Deep Learning-Based Feature Fusion Permanent Scatter Recognition Method

Abstract

This invention discloses a feature fusion-based permanent scatterer (PSS) identification technique based on deep learning. First, a residual attention U-shaped network is constructed to extract spatial features from SAR images, and a C-LSTM network is constructed to extract temporal features from differential interferometric phase. Next, a feature fusion module is constructed to couple spatial and temporal features, and a prediction module is constructed to output dual-channel PSS identification results. Finally, single-channel PSS identification results are obtained by filtering through channel maximum values, and then image patch stitching and geocoding techniques are used to obtain the spatial distribution of PSS within the study area. This invention adopts a data-driven approach, enabling rapid identification of PSS through network training. The PSS identification technique provided by this invention employs a weighted summation of focal loss and residual loss to address the imbalance between positive and negative easy / negative samples and to evaluate the effectiveness of the identification results in downstream tasks. By utilizing the constructed feature fusion network and the mixed loss for PSS identification, this invention can identify a large number of PSS while maintaining a small number of residual points, making it more likely that the identification results will yield accurate surface deformation results when applied to downstream tasks.

Description

Technical Field

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

[0002] The application of Synthetic Aperture Radar Interferometry (InSAR) in land surface deformation measurement can generally be divided into two main categories: Differential InSAR (DInSAR) and Multi-Temporal InSAR (MT-InSAR). DInSAR uses two Synthetic Aperture Radar images covering the same area at different times to calculate the phase difference and thus obtain land surface deformation; its accuracy is easily affected by decoherence and atmospheric delay. MT-InSAR, on the other hand, 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) technique proposed by Ferretti et al. is a representative example in the field of MT-InSAR. One approach uses a threshold set for the amplitude deviation index to filter pixels with stable scattering characteristics in the time series as permanent scatterers, and extracts the corresponding interferometric phase of the permanent scatterers for subsequent parameter solving to improve deformation calculation results. Another approach to overcome spatiotemporal decorrelation in MT-InSAR is the Short Baseline Set InSAR (SBS-InSAR) technique, which uses SAR interferograms with short spatiotemporal baselines for deformation extraction. Since the introduction of PS-InSAR, many researchers have focused on enhancing and extending the PS-InSAR algorithm, but most existing improved algorithms only focus on the spatial features of SAR images, using the amplitude features of SAR images instead of phase stability as the permanent scatterer. The selection of scatterers, i.e., phase stability estimation without phase analysis, presents a challenge. Hooper et al. proposed the Stanford Permanent Scatterer (StaMPS) method, which utilizes the spatial correlation of deformation. First, an amplitude deviation index threshold is set to filter candidate permanent scatterer pixels. Then, permanent scatterers are further selected based on the stability of phase noise. Finally, a three-dimensional spatiotemporal convolution algorithm is applied to analyze the relationship between the interferometric phases corresponding to the permanent scatterers, thereby achieving high-precision deformation monitoring. Permanent scatterers are a key research object in MT-InSAR, and permanent scatterer identification is a crucial step in InSAR data processing. The effectiveness of the selected permanent scatterers limits the accuracy of deformation extraction; therefore, researchers have proposed… A series of model-driven permanent scatterer (PSS) identification algorithms have been developed to improve the accuracy of PSS identification. These algorithms typically require threshold settings to initially filter candidate PSS pixels and various phase filtering algorithms to select PSS pixels with spatiotemporal stability. For example, Navneet et al. proposed a PSS filtering method based on the highest eigenvalue threshold of the coherence matrix, which reduced the false alarm rate of PSS identification in low coherence regions. Wang et al. reduced false identification of PSS based on the phase similarity between adjacent PSS pixels. Although the above algorithms improve the effectiveness of PSS selection, they also increase computational cost and processing time.With the increasing efficiency of deep learning algorithms in image classification and object detection in recent years, more and more researchers are applying data-driven algorithms in remote sensing, including methods that utilize neural networks for permanent scatterer identification. Tiwari et al. proposed two deep learning architectures that correlate interferometric phase histories with the classification of phase-stable or unstable permanent scatterer pixels to improve the density of reliable permanent scatterer pixels. Zhang et al. constructed a dual-channel one-dimensional convolutional neural network to address class imbalance and the model's dependence on the amount of training samples. These algorithms significantly improve the efficiency of permanent scatterer identification; however, quantitative evaluation of the identification results is limited to comparing the density of permanent scatterer pixels within the region of interest, lacking an evaluation of the effectiveness of the identification results in downstream tasks such as phase unwrapping. Summary of the Invention

