Arctic region soil freeze-thaw state inversion method combined with observation data of Chinese Tianmu-1 satellite-borne BDS-R and SMOS radiometer

Abstract

The invention discloses an arctic region soil freeze-thaw state inversion method in combination with observation data of a Chinese Tianmu-1 satellite-borne BDS-R and an SMOS radiometer, namely a method for inversing the arctic region soil freeze-thaw state based on an Attention mechanism and a Multiscale CNN-BiLSTM hybrid model of ResNet, the method is based on the Attention mechanism and the Multiscale CNN-BiLSTM hybrid model of ResNet, the model is mainly composed of three core modules, and the core modules are connected with the Attention mechanism and the Multiscale CNN-BiLSTM hybrid model of ResNet. The hybrid model comprises a ResNet and CNN-based multi-scale feature extraction module, a BiLSTM network-based feature relationship reasoning module and an Attention mechanism-based multi-feature parameter fused full-connection network module, and the hybrid model can significantly improve F / T inversion precision and is helpful to provide data supplement for the arctic region in a severe environment.

Description

Technical Field

[0001] The present invention belongs to the research field of inverting the soil freeze-thaw state in the Arctic region by using GNSS reflection signals, and specifically relates to an inversion method for the soil freeze-thaw state in the Arctic region by combining the observation data of the BDS-R and SMOS radiometers carried by China's Tianmu-1 satellite. Background Art

[0002] Soil freeze / thaw (F / T) is an extremely important indicator of climate change and plays an important role in surface hydrological processes. The F / T state of the Arctic is closely related to the Arctic biome, vegetation distribution, wetland distribution, etc. Monitoring the F / T state in the Arctic region is helpful to study the climate and ecological environment changes in the Arctic region. The traditional method of monitoring the F / T state is to set up in-situ monitoring stations, but the harsh natural environment in the Arctic makes it difficult to set up many in-situ monitoring stations, and it will consume a lot of financial and material resources. The development of remote sensing technology has provided a new idea for monitoring the soil freeze / thaw state. Optical remote sensing can achieve large-scale observations, but its observation results are limited by cloud and weather conditions. Microwave remote sensing can also monitor the F / T state, but its monitoring capability is limited by the long revisit period or spatial resolution. Compared with previous remote sensing technology means, satellite-borne GNSS-R technology has the ability to observe all-weather and all-day, and has higher temporal and spatial resolution. GNSS-R technology has been proven to be able to monitor the F / T process. With the continuous launch of spaceborne GNSS-R small satellites, spaceborne GNSS-R technology has played a huge role in monitoring F / T status. In addition, the feasibility of using CYGNSS to monitor F / T has also been confirmed. However, due to the limited coverage of CYGNSS, freeze-thaw detection results in the Arctic region cannot be obtained. Compared with the US CYGNSS satellite, China's Tianmu-1 constellation has global coverage capabilities, providing valuable opportunities for F / T detection.

[0003] Existing research on inverting F / T based on satellite GNSS-R data mainly focuses on detection algorithms based on seasonal thresholds and detection algorithms based on random forests (RF). The detection algorithm based on seasonal thresholds is a commonly used method for detecting soil freeze-thaw status, but the accuracy of detection depends on the setting of thresholds. Machine learning also plays a great advantage in classification problems. RF can capture the intricate nonlinear relationship between predictor variables and response variables, making RF also have high detection accuracy in detecting soil freeze-thaw status, but the accuracy of a single network model also has certain limitations. The hybrid model can make full use of the advantages of each network model and combine these advantages to enhance the overall performance of the model and improve its inversion and generalization capabilities.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a soil freeze-thaw state inversion method in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on China's Tianmu-1 satellite. The method is based on a hybrid model of Multiscale CNN-BiLSTM (ResNet-MultiscaleCNN-BiLSTM-Att) of the Attention mechanism and ResNet. The model is mainly composed of three core modules, namely a multi-scale feature extraction module based on ResNet and CNN, a feature relationship reasoning module based on the BiLSTM network, and a fully connected network module based on the Attention mechanism for fusing multiple feature parameters. The model can demonstrate the ability of high-precision soil freeze-thaw state inversion in the Arctic region. The hybrid model can significantly improve the F / T inversion accuracy and help provide data supplement for the harsh Arctic environment.

