A polar sea ice thickness inversion method based on multi-source remote sensing data fusion

Abstract

The present invention provides a polar sea ice thickness inversion method based on multi-source remote sensing data fusion. The method extracts features from SAR data, MODIS data, AMSR2 data, and MIRAS data within a target area, constructs image and sequence feature datasets, fuses the image and sequence feature datasets using a designed multimodal data fusion network, obtains a new sequence feature dataset, trains the sequence feature dataset using multiple machine learning regression models, selects the machine learning regression model with the best performance for sea ice thickness inversion, deletes the optical data, and retrains the sea ice thickness inversion model. Cloud mask data is then used to divide the target area into clouded and cloudless areas, and corresponding inversion models are applied to the clouded and cloudless areas, respectively. The present invention fully integrates the advantages of various remote sensing data in sea ice thickness inversion, can realize continuous large-scale sea ice thickness information extraction, and has better universality.

Description

Technical Field

[0001] The present invention relates to the technical field of sea ice thickness inversion, and in particular to a polar sea ice thickness inversion method based on multi-source remote sensing data fusion. Background Art

[0002] Sea ice thickness is a crucial parameter for studying polar environments, climate change, and marine ecosystems. Existing technologies typically collect sea ice thickness data using ships, buoys, and aerial observation platforms. However, due to difficulties in calibrating different sensors and observation methods, existing technologies struggle to achieve large-scale, long-term sea ice thickness monitoring. Furthermore, polar marine navigation and maritime operations urgently require sea ice thickness maps with a spatial resolution better than 1 km.

[0003] With the advancement of remote sensing imaging technology, global sea ice thickness retrieval has become possible. Multi-source remote sensing imaging technology combines the advantages of multiple sensors, including visible light, infrared, and microwave sensors, to overcome the limitations of single data sources and improve the accuracy and reliability of sea ice thickness retrieval. For example, visible light and infrared imagery can provide high-resolution sea ice surface information, while microwave remote sensing (such as radar and passive microwave) can penetrate clouds and darkness to provide all-weather sea ice thickness information. By integrating these multi-source remote sensing data, sea ice thickness can be more accurately retrieved and continuous monitoring over large areas and long periods of time can be achieved.

[0004] Currently, commonly used sea ice thickness retrieval methods rely primarily on data obtained from passive microwave remote sensing, radar altimeters, and laser altimeters. Passive microwave remote sensing (such as SSM / I and AMSR-E) estimates sea ice thickness by measuring the brightness temperature of sea ice, and this method has been widely used worldwide. However, due to the low resolution of passive microwaves and their limited ability to distinguish between thin and thick ice, their retrieval accuracy is limited in complex ice conditions. Radar altimeters (such as CryoSat-2) and laser altimeters (such as ICESat) estimate sea ice thickness by measuring the height difference between the sea ice surface and sea level. This method performs well in thick ice areas, but its retrieval accuracy decreases significantly in thin ice areas or when affected by snow on the ice surface. C-band SAR can provide high-resolution images and capture small-scale sea ice features such as ice ridges, cracks and melt ponds. Because it is very sensitive to the surface roughness, texture and dielectric properties of sea ice, it can distinguish different types of sea ice (such as first-year ice, multi-year ice and new ice). However, due to its limited penetration ability, it has the defects of limited ability to invert the thickness of thin ice and is greatly affected by surface roughness and snow accumulation. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a polar sea ice thickness inversion method based on multi-source remote sensing data fusion to improve the inversion credibility and accuracy.

[0006] Technical solution: A polar sea ice thickness inversion method based on multi-source remote sensing data fusion, including the following steps:

[0007] S1. Collect multi-source remote sensing data and corresponding in-situ sea ice thickness measurement data within the target area;

[0008] S2. Geo-register the multi-source remote sensing data to obtain an image dataset, a sequence dataset, and a physical parameter dataset, respectively. Extract sequence features from the sequence dataset to obtain a sequence feature dataset. Merge the image dataset and the sequence feature dataset into a multimodal fusion dataset, and randomly divide the multimodal fusion dataset into a first training set and a first test set according to a set ratio.

