A joint inversion algorithm for sea surface typhoon wind speed from dual-polarization SAR based on transfer learning and residual network

Abstract

The present invention discloses a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network, and relates to the technical field of marine environmental information monitoring. The present invention comprises the following steps: constructing a sea surface typhoon wind speed joint inversion algorithm model; the model structure comprises: 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 comprises a feature splicing model and a feature fusion model for splicing and fusing features extracted from the medium-low wind speed inversion model and the high wind speed inversion model, as well as an inversion model for typhoon wind speed inversion. The present invention designs a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network. The inversion algorithm can effectively alleviate the problem of small sample learning in high wind speed areas, solve the problem of feature mismatch in cross-wind speed segment migration, and improve the stability and accuracy of typhoon wind speed inversion, especially wind speed inversion in high wind speed areas.

Description

Technical Field

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

[0002] Typhoons are extremely destructive, often accompanied by strong winds, heavy rain, and storm surges. They pose a serious threat to the safety of life and property of coastal ports and residents, and cause enormous economic losses. Accurate estimation of typhoon wind fields is crucial for maritime navigation safety, disaster warning, and typhoon monitoring. However, obtaining high-precision typhoon observation data remains a challenging task.

[0003] With the development of remote sensing technology, satellite remote sensing data has been widely used to monitor the development and evolution of typhoons. Synthetic Aperture Radar (SAR), a remote sensing technology based on radar waves, offers unique advantages in typhoon wind field inversion due to its high resolution and adaptability to strong winds. SAR can obtain sea surface wind speeds exceeding 40 m / s, clearly displaying key features such as sea surface wind streaks, typhoon structure, and spiral rainbands, thus compensating for the lower resolution of other spaceborne sensors.

[0004] Currently, wind field inversion methods based on SAR data primarily rely on geophysical model functions (GMFs), such as CMOD4, CMOD5, and CMOD7. These models can be used to invert wind speeds from C-band co-polarization (VV / HH) SAR imagery. When wind speeds are less than 20 m / s, the inversion error is approximately 2 m / s. However, traditional wind field inversion methods based on the CMOD model suffer from significant errors in extreme weather conditions, especially at high wind speeds. Due to saturation of the scattering coefficient, the inversion accuracy remains limited in high wind speed ranges.

[0005] In recent years, the development of deep learning (DL) methods has provided new technical approaches for typhoon wind field inversion. Deep learning models such as deep neural networks (DNNs), residual neural networks (ResNets), and random forest models possess powerful nonlinear fitting, feature inference, and transfer application capabilities. They are capable of learning and capturing complex features and correlations in SAR imagery, demonstrating strong adaptability and generalization capabilities. These models are not only adaptable to diverse ocean environments and meteorological conditions, but can also integrate multi-source remote sensing data, further improving the stability and accuracy of typhoon wind field estimation.

[0006] However, the inventors believe that the following technical problems still exist when introducing deep learning into SAR technology for typhoon inversion:

[0007] (1) Although the deep learning model can improve the accuracy of wind speed inversion, the inversion error is still large under high wind speed conditions (>33 m / s), and its high wind speed inversion error still exceeds 4 m / s.

[0008] (2) Typhoon observation data is relatively scarce, which makes the deep learning method driven by pure image data restricted by sample distribution during the training process. Especially in high wind speed areas, due to the small number of effective samples, it is difficult for the model to fully learn the influence of all incident angles, wind directions and rainfall conditions, resulting in large inversion errors.

[0009] (3) Existing deep learning models are prone to feature mismatch when migrating across wind speed ranges, which affects the performance of the model.

[0010] Therefore, how to solve the above technical problems is a technical problem that technicians in this field urgently need to solve.

[0011] The information disclosed in this background technology section is only intended to enhance understanding of the overall background of the invention and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the Invention

[0012] In response to the above technical problems, an embodiment of the present invention provides a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network to solve the problems raised in the above background technology.