[0006] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0007] A soil freeze-thaw state inversion method in the Arctic region based on observation data from the BDS-R and SMOS radiometers on China's Tianmu-1 satellite includes the following steps:

[0008] Step S1, obtaining Tianmu-1 / GNOS-M onboard BDS-R data (provided free of charge by Aerospace Tianmu (Chongqing) Satellite Technology Co., Ltd. and Aerospace Science and Industry (Beijing) Space Information Application Co., Ltd.), SMOS L3 F / T data, SMAP L3 F / T data and auxiliary data;

[0009] Step S2, preprocessing the BDS-R data onboard Tianmu-1 / GNOS-M;

[0010] Step S3, calculating the surface reflectivity and dividing the data into a training set and a test set;

[0011] Step S4, construction and training of ResNet-Multiscale CNN-BiLSTM-Att hybrid model;

[0012] Step S5, input the test data set into the trained Resnet-Multiscale CNN-BiLSTM-Att hybrid model to obtain the inverted soil freeze-thaw state results, and compare and evaluate the inversion performance of the model.

[0013] Furthermore, the Tianmu-1 / GNOS-M onboard BDS-R data described in step S1 is Tianmu-1 / GNOS-MBDS-R L1 level data; the SMOS L3 F / T data is SMOS L3 25 km×25km F / T data; the SMAP L3 F / T data is SMAP L3 36 km×36km F / T data; the auxiliary data is the vegetation optical depth data provided by SMAP, and the auxiliary data is obtained from the SMAP radiometer L3 level global daily 36 km EASE grid soil moisture product.

[0014] Furthermore, the preprocessing of the Tianmu-1 / GNOS-M onboard BDS-R data described in step S2 includes calculating the Tianmu-1 BDS-R characteristic observation values, selecting model input parameters, and performing spatiotemporal fusion and data quality control on all acquired data sets.

[0015] In addition, step S2 includes the following sub-steps:

[0016] Step S2.1, calculate the Tianmu-1 BDS-R characteristic observation values ​​including the DDM power average (DDMA) calculated from the power DDM, the leading edge slope (LES) and the leading edge slope (TES) calculated from the bistatic radar cross section (BRCS) image;

[0017] Step S2.2, select the model input parameters including the characteristic observation value of Tianmu-1 BDS-R and the reflectivity at the mirror reflection point (Ddm_sp_reflectivity), GNSS satellite PRN code (Gnss_prn_code), receiver antenna gain at the mirror reflection point (Sp_antenna_gain), UTC time corresponding to the middle moment of DDM acquisition (Ddm_time_utc), DDM effective scattering area (Ddm_effective_area), normalized bistatic radar cross section of DDM mirror reflection point (Ddm_sp_nbrcs), DDM waveform front slope (Ddm_sp_les), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_s ...snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to r), GNSS signal incident angle at the mirror reflection point (Sp_inc_angle), latitude of the mirror reflection point (Sp_lat), longitude of the mirror reflection point (Sp_lon), altitude of the mirror reflection point (Sp_alt), satellite pitch angle (Rx_pitch), satellite roll angle (Rx_roll), azimuth of the mirror reflection point in the satellite body coordinate system (Sp_az_body), azimuth of the mirror reflection point in the orbital coordinate system (Sp_az_orbit), altitude of the mirror reflection point in the orbital coordinate system (Sp_theta_orbit) and altitude of the mirror reflection point in the satellite body coordinate system (Sp_theta_body);

[0018] In step S2.3, the Tianmu-1 / GNOS-M L1 BDS-R variable parameters are spatiotemporally fused with the SMOS L3 F / T data, SMAP L3F / T data, and vegetation optical depth auxiliary data to obtain the matched data set.

[0019] Step S2.4, the quality control of the matched data includes that the observed value must be positive, and it needs to be discarded when the value is Nan; when the satellite roll angle is greater than 30 degrees, the yaw angle is greater than 5 degrees, or the pitch angle is greater than the absolute value of 10 degrees, the data will be discarded; delete the data that detects radio frequency interference (RFI); if the estimated mirror point delay and Doppler shift have high uncertainty, delete the data; if the satellite altitude exceeds the nominal altitude range, delete the data; if the temperature and automatic gain control (AGC) values ​​of the DDM power calibration conditions are beyond the normal range, delete the data; if there are negative values ​​in the BRCS matrix used to calculate NBRCS or the corresponding effective area is invalid, the data will be discarded;

[0020] In addition, data filtering uses basic quality control (QC) flags in Tianmu-1 L1 data. Data filtering is applied using a single QC flag. The settings of a single QC flag include:

