Dual-polarized SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network

Abstract

The invention discloses a dual-polarized SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and a residual network, and relates to the technical field of marine environment information monitoring. The method comprises the following steps: constructing a sea surface typhoon wind speed joint inversion algorithm model; the model structure comprises a pre-training model constructed by a ResNet network model, a medium-low wind speed inversion model and a high wind speed inversion model. The model structure further comprises a feature splicing model and a feature fusion model which are used for splicing and fusing features extracted from the medium-low wind speed inversion model and the high wind speed inversion model, and an inversion model for typhoon wind speed inversion. The invention designs a dual-polarized SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and a residual network, and the inversion algorithm can effectively alleviate the small sample learning problem of a high wind speed region, solves the feature mismatching problem in cross-wind-speed-section transfer, improves typhoon wind speed inversion, and improves the typhoon wind speed inversion efficiency. Especially, the stability and the accuracy of wind speed inversion in a high-wind-speed area are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of marine environmental information monitoring, and particularly relates to a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network. Background Technique

[0002] Typhoons have extremely strong destructive power, often accompanied by strong winds, heavy rains and storm surges, seriously threatening the lives and property safety of coastal ports and residents in coastal areas, and causing huge economic losses. The accurate estimation of the typhoon wind field plays a crucial role in maritime navigation safety, disaster warning and typhoon monitoring. However, obtaining high-precision typhoon observation data is still a very challenging task.

[0003] With the development of remote sensing technology, satellite remote sensing data has been widely used to monitor the development and evolution process of typhoons. Synthetic Aperture Radar (SAR) is a remote sensing technology based on radar waves, and has unique advantages in the research of typhoon wind field inversion due to its high resolution and strong wind adaptability. SAR can obtain sea surface wind speeds higher than 40 m / s, and can clearly display important features such as sea surface wind streaks, typhoon structures, and spiral rain bands, thus making up for the deficiencies of other spaceborne sensors with low resolution.

[0004] At present, the wind field inversion method based on SAR data mainly relies on geophysical model functions (GMFs), such as models like CMOD4, CMOD5, CMOD7, etc., which can be used for wind speed inversion of C-band co-polarization (VV / HH) SAR images. When the wind speed is less than 20 m / s, the inversion error is about 2 m / s. However, the traditional wind field inversion method based on the CMOD model has significant errors under extreme weather conditions, especially when the wind speed is high. Due to the saturation of the scattering coefficient, the inversion accuracy in the high wind speed range is still limited.

[0005] In recent years, the development of deep learning (DL) methods has provided a new technical approach for typhoon wind field inversion. Deep learning models such as deep neural networks (DNN), residual neural networks (ResNet), and random forest network models have powerful nonlinear fitting, feature inference and transfer application functions, can learn and capture complex features and correlations in SAR images, and have strong adaptability and generalization ability. These models can not only adapt to different ocean environments and meteorological conditions, but also combine multi-source remote sensing data to further improve the stability and accuracy of typhoon wind field estimation.

[0006] However, when introducing deep learning into SAR technology for typhoon inversion, the inventor believes that there are still the following technical problems: (1) Although deep learning models can improve the accuracy of wind speed inversion, the inversion error is still large in the case of high wind speeds (>33 m / s), and the high wind speed inversion error still exceeds 4 m / s.

[0007] (2) The observational data of typhoons are relatively scarce, which restricts the deep learning method driven by pure image data during the training process. Especially in the high wind speed area, due to the small number of effective samples, it is difficult for the model to comprehensively learn the influence of all incident angles, wind directions, and rainfall conditions, resulting in a large inversion error. (3) When existing deep learning models are migrated across wind speed segments, feature mismatch is likely to occur, affecting the performance of the models.

[0008] Therefore, how to solve the above technical problems is an urgent technical problem for those skilled in the art at present.

[0009] The information disclosed in this background art section is only intended to enhance the overall understanding of the present invention and should not be regarded as an admission or any form of implication that this information constitutes the prior art already known to those of ordinary skill in the art. Summary of the Invention

[0010] In view of the above technical problems, an embodiment of the present invention provides a joint dual-polarization SAR sea surface typhoon wind speed inversion algorithm based on transfer learning and residual network to solve the problems proposed in the above background art.

