Feature fusion type permanent scatterer identification technology based on deep learning

Abstract

The invention discloses a feature fusion type permanent scatterer recognition technology based on deep learning. The method comprises the following steps: firstly, constructing a residual attention U-shaped network to extract SAR image spatial features, and constructing a C-LSTM network to extract differential interferometric phase time sequence features; a feature fusion module is constructed to couple spatial features and time sequence features, and a prediction module is constructed to output a dual-channel permanent scatterer recognition result; and finally, obtaining a single-channel permanent scatterer identification result through channel maximum value screening, and further obtaining the spatial distribution of the permanent scatterer in the research area by using image block splicing and geocoding technologies. According to the method, a data driving mode is adopted, and rapid recognition of the permanent scatterer can be realized through network training; according to the permanent scatterer identification technology provided by the invention, mixed loss of weighted summation of focus loss and residual point loss is adopted, the problem of imbalance of positive and negative difficult samples is solved, and effectiveness evaluation of an identification result acting on a downstream task is realized; according to the method, permanent scatterer identification is carried out by using the constructed feature fusion type network and the mixed loss, and a relatively small number of residual points can be kept while a relatively large number of permanent scatterers are identified, so that the identification result has relatively large possibility when acting on a downstream task, and an accurate earth surface deformation result can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of microwave remote sensing technology, and in particular to a feature fusion type permanent scatterer identification method based on deep learning. Background Art

[0002] The application of synthetic aperture radar interferometry (InSAR) technology in surface deformation measurement can generally be divided into two categories, namely differential InSAR (DInSAR) and multi-temporal InSAR (MT-InSAR) technology; DInSAR uses two synthetic aperture radar images covering the same area at different times to calculate the phase difference to obtain surface deformation, and its surface deformation measurement accuracy is easily affected by decoherence and atmospheric delay; MT-InSAR uses three or more synthetic aperture radar images covering the same area at different times to obtain high-precision deformation results; the permanent scatterer InSAR (PS-InSAR) technology proposed by Ferretti et al. is a representative technology in the field of MT-InSAR. In one of the works, this method sets a threshold on the amplitude deviation index to screen pixels with stable scattering characteristics in the time series as permanent scatterers, and extracts the corresponding interference phase of the permanent scatterers for subsequent parameter solution to improve the deformation solution result; Another MT-InSAR algorithm that overcomes the temporal and spatial decorrelation is the short baseline set InSAR (SBS-InSAR) technology. This method uses SAR interference pairs with short temporal and spatial baselines for deformation extraction; Since the PS-InSAR technology was proposed, many researchers have been committed to enhancing and expanding the PS-InSAR algorithm, and most of the existing improved algorithms only focus on the spatial characteristics of SAR images, using the amplitude characteristics of SAR images instead of phase stability as the permanent Scatterer selection index, that is, phase stability estimation without phase analysis; To address this problem, Hooper et al. proposed a Stanford Permanent Scatterer (StaMPS) method that uses the spatial correlation of deformation. First, an amplitude deviation index threshold is set to screen out candidate pixels of permanent scatterers. Then, permanent scatterers are further screened out based on the stability of phase noise, and a three-dimensional spatiotemporal convolution algorithm is applied to analyze the relationship between the permanent scatterers and the corresponding interference phases, thereby achieving high-precision deformation monitoring. Permanent scatterers are the key research objects of MT-InSAR, and permanent scatterer identification is a key link in InSAR data processing. The effectiveness of the selected permanent scatterers restricts the accuracy of deformation extraction. Therefore, the researchers proposed A series of model-driven permanent scatterer identification algorithms have been proposed in an attempt to improve the rationality of permanent scatterer identification. Such algorithms usually require threshold setting to preliminarily screen out candidate permanent scatterer pixels, and use a variety of algorithms to perform phase filtering to screen out permanent scatterer pixels with spatiotemporal stability. For example, Navneet et al. proposed a permanent scatterer screening method based on the highest eigenvalue threshold of the coherence matrix, which reduced the false alarm rate of permanent scatterer identification in low coherence areas. Wang et al. reduced the false recognition of permanent scatterers based on the phase similarity between adjacent permanent scatterer pixels. Although the above algorithms improve the effectiveness of permanent scatterer selection, the computational cost and operation time are increased.As deep learning algorithms have demonstrated remarkable efficiency in recent years in fields such as image classification and target detection, a growing number of researchers have begun applying data-driven algorithms to remote sensing, including methods that utilize neural networks to identify persistent scatterers. Tiwari et al. proposed two deep learning architectures that correlate the interferometric phase history with the classification of phase-stable or unstable permanent scatterer pixels to increase the density of reliable permanent scatterer pixels. Zhang et al. constructed a two-channel one-dimensional convolutional neural network to address class imbalance and the model's dependence on the number of training samples. These algorithms significantly improve the efficiency of persistent scatterer identification, but quantitative evaluation of the identification results is limited to comparing the density of permanent scatterer pixels within the region of interest, lacking an assessment of the effectiveness of the identification results in downstream tasks such as phase unwrapping. Summary of the Invention