[0021] If the overall quality is poor, set bit 0 to 1; if any absolute value of the spacecraft roll / pitch / yaw is greater than the threshold, set bit 1 to 1; if the absolute value of the LNA temperature change rate is greater than the threshold, set bit 2 to 1; if the difference between this DDM noise floor and the previous DDM noise floor is greater than the threshold, set bit 3 to 1; if the AGC state of this DDM and the previous DDM is different, set bit 4 to 1; if the difference between the DDM noise floor calculated by two different methods is greater than the threshold, set bit 5 to 1; bit 6: unused; bit 7: unused; if direct_signal_exists in DDM: the absolute value of the difference between the direct signal code phase and the DDM signal code phase is less than 12.5 code chips, set bit 8 to 1; if RFI is detected, then Set bit 9 to 1; Set bit 10 to 1 if the estimated mirror point delay has high uncertainty; Set bit 11 to 1 if the estimated mirror point Doppler has high uncertainty; Set bit 12 to 1 if the spacecraft altitude is outside the rated altitude range; Set bit 13 to 1 if the temperature of the DDM power calibration condition is outside the normal range; Set bit 14 to 1 if the AGC value of the DDM power calibration condition is outside the normal range; Set bit 15 to 1 if the EIRP information of the GNSS satellite is unknown; Set bit 16 to 1 if there are negative values ​​in the BRCS matrix used to calculate NBRCS; Bit 17: Unused; Set bit 18 to 1 if the valid area is invalid; Set bit 19 to 1 if the spacecraft roll / pitch / yaw changes are greater than the threshold.

[0022] Furthermore, the calculation of the surface reflectivity described in step S3 is by using the non-uniform DDM generated by the GNSS-R receiver, and extracting its peak value for discrimination to calculate the surface reflectivity; the method of dividing into training set and test set is to divide the Tianmu-1 / GNOS-M satellite-borne BDS-R data set and the SMOS L3 F / T data set into 70% and 30% respectively.

[0023] The calculation formula of surface reflectivity is as follows:

[0024]

[0025] In the formula, R r and R t are the distances from the mirror point to the receiver and transmitter, respectively, DDM is the peak power of the DDM, N is the noise, and F is the DDM bistatic radar cross section factor, which can be defined as:

[0026]

[0027] Where P t G tis the GNSS effective isotropic radiated power (EIRP), G r is the receiver antenna pattern.

[0028] Furthermore, the ResNet-Multiscale CNN-BiLSTM-Att hybrid model constructed in step S4 mainly includes three modules. The first network module is a multi-scale feature extraction module based on ResNet and CNN, and the input is power DDM and effective scattering area. The convolutional neural network is used to effectively extract spatial feature information related to soil freeze-thaw from power DDM and effective scattering area; the second network module is a feature relationship reasoning module, which uses a bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, BiLSTM) to extract time-related feature information and capture context information, so as to better understand the input sequence; the third network module is a fully connected network module based on the Attention mechanism that integrates multiple feature parameters. The attention mechanism is introduced to emphasize the importance between feature observation variables. Through the fully connected network, the GNSS-R feature observation values ​​and variable parameters are input to classify the soil freeze-thaw state.

[0029] Furthermore, the training process of the ResNet-Multiscale CNN-BiLSTM-Att hybrid model for inverting the soil freeze-thaw state described in step S4 is specifically implemented as follows:

[0030] Step S4.1, input the divided training set as input into the ResNet-Multiscale CNN-BiLSTM-Att hybrid model;

[0031] Step S4.2, input the image power DDM and the effective scattering area into the multi-scale feature extraction module, and input the BDS-R feature observation value and variable parameters into the fully connected network module based on the Attention mechanism to fuse multiple feature parameters;

[0032] Step S4.3, use the sigmoid function and the softmax function as the activation function of the output layer respectively, where the softmax function realizes the classification of F / T and obtains the classification result. The calculation expression of the function is as follows:

[0033]

[0034] In the formula, x is the input value, x i represents the i-th input, is the exponential sum of all inputs.

[0035] Furthermore, the method of comparative evaluation of the inversion performance of the model described in step S5 is to use SMOSL3 F / T data and SMAP L3 F / T data as reference data to evaluate the model inversion results. Accuracy, precision, recall and F-Measure (also known as F-Score) are selected to compare and evaluate the inversion performance of the model. The calculation formulas of the four indicators are as follows:

[0036]

[0037] Where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives.

[0038]

[0039] After adopting the above technical scheme, the present invention has the following beneficial effects compared with the prior art.

[0040] The present invention mainly includes three modules, namely, a multi-scale feature extraction module based on ResNet and CNN, a feature relationship reasoning module based on BiLSTM network, and a fully connected network module based on the Attention mechanism to fuse multiple feature parameters. Among them, the multi-scale feature extraction module based on ResNet and CNN uses a convolutional neural network to effectively extract spatial feature information related to soil freeze-thaw from power DDM (powerDDM) and effective scattering area (effective scattering area); the feature relationship reasoning module is based on the BiLSTM network to perform the reasoning process between feature relationships. The module uses BiLSTM to extract time-related features and capture context information, so as to better understand the input sequence; the fully connected network module is based on the Attention mechanism, emphasizing the importance between feature observation variables. The hybrid model of the present invention can significantly improve the F / T inversion accuracy, which helps to provide data supplement for the harsh Arctic region.