[0013] The present invention provides the following technical solution: a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network, comprising the following steps:

[0014] Construct a joint inversion algorithm model for sea surface typhoon wind speed; its model structure includes: a pre-training model based on the ResNet network model, a medium and low wind speed inversion model, and a high wind speed inversion model;

[0015] The model structure also includes a feature splicing model and a feature fusion model for splicing and fusing features extracted from the medium and low wind speed inversion models and the high wind speed inversion models, as well as an inversion model for typhoon wind speed inversion;

[0016] Obtain model sample data for inverting sea surface typhoon wind speed, divide the model sample data into medium and low wind speed data sets and high wind speed data sets, input them into the pre-trained model for training and feature extraction, and migrate the features to the corresponding low wind speed inversion model and high wind speed inversion model;

[0017] All the sample data of the model are input into the low and medium wind speed inversion model and the high wind speed inversion model for independent training;

[0018] The sea surface typhoon wind speed is inverted using the trained sea surface typhoon wind speed joint inversion algorithm model.

[0019] Preferably, the pre-trained model consists of a fully connected layer and 4 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: ; Physical auxiliary data include: vv_en, vh_en, vv_homo, vh_homo, R, f , dir_u, dir_v, and rain; in, is the VH polarization SAR backscatter coefficient; is the VV polarization SAR backscatter coefficient; is the angle of incidence; is the longitude; is 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 It’s 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 It is the rainfall rate. It should be emphasized that when using pure image data (SAR satellite remote sensing image data) for typhoon inversion, pure image data is pure data (SAR satellite remote sensing image data is mainly obtained by scanning and measuring the ground or sea surface through radar systems). This pure data usually does not consider the underlying physical mechanism or domain knowledge, thus reducing the interpretability and generalization ability of the model.

[0020] 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 driven jointly by data and physical knowledge, solving the problem of poor interpretability and generalization ability caused by relying solely on pure data such as SAR satellite remote sensing image data for model driving.

[0021] Preferably, after the model sample data is acquired, the image data and the physical auxiliary data need to be aligned in time and space to form a unified time series.

[0022] Preferably, after the model sample data is obtained, it needs to be preprocessed, data normalized, 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 a validation set.

[0023] Preferably, in the step of dividing the model sample data into medium and low wind speed data sets and high wind speed data sets, the input features and output features of wind speeds less than 33 m / s are used as medium and low wind speed data sets, and the input features and output features of wind speeds greater than 33 m / s are used as high wind speed data sets.

[0024] Preferably, in the step of inverting the sea surface typhoon wind speed using the trained sea surface typhoon wind speed joint inversion algorithm model, the evaluation indicators of the model training are: RMSE, Bias, MRE and R; among which RMSE is the root mean square error, Bias is the mean deviation, MRE is the mean relative error, and R is the correlation coefficient.

[0025] Preferably, wind speed data is used as the true label as the output feature of the model.

[0026] An embodiment of the present invention provides a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network, which 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. The inversion algorithm can effectively alleviate the small sample learning problem in high wind speed areas, solve the feature mismatch problem in cross-wind speed segment migration, and improve typhoon wind speed inversion, especially the stability and accuracy of wind speed inversion in high wind speed areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is the network framework diagram of the TJ_Resnet typhoon wind speed inversion model in the present invention;

[0028] Figure 2 It is a structural diagram of the basic block in the present invention;

[0029] Figure 3 SHAP honeycomb diagram of input features in the present invention;

[0030] Figure 4 The comparison results of the inversion results of the direct inversion algorithm (Resnet) and the joint inversion algorithm (TJ_Resnet) in the present invention are shown in FIG.

[0031] (4a) is the comparison between the wind speed retrieved by the TJ-Resnet model and the wind speed of the SFMR (full wind speed area);

[0032] (4b) is the comparison between the wind speed retrieved by the TJ-Resnet model and the SFMR wind speed (medium and low wind speed areas);

[0033] (4c) is the comparison between the wind speed retrieved by the TJ-Resnet model and the SFMR wind speed (high wind speed area);

[0034] (4d) is the comparison between the wind speed retrieved by the Resnet model and the wind speed of the SFMR (full wind speed area);

[0035] (4e) is the comparison between the wind speed retrieved by the Resnet model and the SFMR wind speed (medium and low wind speed areas);

[0036] (4f) is the comparison between the wind speed retrieved by the Resnet model and the wind speed of the SFMR (high wind speed area);

[0037] Figure 5 This is a 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. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] In response to 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 solution is as follows:

[0040] The following is combined with Figure 1-5 , and specific implementation methods are further described to illustrate the present invention.