[0003] In view of this, the present invention addresses the shortcomings of existing permanent scatterer identification methods and proposes a feature fusion network architecture, which combines the spatial features and temporal phase features of SAR images to screen permanent scatterers with spatiotemporal stability. In addition, the present invention also proposes a hybrid loss function to solve the imbalance problem of positive and negative difficult and easy samples, and improve the effectiveness of permanent scatterer identification in the downstream phase unwrapping task. This loss function has a reference value for data-driven permanent scatterer identification methods. Experiments show that compared with the traditional StaMPS method, the network trained using the proposed network architecture and loss function achieves fast and effective permanent scatterer identification with a larger number of permanent scatterers, a smaller number of residual points and a shorter screening time.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The feature fusion type permanent scatterer recognition technology based on deep learning provided by the present invention includes the following steps:

[0006] S1: Obtain time series SAR data to construct training and test sets;

[0007] S2: Constructing a residual attention U-network to extract spatial features: Applying the residual U-network and the convolutional block attention module (CBAM) to construct a residual U-network to extract SAR image spatial features and prepare spatial data for feature fusion;

[0008] S3: Construct a C-LSTM network to extract time series features: Apply the convolutional long short-term memory module to build a C-LSTM network to extract the interference phase time series features, and perform time series feature mapping based on three-dimensional convolution to prepare time series data for feature fusion;

[0009] S4: Construct a feature fusion module to couple spatial features and temporal features: Apply the splicing operator to splice spatial features and temporal features in the channel dimension, and construct a feature fusion module based on two-dimensional convolution to obtain joint spatial and temporal features to prepare data for the prediction module;

[0010] S5: Construct a prediction module to predict the spatial distribution of permanent scatterers: Apply the prediction module to output the dual-channel permanent scatterer identification results, and obtain the single-channel permanent scatterer prediction results through channel maximum screening. Finally, use image block stitching and geocoding technology to synthesize the complete spatial distribution of permanent scatterers in the study area.

[0011] Furthermore, the existing model-driven permanent scatterer identification technology relies on phase model construction and iterative solution. The permanent scatterer identification technology does not require the construction of an algorithm model. By training the neural network, it can quickly output the permanent scatterer identification results, greatly improving the identification efficiency.

[0012] Furthermore, the existing data-driven permanent scatterer identification technology only focuses on the extraction of SAR image spatial features, while the technology focuses on both SAR image spatial features and interference phase temporal features, and the obtained permanent scatterer identification results have spatiotemporal stability.

[0013] Furthermore, existing data-driven permanent scatterer identification technologies are unable to evaluate the impact of identification results on downstream tasks. The described technology can provide an evaluation of the effectiveness of identification results on downstream phase unwrapping tasks.

