Method and system for recovering glacier loss velocity from SAR images based on combined network

Abstract

The present invention discloses a method and system for recovering the missing velocity of glaciers from SAR images based on a combined network, and belongs to the technical field of remote sensing image signal processing. The method comprises: acquiring time series SAR images of the study area and performing preprocessing to obtain a glacier velocity field; calculating the mean of each glacier velocity point in the neighborhood, and inputting the mean and the position coordinates of the glacier velocity into an artificial neural network to train the artificial neural network; obtaining the initial recovered glacier velocity using the trained artificial neural network; obtaining the reconstructed glacier velocity using the initial recovered glacier velocity; performing a weighted combination of the initial recovered glacier velocity and the reconstructed glacier velocity as the input of a denoising autoencoder, and finally recovering the missing glacier velocity. The present invention makes full use of the spatiotemporal characteristics of glacier velocity, and can be completed only by relying on time series SAR images, without introducing additional external data, thereby improving the robustness of glacier velocity recovery.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image signal processing and polar remote sensing application technology, and particularly relates to a method and system for recovering glacier loss velocity from SAR images based on a combined network. Background Art

[0002] Glaciers are natural bodies of ice formed by the compaction of accumulated snow over time. They move slowly along the Earth's surface under the influence of gravity. Glacier movement is a key indicator of global climate change. By studying changes in glacier movement speed, we can reconstruct regional and global climate change patterns. Furthermore, glacier movement is closely related to human production and daily life, making the measurement of glacier movement speed of great significance.

[0003] Currently, methods for obtaining glacier velocity using remote sensing imagery primarily include optical remote sensing and synthetic aperture radar (SAR). Although optical remote sensing can provide richer spectral information, optical imagery is often limited by cloud and lighting conditions. SAR, as an active microwave sensor, offers the advantage of all-day, all-weather operation, making it an important tool for monitoring glacier velocity. Pixel offset tracking is the primary method for obtaining glacier velocity based on SAR imagery. However, due to factors such as the glacier's daily freeze-thaw cycle, the surface characteristics of the glacier can change over a short period of time. These changes in characteristics can lead to partial loss of glacier velocity due to decoherence.

[0004] Currently, methods for recovering missing glacier velocities fall into two main categories. One type relies on spatial interpolation, including nearest neighbor methods, kriging, and regression analysis. However, the reliability of the interpolation results depends on the spatial distribution of the missing data and ignores local variations in glacier velocity. Consequently, the recovered glacier velocity results are unstable and can contain significant errors. The other type relies on neural network-based glacier velocity recovery. These methods treat velocity recovery as a supervised training problem, inputting factors related to glacier motion mechanisms and solving them using an artificial neural network. However, existing neural network-based glacier velocity recovery methods require external auxiliary data, such as glacier thickness and glacial terrain slope. This auxiliary data is often extremely difficult to obtain. Even when available, the information obtained is often derived from historical data, which can be time-lagged with the glacier velocity to be recovered. This can interfere with the recovered glacier velocity results, increasing errors in the velocity recovery. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a method and system for recovering the glacier loss velocity from SAR images based on a combined network, which makes full use of the spatiotemporal variation information of the glacier velocity, and uses an artificial neural network (ANN) and a denoising autoencoder (DAE) to recover the glacier velocity: the spatiotemporal information of the glacier velocity is input as a parameter into the ANN for training, the training result is used as the initial value of the DAE, and the DAE and the ANN output result are weightedly combined and used as the input of the DAE again. The error of the reconstructed glacier velocity is minimized through continuous iteration, thereby achieving effective recovery of the glacier loss velocity.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for recovering glacier loss velocity from SAR images based on a combined network, comprising the following steps:

[0008] Step 1: Obtain time series SAR images of the study area and preprocess them to obtain the glacier velocity field;

[0009] Step 2: Calculate the mean of each glacier velocity point in the neighborhood using the glacier velocity field, input the mean and the position coordinates of the glacier velocity point into an artificial neural network, and train the artificial neural network;

[0010] Step 3: Use the trained artificial neural network to restore the missing glacier velocity and obtain the initial restored glacier velocity;

[0011] Step 4: Using the initially recovered glacier velocity as the initial input value of the denoising autoencoder, and outputting the reconstructed glacier velocity;

[0012] Step 5: Perform a weighted combination of the initial recovered glacier velocity and the reconstructed glacier velocity and use them as the input of the denoising autoencoder again, and iterate to minimize the error of the reconstructed glacier velocity, and finally restore the missing glacier velocity.

