotemporal weighted integration method combined with machine learning techniques.

[0060] (d) The land surface temperature dataset is from the “East Asia 0.02° Hourly Seamless Land Surface Temperature Dataset (2016-2021)” (see Dong. A Data Fusion Method for Generating Hourly Seamless LandSurface Temperature from Himawari-8 AHI Data. Remote Sensing 14, 5170. and Zhou. An Improved Temperature and Emissivity Separation Algorithm for the Advanced Himawari Imager. IEEE Transactions on Geoscience and Remote Sensing 58(10), 7105-7124.) (https: / / poles.tpdc.ac.cn / zh-hans / data / ). This dataset uses the iTES algorithm to collect land surface temperature from Himawari-8 / AHI satellite data.

[0061] (e) The land cover dataset comes from Globeland30 data, which provides information on land cover types. By using land cover classification data, the coverage and distribution of land cover types in Northeast China can be obtained. The spatial resolution of the land cover dataset used in this study is 30 meters (http: / / www.globallandcover.com / ).

[0062] (f) Elevation data comes from the SRTM Digital Elevation Model (DEM) 90M dataset. This dataset provides geospatial information about elevation features. SRTM has successfully collected elevation data for more than 80% of the Earth's surface. The spatial resolution of the SRTM DEM 90M dataset is 90 meters (https: / / EarthExplorer.usgs.gov).

[0063] The present invention selects farmland and forest areas in Northeast China (39°-54°N, 116°-135°E) as the experimental area, covering Heilongjiang, Jilin, Liaoning, Inner Mongolia, etc. The above-mentioned datasets were clipped in this experimental area, and the spatial resolution of all data was sampled to the same resolution of 0.25° as the microwave scatterometer.

[0064] Step 2: Create a snow depth inversion dataset

[0065] The time range selected in this invention is from December 1, 2019 to March 31, 2020. The ground observation stations across the country are screened. There are a total of 108 ground observation stations in the selected study area, most of which are located in forest and farmland areas. A relatively scattered part of the observation stations is randomly selected as verification stations, and the remaining stations are used as training and testing stations. Finally, 100 stations are used for training and testing, and 8 stations are used for verification, with a ratio of 8:2. Among the 8 verification stations, 4 stations are forest verification stations, and the other 4 are farmland verification sites. The specific distribution is as follows Figure 2 A total of 12 eigenvalues, including snow density, longitude, latitude, average scattering coefficient (SigmaAve), VV polarization scattering coefficient (SigmaVV), HH polarization scattering coefficient (SigmaHH), dual-polarization scattering coefficient polarization ratio (SigmaPR), dual-polarization scattering coefficient polarization sum (SigmaPM), dual-polarization scattering coefficient polarization sum (SigmaPS), dual-polarization scattering coefficient polarization difference (SigmaPD), surface temperature and elevation (DEM), corresponding to the geographical locations of 108 measured stations, were extracted as initial inputs, and the measured snow depth value was used as output.

[0066] Step 3: Feature Screening Strategy

[0067] Feature screening is important in machine learning. It refers to selecting the most relevant and informative features from the original features, which helps to reduce the data dimension and remove redundant and irrelevant features, thereby simplifying the model, reducing the possibility of overfitting, and improving the generalization ability of the model.

[0068] (a) The present invention first proposes a combined correlation coefficient feature screening (MCCFS) for preliminary screening. The specific flow chart of the MCCFS algorithm is as follows: Figure 2 The correlation coefficient evaluation between features usually includes Pearson correlation coefficient (PCC), Spearman correlation coefficient (SCC) and Kendall correlation coefficient (KCC), which are calculated as follows:

[0069]

[0070] (1) The three correlation coefficients mentioned above are calculated respectively between snow depth and the other 12 characteristics: Pearson correlation coefficient PCC, Spearman correlation coefficient SCC, and Kendall correlation coefficient KCC.

[0071] (2) Calculate the average correlation coefficient ACC (Average correlation coefficient) between snow depth and each feature;

[0072] ACC=(PCC+SCC+KCC) / 3

[0073] (3) Sort the average correlation coefficient ACC from high to low, retain the 10 features with the highest ACC, and finally discard the features with an average correlation coefficient ACC less than 0.1 among the 10 features, and retain the remaining features.

[0074] Finally, it was found that snow density had the highest correlation with snow depth, with an average correlation coefficient of more than 0.6, followed by longitude, latitude, Sigma AVE Sigma HH Sigma VV Sigma PR Sigma PS , surface temperature and altitude, Sigma PM and Sigma PD Features are discarded.