[0003] In view of this, this invention addresses the shortcomings of existing permanent scatterer (PSS) identification methods by proposing a feature fusion network architecture that combines spatial and temporal phase features of SAR images to screen for PSSs with spatiotemporal stability. Furthermore, this invention proposes a hybrid loss function to address the imbalance between positive and negative easy / poor samples, improving the effectiveness of PSS identification in downstream phase unwrapping tasks. This loss function is valuable for data-driven PSS identification methods. Experiments show that the network trained using the proposed architecture and loss function achieves fast and effective PSS identification with a larger number of PSSs, fewer residuals, and shorter screening time compared to the traditional StaMPS method.

[0004] To achieve the above objectives, the present invention provides the following method:

[0005] The feature fusion-based permanent scatterer identification method based on deep learning provided by this invention includes the following steps:

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

[0007] S2: Constructing a residual attention U-shaped network to extract spatial features: Applying a residual U-shaped network and a convolutional block attention module (CBAM) to construct a residual U-shaped network to extract spatial features of SAR images, thus preparing spatial data for feature fusion;

[0008] S3: Constructing a C-LSTM network to extract temporal features: A C-LSTM network is constructed using a convolutional long short-term memory module to extract temporal features of the interference phase, and temporal feature mapping is performed based on three-dimensional convolution to prepare temporal data for feature fusion;

[0009] S4: Construct a feature fusion module to couple spatial and temporal features: Apply splicing operators to splice spatial 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, thus preparing data for the prediction module;

[0010] S5: Construct a prediction module to predict the spatial distribution of permanent scatterers: The prediction module outputs the dual-channel permanent scatterer identification results, and the single-channel permanent scatterer prediction results are obtained by filtering the maximum value of the channels. Finally, the complete spatial distribution of permanent scatterers in the study area is synthesized by using image patch stitching and geocoding technology.

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

[0012] Furthermore, existing data-driven permanent scatterer identification methods only focus on the extraction of spatial features from SAR images, while the method described above simultaneously focuses on both the spatial features of SAR images and the temporal features of the interferometric phase, resulting in permanent scatterer identification results with spatiotemporal stability.

[0013] Furthermore, existing data-driven permanent scatterer identification methods cannot assess the impact of identification results on downstream tasks, while the method described above can provide an effectiveness assessment of the identification results for downstream phase unwrapping tasks.

[0014] Furthermore, the feature fusion-based permanent scatterer identification method utilizes a focal loss function to address the imbalance between 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, For category indexing, and Given the network's output data before passing through the Softmax activation function, the focus loss function is expressed using the following formula:

[0017] ;

[0018] in, For channel category indexing, This represents the value of the tag data in the j-th channel as one-hot encoded, i.e., if the pixel is identified as a permanent scatterer, the category index value is 1, and if the pixel is identified as a non-permanent scatterer, the category index value is 0. The weighting factor representing the balance between positive and negative samples. The focusing factor, which represents the balance between easy and difficult samples, can be manually adjusted based on the ratio of permanent scatterer pixels to non-permanent scatterer pixels in the label data to improve the network's accuracy in recognizing permanent scatterers.

[0019] Furthermore, the feature fusion-based permanent scatterer identification method constructs an interferometric phase triangle network using the Delaunay triangulation method. Using this triangle network, the effectiveness evaluation index, i.e., the number of residual points (RPN), is calculated according to the following formula:

[0020] ;

[0021] in, Represents the triangle index value. This represents the total number of triangles. Represents the vertex index of the triangle; For the triangle The entanglement phase corresponding to each vertex The phase difference is a pair of intertwined phases, which can take the values ​​0 or 1. ; Indicates to Calculate the modulus; the smaller the RPN value, the greater the possibility of obtaining accurate deformation results by using the identified permanent scatterer for phase unwrapping.