[0041] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are part of this application and are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not constitute an improper limitation of the present invention. Obviously, the drawings described below are only some embodiments. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0043] Figure 1 The figure is a flow chart of the method of the present invention.

[0044] Figure 2 It is a framework diagram of the ResNet-Multiscale CNN-BiLSTM-Att hybrid model in an embodiment of the present invention.

[0045] Figure 3 It is the backbone network (ResNet50) of the ResNet-Multiscale CNN-BiLSTM-Att hybrid model in the embodiment of the present invention.

[0046] Figure 4 It is a technical flow chart of F / T detection adopted in the embodiment of the present invention.

[0047] Figure 5 Graph showing the spatial distribution of SMOS and SMAP F / T data in an embodiment of the present invention.

[0048] Figure 6 It is a comparison diagram of the spatial distribution of freeze-thaw inversion results and SMOS F / T data of different inversion models in the embodiments of the present invention.

[0049] Figure 7 It is the daily F / T inversion results of the ResNet-Multiscale CNN-BiLSTM-Att method in different periods and the spatial distribution diagram of SMAP F / T data in an embodiment of the present invention.

[0050] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0052] Example 1

[0053] In order to verify the effectiveness of the method proposed in the present invention, the L1 data products of Tianmu-1 from June 2023 to December 2023, SMOS L325 km×25km F / T data, SMAP L336 km×36kmF / T data and vegetation opacity data provided by SMAP were obtained for experiments, and the experimental results of the present invention were compared with the inversion results of several other machine learning models (i.e., random forest (RF), bagging tree (BT), support vector machine (SVM) and convolutional neural network (CNN)). The basic configuration of the experimental platform for building the Resnet-Multiscale CNN-BiLSTM-Att hybrid model is as follows:

[0054] The experimental environment is Intel Core i7-10700@2.90GHz eight-core CPU, NVIDIA GeForce RTX2080SUPERGPU, 32GB (Critical DDR43000 MHz) running memory, Windows 10 operating system, Matlab 2019b and Python 3.7 as the experimental platform, and tensorflow 2.3.1 and Keras 2.4.3 as the deep learning API.

[0055] The implementation process of a technical solution for inverting soil freeze-thaw status in the Arctic region by combining the observation data of the BDS-R and SMOS radiometers on China's Tianmu-1 spacecraft is as follows: Figure 1-4 As shown, the following steps are included:

[0056] Step S1, obtaining Tianmu-1 / GNOS-M onboard BDS-R data (provided free of charge by Aerospace Tianmu (Chongqing) Satellite Technology Co., Ltd. and Aerospace Science and Industry (Beijing) Space Information Application Co., Ltd.), SMOS L3 F / T data, SMAP L3 F / T data and auxiliary data;

[0057] Step S2, preprocessing the BDS-R data onboard Tianmu-1 / GNOS-M;

[0058] Step S3, calculating the surface reflectivity and dividing the data into a training set and a test set;

[0059] Step S4, construction and training of ResNet-Multiscale CNN-BiLSTM-Att hybrid model;

[0060] Step S5, input the test data set into the trained Resnet-Multiscale CNN-BiLSTM-Att hybrid model to obtain the inverted soil freeze-thaw state results, and compare and evaluate the inversion performance of the model.

[0061] As an implementation manner of this embodiment, in step S1, the Tianmu-1 / GNOS-M onboard BDS-R data is Tianmu-1 / GNOS-M BDS-R L1 data; the SMOS L3 F / T data is SMOS L3 25 km×25 km F / T data; the SMAP L3 F / T data is SMAP L3 36 km×36 km F / T data; the auxiliary data is the vegetation optical depth data provided by SMAP, and the auxiliary data is obtained from the SMAP radiometer L3 level global daily 36 km EASE grid soil moisture product.

[0062] As an implementation method of this embodiment, the preprocessing of the Tianmu-1 / GNOS-M onboard BDS-R data described in step S2 includes calculating the Tianmu-1 BDS-R characteristic observation values, selecting model input parameters, and performing spatiotemporal fusion and data quality control on all acquired data sets.