[0041] The design ideas and optimization steps of the 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:

[0042] Step 1: Acquire Sentinel-1 SAR data containing typhoons (Sentinel-1 is a sentinel satellite launched by the European Space Agency) and SFMR (Stepped-Frequency Microwave Radiometer) airborne radiometer data (wind speed data) that match the typhoons in time and space, rainfall data provided by the Global Precipitation Mission (GPM), and the fifth-generation wind reanalysis data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF-ERA5).

[0043] SFMR data served as the actual wind field, and ERA5 data served as the wind direction input. To ensure data timeliness, linear interpolation was performed on the ERA5 wind field data and GPM rainfall data based on the SAR imaging time to generate data with the same timestamp. SFMR wind field data were selected within ±1 hour of the SAR imaging time.

[0044] Step 2: Extract relevant radar parameters from the SAR image. SAR image preprocessing methods were used, including radiometric calibration, filtering noise suppression (11×11 Lee filtering), thermal noise removal, geometric correction, resampling, and land masking. To suppress the influence of SAR speckle noise and maintain the SAR image's ability to observe fine typhoon structure, the SAR image resolution was resampled to 500 m.

[0045] Radar parameters of the extracted SAR image, including: VH polarization SAR backscatter coefficient , VV polarization SAR backscatter coefficient , angle of incidence ,longitude ,latitude , and calculate the ratio of the radar backscatter coefficients under VV and VH polarizations, that is, the polarization ratio ( PR, vv / vh ). A total of 6 types of radar parameter data are extracted from SAR images.

[0046] Step 3: Extract texture feature parameters from the SAR image. The texture features of the VV- and VH-polarized SAR images are calculated using the gray-level co-occurrence matrix (GLCM). The GLCM reflects comprehensive information about the image's grayscale direction, adjacent spacing, and amplitude of change. It represents the joint probability distribution of the simultaneous occurrence of two grayscale pixels separated by a distance D in the image and reflects the grayscale correlation between adjacent pixels. The texture feature parameters used in this invention are energy, homogeneity, contrast, and correlation, and are extracted separately from the two polarized images.

[0047] A total of 8 types of texture feature parameter data of SAR images 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).

[0048] Step 4: Extract morphological and geographic parameters. Including: the distance of each pixel from the typhoon center ( R ) and the Coriolis parameter ( f First, the IBTrACS optimal path dataset was selected as the actual typhoon position. To match the longitude and latitude of the typhoon center with the model output, the path data was linearly interpolated to the SAR satellite observation time point to extract the longitude and latitude of the typhoon center. After determining the center position, the distance between each pixel position and the longitude and latitude of the typhoon center was calculated. R , in km. Coriolis parameters is the angular velocity of the Earth, It is latitude. Two types of morphological and geographic parameter data are extracted.

[0049] Step 5: Extract additional inputs. This includes relative wind direction from ERA5 data ( φ ) and rainfall rate data provided by GPM rainfall radar ( rain Among them, the wind direction is extracted according to the angle between the u (east-west component of the wind) and v (north-south component) components in the ERA5 data, and the angle between the wind direction and the radar viewing angle is calculated to obtain the relative wind direction ( φ ), usually the north direction is 0°, in order to avoid the wind direction angle jump, you need to φ Convert them into sine and cosine components, namely dir_u and dir_v. A total of three types are extracted. Step 6: Match the SFMR wind field data with the 19 characteristic parameters of the spatiotemporal information. The matching results are screened and outliers are removed to form a two-dimensional matrix of size n × 19, which serves as the feature input dataset, where n is the number of successfully matched SFMR data. The SFMR data are output as separate ground truth values, with a size of n × 1. These two matrices constitute the wind speed dataset.

[0050] Step 7: Use Python programming language to build a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm (TJ_Resnet) based on transfer learning and residual network. The specific steps are as follows:

[0051] (1) Construction of model input and output features

[0052] SAR involves many parameters in the imaging process, and the typhoon dynamic environment is complex. In order to improve the accuracy of typhoon wind field inversion, this paper introduces SAR radar parameter features, SAR texture features, morphological and geographic parameter features, and additional input features, a total of 4 categories, as input features of the model, including 19 features: VH polarization SAR backscatter coefficient , VV polarization SAR backscatter coefficient , angle of incidence ,longitude ,latitude , polarization ratio ( vv / vh ), VV polarization and VH polarization energy (vv_en, vh_en), homogeneity (vv_homo, vh_homo), contrast (vv_con, vh_con) and correlation (vv_cor, vh_cor), distance ( R ), Coriolis parameters ( f ), relative wind direction ( φ )’s sine and cosine components (dir_u, dir_v), rainfall rate ( rain The output feature of the model is the wind speed of the SFMR.