[0022] Furthermore, the feature fusion-based permanent scatterer identification method constructs a residual rate loss function based on an effectiveness evaluation index to improve the effectiveness of the network in identifying permanent scatterers; utilizing the RPN value, the residual rate loss function... Expressed according to the following formula:

[0023] ;

[0024] in, Indicates the index value of the interferometric phase diagram. This indicates the total number of interferometric phase diagrams. This represents the total number of pixels identified as permanent scatterers in the predicted permanent scatterer identification results.

[0025] Furthermore, using the aforementioned focus loss function and residual loss function Hybrid loss function for network training Expressed according to the following formula:

[0026] ;

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

[0028] The feature fusion permanent scatterer identification method based on deep learning provided by this invention was 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 this invention are as follows:

[0030] The present invention provides a feature fusion-based permanent scatterer identification method based on deep learning. This method proposes a feature fusion framework that fully explores the spatial and temporal phase features of SAR images. By using feature fusion operations to achieve spatiotemporal feature coupling, the identified permanent scatterers have spatiotemporal stability, thereby improving the effectiveness of permanent scatterer identification.

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

[0032] The feature fusion permanent scatterer identification method based on deep learning provided by this invention is lightweight. By training the network model, the permanent scatterer identification results of the study area can be obtained quickly, which greatly improves the efficiency of permanent scatterer identification.

[0033] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following description, or may be learned from the practice of the invention. The objectives and other advantages of the invention may be realized and obtained by means of the following description. Attached Figure Description

[0034] Figure 1 This is a flowchart of the permanent scatterer identification process presented by this method;

[0035] Figure 2 This is a schematic diagram of the residual attention U-shaped network structure of the permanent scatterer identification method.

[0036] Figure 3 This is a schematic diagram of the C-LSTM network structure of the permanent scatterer identification method.

[0037] Figure 4This is a schematic diagram of the time-series mapping module of the permanent scatterer identification method.

[0038] Figure 5 This verifies the spatial distribution of permanent scatterers within the study area during the experiment;

[0039] Figure 6 This verifies the spatial distribution of permanent scatterers within local region 1 in the experiment;

[0040] Figure 7 This is to verify the spatial distribution of permanent scatterers in local region 2 during the experiment. Detailed Implementation

[0041] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following description, in conjunction with the accompanying drawings and preferred embodiments, further explains the specific implementation methods, structures, features, and effects of the present invention, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

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

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

[0044] S2: Constructing a residual attention U-shaped network to extract spatial features: Applying a residual U-shaped network and a convolutional block attention module (CBAM) to construct a residual U-shaped network to extract spatial features of SAR images, thus preparing spatial data for feature fusion;

[0045] S3: Constructing a C-LSTM network to extract temporal features: A C-LSTM network is constructed using a convolutional long short-term memory module to extract temporal features of the interference phase, and temporal feature mapping is performed based on three-dimensional convolution to prepare temporal data for feature fusion;

[0046] S4: Construct a feature fusion module to couple spatial and temporal features: Apply splicing operators to splice spatial 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, thus preparing data for the prediction module;

[0047] S5: Construct a prediction module to predict the spatial distribution of permanent scatterers: The prediction module outputs the dual-channel permanent scatterer identification results, and the single-channel permanent scatterer prediction results are obtained by filtering the maximum value of the channels. Finally, the complete spatial distribution of permanent scatterers in the study area is synthesized by using image patch stitching and geocoding technology.

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

[0049] like Figure 2 As shown, the residual attention U-shaped network is used to extract spatial features of three-band SAR image patches. The residual attention U-shaped network includes a convolutional module with 3 kernels, a pooling module, a transposed convolutional module, a convolutional module with 1 kernel, a CBAM attention module, and a residual module. The convolutional module with 3 kernels is used to extract data features. The pooling module is used to remove redundant features to achieve feature dimensionality reduction. The transposed convolutional 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 concatenate shallow and deep features to enhance feature extraction performance. The output of the residual attention U-shaped network includes four dimensions, representing the number of image patches in the batch, the number of channels, the image width, and the image height, respectively.