[0011] The present invention provides the following technical solutions: A joint dual-polarization SAR sea surface typhoon wind speed inversion algorithm based on transfer learning and residual network, including the following steps: Construct a joint sea surface typhoon wind speed inversion algorithm model; its model structure includes: constructing a pre-trained model, a medium and low wind speed inversion model, and a high wind speed inversion model with a ResNet network model; The model structure also includes a feature splicing model and a feature fusion model for splicing and fusing the features extracted from the medium and low wind speed inversion models and the high wind speed inversion models, and an inversion model for performing typhoon wind speed inversion; Obtain model sample data for inverting the sea surface typhoon wind speed, then divide the model sample data into a medium and low wind speed data set and a high wind speed data set, and input them into the pre-trained model for training and feature extraction, and transfer the features to the corresponding low wind speed inversion model and high wind speed inversion model; Input all the sample data of the model into the medium and low wind speed inversion model and the high wind speed inversion model respectively for independent training; Use the trained joint sea surface typhoon wind speed inversion algorithm model to invert the sea surface typhoon wind speed.

[0012] Preferably, the pre-trained model consists of one fully connected layer and four basic blocks; the basic block adopts the ResNet network model, and the activation function uses ReLu to improve the model to avoid gradient disappearance. Preferably, the model sample data includes: image data and physical auxiliary data; the image data includes: ; the physical auxiliary data includes: vv_en, vh_en, vv_homo, vh_homo, R, f , dir_u, dir_v and rain; where, is the VH polarization SAR backscattering coefficient; is the VV polarization SAR backscattering coefficient; is the incident angle; is the longitude; is the latitude; vv / vh is the polarization ratio, vv_en is the energy of VV polarization; vh_en is the energy of VH polarization; vv_homo is the homogeneity of VV polarization; vh_homo is the homogeneity of VH polarization; R is the distance; is the Coriolis parameter; dir_u is the sine component of the wind direction; dir_v is the cosine component of the wind direction; rain is the rainfall rate. It should be emphasized that when using pure image data (SAR satellite remote sensing image data) for typhoon inversion, since pure image data is a kind of pure data (the acquisition of SAR satellite remote sensing image data is mainly a digital data set obtained by scanning and measuring the ground or sea surface through a radar system), this kind of pure data usually does not consider the underlying physical mechanism or domain knowledge, so it will reduce the interpretability and generalization ability of the model; The present invention introduces multiple physical auxiliary data into the model sample data, integrates physical quantity information into the model, and constructs a deep learning model for sea surface typhoon wind speed inversion jointly driven by data and physical knowledge, solving the problems of poor interpretability and poor generalization ability caused by only relying on pure data such as SAR satellite remote sensing image data for model driving.

[0013] Preferably, after the model sample data is obtained, the image data and the physical auxiliary data need to be aligned in time and space into a unified time series.

[0014] Preferably, after the model sample data is obtained, it needs to be preprocessed, data standardized, and the ratio of the training set and the test set is randomly divided, and part of the data in the training set is used as the validation set.

[0015] Preferably, in the step of further dividing the model sample data into a medium and low wind speed data set and a high wind speed data set, the input features and output features with a wind speed less than 33 m / s are used as the medium and low wind speed data set, and the input features and output features with a wind speed greater than 33 m / s are used as the high wind speed data set.

[0016] Preferably, in the step of inversely calculating the sea surface typhoon wind speed by using the trained sea surface typhoon wind speed joint inversion algorithm model, the evaluation indexes for the trained model are: RMSE, Bias, MRE, and R; where RMSE is the root mean square error, Bias is the mean deviation, MRE is the mean relative error, and R is the correlation coefficient.

[0017] Preferably, the wind speed data is used as the true label and used as the output feature of the model.