[0014] Furthermore, the feature fusion type permanent scatterer recognition technology uses the focal loss function to solve the problem of imbalance in the number of positive and negative samples during network training. The network output used to construct the focal loss function is calculated according to the following formula:

[0015] ;

[0016] in, is the category index, and is the output data of the network before passing through the Softmax activation function. Using the output data of the Softmax activation function, the focus loss function is expressed according to the following formula:

[0017] ;

[0018] in, is the category index, if the pixel is identified as a permanent scatterer, the category index value is 1, if the pixel is identified as a non-permanent scatterer, the category index value is 0; Represents the weighting factor for balancing positive and negative samples, The focusing factor represents the balance between difficult and easy samples. The weighting factor and focusing factor can be manually adjusted according to the ratio of permanent scatterer pixels to non-permanent scatterer pixels in the label data to improve the network's recognition accuracy for permanent scatterers.

[0019] Furthermore, the feature fusion type permanent scatterer identification technology uses the Delaunay triangulation method to construct an interference phase triangle network. Using the triangle network, the effectiveness evaluation index, namely the residual point number (RPN), is calculated according to the following formula:

[0020] ;

[0021] in, Represents the triangle index value, represents the total number of triangles, Represents the triangle vertex index value; For triangle The winding phase corresponding to each vertex is is the paired winding phase difference, which can be 0 or ; Express Calculate the modulus value; the smaller the RPN value, the greater the possibility of obtaining accurate deformation results by using the identified permanent scatterers for phase unwrapping.

[0022] Furthermore, the feature fusion type permanent scatterer recognition technology constructs a residual rate loss function based on the effectiveness evaluation index to improve the effectiveness of the network in identifying permanent scatterers; using the RPN value, the residual rate loss function It is expressed as follows:

[0023] ;

[0024] in, represents the interference phase image index value, represents the total number of interferometric phase patterns, Represents the total number of pixels identified as permanent scatterers in the predicted permanent scatterer identification results.

[0025] Furthermore, using the focal loss function And the residual rate loss function , the mixed loss function for network training According to the following formula:

[0026] ;

[0027] in, represents the weight coefficient of the focal loss function, Represents the weight coefficient of the residual rate loss function; this hybrid loss function can solve the imbalance problem of positive and negative difficult and easy samples and realize the effectiveness evaluation of the permanent scatterer identification results on the downstream phase unwrapping task.

[0028] The deep learning-based feature fusion type permanent scatterer recognition technology provided by the present invention is trained and tested on a platform equipped with a 13th Gen Intel(R) Core(TM) i9-13900K central processing unit and an NVIDIA GeForce RTX 4090 graphics card.

[0029] The beneficial effects of the present invention are:

[0030] The present invention provides a deep learning-based feature fusion type permanent scatterer identification technology. This technology proposes a feature fusion framework that fully exploits the spatial features and temporal phase features of SAR images, and realizes spatiotemporal feature coupling with the help of feature fusion operations, so that the identified permanent scatterers have spatiotemporal stability, thereby improving the effectiveness of permanent scatterer identification.

[0031] The present invention proposes a hybrid loss function that combines focal loss and residual point rate to constrain network training. The network trained under the constraint of this loss function can not only effectively solve the problems of imbalance between positive and negative samples and imbalance between difficult and easy samples, but also further realize the effectiveness evaluation of the permanent scatterer recognition results on downstream tasks.

[0032] The deep learning-based feature fusion type permanent scatterer identification technology provided by the present invention has the characteristics of being lightweight. By training the network model, the permanent scatterer identification results in the study area can be quickly obtained, which greatly improves the efficiency of permanent scatterer identification.

[0033] Other advantages, objectives and features of the present invention will be described in part in the following description, and to some extent, will be obvious to those skilled in the art based on an examination of the following, or can be taught from the practice of the present invention, and the objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is the permanent scatterer identification flow chart presented by this technology;

[0035] Figure 2 This is a schematic diagram of the residual attention U-shaped network structure of this permanent scatterer recognition technology;

[0036] Figure 3 This is a schematic diagram of the C-LSTM network structure of this permanent scatterer recognition technology;

[0037] Figure 4This is a schematic diagram of the temporal mapping module of the permanent scatterer identification technology;

[0038] Figure 5 It is the spatial distribution of permanent scatterers in the study area during the verification experiment;

[0039] Figure 6 It is the spatial distribution of permanent scatterers in the local area 1 in the verification experiment;

[0040] Figure 7 It is to verify the spatial distribution of permanent scatterers in the local area 2 in the experiment. DETAILED DESCRIPTION

[0041] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are further described below in combination with the accompanying drawings and preferred embodiments, so that those skilled in the art can better understand the present invention and implement it. However, the embodiments given are not intended to limit the present invention.