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

[0051] like Figure 3 As shown, the input of the C-LSTM module includes cell state. Hidden state Differential interferometric phase data The C-LSTM module output includes cell state. Hidden state Based on the input hidden state and differential interference phase data, the forget gate is solved using the following formula. Input gate Output gate Candidate cell status and output cell state and hidden state :

[0052] ;

[0053] in, This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. For convolution operators, For Hadamard product operation; , , , Let represent the weight matrices of the forget gate, input gate, output gate, and candidate cell state gate, respectively. , , , These represent the bias matrices for the forget gate, input gate, output gate, and candidate cell state gate, respectively; the bias matrices mentioned above are in... Figure 3 The merger 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 processed by the temporal mapping; the temporal mapping module includes a 3D convolutional layer with 3 kernels, a 3D batch normalization layer, a ReLU activation layer, a 3D convolutional layer with 1 kernel, and a time-step averaging layer; the 3D convolutional layer with 3 kernels is used to enhance the stability of the temporal mapping; the 3D batch normalization layer is used to avoid the vanishing and exploding gradient phenomena of temporal features; the ReLU activation layer is used to enhance the nonlinearity of the network; the 3D convolutional layer with 1 kernel is used for temporal feature mapping; the time-step averaging layer is used to integrate temporal data and achieve data dimensionality reduction; the output of the temporal mapping module has the same output dimension as the residual attention U-shaped network.

[0055] In this embodiment, the output of the residual attention U-shaped network and the output of the temporal mapping module are concatenated in the channel dimension and fed into the feature fusion module. The feature fusion module includes a two-dimensional convolutional layer with 3 kernels, a two-dimensional batch normalization layer, a ReLU activation layer, and a two-dimensional convolutional layer with 1 kernel. The two-dimensional convolutional layer with 3 kernels is used to maintain the spatial stability of the fused features. The two-dimensional batch normalization layer is used to avoid the vanishing and exploding of spatial feature gradients. The two-dimensional convolutional layer with 1 kernel is used for spatial feature mapping. In this embodiment, the feature fusion module is used to map the doubled number of channels after concatenation to the number of channels before concatenation, thereby achieving deep fusion of image spatial features and phase temporal features.

[0056] In this embodiment, the output of the feature fusion module is used as the input of the prediction layer. The prediction layer outputs a dual-channel binarized permanent scatterer prediction result, which is then compared with the label data encoded by one-hot encoding to calculate a loss function, which is used to constrain the direction of network training. The prediction layer includes a two-dimensional convolutional layer with a kernel of 1, a two-dimensional batch normalization layer, and a Softmax activation layer. The Softmax activation layer is used to normalize the range of the prediction result to between 0 and 1.

[0057] This embodiment obtains a single-channel permanent scatterer selection image patch by filtering the normalized two-channel prediction results through the maximum channel value, and outputs the complete permanent scatterer selection results within the study area through image patch stitching and geocoding; the total number of network parameters in this embodiment is about 79MB, which has the characteristics of lightweight; the network in this embodiment can achieve a permanent scatterer recognition time far lower than StaMPS after training.

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

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

[0060] In this embodiment, the data used to construct the training set consists of three parts: three-band SAR images, differential interferometric phase images, and label data. The three-band SAR images are formed by superimposing the average amplitude, amplitude deviation index, and coherence coefficient of 10 auxiliary images from the time-series SLC images in the band dimension. The differential interferometric phase images are single-channel five-dimensional data with a time step of 10. Existing research has shown that 10 differential interferometric 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 composed 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 labels, guiding the network training direction. 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 the label data.

[0061] After acquiring the data used for training set construction, the edge parts of the differential interferometric phase image without data are removed and shaped into multiples of 128 in both width and height for image block segmentation. In the image block segmentation operation, the three-band SAR image, differential interferometric phase image and label data are segmented from left to right and from top to bottom into image blocks with a width and height of 128 for network batch processing. The network computing efficiency is improved by parallel computing using GPU.