[0009] S3. Construct a multimodal data fusion network, input the first training set into the multimodal data fusion network for training, and obtain a trained multimodal data fusion network;

[0010] S4. Input the first test set into the trained multimodal data fusion network for feature fusion to obtain a fused sequence feature dataset, merge the fused sequence feature dataset with the corresponding physical parameter dataset into a sea ice thickness inversion dataset, and randomly divide the sea ice thickness inversion dataset into a second training set and a second test set according to a set ratio;

[0011] S5. Divide the second training set into several training subsets, input each training subset into several different machine learning regression models for training, obtain sea ice thickness inversion data, compare the data with the corresponding in-situ sea ice thickness measured data, obtain performance evaluation indicators, and select the machine learning regression model with the best performance evaluation indicator as the sea ice thickness inversion model;

[0012] S6. Input the second test set into the sea ice thickness inversion model and output corresponding sea ice thickness inversion data.

[0013] Specifically, the multi-source remote sensing data include synthetic aperture radar SAR data, moderate resolution imaging spectrometer MODIS data, advanced microwave scanning radiometer 2 AMSR2 data and L-band synthetic aperture imaging radiometer MIRAS data; the in-situ sea ice thickness measurement data are IceBridge airborne detection data.

[0014] Specifically, step S2 includes:

[0015] After preprocessing, the multi-source remote sensing data were georeferenced according to the latitude and longitude information of the IceBridge airborne detection data, and projection transformation was performed to uniformly transform the multi-source remote sensing data into a polar position projection. Sequence features were extracted from SAR data, MODIS data, AMSR2 data, and MIRAS data to obtain sequence feature datasets, including SAR sequence feature datasets, MODIS sequence feature datasets, AMSR2 sequence feature datasets, MIRAS sequence feature datasets, and physical parameter sequence feature datasets. The SAR data were cropped to obtain a SAR image dataset, and salinity data and snow depth data were extracted from the AMSR2 data and MIRAS data to construct a physical parameter dataset.

[0016] Specifically, the multimodal data fusion network includes three modalities. The SAR image dataset is input into the first modality for processing. The first modality adopts the U-Net backbone network structure and integrates the CBAM attention mechanism to extract image semantic features; the AMSR2 sequence feature dataset and the MIRAS sequence feature dataset are input into the second modality for processing, and the SAR sequence feature dataset and the MODIS sequence feature dataset are input into the third modality for processing. The second and third modalities both use one-dimensional convolutional neural networks to extract sequence semantic features. The sequence semantic features of the second and third modalities are spliced ​​and fused to obtain fused sequence semantic features. Then, the cross-Transformer attention mechanism is used to fuse the fused sequence semantic features and image semantic features to obtain global semantic features.

[0017] Specifically, the first modality obtains three levels of image semantic features through three-layer downsampling operations, which are low-level image semantic features, intermediate image semantic features and high-level image semantic features in the order of downsampling; the second and third modalities obtain three levels of sequence semantic features through three-layer downsampling operations, which are low-level sequence semantic features, intermediate sequence semantic features and high-level sequence semantic features in the order of downsampling; the sequence semantic features of the same level in the second and third modalities are spliced ​​and fused to obtain three levels of fused sequence semantic features, the fused sequence semantic features and the image semantic features of the corresponding levels are fused to obtain three levels of global semantic features, and a total loss function weighted by the three-level global semantic feature loss function is constructed to train the multimodal data fusion network.

[0018] Specifically, the total loss function calculation formula is:

[0019] ,

[0020] Where: is the total loss function, 、 、 is the feature weight, is the low-level global semantic feature loss function, is the mid-level global semantic feature loss function, is the high-level global semantic feature loss function, represents any global semantic feature loss function, is the Huber loss function, is the coefficient of determination.