[0063] In addition, step S2 includes the following sub-steps:

[0064] Step S2.1, calculate the Tianmu-1 BDS-R characteristic observation values ​​including the DDM power average (DDMA) calculated from the power DDM, the leading edge slope (LES) and the leading edge slope (TES) calculated from the bistatic radar cross section (BRCS) image;

[0065] Step S2.2, select the model input parameters including the characteristic observation value of Tianmu-1 BDS-R and the reflectivity at the mirror reflection point (Ddm_sp_reflectivity), GNSS satellite PRN code (Gnss_prn_code), receiver antenna gain at the mirror reflection point (Sp_antenna_gain), UTC time corresponding to the middle moment of DDM acquisition (Ddm_time_utc), DDM effective scattering area (Ddm_effective_area), normalized bistatic radar cross section of DDM mirror reflection point (Ddm_sp_nbrcs), DDM waveform front slope (Ddm_sp_les), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_s ...snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_snbrcs), DDM mirror reflection point signal-to r), GNSS signal incident angle at the mirror reflection point (Sp_inc_angle), latitude of the mirror reflection point (Sp_lat), longitude of the mirror reflection point (Sp_lon), altitude of the mirror reflection point (Sp_alt), satellite pitch angle (Rx_pitch), satellite roll angle (Rx_roll), azimuth of the mirror reflection point in the satellite body coordinate system (Sp_az_body), azimuth of the mirror reflection point in the orbital coordinate system (Sp_az_orbit), altitude of the mirror reflection point in the orbital coordinate system (Sp_theta_orbit) and altitude of the mirror reflection point in the satellite body coordinate system (Sp_theta_body);

[0066] In step S2.3, the Tianmu-1 / GNOS-M L1 BDS-R variable parameters are spatiotemporally fused with the SMOS L3 F / T data, SMAP L3F / T data, and vegetation optical depth auxiliary data to obtain the matched data set.

[0067] Step S2.4, the quality control of the matched data includes that the observed value must be positive, and it needs to be discarded when the value is Nan; when the satellite roll angle is greater than 30 degrees, the yaw angle is greater than 5 degrees, or the pitch angle is greater than the absolute value of 10 degrees, the data will be discarded; delete the data that detects radio frequency interference (RFI); if the estimated mirror point delay and Doppler shift have high uncertainty, delete the data; if the satellite altitude exceeds the nominal altitude range, delete the data; if the temperature and automatic gain control (AGC) values ​​of the DDM power calibration conditions are beyond the normal range, delete the data; if there are negative values ​​in the BRCS matrix used to calculate NBRCS or the corresponding effective area is invalid, the data will be discarded;

[0068] In addition, data filtering was applied using basic quality control (QC) flags in the Tianmu-1 L1 data using a single QC flag bit (the setting of the flag bit is shown in Table 1).

[0069] Table 1 Setting of Tianmu-1 data identification bit

[0070]

[0071] As an implementation of this embodiment, in step S3, the surface reflectivity is calculated by using the non-uniform DDM generated by the GNSS-R receiver and extracting its peak value for discrimination to calculate the surface reflectivity; the method of dividing into training set and test set is to divide the Tianmu-1 / GNOS-M satellite-borne BDS-R data set and the SMOS L3 F / T data set into 70% and 30% respectively. The surface reflectivity calculation formula is as follows:

[0072]

[0073] In the formula, R r and R t are the distances from the mirror point to the receiver and transmitter, respectively, DDM is the peak power of the DDM, N is the noise, and F is the DDM bistatic radar cross section factor, which can be defined as:

[0074]

[0075] Where P t G t is the GNSS effective isotropic radiated power (EIRP), G r is the receiver antenna pattern.

[0076] As an implementation method of this embodiment, in step S4, the constructed ResNet-MultiscaleCNN-BiLSTM-Att hybrid model mainly includes three modules. The first network module is a multi-scale feature extraction module based on ResNet and CNN, and the input is power DDM and effective scattering area. The convolutional neural network is used to effectively extract spatial feature information related to soil freeze-thaw from power DDM and effective scattering area; the second network module is a feature relationship reasoning module, which uses a bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, BiLSTM) to extract time-related feature information and capture context information, so as to better understand the input sequence; the third network module is a fully connected network module based on the Attention mechanism that integrates multiple feature parameters. The attention mechanism is introduced to emphasize the importance between feature observation variables. Through the fully connected network, the GNSS-R feature observation values ​​and variable parameters are input to classify the soil freeze-thaw state.

[0077] As an implementation of this embodiment, in step S4, the training process of the ResNet-MultiscaleCNN-BiLSTM-Att hybrid model for inverting the soil freeze-thaw state is specifically implemented as follows:

[0078] Step S4.1, input the divided training set as input into the Resnet-Multiscale CNN-BiLSTM-Att hybrid model;

[0079] Step S4.2, input the image power DDM and the effective scattering area into the multi-scale feature extraction module, and input the BDS-R feature observation value and variable parameters into the fully connected network module based on the Attention mechanism to fuse multiple feature parameters;

[0080] Step S4.3, use the sigmoid function and the softmax function as the activation function of the output layer respectively, where the softmax function realizes the classification of F / T and obtains the classification result. The calculation expression of the function is as follows:

[0081]

[0082] In the formula, x is the input value, x i represents the i-th input, is the exponential sum of all inputs.