[0075] (b) The correlation coefficient expresses the correlation between a single feature and snow depth, while the relationship between features cannot be reflected in the above feature screening. Some features that have a small effect on the machine learning snow depth inversion model may overlap with the effects of features that have a large effect on the model. Some features may even reduce the positive impact of features that have a large effect on the model. Therefore, it is necessary to combine ablation experiments to remove the features that have repeated effects on the model, and finally reduce feature redundancy and obtain the best input features. The ablation experiment in this step uniformly uses the random forest algorithm, and the ratio of the number of training set data to the test set data is 8:2. The specific process is as follows: Figure 3 , the initial value of N is 10.

[0076] (1) Ablation experiment feature screening The initial input features consist of 10 features obtained from the MCCFS algorithm. N1 = 10, the input features are snow density, longitude, latitude, Sigma AVE Sigma HH Sigma VV Sigma PR Sigma PS , surface temperature and elevation, named ①-⑩ in turn. Table 1 shows different feature combinations:

[0077] Table 1 N1 = 10 different feature combinations

[0078]

[0079] The comparison results of the test set with N1=10 show that the evaluation indicators of M6, M7, M8 and M0 are roughly the same. This observation shows that the feature Sigma HH Sigma PR and Sigma PS The impact on the snow depth inversion result is small. Therefore, these features can be excluded and N2 = N1-3.

[0080] (2) N2 = 7, input features are snow density, longitude, latitude, Sigma AVE Sigma VV , surface temperature and elevation, named ①-⑦ in turn. Table 2 shows different feature combinations:

[0081] Table 2 N2 = 7 different feature combinations

[0082]

[0083] The comparison results of the test set with N2=7 show that the snow depth inversion results of M4 are better than those of M0. AVE The impact on the snow depth inversion result is small, and removing this feature can even improve the accuracy of snow depth inversion. AVE can be eliminated, N3=N2-1.

[0084] (3) N3 = 6, input features are snow density, longitude, latitude, Sigma VV , surface temperature and elevation, named ①-⑥ in sequence. Table 3 shows different feature combinations:

[0085] Table 3 N3=6 different feature combinations

[0086]

[0087] The comparison of the test set with N3=6 shows that the evaluation indicators of snow depth inversion in M1 to M6 are not better than those in M0. VV , longitude, latitude, snow density, surface temperature and elevation can improve the snow inversion results. The effects of improving snow inversion accuracy from high to low are: snow density, surface temperature, Sigma VV , latitude, elevation and longitude. Therefore, these 6 features are determined as the final input features of the model.

[0088] The comparison of the M0 test set results in the three experiments with N1=10, N2=7 and N3=6 is shown in the figure. Figure 4 shown.

[0089] (4) Snow depth inversion was performed using the original features, the features obtained by MCCFS, and the features obtained by the ablation experiment. The results of the training set and the test set are shown in the figure below. Figure 5As shown in the figure. When the input in the test set is the original feature, the RMSE and MAE of the snow depth inversion result are 3.72cm and 1.98cm respectively; after MCCFS, the corresponding indicators are slightly improved, and the RMSE and MAE are 3.70cm and 1.97cm respectively; finally, feature selection is performed through ablation experiments, and the RMSE, MAE and R2 values ​​are 3.60cm, 1.88cm and 0.85 respectively, indicating that it is necessary to perform feature selection through designed ablation experiments.

[0090] Step 4: Snow depth inversion based on optimal features and machine learning

[0091] In the present invention, the input features are the features finally retained after the MCCFS and ablation experiment feature screening in step three, and the four machine learning regression models of SVR, KNN, XGBoost and random forest (RF) are used to perform snow depth inversion to obtain the snow depth results of the study area. Among them, SVR and KNN belong to single machine learning algorithms, and random forest and XGBoost belong to integrated machine learning algorithms, which complete the initial learning task by building and combining multiple learners. The integrated machine learning algorithm is divided into Bagging and Boosting according to whether there is a dependency relationship between each classifier. There is no dependency between bagging classifiers, and they can be run in parallel. This type of algorithm uses the Bootstraping method to extract n training samples from the original sample set, and performs k rounds of extraction to obtain k training sets. These k training sets are used for training respectively to obtain k models. Finally, under the premise of ensuring that the weight of each model is the same, these k models are used to produce results together. The most representative algorithm of bagging is random forest. Boosting requires that there is a mutual dependence between each classifier and that the learners must be serial. This type of algorithm first trains a base learner from the original data set, and then trains the next learner based on the performance of the base learner. Each time training, the previous learner is optimized, and k base learners are trained in turn. Finally, the results of these k models are added together to produce the result. The most representative algorithm of Boosting is XGBoost.