[0021] Specifically, in step S5, the machine learning regression models include random forest RF, support vector machine SVM, multi-layer perceptron MLP, extreme boosting tree XGBoost, gradient boosting tree LightGBM and classification and regression tree ordered boosting CatBoost.

[0022] Optionally, in step S2, sequence feature extraction includes:

[0023] The multi-source remote sensing data were resampled to the same resolution as the SAR data using cubic linear interpolation. The longitude and latitude information of the IceBridge airborne detection data were mapped to the resampled multi-source remote sensing data. The data were extracted using a point-to-point mapping method to obtain a sequence feature dataset.

[0024] Optionally, in step S2, sequence feature extraction includes:

[0025] A nearest neighbor tree is established, and a search tree is established for the multi-source remote sensing data based on the latitude and longitude information of the IceBridge airborne detection data. The multi-source remote sensing data is mapped one-to-one with the latitude and longitude information of the IceBridge airborne detection data to obtain a sequence feature data set.

[0026] Preferably, the above step S6 is replaced by:

[0027] S61. Delete the MODIS data from the multi-source remote sensing data and repeat steps S2 to S5 to obtain a sea ice thickness inversion model without optical image data; input the multi-source remote sensing data to be tested in the target area, extract the cloud mask data in the target area from the MODIS data, use the cloud mask data to divide the target area into a cloudy area and a cloudless area, input a second test set, apply the sea ice thickness inversion model without optical image data in the cloudy area, apply the sea ice thickness inversion model obtained in step S5 in the cloudless area, and output corresponding sea ice thickness inversion data.

[0028] Beneficial effects: Compared with the prior art, the present invention has the following significant effects:

[0029] 1. In terms of raw data selection, the present invention combines synthetic aperture radar data with other remote sensing data to overcome their limitations. At the same time, a multimodal data fusion network model based on deep learning is proposed. It fully considers the large scale differences between different remote sensing image data, and orderly fuses the input image features and sequence features of different scales. It not only considers the spatial relationship between point features, but also integrates features such as image texture. It effectively alleviates the shortcomings of SAR in thin ice recognition and enhances the robustness and prediction accuracy of the model under complex ice conditions (such as periglacial areas and melt periods).

[0030] 2. The present invention uses multiple original machine learning regression models to invert and predict the data after feature fusion, and uses performance evaluation indicators to select the machine learning regression model with the best indicators for sea ice thickness inversion. The multi-model fusion prediction method can effectively reduce the adaptability problem existing in single model prediction.

[0031] 3. The present invention adaptively fuses multi-source remote sensing data according to cloud mask conditions. By using different models in cloudy and cloudless areas, it is more in line with the natural conditions of actual scenes, enhancing the applicability and accuracy of the sea ice thickness inversion method, and effectively improving both local and overall accuracy.

[0032] 4. The nearest neighbor tree method is used to establish a mapping between multi-source remote sensing data and longitude and latitude. Considering the spatial scale differences between different remote sensing data, the nearest neighbor tree can better achieve data matching compared to the traditional upsampling method. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the method of the present invention.

[0034] Figure 2 It is a flow chart of the multimodal data fusion network of the present invention.

[0035] Figure 3 It is a flow chart of feature fusion and sea ice inversion output of the present invention.

[0036] Figure 4 This is a statistical graph of the sea ice inversion results in cloudy areas using the random forest model in this invention.

[0037] Figure 5 This is a statistical chart of the sea ice inversion results in cloudy areas using the XGBoost model in this paper.

[0038] Figure 6 This is a statistical chart of the sea ice inversion results in cloudy areas using the LightGBM model in this paper.

[0039] Figure 7 This is a statistical chart of the sea ice inversion results in cloudy areas using the CatBoost model in this invention.

[0040] Figure 8 This is a statistical graph of the sea ice inversion results in cloudy areas using the SVM model in the present invention.