[0013] In a second aspect, the present invention provides a SAR image glacier loss rate recovery system based on a combined network, comprising:

[0014] The preprocessing module is used to obtain and preprocess the time series SAR images of the study area and extract the glacier velocity field data;

[0015] A training module, used for inputting the glacier velocity field data into an artificial neural network and training the artificial neural network;

[0016] An initial calculation module is used to restore the missing glacier velocity using the trained artificial neural network to obtain the initial restored glacier velocity;

[0017] a reconstruction calculation module, configured to use the initially recovered glacier velocity as an initial input value of a denoising autoencoder and output a reconstructed glacier velocity;

[0018] An iterative calculation module is used to perform a weighted combination of the initially restored glacier velocity and the reconstructed glacier velocity and use them as the input of the denoising autoencoder, and to minimize the error of the reconstructed glacier velocity through iteration, so as to finally restore the missing glacier velocity.

[0019] In a third aspect, the present invention provides an electronic device comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned combined network-based SAR image glacier loss rate recovery method.

[0020] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned method for recovering the glacier loss rate of SAR images based on a combined network.

[0021] The beneficial effects of the present invention are:

[0022] This method fully utilizes the spatiotemporal characteristics of glacier velocity and combines ANN with DAE to recover the missing glacier velocity. ANN is used to capture the spatial correlation of glacier velocity, while DAE is used to deal with noise and missing values ​​in time series data. This method can not only handle local variations in glacier velocity, but also continuously optimize the recovery results during the iterative process, ensuring that the final recovered glacier velocity is more accurate.

[0023] The present invention can be completed only by relying on time series SAR images without introducing additional external data, thereby improving the robustness of glacier velocity recovery.

[0024] This method uses a combined neural network approach to perform supervised and unsupervised learning of glacier velocity. The results obtained by the ANN and DAE are weighted, and the optimal solution of the combination is found in an iterative process. This method effectively improves the accuracy of glacier velocity recovery. It is applicable to various types of SAR image data and works effectively in different glacial regions and climate conditions. By adjusting the parameters of the ANN and DAE, it can adapt to different data characteristics and recovery requirements, demonstrating its broad applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1This is a flow chart of the method for recovering glacier loss velocity from SAR images based on a combined network of the present invention;

[0026] Figure 2 This is the logic flow chart of the method for recovering the glacier loss velocity from SAR images based on a combined network of the present invention;

[0027] Figure 3 (a) shows the main SAR image after registration;

[0028] Figure 3 (b) shows the SAR auxiliary image after registration;

[0029] Figure 4 Schematic diagram of the glacier velocity field with missing glaciers;

[0030] Figure 5 Schematic diagram of the glacier velocity field restored based on the method of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described below with reference to the accompanying drawings and examples.

[0032] like Figure 1 As shown, the present invention provides a method for recovering the glacier loss rate from SAR images based on a combined network, which includes the following steps:

[0033] Step 1: Obtain time series SAR images of the study area and preprocess them to extract glacier velocity and obtain the glacier velocity field;

[0034] Step 2: Calculate the mean of each glacier velocity point in the neighborhood using the obtained glacier velocity field, input the mean and the position coordinates of the glacier velocity into the ANN, and train the ANN;

[0035] Step 3: Use the trained ANN to restore the missing glacier velocity and obtain the initial restored glacier velocity;

[0036] Step 4: Use the initially recovered glacier velocity as the initial input value of DAE and output the reconstructed glacier velocity;

[0037] Step 5: Perform a weighted combination of the initial restored glacier velocity and the reconstructed glacier velocity and use them as the input of the DAE again. Then, the error of the reconstructed glacier velocity is minimized through iteration, and the missing glacier velocity is finally restored.

