A SAR pixel offset tracking deformation estimation method and system based on an iterative adaptive window

Abstract

The application discloses a SAR pixel offset tracking deformation estimation method and system based on an iterative adaptive window, and comprises the following steps: acquiring two scenes of SAR images of a target area at different times; constructing an elliptical registration window according to the resolution ratio of the range direction and the azimuth direction of the SAR images; calculating the similarity plane of the master-slave image registration window; constructing a robustness index of offset estimation; judging whether the stability index reaches a threshold value; if the stability index threshold value is not reached, taking the elliptical registration window as an initial window, expanding in multiple directions, taking the registration window with the highest similarity as the optimal expansion direction, until the robustness index of the current registration window reaches the threshold value; and then acquiring the maximum position of the two-dimensional similarity plane of the optimal registration window, so as to determine the two-dimensional pixel offset. The application realizes adaptive optimization of the optimal shape and size of the offset estimation registration window, improves the precision of deformation estimation, and retains the deformation detail information.

Description

Technical Field

[0001] This invention belongs to the field of surface deformation monitoring technology, and relates to a SAR pixel offset tracking deformation estimation method and system based on iterative adaptive window. Background Technology

[0002] my country's complex geological conditions and frequent tectonic activity contribute to the frequent occurrence of geological disasters. Deformation monitoring of geological disasters provides crucial decision-making support for disaster forecasting and early warning, making it a fundamental and vital task in national geological disaster prevention and control. Synthetic Aperture Radar Interferometry (D-InSAR) technology, developed in recent years, has been widely applied in geological disaster monitoring due to its all-weather and high-precision characteristics. However, D-InSAR is susceptible to decoherence due to factors such as vegetation and atmospheric delay, and is limited to scenarios with small deformation gradients due to phase gradient limitations. SAR Pixel Offset Tracking (POT) technology, based on the principle of dense image matching, can acquire the two-dimensional deformation field of a target area and has been widely used in monitoring scenarios with large deformation gradients, such as landslides, glaciers, earthquakes, and mine subsidence. Compared to D-InSAR, this method is unaffected by deformation gradients and decoherence, and its measurement accuracy is determined by image resolution (approximately 1 / 10 to 1 / 30 of a pixel). Complementing D-InSAR, this method has significant application potential in the field of surface deformation monitoring.

[0003] Pixel offset tracking methods primarily utilize the amplitude information of SAR images, based on the principle of corresponding point matching, to obtain the two-dimensional deformation fields of the master and slave images in the azimuth and range directions. Traditional SAR pixel offset tracking methods mainly use the normalized cross correlation (NCC) method to evaluate the similarity of two regular rectangular matching windows of the master and slave images, thereby achieving registration of the corresponding points at the center of the matching windows. First, pixel offset tracking methods estimate the statistical similarity of the two matching windows instead of the similarity of the two corresponding center pixels. This matching strategy is feasible in homogeneous motion regions (i.e., the motion of local areas is basically the same), but in heterogeneous motion regions (i.e., there are large differences in motion in local areas), pixels within the matching window often have different motion characteristics (stable or moving, fast or slow). Typical heterogeneous motion regions include landslide boundary areas, local collapse areas of landslide bodies, and other areas with large gradient changes. Therefore, the regular rectangular matching window may lead to differences in the motion attributes of pixels within the matching window, making it difficult to accurately estimate the offset of the center pixel of the window. Secondly, traditional SAR pixel offset tracking methods typically use the same registration window size for offset estimation within the same image pair, relying on subjective experience. While larger registration windows can achieve higher accuracy in homogeneous motion regions, they also increase the likelihood of heterogeneous motion characteristics in pixels within the window. In heterogeneous motion regions, this can lead to unstable or overly smoothed estimation results. To ensure accurate registration, smaller registration windows should be used for offset estimation to preserve deformation details. However, within the same image pair, the degree of motion heterogeneity varies across different surface regions, and the noise levels of SAR images also differ. Therefore, it is not feasible to simply use the same registration window size for offset estimation across all regions. Instead, an optimal registration window size must be adaptively adopted based on the degree of motion heterogeneity and noise levels in different regions.