[0041] Figure 9 This is a statistical chart of the sea ice inversion results in cloudy areas using the MLP model in this invention.

[0042] Figure 10 This is a statistical graph of the sea ice inversion results in cloudless areas using the random forest model in this invention.

[0043] Figure 11 This is a statistical chart of the sea ice inversion results in cloudless areas using the XGBoost model in this invention.

[0044] Figure 12 This is a statistical chart of the sea ice inversion results in cloudless areas using the LightGBM model in this paper.

[0045] Figure 13 This is a statistical chart of the sea ice inversion results in cloudless areas using the CatBoost model in this invention.

[0046] Figure 14 This is a statistical chart of the sea ice inversion results in cloudless areas using the MLP model in this invention.

[0047] Figure 15 This is a statistical graph of the sea ice inversion results in cloudless areas using the SVM model in the present invention.

[0048] Figure 16 This is a comparison chart of the sea ice thickness inversion results of the present invention. DETAILED DESCRIPTION

[0049] A preferred embodiment of the present invention is further described below with reference to the accompanying drawings.

[0050] Example 1

[0051] See also Figure 1 As shown, this embodiment provides a polar sea ice thickness inversion method based on multi-source remote sensing data fusion, including the following steps:

[0052] S1. Collect multi-source remote sensing data and corresponding in-situ sea ice thickness data in the polar regions.

[0053] In this example, the IceBridge airborne data acquired in the Arctic by NASA's Operation IceBridge was used as validation data. This data includes in-situ sea ice thickness measurements and latitude and longitude information. Multi-source remote sensing data was collected concurrently with the IceBridge data, including C-SAR data, Moderate Resolution Imaging Spectroradiometer (MODIS) data, Advanced Microwave Scanning Radiometer 2 (AMSR2) data, and MIRAS data from the SMOS satellite.

[0054] S2. After preprocessing the multi-source remote sensing data, georeference it based on the latitude and longitude information of the IceBridge airborne detection data and perform projection transformation. The multi-source remote sensing data is uniformly transformed to the polar position projection. Sequence feature extraction is performed on the C-SAR data, MODIS data, AMSR2 data, and MIRAS data to obtain sequence feature datasets, including C-SAR sequence feature datasets, MODIS sequence feature datasets, AMSR2 sequence feature datasets, MIRAS sequence feature datasets, and physical parameter sequence feature datasets. The C-SAR data is cropped to obtain a C-SAR image dataset. Salinity data and snow depth data are extracted from the AMSR2 data and MIRAS data to construct a physical parameter dataset. The image dataset and sequence feature dataset are merged into a multimodal fusion dataset, and the multimodal fusion dataset is randomly divided into a first training set and a first test set according to a set ratio.

[0055] In this example, IceBridge airborne data is first mapped onto C-SAR data based on its latitude and longitude. Following the aircraft's flight sequence, parameters such as sea ice thickness and snow cover are sequentially determined to detect outliers and excessive ice thickness (C-SAR has limited penetration, detecting only sea ice thicknesses ≤ 2 meters). Samples where these parameters exceed the sensor's physical inversion limits are removed. A 25×25 pixel rectangular region centered around the pixel where the sample point is located is then extracted from the data sample after outliers have been removed, generating a C-SAR image dataset. This 25×25×number of bands data is stored, along with the value of the central sample point. Because MODIS, AMSR2, and MIRAS data exhibit significant spatial differences from C-SAR and IceBridge airborne data, this example uses a nearest neighbor tree (DT-Tree) approach to construct a search tree based on latitude and longitude, mapping the data one-to-one and collecting the corresponding location information as sequence features. In this embodiment, 12-dimensional SAR features, 26-dimensional AMSR2 features, 26-dimensional MIRAS features, and 32-dimensional MODIS features are extracted based on the empirical index.