[0083] As an implementation of this embodiment, in step S5, the method of comparing and evaluating the inversion performance of the model is to use SMOS L3 F / T data and SMAP L3 F / T data as reference data to evaluate the model inversion results. Accuracy, precision, recall and statistic F-Measure (also known as F-Score) are selected to compare and evaluate the inversion performance of the model. The calculation formulas of the four indicators are as follows:

[0084]

[0085] Where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives.

[0086]

[0087] The trained ResNet-Multiscale CNN-BiLSTM-Att model is used to invert the soil freeze-thaw state in the Arctic region, and the SMOS F / T data and SMAP F / T data are used as reference data. In order to ensure the accuracy of the evaluation model detection performance, it is necessary to analyze the spatiotemporal distribution characteristics between the true values ​​of the two types of data, SMOS and SMAP, to ensure that they have good consistency in spatiotemporal distribution. Figure 5 The spatial distribution of SMOS F / T data and SMAP F / T data is shown, and it can be concluded that SMOS data and SMAP data have good consistency in spatiotemporal distribution. The experimental results of the present invention are compared with the inversion results of several other machine learning models (i.e., random forest (RF), bagging tree (BT), support vector machine (SVM) and convolutional neural network (CNN)). Table 2, Table 3 and Table 4 respectively give the performance comparison of different model inversion F / T with SMOS and SMAP F / T data and the daily F / T inversion results of ResNet-Multiscale CNN-BiLSTM-Att model in different periods compared with SMAP F / T data. The following conclusions can be drawn from the table:

[0088] (1) When SMOS F / T data is used as reference data, the proposed ResNet-MultiscaleCNN-BiLSTM-Att model has the best inversion performance compared with four traditional machine learning methods, namely RF, BT, SVM and CNN, and is the best in terms of accuracy, precision, recall and F-Score, which are 98.7%, 98.0%, 89.5% and 93.6% respectively. When SMAP F / T data is used as reference data, the proposed model has the best inversion performance compared with RF, BT, SVM and CNN, and is the best in terms of accuracy, precision, recall and F-Score, which are 97.2%, 96.5%, 88.0% and 92.1% respectively. In terms of F-Score IR, the proposed model is 9.9%, 10.5%, 11.9% and 10.6% relative to RF, BT, SVM and CNN, respectively.

[0089] (2) When SMAP F / T data is used as reference data, the highest accuracy and F-Score of the daily F / T inversion results of the ResNet-MultiscaleCNN-BiLSTM-Att model proposed in the present invention compared with the SMAP F / T data are 97.2% and 89.8%, respectively, and the lowest accuracy and F-Score are 81.8% and 71.3%, respectively, indicating that the model proposed in this paper has better inversion performance. In terms of accuracy, the accuracy of the daily results of different periods inverted by the ResNet-Multiscale CNN-BiLSTM-Att model is generally high, which shows that the gap between the inversion results of the proposed model and the SMAP soil freeze-thaw data is relatively small, indicating that the model proposed in the present invention performs well in freeze-thaw detection and has strong reliability and robustness.

[0090] Table 2 Performance comparison of different model inversion F / T and SMOS F / T data

[0091]

[0092] Table 3 Performance comparison of different model inversion F / T and SMAP F / T data

[0093]

[0094] Table 4 Comparison of daily F / T inversion results of ResNet-Multiscale CNN-BiLSTM-Att method in different periods with SMAP F / T data

[0095]

[0096]

[0097] In order to more intuitively see the results of F / T inversion of each model and the differences between the F / T inversion results of different models, Figure 5 and attached Figure 6 The spatial distribution comparison of the inversion F / T results of different inversion models and the SMOS F / T data and SMAP F / T data is given in the Appendix. Figure 7 The daily F / T inversion results of the ResNet-Multiscale CNN-BiLSTM-Att model in different periods and the spatial distribution map of SMAP F / T are given. These results show that the model proposed in the present invention performs better than the RF, BT, SVM and CNN models in freeze-thaw detection, and the model has higher inversion accuracy.

[0098] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this patent can make some changes or modify the technical contents suggested above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the solution of the present invention.