[0038] like Figure 2Figure 2 shows a logic flow chart of a combined network-based method for recovering glacier loss rates from SAR images. First, the acquired time-series SAR images are preprocessed, specifically by filtering and registering the SAR images. Since speckle noise is often present in SAR images, this noise is first removed by filtering. Common filtering methods include mean filtering and Lee filtering. The time-series SAR images are then registered. SAR image registration can be achieved using grayscale-based registration methods, feature-based registration methods, or by using relevant software (such as SNAP and ENVI).

[0039] The normalized cross-correlation algorithm is applied to the registered SAR image pair. The normalized cross-correlation function used in the normalized cross-correlation algorithm is as follows:

[0040] (1)

[0041] Where, is the template window in the SAR main image, and Template window in the SAR main image The size, is the search window in the SAR auxiliary image, and the search window in the SAR auxiliary image The size is larger than the template window size; Represents the template window in the SAR main image and the search window in the SAR auxiliary image Template window in the SAR main image The corresponding row elements, Table SAR template window in the main image and the search window in the SAR auxiliary image Template window in the SAR main image The corresponding Column elements, It is the template window in the SAR main image The average value of the pixels in is the search window in the SAR auxiliary image The average value of the pixels in the range corresponding to the template window in the SAR main image. Search window in SAR auxiliary image The glacier velocity is estimated by calculating the normalized cross-correlation function value after each movement and finding the pixel offset number corresponding to the maximum normalized cross-correlation function value:

[0042] (2)

[0043] in, and are the estimated glacier velocities in azimuth and range directions, is the number of pixel offsets in azimuth obtained by normalizing the cross-correlation function, is the distance to pixel offset number obtained by normalizing the cross-correlation function, is the azimuthal resolution of pixels, is the distance resolution of pixels, The time interval between the acquisition time of the SAR main image and the auxiliary image;

[0044] Estimate the glacier velocity for all SAR image pairs and obtain the corresponding glacier velocity field.

[0045] Generally speaking, the glacier velocity at a certain point is usually related to the velocities of other glaciers in the neighborhood. Therefore, the estimated glacier velocity in the azimuth and range directions for each point in the glacier velocity field can be expressed as:

[0046] (3)

[0047] in, is the distribution function describing the glacier velocity field, and are the spatial coordinates of the velocity point in the range and azimuth directions, and is the average velocity in the range and azimuth directions within the velocity point's neighborhood. An artificial neural network (ANN) was modeled and trained using the acquired glacier velocities. An ANN is a computational model inspired by biological neural networks and consists of an input layer, a hidden layer, and an output layer. Each layer contains several nodes, connected by weighted connections. The ANN learns and predicts by adjusting these weights. Parameters related to the missing glacier velocity were input into the trained network, including the spatial coordinates of the missing velocity point in the range and azimuth directions and the average velocity in the range and azimuth directions within the velocity point's neighborhood, to obtain the initial recovered glacier velocity.

[0048] Then DAE is constructed. DAE is an unsupervised deep learning method consisting of an encoder and a decoder. Its basic principle is to automatically model the distribution of data through deep learning.

[0049] First, we obtain the glacier velocity time series and fill the missing parts of the glacier velocity time series with the glacier velocity obtained by the ANN. We add noise to the input data and learn the mapping relationship:

[0050] (4)

[0051] in, represents the noise function, represents the initial recovery speed of the glacier, Represents the generated noisy data; the encoder is used to Transformed into latent variables :

[0052] (5)

[0053] Where, and are the encoder weights and biases, respectively, Is the activation function; use the decoder to transform the hidden variables back to the original data :

[0054] (6)

[0055] Where, and are decoder weights and biases respectively. With the goal of minimizing the error function, optimize the parameters , , , ,Right now:

[0056] (7)

[0057] Where, is the error function, is the length of the time series, Indicates the The initial recovery speed of the glacier at a time point, Indicates the The glacier velocity after encoding and decoding at each time point.

[0058] Finally, the glacier velocity reconstructed by DAE and the initial restored glacier velocity obtained by ANN are weighted:

[0059] (8)

[0060] Where, is the new weighted glacier velocity, and Represents the initial recovery glacier speed and reconstructed glacier velocity The weight of ; The weighted new glacier velocity is used to fill in the missing glacier velocity of DAE and iterate until the error function Converges to a minimum.