[0056] S3. Construct a multimodal data fusion network, input the first training set into the multimodal data fusion network for training, and obtain a trained multimodal data fusion network.

[0057] Please refer to Figure 2 As shown, in this embodiment, the multimodal data fusion network includes three modalities, and the C-SAR image dataset is input into the first modality for processing; for sequence features, due to the difference in spatial resolution of different data, the AMSR2 sequence feature dataset and the MIRAS sequence feature dataset are input into the second modality for processing, and the C-SAR sequence feature dataset and the MODIS sequence feature dataset are input into the third modality for processing.

[0058] The first modality adopts the U-Net backbone network structure and a three-layer downsampling structure. ResBlock is used to extract spatial features and then spliced. The image features enhanced by the CBAM attention mechanism are integrated to extract image semantic features layer by layer. The downsampling order is low-level image semantic features, intermediate image semantic features, and high-level image semantic features.

[0059] Both the second and third modalities use a one-dimensional convolutional neural network (1D-CNN) to extract sequence semantic features. Since the C-SAR and MODIS sequence features have spatial continuity, based on the 1D-CNN extraction and expansion features, the Transformer module is used to capture spatial global features. The second and third modalities gradually increase the number of convolution layers through three layers to enhance the abstract depth of the extracted semantic features, obtaining three levels of sequence semantic features, which are low-level sequence semantic features, intermediate sequence semantic features, and high-level sequence semantic features in order of the number of semantic feature layers. The sequence semantic features of the same level in the second and third modalities are spliced ​​and fused to obtain three levels of fused sequence semantic features. The obtained image semantic features are reduced to 1-dimensional sequence features using Global Average Pooling. The fused sequence semantic features and the image semantic features of the corresponding level are then fused using the Cross-Transformer attention mechanism fusion module to obtain 3-level global semantic features. The fully connected layer output is added to the global semantic features to assist in calculating the weighted total loss function of the 3-level global semantic feature loss function, so as to optimize the training of the multimodal data fusion network. Please refer to Figure 3As shown in the figure, the fused global semantic features are concatenated with the image modality and fed into the Transformer module. A multi-head attention mechanism is used to capture global features. Finally, a Flatten layer is used to flatten the output into one dimension. This layer connects the feature output layer and the final output layer. The feature output layer extracts high-level features from the input data, which are then fed to the final output layer for sea ice thickness prediction. The final output layer, through error feedback, transmits loss information to the feature output layer, guiding it to optimize feature extraction. Ultimately, the entire model improves feature extraction quality through continuous weight adjustment.

[0060] The feature output layer outputs the new fused sequence features. The output sequence feature dimension is set to the number of input sequence feature dimensions multiplied by a coefficient greater than 1, such as 1.2. The feature output layer and the final output layer are configured via two fully connected layers. The feature output layer sets the number of searchable units and uses a linear activation function to adjust the number of features and feature output. The outputs of the two branches are adaptively fused through a weight learning mechanism. The weight learning layer contains two fully connected layers that learn weights for the CNN and Transformer branches respectively. The sigmoid function is used to constrain the weight values ​​between 0 and 1 to dynamically adjust the fusion ratio of local and global features. The fused features are further adjusted and outputted through the feature output layer and the final output layer.

[0061] The total loss function calculation formula is:

[0062] ,

[0063] Where: is the total loss function, 、 、 is the feature weight, in this embodiment, the feature weight is set to 1, is the low-level global semantic feature loss function, is the mid-level global semantic feature loss function, is the high-level global semantic feature loss function, represents any global semantic feature loss function, is the Huber loss function, is the coefficient of determination.

[0064] S4. Input the first test set into the multimodal data fusion network trained in step S3 for feature fusion to obtain a fused sequence feature dataset, merge the fused sequence feature dataset with the corresponding physical parameter dataset into a sea ice thickness inversion dataset, and randomly divide the sea ice thickness inversion dataset into a second training set and a second test set according to a set ratio.