[0004] In summary, existing SAR pixel offset tracking methods based on the same regular rectangular registration window are difficult to accurately estimate two-dimensional surface deformation in complex regions. Therefore, adaptive optimization of the shape and size of the registration window in the SAR pixel offset tracking method is needed. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a SAR pixel offset tracking deformation estimation method based on an iterative adaptive registration window. This method achieves adaptive optimization of the optimal shape and size of the offset estimation registration window, thereby improving the accuracy of deformation estimation while preserving detailed deformation information.

[0006] Another objective of this invention is to provide a SAR pixel offset tracking deformation estimation system based on an iterative adaptive registration window.

[0007] The technical solution adopted in this invention is a SAR pixel offset tracking deformation estimation method based on iterative adaptive window, comprising the following steps:

[0008] Step S1: Acquire two SAR images of the target area at different times and preprocess the SAR images;

[0009] Step S2: Construct an elliptical registration window based on the resolution ratio of the range direction to the azimuth direction of the SAR image;

[0010] Step S3: Calculate the similarity plane of the master-slave image registration window based on the amplitude information using maximum likelihood similarity;

[0011] Step S4: Construct a robustness index for offset estimation by the ratio of the range to the mean absolute deviation of the similarity plane;

[0012] Step S5: Determine if the stability index has reached the threshold. If not, use an elliptical registration window as the initial window and expand in multiple directions. Calculate the similarity between the master and slave images for each expanded candidate window. Select the registration window with the highest similarity as the optimal expansion direction and use it as the current registration iteration template until the robustness index of the current registration window reaches the threshold. Then, obtain the maximum position of the two-dimensional similarity plane of the optimal registration window, thereby determining the azimuth and range offsets of this position relative to the center of the master image, i.e., the two-dimensional pixel offset. This method does not rely on prior deformation range information but iterates itself during the offset calculation process, enabling it to adaptively generate windows with optimal shapes and sizes.

[0013] Furthermore, step S1 includes the following steps:

[0014] Step S11: Decompress the acquired raw SAR data, convert the data format, and correct the data.

[0015] Step S12: The SAR data files are stitched and mosaicked to form a complete dataset of Earth's surface cover.

[0016] Step S13: SAR data is filtered to reduce the impact of speckle noise.

[0017] Furthermore, in step S2, the major axis a of the elliptical registration window is parallel to the image distance direction, and the minor axis b is parallel to the image azimuth direction, where a:b=P rng :P azi P rng This represents the range resolution of the SAR image. This represents the azimuth resolution of the SAR image.

[0018] Furthermore, step S3 includes the following steps:

[0019] Within a fixed search area in the image, the registration window moves by one pixel in both the range and azimuth directions. Based on the amplitude information, the similarity index between the master and slave images at that location is calculated using the maximum likelihood estimation criterion. This process is repeated throughout the entire search window to obtain a similarity plane.

[0020] Furthermore, the method for calculating the similarity index between the master and slave images is as follows:

[0021] (1)

[0022] in, This represents the maximum likelihood similarity index. It is the pixel amplitude value within the main image registration window. It is the pixel amplitude value within the image registration window. The number of pixels in the registration window; i and j represent the azimuth and range numbers of the corresponding pixels in the registration window, respectively.

[0023] Furthermore, step S4 includes the following steps:

[0024] Step S41: Calculation of the range R of the similarity plane:

[0025] (3)

[0026] Step S42: Calculation of the mean absolute deviation of the similarity plane:

[0027] (4)

[0028] Step S43: Construct a robustness metric for offset estimation Q:

[0029] (5)

[0030] in, for Location similarity index .

[0031] Furthermore, the aforementioned Method for determining:

[0032] First, determine the maximum range and azimuth deformation values ​​of the target area. Divide the maximum range deformation value by the image range resolution to obtain the maximum range offset, and divide the maximum azimuth deformation value by the image azimuth resolution to obtain the maximum azimuth offset. Set the initial values ​​using these maximum range and azimuth offsets. .