[0053] (2) Feature importance analysis

[0054] The SHAP analysis method is used to sort the importance of the n×19 input features in the wind speed dataset matched in step 6 to the model, and the input feature combination with the highest model inversion accuracy is selected. After analysis, the final model has 15 input features, namely: VH polarization SAR backscatter coefficient , angle of incidence ,longitude ,latitude , polarization ratio ( vv / vh ), VV polarization and VH polarization energy (vv_en, vh_en), homogeneity (vv_homo, vh_homo), distance ( R ), Coriolis parameters ( f ), the sine and cosine components of wind direction (dir_u, dir_v), rainfall rate ( rain ).

[0055] (3) Training set, validation set and test set division

[0056] The n×15 input features in the wind speed dataset selected through importance analysis in (2) above are randomly divided into a training set and a test set in a ratio of 8:2. The test set is used to independently verify the wind field inversion effect of the model. 20% of the data in the training set is used as a validation set to verify whether the model has good convergence. The n×1 true value is processed in the same way.

[0057] (3) Data standardization

[0058] When the values of the input parameters differ greatly, if the original values are used directly for analysis, the effect of the parameters with higher values in the model will be highlighted, and the effect of the parameters with lower values will be relatively weakened. Therefore, in order to ensure the reliability of the results, it is necessary to standardize the original input data. The present invention adopts the Min-Max standardization method, that is, in , are the values before and after conversion, and are the minimum and maximum values of the data, respectively.

[0059] (4) Division of dual wind speed area dataset

[0060] Taking the SFMR data of 33 m / s as the boundary, the input features and output features of wind speeds less than 33 m / s are regarded as the medium and low wind speed dataset, and the input features and output features of wind speeds greater than 33 m / s are regarded as the high wind speed dataset.

[0061] (5) Model structure

[0062] The TJ_Resnet typhoon wind speed inversion model constructed by the present invention includes a pre-training module, a low / high wind speed inversion module, a feature splicing module, a self-attention network module, a feature fusion module and an inversion model. The network framework is as follows Figure 1 shown.

[0063] First, the model adopts the two basic ideas of "cross-layer connection" and "cross-layer addition". A pre-training model is built based on the ResNet deep learning network framework. It is trained using medium and low wind speed datasets and high wind speed datasets respectively, and the corresponding features are extracted. The pre-training model consists of a fully connected layer and 4 layers of basic blocks. The basic block is improved based on the ResNet network. The activation function uses ReLu to avoid gradient disappearance. The structure is as follows Figure 2 shown.

[0064] 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, thereby obtaining the feature F1 of the low- and medium-wind speed network model and the feature F2 of the high-wind speed network model.

[0065] Then, feature F1 and feature F2 are concatenated, and the self-attention network module is used to calculate the attention weight of each feature element and output the weighted sum. This allows for more accurate learning of the feature output results, suppressing the learning of irrelevant features, and improving the training speed and efficiency of the model. The feature fusion module is then used to fuse the feature parameters.

[0066] Finally, the wind speed is obtained by inputting it into the inversion module composed of a fully connected layer.

[0067] The model uses the mean absolute error (MAE) as the loss function, representing the average difference between the inverted wind speed and the true wind speed. This loss function makes the inversion results less susceptible to outliers, more inclusive, and more conducive to model convergence. The optimizer uses Adam (Adaptive Moment Estimation) and a learning rate of 0.0001 to help stabilize training.