[0018] The dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network provided by the embodiment of the present invention has the following beneficial effects: The present invention designs a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network. This inversion algorithm can effectively alleviate the small sample learning problem in the high wind speed area, solve the feature mismatch problem in cross-wind speed segment transfer, and improve the stability and accuracy of typhoon wind speed inversion, especially the wind speed inversion in the high wind speed area. Description of the Drawings

[0019] Figure 1 It is the network framework diagram of the TJ_Resnet typhoon wind speed inversion model in the present invention; Figure 2 It is the structural diagram of the basic block in the present invention; Figure 3 It is the SHAP honeycomb diagram of the input features in the present invention; Figure 4 It is the comparison result of the inversion results of the direct inversion algorithm (Resnet) and the joint inversion algorithm (TJ_Resnet) in the present invention; where; (4a) is the comparison of the wind speed inverted by the TJ-Resnet model and the SFMR wind speed (full wind speed area); (4b) is the comparison of the wind speed inverted by the TJ-Resnet model and the SFMR wind speed (medium and low wind speed area); (4c) is the comparison of the wind speed inverted by the TJ-Resnet model and the SFMR wind speed (high wind speed area); (4d) is the comparison of the wind speed inverted by the Resnet model and the SFMR wind speed (full wind speed area); (4e) is the comparison of the wind speed inverted by the Resnet model and the SFMR wind speed (medium and low wind speed area); (4f) Comparison of the wind speed inverted by the Resnet model and the SFMR wind speed (high wind speed area); Figure 5 This is the flow chart of the dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm (TJ_Resnet) based on transfer learning and residual network of the present invention. Specific implementation manners

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the protection scope of the present invention.

[0021] In view of the problems mentioned in the above background technology, the embodiments of the present invention provide a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network to solve the above technical problems. The technical solutions are as follows: Next, in combination with the attached Figures 1 - 5 , and specific implementation manners, the present invention will be further described.

[0022] The design idea and optimization steps of a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network provided by the present invention are as follows: Step 1: Obtain Sentinel-1 SAR data containing typhoons (Sentinel-1 is a sentinel satellite launched by the European Space Agency), SFMR (Stepped-Frequency Microwave Radiometer) airborne radiometer data (wind speed data) that is spatio-temporally matched with it, rainfall data provided by the Global Precipitation Measurement mission (GPM), and the fifth-generation wind field reanalysis data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF-ERA5).

[0023] Among them, the SFMR data is used as the real wind field, and the ERA5 data is used as the wind direction input. To ensure the timeliness of the data, based on the SAR imaging time, linear interpolation is performed on the ERA5 wind field data and GPM rainfall data to generate data with the same timestamp; the SFMR wind field data selects data samples within the range of ±1 h of the SAR imaging time.

[0024] Step 2: Extract relevant radar parameters of the SAR image. Preprocess the SAR remote sensing image according to the preprocessing method of SAR remote sensing images, including radiometric calibration, filtering and noise suppression (11×11 Lee filter), thermal noise removal, geometric correction, resampling, and land masking. In order to suppress the influence of SAR speckle noise and maintain the advantage of the SAR image in observing the fine typhoon structure, the resolution of the SAR remote sensing image is resampled to 500 m.

[0025] The radar parameters of the extracted SAR image include: VH polarization SAR backscattering coefficient , VV polarization SAR backscattering coefficient , incident angle , longitude , latitude , and calculate the ratio of the radar backscattering coefficients under VV and VH polarizations, that is, the polarization ratio ( PR, vv / vh ). A total of 6 types of radar parameter data of the SAR image are extracted.

[0026] Step 3: Extract texture feature parameters of the SAR image. Use the gray-level co-occurrence matrix (GLCM) to calculate the texture features of VV polarization and VH polarization SAR images. The gray-level co-occurrence matrix can reflect the comprehensive information of the image gray level about direction, adjacent interval, and change amplitude, representing the joint probability distribution of two gray pixels with a distance of D in the image, and reflecting the gray correlation of adjacent pixels. The texture feature parameters used in the present invention are energy, homogeneity, contrast, and correlation, and are extracted separately on the two polarization images.

[0027] A total of 8 types of texture feature parameter data of the SAR image are extracted, namely vv_en (energy of the image under VV polarization), vh_en (energy of the image under VH polarization), vv_homo (homogeneity of the image under VV polarization), vh_homo (homogeneity of the image under VH polarization), vv_con (contrast of the image under VV polarization), vh_con (contrast of the image under VH polarization), vv_cor (contrast of the image under VV polarization), and vh_cor (contrast of the image under VH polarization).