[0065] S5. Divide the second training set into several training subsets, input each training subset into several different machine learning regression models for training, obtain sea ice thickness inversion data, compare it with the corresponding in-situ sea ice thickness measured data, obtain performance evaluation indicators, and select the machine learning regression model with the best performance evaluation indicator as the sea ice thickness inversion model.

[0066] In this embodiment, the machine learning regression models used include random forest RF, support vector machine SVM, multilayer perceptron MLP, extreme boosting tree XGBoost, gradient boosting tree LightGBM and classification and regression tree ordered boosting CatBoost. For each machine learning regression model, grid search and random search are used to optimize the hyperparameters.

[0067] The typical performance evaluation index is to standardize the mean absolute error (MAE) and the root mean square error (RMSE) and then compare them with the coefficient of determination (R). 2 Perform weighted averaging. The larger the performance evaluation index is, the better the performance of the corresponding machine learning regression model is.

[0068] S6. Delete the MODIS data from the multi-source remote sensing data and repeat steps S2 to S5 to obtain a sea ice thickness inversion model without optical image data; input the multi-source remote sensing data to be tested in the target area, extract cloud mask data in the target area from the MODIS data, use the cloud mask data to divide the target area into a cloudy area and a cloudless area, input a second test set, apply the sea ice thickness inversion model without optical image data in the cloudy area, and apply the sea ice thickness inversion model obtained in step S5 in the cloudless area, and output corresponding sea ice thickness inversion data.

[0069] In practical scenarios, the use of MODIS data requires consideration of the impact of clouds and ice fog on the reflectivity of each band. In this example, the QA band in the MODIS data is converted to a binary representation of cloud status. NDCI is calculated using a threshold of 0.5 to extract cloud mask data. For cloudy areas, MODIS data is not included in the fusion of multi-source remote sensing data. Therefore, additional training is required for the recognition scheme based solely on C-SAR, AMSR2, and MIRAS data. Through steps S2 to S5, a sea ice thickness inversion model without optical imagery is obtained.

[0070] To demonstrate the improved effect of the method of the present invention, under the same test conditions, the above-mentioned machine learning regression model was used to invert the sea ice thickness of the multi-source remote sensing data after feature fusion under cloudy conditions and cloudless conditions, respectively. The same machine learning regression model was further used to invert the sea ice thickness of the multi-source remote sensing data after feature fusion and the multi-source remote sensing data without feature fusion.

[0071] Please refer to Figures 4 to 9 As shown, Figures 4 to 9 The inversion results of sea ice thickness directly modeled under cloudy conditions and the sea ice thickness modeled after deep learning fusion, that is, the results of sea ice thickness inversion without optical data, Figures 4 to 9 The blue dashed line in the figure represents the ideal comparison line (ICL). Figures 4 to 9 In each figure, sub-graph a is the direct modeling result graph, and sub-graph b is the deep learning modeling result graph. Figures 4 to 9 It can be seen that when the sea ice thickness is within the range of 2m, the sea ice inversion thickness is closer to the true value. At this time, the inversion accuracy is higher and the deviation from the data is smaller. As the sea ice thickness increases, the data deviation begins to gradually increase. The XGBoost model is used to perform the sea ice thickness inversion task in this case to obtain the best results. The mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R 2 ) and correlation coefficient (r) can reach 0.506, 0.571, 0.891 and 0.945 respectively.