[0092] In the present invention, the random forest model parameters n_estimators = 10, criterion = 'mse', min_samples_split = 2, min_samples_leaf = 1, bootstrap = True, etc. are set; the SVR model parameters kernel = 'rbf', gamma = 0.1, C = 100, epsilon = 0.1, etc. are set; the XGBoost model parameters max_depth = 2, learning_rate = 0.1, n_estimators = 20000, subsample = 0.8, etc. are set; the KNN model parameters n_neighbors = 5 are set, and the remaining parameters are set to the default values.

[0093] According to the analysis results of step 3, six input features were selected, including snow density, surface temperature, SigmaVV, latitude, elevation and longitude. Then, the training data were input into the SVR, KNN, RF and XGBoost models respectively. The snow depth inversion results and fitting curves of the training and test sets of the four models are shown in Figure 2. Figure 6 Among these models, the RF model has the best snow depth inversion results, with RMSE, MAE and R 2 They are 3.54 cm, 1.74 cm, and 0.86, respectively. The slope of the fitting curve of the RF model is 0.83 on the training set and 0.95 on the test set. In contrast, the slope values ​​of the fitting curves of the training and test sets of the other three models are smaller. Both the snow depth inversion values ​​and the fitting curves show that the RF model performs better than the other models, followed by the XGBoost model. The SVR and KNN models show more obvious differences in the snow depth inversion results and fitting curves, indicating that their prediction errors are larger.

[0094] Experimental results:

[0095] The above RF snow depth inversion results are compared with the existing passive microwave snow depth products. The passive microwave snow product used is the "China 1980-2020 Snow Water Equivalent 25 km Daily Product" (http: / / www.ncdc.ac.cn) of the National Cryosphere Desert Data. The comparison between the passive microwave snow product and the machine learning RF model in the training set and the test set is shown in the figure below. Figure 7 ,from Figure 7 As can be seen from the figure, the RMSE of the machine learning RF model on the training set is 1.21cm, the MAE is 0.63cm, and the R 2 is 0.98, RMSE is 3.54cm, MAE is 1.74cm, R 2 The passive microwave snow product has an RMSE of 7.52 cm and a MAE of 4.92 cm on the training set. 2is 0.12, RMSE is 7.74cm, MAE is 5.14cm, R 2 The snow depth inversion results obtained using the RF model are significantly better than those obtained using the passive microwave snow accumulation product.

[0096] The present invention proposes to use MCCFS and ablation experiment feature screening to perform optimal feature screening, and then use a machine learning model to perform snow depth inversion, making full use of the relationship between snow depth and features and the relationship between features. Four machine learning methods are used for snow depth inversion and comparison, and it can be concluded that the RF model has the best result. The feasibility of the RF model in snow depth inversion is verified using an independent verification site. Finally, comparison with the passive microwave snow depth product further illustrates the advantages of the RF model.