[0028] Step 4: Extract morphological and geographical parameters. Include: the distance of each pixel point from the typhoon center ( R ) and the Coriolis parameter ( f). First, select the IBTrACS optimal path dataset as the true typhoon location. To match the typhoon center longitude and latitude with the model output, after linearly interpolating the path data to the SAR satellite observation time points, extract the typhoon center longitude and latitude. After determining the center position, calculate the distance between each pixel position and the typhoon center longitude and latitude position R , with the unit of km. The Coriolis parameter is the angular velocity of the earth, and

[0029] is the latitude. A total of 2 morphological and geographical parameter data are extracted φ ). Step 5: Extract other additional inputs. Include the relative wind direction obtained from ERA5 data ( rain ) and the rainfall rate data provided by GPM rainfall radar ( φ ). Among them, extract the wind direction according to the included angle between the u (east-west component of the wind) and v (north-south component) components in ERA5 data, and calculate the included angle between the wind direction and the radar viewing angle to obtain the relative wind direction ( φ ). Usually, the due north direction is 0°. To avoid wind direction angle jumps, it is necessary to

[0030] Step 6: Match 19 characteristic parameters of the SFMR wind field data with the corresponding spatio-temporal information, and screen and eliminate outliers from the matching results to form a two-dimensional matrix of size n×19 as the characteristic input dataset, where n is the number of successfully matched SFMR data. The SFMR data is output separately as the true value, with a size of n×1. These two matrices form the wind speed dataset (1) Construction of model input and output features There are many parameters involved in the SAR imaging process, and the typhoon dynamic environment is complex. To improve the accuracy of typhoon wind field inversion, the present invention introduces 4 types of SAR radar parameter features, SAR texture features, morphological and geographical parameter features, and additional input features as the input features of the model, including 19 features: VH polarization SAR backscattering coefficient , VV polarization SAR backscattering coefficient , incident angle , longitude , latitude , polarization ratio ( vv / vh), the energies of VV polarization and VH polarization (vv_en, vh_en), the homogeneities (vv_homo, vh_homo), the contrasts (vv_con, vh_con), and the correlations (vv_cor, vh_cor), the distance ( R ), the Coriolis parameter ( f ), the sine and cosine components of the relative wind direction ( φ ), (dir_u, dir_v), and the rainfall rate ( rain ). The output feature of the model is the wind speed of SFMR.

[0031] (2) Feature importance analysis Use the SHAP analysis method to rank the importance contributions of the n×19 input features in the well-matched wind speed dataset in step 6 to the model, and screen out the input feature combination that makes the model inversion accuracy the highest. After analysis, there are 15 input features in the final model, namely: the VH polarization SAR backscattering coefficient , the incident angle , the longitude , the latitude , the polarization ratio ( vv / vh ), the energies of VV polarization and VH polarization (vv_en, vh_en), the homogeneities (vv_homo, vh_homo), the distance ( R ), the Coriolis parameter ( f ), the sine and cosine components of the wind direction (dir_u, dir_v), and the rainfall rate ( rain ).

[0032] (3) Division of the training set, validation set, and test set Randomly divide the n×15 input features in the wind speed dataset screened by the importance analysis in (2) above into a training set and a test set according to a ratio of 8:2. The test set is used to independently test the wind field inversion effect of the model. Take 20% of the data in the training set as the validation set to verify whether the model has good convergence. The n×1 true values are also processed in the same way.

[0033] (3) Data standardization When the numerical values of the input parameters vary greatly, if the original values are directly used for analysis, it will highlight the role of the parameters with higher numerical values in the model and relatively weaken the role of the parameters with lower numerical values. Therefore, in order to ensure the reliability of the results, it is necessary to perform standardization processing on the original input data. The Min-Max standardization method adopted in the present invention is, that is where , are the values before and after conversion respectively, and They are the minimum and maximum values of the data respectively.

[0034] (4)Division of the dual wind speed region dataset Taking the SFMR data of 33 m / s as the boundary, the input features and output features with wind speed less than 33 m / s are used as the medium and low wind speed dataset, and the input features and output features with wind speed greater than 33 m / s are used as the high wind speed dataset.

[0035] (5)Model structure The TJ_Resnet typhoon wind speed inversion model constructed in the present invention consists of a pre-training module, a low / high wind speed inversion module, feature splicing, a self-attention network module, a feature fusion module, and an inversion model. The network framework is as Figure 1 shown.