[0068] (6) Accuracy assessment

[0069] The selected accuracy evaluation indices are root mean square error (RMSE), mean bias (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 inverted wind field data. The calculation formula is as follows: ;

[0070] The average bias can measure the difference in the systematic deviation between two data sets. The calculation formula is: ;

[0071] The mean relative error (MRE) can be expressed as the ratio of the deviation between the inverted wind speed and the SFMR data to the true value. A larger mean relative error indicates a more dispersed distribution of the deviation between the two data. The calculation formula is: ;

[0072] By calculating the correlation coefficient between the inverted wind speed and the SFMR data, we can accurately show the degree of correlation between the two and show the linear relationship between the two data. The correlation coefficient calculation formula is as follows: ;

[0073] in, S and I represent the inverted wind speed value and SFMR wind speed value, respectively. Specific embodiments

[0074] This embodiment uses Sentinel-1 A / B SAR remote sensing data and SFMR airborne radiometer wind speed products as examples to illustrate the dual-region sea surface typhoon wind speed joint inversion method (TJ_Resnet) based on transfer learning and residual network of the present invention.

[0075] First, install Python programming software on the user terminal. You also need to equip it with Python 3.10, tensoR, flow 2.16.1, and keras 3.0.5 environment packages. (tensoR, flow, and keras are open-source software libraries for deep learning. TensoR and flow provide a flexible platform for building, training, and deploying various complex neural network models in multiple programming languages. Keras is an advanced neural network API that provides a simpler and more user-friendly interface based on tensoR and flow, allowing users to quickly build and run common neural network models.)

[0076] Step 1: Feature importance analysis.

[0077] During training, various input parameters affect the model's training efficiency and ultimate performance. A model with too few input parameters cannot capture complex patterns and features in the data, resulting in poor model performance and inability to predict new data. A model with too many input parameters may overfit the training set data, failing to generalize to new data, increasing model complexity and increasing training and inference time. Therefore, to investigate the contribution of different input parameters to the performance of the TJ_Resnet model, we calculated the SHAP (SHapley Additive exPlanations) values of various input features and performed feature importance analysis to assess and screen the degree to which different input features contribute positively or negatively to the model.

[0078] There are 19 input features, including: VH polarization SAR backscatter coefficient , VV polarization SAR backscatter coefficient , angle of incidence ,longitude ,latitude , polarization ratio ( vv / vh ), VV polarization and VH polarization energy (vv_en, vh_en), homogeneity (vv_homo, vh_homo), contrast (vv_con, vh_con) and correlation (vv_cor, vh_cor), distance ( R ), Coriolis parameters ( f ), relative wind direction ( φ )’s sine and cosine components (dir_u, dir_v), rainfall rate ( rain ). The SHAP diagrams of the 19 input features are drawn as honeycomb diagrams, such as Figure 3The color of the point represents the size of the eigenvalue, usually represented by a gradient color from low to high eigenvalues, the horizontal axis is the SHAP value, and the vertical axis from high to low shows the importance of each input factor in the typhoon wind field inversion.

[0079] pass Figure 3 It can be found that the VH polarization backscattering coefficient The distance (R) contributes most to the model's importance. To more accurately analyze the impact of different parameters on model training accuracy, seven sets of experiments were conducted, sequentially inputting different input parameters into the model for training. The goal was to identify the optimal input parameter combination with the highest training accuracy. The model input parameters are shown in Table 1. Model 1 only inputs the variables required for inversion using traditional physical algorithms, namely, backscatter coefficient, relative wind direction, and angle of incidence; model 2 inputs all radar parameter variables for SAR images; model 3 adds rainfall rate and Coriolis force; models 4 through 7 introduce the impact of texture feature parameters on image accuracy. Model 6 explores whether the input of polarization ratio improves the accuracy of the model training results.

[0080] Table 1 List of model input parameters

[0081]

[0082] The wind speed dataset used in the experiment contains 18,395 valid data pairs, 80% of which are training sets (20% of which are validation sets) and 20% are test sets. The training results of seven experiments on the test set are shown in Table 2. These results show that the model 6 input parameter combination achieves the best accuracy and performance on the test set, and is the optimal parameter combination among multiple data sets.

[0083] Since the present invention focuses on solving the problem of poor inversion accuracy of typhoon high wind speed, the input parameters of model6 are selected as the optimal parameter combination, that is, the input VH polarization SAR backscatter coefficient , VV polarization SAR backscatter coefficient , angle of incidence ,longitude ,latitude , polarization ratio ( vv / vh ), VV polarization and VH polarization energy (vv_en, vh_en), homogeneity (vv_homo, vh_homo), distance ( R ), Coriolis parameters ( f ), the sine and cosine components of wind direction (dir_u, dir_v), rainfall rate ( rain ), a total of 15 types.