[0042] like Figure 1 As shown, the feature fusion type permanent scatterer recognition technology based on deep learning provided in this embodiment includes the following steps:

[0043] S1: Obtain time series SAR data to construct training and test sets;

[0044] S2: Constructing a residual attention U-network to extract spatial features: Applying the residual U-network and the convolutional block attention module (CBAM) to construct a residual U-network to extract SAR image spatial features and prepare spatial data for feature fusion;

[0045] S3: Construct a C-LSTM network to extract time series features: Apply the convolutional long short-term memory module to build a C-LSTM network to extract the interference phase time series features, and perform time series feature mapping based on three-dimensional convolution to prepare time series data for feature fusion;

[0046] S4: Construct a feature fusion module to couple spatial features and temporal features: Apply the splicing operator to splice spatial features and temporal features in the channel dimension, and construct a feature fusion module based on two-dimensional convolution to obtain joint spatial and temporal features to prepare data for the prediction module;

[0047] S5: Construct a prediction module to predict the spatial distribution of permanent scatterers: Apply the prediction module to output the dual-channel permanent scatterer identification results, and obtain the single-channel permanent scatterer prediction results through channel maximum screening. Finally, use image block stitching and geocoding technology to synthesize the complete spatial distribution of permanent scatterers in the study area.

[0048] In this embodiment, permanent scatterer identification is performed in the following manner:

[0049] like Figure 2 As shown, the residual attention U-type network is used to extract the spatial features of three-band SAR image blocks; the residual attention U-type network includes a convolution module with a convolution kernel of 3, a pooling module, a transposed convolution module, a convolution module with a convolution kernel of 1, a CBAM attention module and a residual module; the convolution module with a convolution kernel of 3 is used to extract the features of the data; the pooling module is used to remove redundant features to achieve feature dimensionality reduction; the transposed convolution module is used for feature mapping; the CBAM attention module is used to focus on the channel features and spatial features of the network; the residual module is used to splice shallow features and deep features to enhance feature extraction performance; the output of the residual attention U-type network contains 4 dimensions, which respectively represent the number of batch image blocks, the number of channels, the image width and the image height.

[0050] The C-LSTM network is used to extract the time series phase features of the differential interference image; the C-LSTM network includes a C-LSTM module and a time series mapping module; the C-LSTM module is used to extract the features of the time series data; the time series mapping module is used to map the time steps of the time series data; the C-LSTM network provided in this embodiment is set with two C-LSTM modules.

[0051] like Figure 3 As shown, the C-LSTM module input contains the cell state , hidden state , differential interferometry phase data ; The C-LSTM module output contains the cell state , hidden state ; According to the input hidden state and differential interference phase data, solve the forget gate according to the following formula , input gate , output gate , candidate cell states and the output cell state and hidden state :

[0052] ;

[0053] in, represents the Sigmoid activation function, represents the hyperbolic tangent activation function, is the convolution operator, is the Hadamard product operation; 、 、 、 Represent the weight matrices of the forget gate, input gate, output gate, and candidate cell state gate, respectively, and 、 、 、 They represent the bias matrices of the forget gate, input gate, output gate and candidate cell state gate respectively; the above bias matrices are Figure 3 The merge is represented as .