Claims

1. A snow depth inversion method based on machine learning and feature screening strategy for spaceborne microwave scatterometers, characterized in that: The steps of this method are as follows: Step 1: Data acquisition and preprocessing: (1) Obtain the following data: a. Obtain satellite data of the L1B level of the microwave scatterometer of the Ocean No. 2B satellite, split each L1B level data according to the different incident angles, and generate the daily HH polarization scattering coefficient and HH polarization scattering coefficient; then, use the HH polarization scattering coefficient and VV polarization scattering coefficient to calculate the polarization ratio, polarization sum, polarization difference and polarization product of the dual-polarization scattering coefficient; Sigma PR =Sigma VV / Sigma HH (1) Sigma PM =Sigma VV ×Sigma HH (2) Sigma PS =Sigma VV +Sigma HH (3) Sigma PD =Sigma VV -Sigma HH (4) Sigma in formula (1) to formula (4) VV represents the VV polarization scattering coefficient, Sigma HH represents the HH polarization scattering coefficient, Sigma PR Represents the polarization ratio of the dual-polarization scattering coefficient, Sigma PM Represents the polarization product of the dual-polarization scattering coefficient, Sigma PS represents the polarization sum of the dual-polarization scattering coefficients, Sigma PD represents the polarization difference of dual-polarization scattering coefficient; b. Obtain the snow density data from the "Daily Snow Density Dataset in China (2013-2020)" provided by the National Cryosphere and Desert Data Center; c. Obtain the surface temperature data from the East Asia 0.02° Hourly Seamless Surface Temperature Dataset (2016-2021); d. Obtain the coverage and distribution of land cover types in the study area from the Globeland30 data; e. Obtain elevation data from the SRTM digital elevation model 90M dataset; (2) The above data are cropped, spliced ​​and resampled according to the selected research area to obtain a multi-feature dataset with the same spatial resolution; Step 2: Create a snow depth inversion dataset: The snow depth data measured by ground observation stations are used as label data. This dataset includes the station number, station level information, and longitude, latitude, altitude, year, month, day, and snow depth attributes of each ground observation station. A portion of observation sites in the study area are randomly selected as validation sites, and the remaining sites are used as training and testing sites. The validation sites are divided into validation sites of different land cover types according to the area ratio of the land cover type; The snow density, longitude, latitude, average scattering coefficient, VV polarization scattering coefficient, HH polarization scattering coefficient, dual-polarization scattering coefficient polarization ratio, dual-polarization scattering coefficient polarization product, dual-polarization scattering coefficient polarization sum, dual-polarization scattering coefficient polarization difference, surface temperature and elevation corresponding to the geographical location of the ground observation station are extracted as initial inputs, and the measured snow depth value is used as output to obtain the snow depth inversion data set; the average scattering coefficient calculation formula is as follows: (5) where SigmaAve represents the average scattering coefficient, n is the number of orbits that scan the study area on that day, and Sigma i represents the backscatter coefficient of each swept track data; Step 3: Feature screening: (a) Correlation coefficient feature screening: (1) Calculate the three correlation coefficients of snow depth and other 12 characteristics: Pearson correlation coefficient PCC, Spearman correlation coefficient SCC, and Kendall correlation coefficient KCC; The calculation methods of Pearson correlation coefficient PCC, Spearman correlation coefficient SCC and Kendall correlation coefficient KCC are as follows: (6) Where PCC represents the Pearson correlation coefficient, and denote the mean values ​​of Y and X respectively, and n is the number of features used for evaluation; (7) where SCC denotes the Spearman correlation coefficient, d i represents the difference in the rank value of the i-th data pair, and n is the number of features used for evaluation; (8) where KCC represents the Kendall correlation coefficient, C represents the number of pairs of elements with consistency in XY, with each pair of two elements, and D represents the number of pairs of elements with inconsistency in XY; (2) Calculate the average correlation coefficient ACC between snow depth and each feature: ACC = (PCC + SCC + KCC) / 3; (3) Sort the average correlation coefficient ACC from high to low, retain the N features with the highest average correlation coefficient ACC, and finally discard the features with ACC less than M among the N features, and retain the remaining features, where M and N are obtained through experiments and the significance of the correlation coefficient; (b) Ablation experiment feature screening: In this step, the random forest algorithm is used to invert the snow depth data, and the ratio of the training set data to the test set data is 8:2; (1) First, the N features obtained by screening in step (a) are used as the initial input features to obtain the snow depth inversion result A0. The N features are eliminated in sequence to obtain N groups of N-1 input features, each of which has different input features. The inversion results A1 to A1 are obtained by training them respectively. N ; (2) The inversion results A1~A N Compared with A0, if the evaluation index of snow depth inversion results of a certain group of results is lower than that of A0, it means that the corresponding eliminated feature has a positive effect on snow depth inversion and is retained; If the evaluation index of the snow depth inversion result of a certain group of results is unchanged or increased when compared with A0, it means that the corresponding feature quantity that is eliminated has no effect on the snow depth inversion and may even reduce the snow depth inversion accuracy. The number of these features is recorded as Na, and they are all discarded, and N = N-Na; (3) Repeat the above steps until N no longer changes. The features finally retained are the optimal input features, which are used as the input features for subsequent snow depth inversion experiments. Step 4: Snow depth inversion based on optimal features and machine learning: The features finally retained after feature screening in step 3 were used as input features. Four machine learning regression models, namely support vector regression model, K-nearest neighbor algorithm model, extreme gradient boosting model and random forest model, were used to invert snow depth. The snow depth results of the study area were obtained. The root mean square error, mean absolute error and determination coefficient of the inversion results were analyzed to screen out the machine learning regression model with the best results. Step 5: Take the features finally retained after feature screening in step 3 as input features, input them into the machine learning regression model with the best results screened out in step 4, and invert to obtain the final snow depth value.

Images