[0061] Without loss of generality, when obtaining glacier velocity from time series SAR images, in addition to the normalized cross-correlation method, other methods such as interferometry and optical flow can also be used.

[0062] At this point, the recovery of the missing glacier speed is completed.

[0063] Example

[0064] First, this example selects the Petermann Glacier in Greenland as the study area. Time series SAR images of this area are downloaded and filtered and registered. Figures 3(a) and 3(b) show a pair of preprocessed SAR images. Figure 3(a) is the registered primary SAR image, and Figure 3(b) is the registered secondary SAR image. The normalized cross-correlation algorithm is used for each pair of SAR images according to formulas (1) and (2) to estimate the glacier velocity of the SAR image pair. Figure 4 The figure shows an example of a missing glacier velocity field. The circles in the figure indicate the missing areas of glacier velocity, and the velocity units are meters per day (m / d). First, the parameters related to the glacier velocity are input into the ANN and the network is trained using formula (3). The parameters related to the missing glacier velocity are input into the trained network to obtain the initial restored glacier velocity. Then, the results obtained by the ANN are used as the filling values ​​for the missing parts of the glacier velocity time series. The DAE is trained according to formulas (4)-(7) to obtain the reconstructed missing glacier velocity. Finally, the missing glacier velocity reconstructed by the DAE and the initial restored glacier velocity obtained by the ANN are weighted according to formula (8), and the DAE training process is repeated until the error of the reconstructed glacier velocity is minimized. Figure 5 This is the glacier velocity field restored using the method of the present invention. It can be seen that the missing glacier velocity has been completely restored.

[0065] On the other hand, the present invention provides a SAR image glacier loss rate recovery system based on a combined network, which includes various modules capable of implementing various steps of the aforementioned method, specifically including:

[0066] The preprocessing module is used to obtain and preprocess the time series SAR images of the study area and extract the glacier velocity field data;

[0067] A training module, used for inputting the glacier velocity field data into an artificial neural network and training the artificial neural network;

[0068] An initial calculation module is used to restore the missing glacier velocity using the trained artificial neural network to obtain the initial restored glacier velocity;

[0069] a reconstruction calculation module, configured to use the initially recovered glacier velocity as an initial input value of a denoising autoencoder and output a reconstructed glacier velocity;

[0070] An iterative calculation module is used to perform a weighted combination of the initially restored glacier velocity and the reconstructed glacier velocity and use them as the input of the denoising autoencoder, and to minimize the error of the reconstructed glacier velocity through iteration, so as to finally restore the missing glacier velocity.

[0071] In a third aspect, the present invention provides an electronic device comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned combined network-based SAR image glacier loss rate recovery method.

[0072] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned method for recovering the glacier loss rate of SAR images based on a combined network.

[0073] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for recovering glacier loss velocity from SAR images based on a combined network, characterized by: The steps include: Step 1: Obtain time series SAR images of the study area and preprocess them to obtain the glacier velocity field; Step 2: Calculate the mean of each glacier velocity point in the neighborhood using the glacier velocity field, input the mean and the position coordinates of the glacier velocity point into an artificial neural network, and train the artificial neural network. This includes: modeling the artificial neural network and training it using the acquired glacier velocity field data, wherein the estimated glacier velocity in the azimuth and range directions of each glacier velocity point in the glacier velocity field can be expressed as: （3） Where, and are the estimated glacier velocities in azimuth and range directions, is the distribution function describing the glacier velocity field, and are the spatial coordinates of the velocity point in the range and azimuth directions, and is the average velocity in the range and azimuth directions within the velocity point neighborhood; Step 3: Using the trained artificial neural network to recover the missing glacier velocity to obtain the initial recovered glacier velocity; including: inputting the spatial coordinates of the missing velocity point in the range and azimuth directions and the average velocity in the range and azimuth directions within the neighborhood of the velocity point into the trained artificial neural network to obtain the initial recovered glacier velocity; Step 4: Using the initially recovered glacier velocity as the initial input value of the denoising autoencoder to output the reconstructed glacier velocity; including: Constructing the denoising autoencoder, including an encoder and a decoder; Get the glacier velocity time series and fill the missing parts in the glacier velocity time series with the initial recovered glacier velocity, add noise to the input data and learn the mapping relationship: （4） Where, represents the noise function, represents the initial recovered glacier velocity data, represents the generated noisy data; Use the encoder to generate noisy data Transformed into latent variables : （5） Where, and are the encoder weights and biases, respectively, is the activation function; Use the decoder to transform the hidden variables Transformed into reconstructed glacier velocity : （6） Where, and are decoder weights and biases respectively; Taking minimizing the error function as the training goal, optimize the parameters , , , ,Right now: （7） Where, is the error function, is the length of the time series, Indicates the The initial recovery speed of the glacier at a time point, Indicates the The glacier velocity after encoding and decoding at each time point; Step 5: Weighting the initial recovered glacier velocity and the reconstructed glacier velocity and combining them as the input of the denoising autoencoder, and iteratively minimizing the error of the reconstructed glacier velocity to finally restore the missing glacier velocity, including: weighting the glacier velocity reconstructed by the denoising autoencoder and the initial recovered glacier velocity obtained by the artificial neural network: （8） Where, is the weighted new glacier velocity, and Represents the initial recovery glacier speed and reconstructed glacier velocity The weight of ; The new glacier velocity after weighting Used to fill in the missing glacier velocity of the denoising autoencoder and iterate until the error function Converges to a minimum.