[0072] Please refer to Figures 10 to 15 As shown, Figures 10 to 15 The inversion results of sea ice thickness modeled directly under cloudless conditions and the inversion results of sea ice thickness modeled after deep learning fusion are both the results of sea ice thickness inversion with optical data. Figures 10 to 15 The middle blue dotted line represents the ideal comparison line (ICL). Figures 10 to 15 In each figure, sub-graph a is the direct modeling result graph, and sub-graph b is the deep learning modeling result graph. Figures 10 to 15 It can be seen that after the introduction of optical data, the accuracy of sea ice thickness inversion is further improved, and the improvement is particularly obvious when the sea ice thickness is within the range of 2m. After adopting the deep learning method, all kinds of parameters have been significantly improved compared with the direct modeling of sea ice thickness model. Table 1 below is the statistical indicator parameters of each method under the sea ice thickness of 0-2m. Among them, the XGBoost model is used to perform the sea ice thickness inversion task in this case to obtain the best results, with the mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R 2 ) and correlation coefficient (r) reached 0.073, 0.014, 0.911, and 0.959, respectively. Compared with the case without optical data, the introduction of optical data further improves the accuracy of sea ice thickness retrieval using this method.

[0073] Table 1

[0074]

[0075] Please refer to Figure 16As shown in the figure, the random forest RF model is also used to invert the sea ice thickness of the multi-source remote sensing data after feature fusion and the multi-source remote sensing data without feature fusion, and the inversion results are compared. Figure 16 Sub-image a is the inversion result obtained from multi-source remote sensing data without feature fusion. Figure 16 Sub-figure b is the inversion result obtained by multi-source remote sensing data after feature fusion. The inversion result obtained by multi-source remote sensing data after feature fusion is higher in mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R 2 ) and correlation coefficient (r) are significantly better than the inversion results obtained from multi-source remote sensing data without feature fusion, among which MAE is reduced by 0.022, RMSE is reduced by 0.137, and R 2 It increased by 0.035 and r increased by 0.018.

Claims

1. A polar sea ice thickness inversion method based on multi-source remote sensing data fusion, characterized in that: The following steps are involved: S1. Collect multi-source remote sensing data and corresponding in-situ sea ice thickness measurement data within the target area; S2. Geo-register the multi-source remote sensing data to obtain an image dataset, a sequence dataset, and a physical parameter dataset, respectively. Extract sequence features from the sequence dataset to obtain a sequence feature dataset. Merge the image dataset and the sequence feature dataset into a multimodal fusion dataset, and randomly divide the multimodal fusion dataset into a first training set and a first test set according to a set ratio. S3. Construct a multimodal data fusion network, input the first training set into the multimodal data fusion network for training, and obtain a trained multimodal data fusion network; S4. Input the first test set into the trained multimodal data fusion network for feature fusion to obtain a fused sequence feature dataset, merge the fused sequence feature dataset with the corresponding physical parameter dataset into a sea ice thickness inversion dataset, and randomly divide the sea ice thickness inversion dataset into a second training set and a second test set according to a set ratio; S5. Divide the second training set into several training subsets, input each training subset into several different machine learning regression models for training, obtain sea ice thickness inversion data, compare the data with the corresponding in-situ sea ice thickness measured data, obtain performance evaluation indicators, and select the machine learning regression model with the best performance evaluation indicator as the sea ice thickness inversion model; S6. Input the second test set into the sea ice thickness inversion model and output corresponding sea ice thickness inversion data.

2. The polar sea ice thickness inversion method according to claim 1, characterized in that: The multi-source remote sensing data include synthetic aperture radar SAR data, moderate resolution imaging spectrometer MODIS data, advanced microwave scanning radiometer 2 AMSR2 data and L-band synthetic aperture imaging radiometer MIRAS data; the in-situ sea ice thickness measured data are IceBridge airborne detection data.

3. The polar sea ice thickness inversion method according to claim 2, characterized in that: The step S2 comprises: After preprocessing, the multi-source remote sensing data were georeferenced according to the latitude and longitude information of the IceBridge airborne detection data, and projection transformation was performed to uniformly transform the multi-source remote sensing data into a polar position projection. Sequence features were extracted from SAR data, MODIS data, AMSR2 data, and MIRAS data to obtain sequence feature datasets, including SAR sequence feature datasets, MODIS sequence feature datasets, AMSR2 sequence feature datasets, MIRAS sequence feature datasets, and physical parameter sequence feature datasets. The SAR data were cropped to obtain a SAR image dataset, and salinity data and snow depth data were extracted from the AMSR2 data and MIRAS data to construct a physical parameter dataset.