[0036] First of all, the model adopts two basic ideas of "cross-layer connection" and "cross-layer addition", constructs a pre-training model based on the ResNet deep learning network framework, trains it using the medium and low wind speed dataset and the high wind speed dataset respectively, and extracts the corresponding features. The pre-training model consists of a fully connected layer and 4 basic blocks. Among them, the basic block is improved based on the ResNet network, and the activation function uses ReLu to avoid gradient disappearance. The structure is as Figure 2 shown.

[0037] Secondly, the features are transferred to the corresponding low wind speed inversion module and high wind speed inversion module. On this basis, the feature parameters are fine-tuned to improve the prediction speed and performance of the feature extraction module, and then the feature F1 of the medium and low wind speed network model and the feature F2 of the high wind speed network model are obtained.

[0038] Then, the feature F1 and the feature F2 are spliced, and the self-attention network module is used to calculate the attention weights of each feature element and weighted sum output to perform more accurate learning on the result of the feature output, suppress the learning of irrelevant features, and improve the training speed and efficiency of the model. And the feature fusion module is used to fuse the feature parameters.

[0039] Finally, it is input into the inversion module composed of a fully connected layer to invert the wind speed.

[0040] The loss function used in the model (Loss Function) is the mean absolute error (Mean Absolute Error, MAE), which represents the average value of the distance between the inverted wind speed and the true wind speed. Using the loss function, the result of the inversion is less affected by outliers, more inclusive, and more conducive to the convergence of the model. The optimizer uses Adam (Adaptive Moment Estimation), and the learning rate is set to 0.0001, which helps to stabilize the training.

[0041] (6)Precision evaluation The selected precision evaluation indices are Root Mean Square Error (RMSE), Bias, Standard Deviation (STD), and Correlation Coefficient (R). The Root Mean Square Error (RMSE) can well identify large and small deviations and show the precision of the retrieved wind field data. The calculation formula is as follows: ; The Bias can measure the difference in systematic deviation between two types of data. The calculation formula is: ; The Mean Relative Error (MRE) can represent the proportion of the deviation between the retrieved wind speed and the SFMR data relative to the true value. The larger the Mean Relative Error, the more dispersed the deviation distribution between the two types of data. The calculation formula is: ; By calculating the correlation coefficient between the retrieved wind speed and the SFMR data, the correlation degree between the two can be accurately shown, demonstrating the linear relationship between the two types of data. The calculation formula for the correlation coefficient is as follows: ; Among them, S and I represent the retrieved wind speed value and the SFMR wind speed value respectively. Specific embodiments

[0042] In this embodiment, taking the SAR remote sensing data of Sentinel satellites (Sentinel-1 A / B) and the wind speed products of the SFMR airborne radiometer as examples, the dual-region sea surface typhoon wind speed joint inversion method (TJ_Resnet) based on transfer learning and residual network of the present invention is described.

[0043] First, install the python programming software on the user terminal, and it is necessary to configure the Python-3.10, tensoR, flow-2.16.1, keras-3.0.5 environment packages (tensoR, flow, and keras are open-source software libraries for deep learning; tensoR, flow provide a flexible platform for building, training, and deploying various complex neural network models, and it supports multiple programming languages; keras is a high-level neural network API that can provide a more concise and friendly interface based on tensoR, flow, allowing users to quickly build and run common neural network models).

[0044] Step 1: Feature importance analysis.

[0045] During the training process, various parameter inputs will affect the training efficiency and final performance of the model. If the model parameter input is small, complex patterns and features in the data cannot be captured, resulting in poor model performance and inability to predict new data well. If the model input parameters are numerous, it may overfit the training set data, be unable to generalize to new data, and increase the complexity of the model, leading to an increase in the training and inference time of the model. Therefore, in order to study the contribution of different input parameters to the performance of the TJ_Resnet model, the SHAP (SHapley Additive exPlanations) values of different input features are calculated, and feature importance analysis is carried out on them to evaluate and screen the positive or negative degrees provided by different input features to the model.