[0084] Table 2 Effects of different input parameters on model accuracy

[0085]

[0086] Step 2: Accuracy analysis. The present invention performs initial training in the low and medium wind speed ranges and the high wind speed range, uses the transfer learning method to achieve feature transfer, and uses the self-attention network to extract features more accurately, so that the model can achieve good inversion performance in both wind speed ranges. Figure 4 As shown in the figure, the overall root mean square error (RMSE) and mean bias (MB) of the joint inversion algorithm are 1.40 m / s and -0.05 m / s, respectively, with a correlation coefficient of 0.99. The RMSE in high-wind-speed regions is reduced to 2.55 m / s. Compared with the inversion results directly using the full wind speed dataset, the RMSE in high-wind-speed regions is reduced by 0.65 m / s, improving the accuracy of typhoon wind speed inversion.

[0087] Step 3: Comparison with other classic inversion models. To verify the inversion performance of the typhoon wind speed inversion model in this example, the inversion results were compared with those of classic neural network models—deep neural network (DNN) and random forest (RandomForest). The comparison results on the test set are shown in Table 3. Compared with the DNN model and random forest model, the root mean square error of the inversion of the TJ_Resnet algorithm of the present invention was reduced by 0.6 m / s to 1.58 m / s.

[0088] Table 3 Comparison of wind speed retrieval accuracy of different neural network models

[0089]

[0090] In summary, the present invention proposes a dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual networks. A multi-feature combination analysis, including SAR radar parameters, texture parameters, morphological and geographic parameters, and physical environmental factors, is performed. The optimal parameter combination suitable for SAR sea surface typhoon wind field inversion in extreme weather conditions is determined using feature importance analysis. The inversion results for low- to medium-speed wind speed regions and high-speed wind speed regions are jointly modeled using transfer learning and feature fusion methods to achieve complementary fusion of dual-polarization SAR data. This inversion algorithm can effectively alleviate the small sample learning problem in high-speed wind regions, resolve the feature mismatch problem in cross-wind speed segment migration, and improve the stability and accuracy of typhoon wind speed inversion (especially in high-speed regions).

[0091] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network, characterized by: The following steps are involved: Construct a joint inversion algorithm model for sea surface typhoon wind speed; its model structure includes: a pre-training model based on the ResNet network model, a medium and low wind speed inversion model, and a high wind speed inversion model; The model structure also includes a feature splicing model and a feature fusion model for splicing and fusing features extracted from the medium and low wind speed inversion models and the high wind speed inversion models, as well as an inversion model for typhoon wind speed inversion; Obtain model sample data for inverting sea surface typhoon wind speed, divide the model sample data into medium and low wind speed data sets and high wind speed data sets, input them into the pre-trained model for training and feature extraction, and migrate the features to the corresponding low wind speed inversion model and high wind speed inversion model; All the sample data of the model are input into the low and medium wind speed inversion model and the high wind speed inversion model for independent training; The sea surface typhoon wind speed is inverted using the trained sea surface typhoon wind speed joint inversion algorithm model; Model sample data includes: image data and physical auxiliary data; image data includes: σ VH , σ VV , θ, lon, lat, vv / vh; physical auxiliary data include: vv_en, vh_en, vv_homo, vh_homo, R, f, dir_u, dir_v and rain; Among them, σ VH is the VH polarization SAR backscatter coefficient; σ VV is the VV polarization SAR backscatter coefficient; θ is the angle of incidence; lon is the longitude; lat 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 range; f is the Coriolis parameter; dir_u is the sine component of wind direction; dir_v is the cosine component of wind direction; rain is the rainfall rate.

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

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

4. The dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network according to claim 1 is characterized in that: After the model sample data is obtained, it needs to be preprocessed, data normalized, and the ratio of the training set and test set needs to be randomly divided, and part of the data in the training set needs to be used as the validation set.

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

6. The dual-polarization SAR sea surface typhoon wind speed joint inversion algorithm based on transfer learning and residual network according to claim 1 is characterized in that: In the step of inverting the sea surface typhoon wind speed using the trained sea surface typhoon wind speed joint inversion algorithm model, the evaluation indicators of the model training are: RMSE, Bias, MRE and R; among which RMSE is the root mean square error, Bias is the mean deviation, MRE is the mean relative error, and R is the correlation coefficient.

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

Images