4. The polar sea ice thickness inversion method according to claim 3, characterized in that: The multimodal data fusion network includes three modalities. The SAR image dataset is input into the first modality for processing. The first modality adopts the U-Net backbone network structure and integrates the CBAM attention mechanism to extract image semantic features. The AMSR2 sequence feature dataset and the MIRAS sequence feature dataset are input into the second modality for processing. The SAR sequence feature dataset and the MODIS sequence feature dataset are input into the third modality for processing. The second and third modalities both adopt one-dimensional convolutional neural networks to extract sequence semantic features. The sequence semantic features of the second and third modalities are spliced ​​and fused to obtain fused sequence semantic features. Then, the cross Transformer attention mechanism is used to fuse the fused sequence semantic features and the image semantic features to obtain global semantic features.

5. The polar sea ice thickness inversion method according to claim 4, characterized in that: The first modality obtains three levels of image semantic features through a three-layer downsampling operation, which are low-level image semantic features, intermediate-level image semantic features, and high-level image semantic features in the order of downsampling; the second modality and the third modality obtain three levels of sequence semantic features through a three-layer downsampling operation, which are low-level sequence semantic features, intermediate-level sequence semantic features, and high-level sequence semantic features in the order of downsampling; The sequence semantic features of the same level in the second and third modalities are concatenated and fused to obtain three-level fused sequence semantic features. The fused sequence semantic features are fused with the image semantic features of the corresponding levels to obtain three-level global semantic features. A total loss function weighted by the three-level global semantic feature loss functions is constructed to train the multimodal data fusion network.

6. The polar sea ice thickness inversion method according to claim 5, characterized in that: The total loss function calculation formula is: ， Where: is the total loss function, 、 、 is the feature weight, is the low-level global semantic feature loss function, is the mid-level global semantic feature loss function, is the high-level global semantic feature loss function, represents any global semantic feature loss function, is the Huber loss function, is the coefficient of determination.

7. The polar sea ice thickness inversion method according to claim 1, characterized in that: In step S5, the machine learning regression model includes random forest RF, support vector machine SVM, multi-layer perceptron MLP, extreme boosting tree XGBoost, gradient boosting tree LightGBM and classification and regression tree ordered boosting CatBoost.

8. The polar sea ice thickness inversion method according to claim 3, characterized in that: In step S2, sequence feature extraction includes: The multi-source remote sensing data were resampled to the same resolution as the SAR data using cubic linear interpolation. The longitude and latitude information of the IceBridge airborne detection data were mapped to the resampled multi-source remote sensing data. The data were extracted using a point-to-point mapping method to obtain a sequence feature dataset.

9. The polar sea ice thickness inversion method according to claim 3, characterized in that: In step S2, sequence feature extraction includes: A nearest neighbor tree is established, and a search tree is established for the multi-source remote sensing data based on the latitude and longitude information of the IceBridge airborne detection data. The multi-source remote sensing data is mapped one-to-one with the latitude and longitude information of the IceBridge airborne detection data to obtain a sequence feature data set.

10. The polar sea ice thickness inversion method according to claim 2, characterized in that: The step S6 is replaced by: S61. Delete the MODIS data from the multi-source remote sensing data and repeat steps S2 to S5 to obtain a sea ice thickness inversion model without optical image data; input the multi-source remote sensing data to be tested in the target area, extract the cloud mask data in the target area from the MODIS data, use the cloud mask data to divide the target area into a cloudy area and a cloudless area, input a second test set, apply the sea ice thickness inversion model without optical image data in the cloudy area, apply the sea ice thickness inversion model obtained in step S5 in the cloudless area, and output corresponding sea ice thickness inversion data.

Images