[0046] There are a total of 19 input features, including: VH polarization SAR backscattering coefficient , VV polarization SAR backscattering coefficient , incident angle , longitude , latitude , polarization ratio ( vv / vh ), energies of VV polarization and VH polarization (vv_en, vh_en), homogeneity (vv_homo, vh_homo), contrast (vv_con, vh_con) and correlation (vv_cor, vh_cor), distance ( R ), Coriolis parameter ( f ), sine and cosine components (dir_u, dir_v) of the relative wind direction ( φ ), rainfall rate ( rain ). The SHAP diagrams of the 19 input features are plotted as honeycomb diagrams, as shown in Figure 3 . The color of the points represents the magnitude of the feature values, usually represented by a gradient color indicating the feature values from low to high. The abscissa is the SHAP value, and the ordinate shows the importance of each input factor in the typhoon wind field inversion from high to low.

[0047] It can be found through Figure 3 that the VH polarization backscattering coefficient The sum of distances (R) contributes the most significantly to the model. To more precisely analyze the impact of different parameters on the training accuracy of the model, 7 groups of experiments were set up. Different input parameters were successively input into the model for training, aiming to find the optimal combination of input parameters with the highest training accuracy. The input parameters of the model are shown in Table 1. Among them, model1 only inputs the variables required for inversion by traditional physical algorithms, namely the backscattering coefficient, relative wind direction, and incident angle; model2 inputs all the radar parameter variables of SAR images; model3 adds the rainfall rate and Coriolis force; model4 - model7 introduce the impact of texture feature parameters on image accuracy. Among them, model6 explores whether the input of the polarization ratio can improve the accuracy of the training results of the model.

[0048] Table 1 List of Model Input Parameters

[0049] The "wind speed dataset" used in the experiment contains a total of 18,395 valid data pairs, of which 80% are the training set (20% of the training set is the validation set), and 20% is the test set. The training results of the 7 groups of experiments on the test set are shown in Table 2. The results show that the input parameter combination of model6 has the best accuracy and performance on the test set, and it is the optimal parameter combination among multiple groups of data.

[0050] Since the present invention focuses on solving the problem of poor inversion accuracy of high typhoon wind speeds, the input parameters of model6 are selected as the optimal parameter combination, that is, inputting the VH polarization SAR backscattering coefficient 、the VV polarization SAR backscattering coefficient 、the incident angle 、longitude 、latitude 、polarization ratio ( vv / vh ), the energies of VV polarization and VH polarization (vv_en, vh_en), homogeneity (vv_homo, vh_homo), distance ( R ), Coriolis parameter ( f ), the sine and cosine components of the wind direction (dir_u, dir_v), rainfall rate ( rain ), a total of 15 kinds.

[0051] Table 2 Impact of Different Input Parameters on Model Accuracy

[0052] Step 2: Accuracy analysis. The present invention conducts initial training in the low - to - medium wind speed range and the high wind speed range respectively, uses the method of transfer learning to achieve feature transfer, and uses a self - attention network to more precisely extract features, enabling the model to achieve better inversion performance in both wind speed ranges. Therefore, as Figure 4As shown, the root mean square error (RMSE) and mean bias (MB) of the overall joint inversion algorithm are 1.40 m / s and -0.05 m / s respectively, the correlation coefficient reaches 0.99, and the RMSE in the high wind speed area is reduced to 2.55 m / s. Compared with the inversion results directly using the full wind speed dataset, the RMSE in the high wind speed area is reduced by 0.65 m / s, improving the accuracy of typhoon wind speed inversion.

[0053] Step 3: Compare with other classical inversion models. To verify the inversion effect of the typhoon wind speed inversion model in this implementation case, the inversion results are compared with the inversion results of classical neural network models - Deep Neural Network (DNN) and Random Forest. Among them, the comparison results of the test set are shown in Table 3. Compared with the DNN model and the Random Forest model, the root mean square error of the inversion of the TJ_Resnet algorithm of the present invention is reduced by 0.6 m / s to 1.58 m / s.

[0054] Table 3 Comparison of the accuracy of wind speed inversion by different neural network models

[0055] In summary, the present invention proposes a joint inversion algorithm for typhoon wind speed on the sea surface by dual-polarization SAR based on transfer learning and residual network. Through the combined analysis of multiple features such as SAR radar parameters, texture parameters, morphological and geographical parameters, and physical environment factors, the optimal parameter combination suitable for the inversion of the SAR sea surface typhoon wind field under extreme weather is determined by the feature importance analysis method. The inversion results in the medium and low wind speed areas and the high wind speed areas are jointly modeled by the methods of transfer learning and feature fusion to achieve the complementary fusion of dual-polarization SAR data. This inversion algorithm can effectively alleviate the small sample learning problem in the high wind speed area, solve the feature mismatch problem in the cross-wind speed segment migration, and improve the stability and accuracy of typhoon wind speed inversion (especially in the high wind speed area).