2. The method for recovering glacier loss rate from SAR images based on a combined network according to claim 1, characterized in that: In the step 1, preprocessing the time series SAR images includes filtering and registering the time series SAR images.

3. The method for recovering glacier loss rate from SAR images based on a combined network according to claim 2, characterized in that: In step 1, a normalized cross-correlation algorithm is applied to the registered SAR image pair, wherein the normalized cross-correlation function as follows: （1） Where, is the template window in the SAR main image, and Template window in the SAR main image The size, is the search window in the SAR auxiliary image, and the search window in the SAR auxiliary image The size is larger than the template window in the SAR main image size; Represents the template window in the SAR main image and the search window in the SAR auxiliary image Template window in the SAR main image The corresponding row elements, Represents the template window in the SAR main image and the search window in the SAR auxiliary image Template window in the SAR main image The corresponding Column elements, It is the template window in the SAR main image The average value of the pixels in is the search window in the SAR auxiliary image The average value of the pixels in the range corresponding to the template window in the SAR main image; when searching, the template window in the SAR main image Search window in SAR auxiliary image The glacier velocity is estimated by calculating the normalized cross-correlation function value after each movement and finding the pixel offset number corresponding to the maximum normalized cross-correlation function value: （2） Where, and are the estimated glacier velocities in azimuth and range directions, The normalized cross-correlation function The number of pixel offsets in the azimuth direction obtained, The normalized cross-correlation function The distance obtained is offset by the number of pixels. is the azimuthal resolution of pixels, is the distance resolution of pixels, The time interval between the acquisition time of the SAR main image and the auxiliary image; Estimate the glacier velocity for all SAR image pairs and obtain the corresponding glacier velocity field.

4. A SAR image glacier loss rate recovery system based on a combined network, used to implement the method according to any one of claims 1 to 3, characterized in that: include: The preprocessing module is used to obtain and preprocess the time series SAR images of the study area and extract the glacier velocity field data; A training module, used for inputting the glacier velocity field data into an artificial neural network and training the artificial neural network; An initial calculation module is used to restore the missing glacier velocity using the trained artificial neural network to obtain the initial restored glacier velocity; a reconstruction calculation module, configured to use the initially recovered glacier velocity as an initial input value of a denoising autoencoder and output a reconstructed glacier velocity; An iterative calculation module is used to perform a weighted combination of the initially restored glacier velocity and the reconstructed glacier velocity and use them as the input of the denoising autoencoder, and to minimize the error of the reconstructed glacier velocity through iteration, so as to finally restore the missing glacier velocity.

5. An electronic device, characterized in that: include: one or more processors; a memory for storing one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the method for recovering glacier loss velocity in SAR images based on a combined network as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement a method for recovering glacier loss velocity from SAR images based on a combined network as described in any one of claims 1 to 3.

Images