[0054] like Figure 4 As shown, the input of the temporal mapping module is a temporal feature map; the output of the temporal mapping module is a feature map through temporal mapping; the temporal mapping module includes a three-dimensional convolution layer with a convolution kernel of 3, a three-dimensional batch normalization layer, a ReLU activation layer, a three-dimensional convolution layer with a convolution kernel of 1 and a time step averaging layer; the three-dimensional convolution layer with a convolution kernel of 3 is used to enhance the stability of temporal mapping; the three-dimensional batch normalization layer is used to avoid the temporal feature gradient disappearance and gradient explosion phenomena; the ReLU activation layer is used to enhance the nonlinear ability of the network; the three-dimensional convolution layer with a convolution kernel of 1 is used for temporal feature mapping; the time step averaging layer is used to integrate temporal data and realize data dimensionality reduction; the output of the temporal mapping module is consistent with the output dimension of the residual attention U-type network.

[0055] In this embodiment, the output of the residual attention U-type network and the output of the timing mapping module are spliced ​​in the channel dimension and sent to the feature fusion module; the feature fusion module includes a two-dimensional convolution layer with a convolution kernel of 3, a two-dimensional batch normalization layer, a ReLU activation layer, and a two-dimensional convolution layer with a convolution kernel of 1; the two-dimensional convolution layer with a convolution kernel of 3 is used to maintain the spatial stability of the fusion feature; the two-dimensional batch normalization layer is used to avoid the vanishing and gradient explosion phenomena of spatial feature gradients; the two-dimensional convolution layer with a convolution kernel of 1 is used for spatial feature mapping; in this embodiment, the feature fusion module is used to map the doubled number of channels after splicing to the number of channels before splicing, thereby realizing the deep fusion of image spatial features and phase and timing features.

[0056] This embodiment uses the output of the feature fusion module as the input of the prediction layer, outputs the dual-channel binary permanent scatterer prediction result through the prediction layer, and calculates the loss function with the one-hot encoded label data to constrain the network training direction; the prediction layer includes a two-dimensional convolution layer with a convolution kernel of 1, a two-dimensional batch normalization layer, and a Softmax activation layer; the Softmax activation layer is used to normalize the prediction result value range to between 0 and 1.

[0057] In this embodiment, the normalized two-channel prediction results are filtered through the channel maximum value to obtain a single-channel permanent scatterer selection image block, and the complete permanent scatterer selection results in the study area are output through image block splicing and geocoding; the total parameter volume of the network described in this embodiment is approximately 79MB, which has a lightweight feature; after training, the network described in this embodiment can achieve a permanent scatterer recognition time that is much lower than StaMPS.

[0058] In this embodiment, the training set is constructed in the following manner:

[0059] Twenty-one time-series SLC images acquired by the Sentinel-1A satellite covering three study areas with different land types were selected for dataset construction and processing. The land cover types are mainly divided into developed areas and undeveloped areas, consisting of buildings, vegetation, water areas, bare land, and cultivated land. The proportions of developed and undeveloped areas in different study areas vary. The time-series SLC images operate in interferometric wide-scan (IW) imaging mode and VV+VH fluctuation mode, with a time baseline span of 12 to 144 days. The Sentinel data processing platform SNAP platform was used to perform preprocessing operations such as orbit correction, registration, and removal of flat land phase and terrain phase on the 21 time-series SLC images to obtain 20 differential interferometry images. In the registration operation, one image was selected as the primary image, and the remaining 20 images were selected as auxiliary images.

[0060] The data used to construct the training set in this embodiment consists of three parts: three-band SAR images, differential interferometry phase images, and label data. The three-band SAR images are composed of the average amplitude, amplitude deviation index, and coherence coefficient of 10 auxiliary images in the time-series SLC images superimposed in the band dimension. The differential interferometry phase is single-channel five-dimensional data with a time step of 10. Studies have shown that 10 differential interferometry images are sufficient for the network to select permanent scatterers, and using more differential interferograms can further improve the network's generalization ability. The label data is a single-channel binary image consisting of 0 and 1, where 0 indicates that the pixel is a non-permanent scatterer and 1 indicates that the pixel is a permanent scatterer. The label data is used to calculate the loss with the permanent scatterer prediction results through one-hot labeling to guide the direction of network training. In this embodiment, the data output from the preprocessing operation using the SNAP platform is sent to the StaMPS toolkit for permanent scatterer selection to obtain label data.