[0062] exist Figure 5 middle, Figure 5 (a) and Figure 5 (b) The results of permanent scatterer identification using the StaMPS algorithm and the method described above are shown respectively within the study area. The study area has dimensions of 1664×7680 and contains 780 image blocks with a width and height of 128. In terms of the number of identified permanent scatterers, the StaMPS algorithm identified 471,490 permanent scatterers, while the method described above identified 498,407, 26,917 more than StaMPS. Furthermore, based on the RPN values ​​calculated from the identification results of the two methods, the StaMPS algorithm identified 30,130 residual points, while the method described above identified 26,705, 3,425 fewer residual points than StaMPS. This demonstrates that the method can identify a large number of permanent scatterers while maintaining a small number of residual points, thus proving its effectiveness. Moreover, in terms of the distribution of permanent scatterers, both the StaMPS algorithm and the method described above exhibit certain clustering characteristics. Figure 5 In the selection process of the permanent scatterer identification result shown, the StaMPS algorithm took 4860 seconds; given that the network can quickly obtain the permanent scatterer identification result through training, its prediction process only took 29 seconds, proving that the method is efficient; Figure 5 In the image, local regions 1 and 2 are selected using red boxes for further analysis.

[0063] exist Figure 6 In the middle, it is magnified and displayed. Figure 5 The permanent scattering object identification results are shown in the northeastern local area 1 selected by the red box in the study area; it can be seen that, in Figure 6 Within a local region, both the StaMPS algorithm and the aforementioned method identify permanent scatterers with good aggregation; while located in Figure 6 In the permanent scattering object identification results of the central and southeastern regions, the permanent scattering objects identified by the StaMPS algorithm not only appeared in areas with buildings, but were also scattered in nearby farmland areas; the permanent scattering object identification results of the proposed method were more obviously clustered in areas with buildings, and its distribution clustering was better than that of the StaMPS algorithm.