[0056] Although the specific implementation manner of the present invention is described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that based on the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative labor are still within the protection scope of the present invention.

Claims

1. A joint inversion algorithm for typhoon wind speed over the sea surface using dual-polarization SAR based on transfer learning and residual network, characterized in that, It includes the following steps: Construct a joint inversion algorithm model for sea surface typhoon wind speed; the model structure includes: constructing a pre-training model, a medium-low wind speed inversion model, and a high wind speed inversion model using a ResNet network model; The model structure also includes a feature splicing model and a feature fusion model for splicing and fusing the features extracted from the medium-low wind speed inversion model and the high wind speed inversion model, and an inversion model for typhoon wind speed inversion; Obtain model sample data for sea surface typhoon wind speed inversion, then divide the model sample data into a medium-low wind speed data set and a high wind speed data set, and input them into the pre-training model for training and feature extraction, and transfer the features to the corresponding low wind speed inversion model and high wind speed inversion model; Input all the sample data of the model into the medium-low wind speed inversion model and the high wind speed inversion model for independent training respectively; Use the trained joint inversion algorithm model for sea surface typhoon wind speed to invert the sea surface typhoon wind speed.

2. The joint inversion algorithm for dual-polarization SAR sea surface typhoon wind speed based on transfer learning and residual network according to claim 1, characterized in that, The pre-training model consists of one fully connected layer and 4 basic blocks; the basic blocks use the ResNet network model, and the activation function uses ReLu to improve the model to avoid gradient disappearance.

3. The joint inversion algorithm for dual-polarization SAR sea surface typhoon wind speed based on transfer learning and residual network according to claim 1, wherein, The model sample data includes: image data and physical auxiliary data; the image data includes: ; the physical auxiliary data includes: vv_en, vh_en, vv_homo, vh_homo, R、 f , dir_u, dir_v, and rain; Among them, is the VH polarization SAR backscattering coefficient; is the VV polarization SAR backscattering coefficient; is the incident angle; is the longitude; is the latitude; vv / vh is the polarization ratio, vv_en is the energy of VV polarization; vh_en is the energy of VH polarization; vv_homo is the homogeneity of VV polarization; vh_homo is the homogeneity of VH polarization; R is the distance; is the Coriolis parameter; dir_u is the sine component of the wind direction; dir_v is the cosine component of the wind direction; rain is the rainfall rate.

4. The joint inversion algorithm for typhoon wind speed over the sea surface using dual-polarization SAR based on transfer learning and residual network according to claim 1, characterized in that, After obtaining the model sample data, the image data and physical auxiliary data need to be aligned in time and space into a unified time series.

5. The joint inversion algorithm for dual-polarization SAR sea surface typhoon wind speed based on transfer learning and residual network according to claim 1, characterized in that, After obtaining the model sample data, it is necessary to perform preprocessing, data standardization processing, randomly divide the ratio of the training set and the test set, and use part of the data in the training set as the validation set.

6. The joint inversion algorithm for dual-polarization SAR sea surface typhoon wind speed based on transfer learning and residual network according to claim 1, characterized in that In the step of dividing the model sample data into a medium-low wind speed data set and a high wind speed data set, the input features and output features with a wind speed less than 33 m / s are used as the medium-low wind speed data set, and the input features and output features with a wind speed greater than 33 m / s are used as the high wind speed data set.

7. The joint inversion algorithm for typhoon wind speed over the sea surface using dual-polarization SAR based on transfer learning and residual network according to claim 1, wherein, In the step of using the trained joint inversion algorithm model for sea surface typhoon wind speed to invert the sea surface typhoon wind speed, the evaluation indexes for the trained model are: RMSE, Bias, MRE, and R; among them, RMSE is the root mean square error, Bias is the mean deviation, MRE is the mean relative error, and R is the correlation coefficient.

8. The joint inversion algorithm for dual-polarization SAR sea surface typhoon wind speed based on transfer learning and residual network according to claim 1, characterized in that, The wind speed data is used as the true label and used as the output feature of the model.

Images