[0061] After the data for training set construction is acquired, the edge portions where no data exists in the differential interferometer phase image are removed and reshaped into image blocks with a width and height both being multiples of 128 for image block segmentation. In the image block segmentation operation, the three-band SAR image, the differential interferometer phase image, and the label data are segmented from left to right and from top to bottom into image blocks with a width and height both being equal to 128 for network batch processing, and the network computing efficiency is improved through GPU parallel computing.

[0062] exist Figure 5 middle, Figure 5 (a) with Figure 5 (b) shows the results of permanent scatterer identification in the study area using the StaMPS algorithm and the technology respectively; the dimension of the study area is 1664×7680, and it contains 780 image blocks with a width and height of 128; in terms of the number of permanent scatterers identified, the StaMPS algorithm identified 471,490 permanent scatterers, while the technology identified 498,407 permanent scatterers, which is 26,917 more permanent scatterers than StaMPS; and based on the RPN values ​​calculated for the two permanent scatterer identification results, it was found that the number of residual points corresponding to the StaMPS algorithm identification result was 30,130, while the number of residual points corresponding to the technology identification result was 26,705, which is 3,425 less than the StaMPS algorithm, proving that the technology can identify a large number of permanent scatterers while maintaining a small amount of residual points, and is effective; and in terms of the distribution of permanent scatterers, the permanent scatterers identified by the StaMPS algorithm and the technology both have certain aggregation characteristics; as shown in Figure 5 In the selection process of the permanent scatterer identification results shown in the figure, the StaMPS algorithm took 4860 seconds; given that the network can quickly obtain the permanent scatterer identification results through training, its prediction process only took 29 seconds, proving that the technology is highly efficient; Figure 5 In the figure, local area 1 and local area 2 are selected with red boxes for further analysis.

[0063] exist Figure 6 In the figure, the enlarged Figure 5 The permanent scatterer identification results in the northeast local area 1 selected by the red box in the study area are shown; it can be seen that Figure 6 In the local area of ​​, both the StaMPS algorithm and the above technique identify permanent scatterers with good aggregation; Figure 6 In the permanent scatterer identification results in the central and southeastern regions, the permanent scatterers identified by the StaMPS algorithm appear not only in areas with buildings but also scattered in nearby farmland areas. The permanent scatterer identification results of the proposed technology are more clearly concentrated in areas with buildings, and its distribution and clustering are better than those of the StaMPS algorithm.

[0064] exist Figure 7 In the figure, the enlarged Figure 5 The identification results of permanent scatterers in the central local area 2 selected by the red box in the study area are shown; Figure 7In the local area, the area selected by the red box is the area with more buildings such as houses, and the area selected by the red oval is the green area or cultivated land area; it can be seen that the permanent scatterers identified by the StaMPS algorithm are concentrated on some houses, while the permanent scatterer recognition technology can extract permanent scatterers on scattered buildings, and the distribution is more uniform; the reason for this phenomenon is that the use of the feature fusion network can learn more information with spatiotemporal stability, which not only considers the spatial distribution of permanent scatterers but also their temporal phase characteristics, so that the recognition results are more evenly distributed while being clustered; there are no high coherence points in the green or cultivated land area where no buildings are concentrated, so the permanent scatterers identified in this area are misclassified pixels. The StaMPS algorithm misclassifies some pixels located in this area, while the permanent scatterer recognition technology does not misclassify pixels and performs well in extracting permanent scatterer features.