[0064] exist Figure 7 In the middle, it is magnified and displayed. Figure 5 The permanent scatterer identification results are shown in the central local region 2 selected by the red box in the study area; Figure 7Within the local area, the area selected by the red box is an area with more buildings such as houses, and the area selected by the red ellipse is a green or farmland area. It can be seen that the permanent scatterers identified by the StaMPS algorithm are clustered on some buildings, while the permanent scatterer identification method can extract permanent scatterers on scattered buildings, and the distribution is more uniform. The reason for this phenomenon is that the feature fusion network can learn more information with spatiotemporal stability, considering not only the spatial distribution of permanent scatterers but also their temporal phase characteristics, making the identification results more uniform in distribution while still having clustering. In the green or farmland areas where there is no clustering of buildings, there are no high coherence points. Therefore, the permanent scatterers identified in this area are misclassified pixels. The StaMPS algorithm misclassified some pixels located in this area, while the permanent scatterer identification method did not misclassify any pixels and performed well in extracting permanent scatterer features.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some more or modified equivalent embodiments based on the above-disclosed technical content without departing from the scope of the technical solution of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A deep learning-based feature fusion type permanent scatterer identification method, characterized in that: The method comprises the following steps: S1: obtaining time-series SAR data for constructing a training set and a test set; S2: constructing a residual attention U-shaped network to extract spatial features: a residual U-shaped network and a convolution block attention module CBAM are applied to construct a residual U-shaped network to extract spatial features of a SAR image, and spatial data is prepared for feature fusion; S3: constructing a C-LSTM network to extract time-series features: a convolution long short-term memory module is applied to construct a C-LSTM network to extract time-series features of an interference phase, and time-series features are mapped based on three-dimensional convolution, and time-series data is prepared for feature fusion; The C-LSTM network comprises a C-LSTM module and a time-series mapping module; The C-LSTM module is used for extracting features of time-series data; The time-series mapping module is used for mapping time steps of time-series data; The C-LSTM module inputs include cell state , hidden state , differential interference phase data ; The C-LSTM module outputs include cell states , hidden states ; According to the input hidden state and the differential interference phase data, the forget gate is solved according to the following formula , the input gate , the output gate , the candidate cell state and the output cell state and the hidden state : ； wherein, denotes a Sigmoid activation function, denotes a hyperbolic tangent activation function, is a convolution operator, is a Hadamard product operation; , , , denote the weight matrices of the forget gate, input gate, output gate and candidate cell state gate, respectively, and , , , denote the bias matrices of the forget gate, input gate, output gate and candidate cell state gate, respectively; S4: constructing a feature fusion module to couple spatial features and time-series features: a concatenation operator is applied to concatenate spatial features and time-series features in a channel dimension, a feature fusion module is constructed based on two-dimensional convolution, spatial time-series joint features are obtained, and data is prepared for a prediction module; The feature fusion module comprises 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 1 is used for mapping the number of concatenated channels to the number of channels before concatenation, realizing deep fusion of spatial features and time-series features; S5: constructing a prediction module to predict the spatial distribution of permanent scatterers: a prediction module is applied to output a two-channel permanent scatterer identification result, a single-channel permanent scatterer prediction result is obtained by screening the maximum value of the channel, and finally an image block is spliced and geographic coding technology is used to synthesize a complete permanent scatterer spatial distribution in a study area; The technology uses the Delaunay triangulation method to construct an interference phase triangle network, and an effectiveness evaluation index, i.e., a residual point number (RPN), is calculated according to the following formula: ； in, Represents the triangle index value. Represents the total number of triangles. Represents the vertex index of the triangle; For the triangle The entanglement phase corresponding to each vertex The phase difference is a pair of intertwined phases, which can take the values ​​0 or 1. ; Indicates to Calculate the modulus; the smaller the RPN value, the greater the possibility of obtaining accurate deformation results by using the identified permanent scatterer for phase unwrapping. The method improves the effectiveness of the network in identifying permanent scatterers based on the construction of a residual rate loss function based on an effectiveness evaluation index, and the residual rate loss function is is expressed according to the following formula: ； wherein, represents an interference phase map index value, represents a total number of interference phase maps, represents a total number of pixels identified as permanent scatterers in the predicted permanent scatterer identification result. 2.The deep learning-based feature fusion type permanent scatterer identification method of claim 1, wherein: The method uses a focal loss function to solve the imbalance problem between positive and negative samples in the network training process; the network output for constructing the focal loss function is calculated according to the following formula: ； wherein, is a class index, and is the output data before the Softmax activation function by the network. 3.The deep learning-based feature fusion type permanent scatterer identification method of claim 2, wherein: The focal loss function uses the output data of the network before the Softmax activation function, which is expressed according to the following formula: ； wherein, is a channel category index, represents the value of the label data in one-hot encoding on the jth channel, that is, if the pixel is identified as a permanent scatterer, the category index value is 1, and if the pixel is identified as a non-permanent scatterer, the category index value is 0; represents a weighting factor for balancing positive and negative samples, represents a focusing factor for balancing easy and difficult samples. The weighting factor and the focusing factor can be manually adjusted according to the proportion of permanent scatterer pixels and non-permanent scatterer pixels in the label data to improve the identification accuracy of the network for permanent scatterers. 4.The deep learning-based feature fusion type permanent scatterer identification method of claim 2, wherein: According to the focal loss function and the residual rate loss function , the hybrid loss function for network training is expressed as follows: ； wherein, represents a focus loss function weight coefficient, represents a residual rate loss function weight coefficient; the mixed loss function can solve the imbalance problem of positive and negative difficult samples and realize the effectiveness evaluation of the identification result of the permanent scatterer on the downstream phase unwrapping task.

5. A deep learning-based feature fusion type permanent scatterer identification system, wherein the system is trained and tested on a platform equipped with a 13th Gen Intel(R) Core(TM) i9-13900K central processing unit and a NVIDIA GeForce RTX 4090 graphics card, characterized in that, The processor executes the program to realize the method of any one of claims 1 to 4.

Images