Claims

1. A soil freeze-thaw state inversion method in the Arctic region based on the observation data of BDS-R and SMOS radiometers on China's Tianmu-1 satellite, characterized in that: The following steps are involved: Step S1, obtaining Tianmu-1 / GNOS-M onboard BDS-R data (provided free of charge by Aerospace Tianmu (Chongqing) Satellite Technology Co., Ltd. and Aerospace Science and Industry (Beijing) Space Information Application Co., Ltd.), SMOS L3 F / T data, SMAP L3 F / T data and vegetation optical depth auxiliary data; Step S2, preprocessing the BDS-R data onboard Tianmu-1 / GNOS-M; Step S3, calculating the surface reflectivity and dividing the data into a training set and a test set; Step S4, construction and training of ResNet-Multiscale CNN-BiLSTM-Att hybrid model; Step S5, input the test data set into the trained ResNet-Multiscale CNN-BiLSTM-Att hybrid model to obtain the inverted soil freeze-thaw state results, and compare and evaluate the inversion performance of the model.

2. The method for inverting the soil freeze-thaw state in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite of China according to claim 1, characterized in that: In step S1: The Tianmu-1 / GNOS-M onboard BDS-R data are Tianmu-1 / GNOS-M BDS-R L1-level data; the SMOS L3 F / T data are SMOS L3 25 km×25km F / T data; the SMAP L3 F / T data are SMAP L3 36km×36km F / T data; the auxiliary data are the vegetation optical depth data provided by SMAP, which are obtained from the SMAP radiometer L3-level global daily 36km EASE grid soil moisture product.

3. The method for inverting the soil freeze-thaw state in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite of China according to claim 1, characterized in that: In step S2, the preprocessing of the Tianmu-1 / GNOS-M onboard BDS-R data includes calculating the Tianmu-1 BDS-R characteristic observation values, selecting the model input parameters, and performing spatiotemporal fusion and data quality control on all acquired data sets.

4. The method for inverting the soil freeze-thaw state in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite of China according to claim 3 is characterized in that: The preprocessing of Tianmu-1 / GNOS-M onboard GNSS-R data in step S2 includes the following sub-steps: Step S2.1, calculate the Tianmu-1 BDS-R characteristic observation values ​​including the DDM power average (DDMA) calculated from the power DDM, the leading edge slope (LES) and the leading edge slope (TES) calculated from the bistatic radar cross section (BRCS) image; Step S2.2, select the model input parameters including the characteristic observation value of Tianmu-1 BDS-R and the reflectivity at the mirror reflection point (Ddm_sp_reflectivity), GNSS satellite PRN code (Gnss_prn_code), receiver antenna gain at the mirror reflection point (Sp_antenna_gain), UTC time corresponding to the middle moment of DDM acquisition (Ddm_time_utc), DDM effective scattering area (Ddm_effective_area), normalized bistatic radar cross section of DDM mirror reflection point (Ddm_sp_nbrcs), DDM waveform front slope (Ddm_sp_les), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_les), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_les), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_les), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_les), DDM mirror reflection point signal-to-noise ratio (Ddm_sp_snbrcs), DDM waveform front slope (Ddm_sp_snbrcs ... r), GNSS signal incident angle at the mirror reflection point (Sp_inc_angle), latitude of the mirror reflection point (Sp_lat), longitude of the mirror reflection point (Sp_lon), altitude of the mirror reflection point (Sp_alt), satellite pitch angle (Rx_pitch), satellite roll angle (Rx_roll), azimuth of the mirror reflection point in the satellite body coordinate system (Sp_az_body), azimuth of the mirror reflection point in the orbital coordinate system (Sp_az_orbit), altitude of the mirror reflection point in the orbital coordinate system (Sp_theta_orbit) and altitude of the mirror reflection point in the satellite body coordinate system (Sp_theta_body); Step S2.3, spatially and temporally fuse the Tianmu-1 / GNOS-ML1BDS-R variable parameters with the SMOSL3F / T data, SMAPL3F / T data, and auxiliary data to obtain the matched data set; Step S2.4, the quality control of the matched data includes that the observed value must be positive, and it needs to be discarded when the value is Nan; when the satellite roll angle is greater than 30 degrees, the yaw angle is greater than 5 degrees, or the pitch angle is greater than the absolute value of 10 degrees, the data will be discarded; the detected radio frequency interference (RFI) data will be deleted; if the estimated mirror point delay and Doppler shift have high uncertainty, the data will be deleted; if the satellite altitude exceeds the nominal altitude range, the data will be deleted; if the temperature and automatic gain control values ​​of the DDM power calibration conditions are beyond the normal range, the data will be deleted; If there are negative values ​​in the BRCS matrix used to calculate NBRCS or the corresponding valid area is invalid, the data will be discarded; In addition, data filtering uses basic quality control (QC) flags in Tianmu-1L1 data and data filtering is applied using a single QC flag bit.