[0065] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any person skilled in the art can make some more or modify the equivalent embodiments with equivalent changes by using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. Feature fusion type permanent scatterer recognition technology based on deep learning, characterized by: The steps include: S1: Obtain time series SAR data for constructing training and test sets; S2: Construct a residual attention U-type network to extract spatial features: Apply the residual U-type network and the convolutional block attention module (CBAM) to construct a residual U-type network to extract the spatial features of SAR images and prepare spatial data for feature fusion; S3: Construct a C-LSTM network to extract time series features: Apply the convolutional long short-term memory module to construct a C-LSTM network to extract the interference phase time series features, and perform time series feature mapping based on three-dimensional convolution to prepare time series data for feature fusion; S4: Construct a feature fusion module to couple spatial features and temporal features: Apply the splicing operator to splice spatial features and temporal features in the channel dimension, and construct a feature fusion module based on two-dimensional convolution to obtain spatial and temporal joint features to prepare data for the prediction module; S5: Construct a prediction module to predict the spatial distribution of permanent scatterers: Apply the prediction module to output the dual-channel permanent scatterer identification results, and obtain the single-channel permanent scatterer prediction results through channel maximum screening, and finally use image block stitching and geocoding technology to synthesize the complete spatial distribution of permanent scatterers in the study area.

2. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: The existing model-driven permanent scatterer identification technology relies on phase model construction and iterative solution. The permanent scatterer identification technology does not need to build an algorithm model. It can quickly output permanent scatterer identification results by training a neural network, greatly improving the identification efficiency.

3. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: The existing data-driven permanent scatterer identification technology only focuses on the extraction of SAR image spatial features, while the technology focuses on both SAR image spatial features and interference phase timing features, and the obtained permanent scatterer identification results have spatiotemporal stability.

4. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: Existing data-driven permanent scatterer identification technology cannot evaluate the impact of the identification results on downstream tasks. The technology can provide an evaluation of the effectiveness of the identification results on the downstream phase unwrapping task.

5. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: The technology uses the focal loss function to solve the problem of imbalance in the number of positive and negative samples during network training; the network output used to construct the focal loss function is calculated according to the following formula: ； in, is the category index, and It is the output data of the network before passing through the Softmax activation function.

6. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: The focal loss function uses the output data of the network before passing through the Softmax activation function and is expressed as follows: ； in, is the category index, if the pixel is identified as a permanent scatterer, the category index value is 1, if the pixel is identified as a non-permanent scatterer, the category index value is 0; Represents the weighting factor for balancing positive and negative samples, The focusing factor represents the balance between difficult and easy samples. The weighting factor and focusing factor can be manually adjusted according to the ratio of permanent scatterer pixels to non-permanent scatterer pixels in the label data to improve the network's recognition accuracy for permanent scatterers.

7. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: The technology uses the Delaunay triangulation method to construct an interferometric phase triangle network, and the effectiveness evaluation indicator is the residual point number (RPN), which is calculated according to the following formula: ； in, Represents the triangle index value, represents the total number of triangles, Represents the triangle vertex index value; For triangle The winding phase corresponding to each vertex is is the paired winding phase difference, which may be 0 or ; Express Calculate the modulus value; the smaller the RPN value, the greater the possibility of obtaining accurate deformation results by phase unwrapping using the identified permanent scatterer.

8. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: The technology constructs a residual rate loss function based on the effectiveness evaluation index to improve the effectiveness of the network in identifying permanent scatterers. The residual rate loss function It is expressed as follows: ； in, represents the interference phase pattern index value, represents the total number of interference phase patterns, Represents the total number of pixels identified as permanent scatterers in the predicted permanent scatterer identification results.

9. The feature fusion type permanent scatterer recognition technology based on deep learning as claimed in claim 1, characterized in that: According to the focal loss function And the residual rate loss function , a hybrid loss function for network training According to the following formula: ； in, represents the weight coefficient of the focal loss function, Represents the weight coefficient of the residual rate loss function; this hybrid loss function can solve the imbalance problem of positive and negative difficult and easy samples and realize the effectiveness evaluation of the permanent scatterer identification results on the downstream phase unwrapping task.

10. The feature fusion type permanent scatterer recognition technology based on deep learning provided by the present invention is trained and tested on a platform equipped with a 13th Gen Intel(R) Core(TM) i9-13900K CPU and an NVIDIA GeForce RTX 4090 graphics card, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images