5. The method for inverting the soil freeze-thaw state in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite according to claim 4, characterized in that: The settings of a single QC flag include: If the overall quality is poor, set bit 0 to 1; if any absolute value of the spacecraft roll / pitch / yaw is greater than the threshold, set bit 1 to 1; if the absolute value of the LNA temperature change rate is greater than the threshold, set bit 2 to 1; if the difference between this DDM noise floor and the previous DDM noise floor is greater than the threshold, set bit 3 to 1; if the AGC state of this DDM and the previous DDM is different, set bit 4 to 1; if the difference between the DDM noise floor calculated by two different methods is greater than the threshold, set bit 5 to 1; bit 6: unused; bit 7: unused; if direct_signal_exists in DDM: the absolute value of the difference between the direct signal code phase and the DDM signal code phase is less than 12.5 code chips, set bit 8 to 1; if RFI is detected, then Set bit 9 to 1; Set bit 10 to 1 if the estimated mirror point delay has high uncertainty; Set bit 11 to 1 if the estimated mirror point Doppler has high uncertainty; Set bit 12 to 1 if the spacecraft altitude is outside the rated altitude range; Set bit 13 to 1 if the temperature of the DDM power calibration condition is outside the normal range; Set bit 14 to 1 if the AGC value of the DDM power calibration condition is outside the normal range; Set bit 15 to 1 if the EIRP information of the GNSS satellite is unknown; Set bit 16 to 1 if there are negative values ​​in the BRCS matrix used to calculate NBRCS; Bit 17: Unused; Set bit 18 to 1 if the valid area is invalid; Set bit 19 to 1 if the spacecraft roll / pitch / yaw changes are greater than the threshold.

6. The method for inverting the soil freeze-thaw state in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite according to claim 1, characterized in that: The calculation of the surface reflectivity described in step S3 is performed by using the non-uniform DDM generated by the GNSS-R receiver and extracting its peak value for discrimination to calculate the surface reflectivity; the method of dividing into training set and test set is to divide the Tianmu-1 / GNOS-M satellite-borne BDS-R data set and the SMOSL3F / T data set into 70% and 30% respectively; The calculation formula of surface reflectivity is as follows: In the formula, R r and R t are the distances from the mirror point to the receiver and transmitter, respectively, DDM is the peak power of the DDM, N is the noise, and F is the DDM bistatic radar cross section factor, defined as: Where P t G t is the GNSS effective isotropic radiated power (EIRP), G r is the receiver antenna pattern.

7. The method for inverting the freeze-thaw state of soil in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite of China according to claim 1, characterized in that: The ResNet-MultiscaleCNN-BiLSTM-Att hybrid model constructed in step S4 mainly includes three network modules. The first network module is a multi-scale feature extraction module based on ResNet and CNN, with power DDM and effective scattering area as input; the second network module is a feature relationship reasoning module, which uses a bidirectional long short-term memory network to extract time-related feature information and capture context information to better understand the input sequence; the third network module is a fully connected network module based on the Attention mechanism that integrates multiple feature parameters. The attention mechanism is introduced to emphasize the importance of feature observation variables, and the GNSS-R feature observation values ​​and variable parameters are input through the fully connected network to classify the soil freeze-thaw state.

8. The method for inverting the soil freeze-thaw state in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite according to claim 7, characterized in that: The specific implementation steps of the training process of the ResNet-MultiscaleCNN-BiLSTM-Att hybrid model are as follows: Step S4.1, input the divided training set as input into the ResNet-MultiscaleCNN-BiLSTM-Att hybrid model; Step S4.2, input the image power DDM and the effective scattering area into the multi-scale feature extraction module, and input the GNSS-R feature observation value and variable parameters into the fully connected network module that fuses multiple feature parameters based on the Attention mechanism; Step S4.3, use the sigmoid function and the softmax function as the activation function of the output layer respectively, where the softmax function realizes the classification of F / T and obtains the classification result. The calculation expression of the function is as follows: In the formula, x is the input value, x i represents the i-th input, is the exponential sum of all inputs.

9. The method for inverting the soil freeze-thaw state in the Arctic region based on the observation data of the BDS-R and SMOS radiometers on the Tianmu-1 satellite of China according to claim 1, characterized in that: In step S5, the inversion performance of the model is compared and evaluated by using SMOSL3F / T data and SMAPL3F / T data as reference data to evaluate the inversion results of the model. Accuracy, precision, recall and statistic F-Measure (also known as F-Score) are selected to compare and evaluate the inversion performance of the model. The calculation formulas of the four indicators are as follows: Where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives.

Images