[0033] Furthermore, in step S5, the angular iteration step size for multi-directional expansion using an elliptical registration window as the initial window is... to The radial iteration step size ranges from 1 to 32 pixels, balancing computational complexity and accuracy.

[0034] A SAR pixel offset tracking deformation estimation system based on iterative adaptive window, employing the aforementioned SAR pixel offset tracking deformation estimation method based on iterative adaptive window, includes:

[0035] The SAR image preprocessing module is used to preprocess two SAR images of the target area at different times to obtain master and slave images.

[0036] The window similarity calculation module is used to calculate the similarity plane of the master-slave image registration window based on the magnitude information through maximum likelihood similarity.

[0037] The robustness index construction module is used to construct a robustness index for offset estimation by the ratio of the range to the mean absolute deviation of the similarity plane.

[0038] The adaptive expansion module is used to expand in multiple directions with an elliptical registration window as the initial window. It calculates the similarity between the master and slave images of each expanded candidate window, takes the registration window with the highest similarity as the optimal expansion direction, and determines the registration window with the optimal expansion direction as the current registration iteration template until the robustness index of the current registration window reaches the threshold, and obtains the optimal registration window.

[0039] The output module obtains the position of the maximum value of the two-dimensional similarity plane of the optimal registration window, and then outputs the two-dimensional pixel offset.

[0040] The beneficial effects of this invention are:

[0041] 1. This invention designs an elliptical window that can adaptively expand in multiple directions. By controlling the angular iteration step size and the radial iteration step size, the registration window can adaptively approach any shape. Furthermore, through iterative optimization, the registration window is adaptively optimized during the offset estimation process. This ensures that, within the same image pair, the optimal registration window size and shape are adaptively adopted based on the degree of motion heterogeneity and noise level in different surface regions. This maximizes the accuracy and resolution of deformation estimation results for different regions, especially in motion-suppressed areas, providing crucial support for the analysis and interpretation of geological hazards.

[0042] 2. Unlike the traditional normalized cross-correlation coefficient method, this invention uses SAR amplitude imagery as input information and introduces a maximum likelihood estimation criterion to calculate the similarity between two registration windows.

[0043] 3. This invention proposes a robustness index for offset estimation in deformation monitoring of geological hazards such as landslides under complex environments. The ratio of the range (R) to the mean absolute deviation (MAD) of similarity planes is used to measure the registration robustness during the offset estimation process. Based on a preset robustness index threshold, the size and shape of the registration window are iteratively changed to achieve optimal registration window estimation for 2D offset deformation estimation in different regions. By iteratively obtaining the optimal size and shape of the registration window for SAR offset estimation in different surface regions, a high-precision 2D deformation field of the surface can be obtained, providing important technical support for deformation monitoring of geological hazards such as landslides. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of the estimation method according to an embodiment of the present invention.

[0046] Figure 2 This is a schematic diagram illustrating the principle of elliptical registration window and offset estimation in an embodiment of the present invention.

[0047] Figure 3 This is a schematic diagram of the iterative adaptive registration window in the motion suppression region according to an embodiment of the present invention.

[0048] Figure 4 This is a schematic diagram of the master-slave SAR image amplitude of the Baige landslide area according to an embodiment of the present invention.

[0049] Figure 5 This is a schematic diagram of the radar azimuth and range offset estimation results of the Baige landslide obtained in an embodiment of the present invention; where (a) represents the range estimation result and (b) represents the azimuth estimation result.

[0050] Figure 6 This is a schematic diagram of the estimation results of the Baige landslide offset under different window sizes using the traditional normalized cross-correlation coefficient method; where (a) represents the distance estimation result of the 32×32 window, (b) represents the distance estimation result of the 64×64 window, (c) represents the distance estimation result of the 128×128 window, (d) represents the azimuth estimation result of the 32×32 window, (e) represents the azimuth estimation result of the 64×64 window, and (f) represents the azimuth estimation result of the 128×128 window.

[0051] Figure 7This is a schematic diagram of the estimated offset of the Baige landslide obtained by prior art 1; where (a) represents the distance estimation result and (b) represents the azimuth estimation result. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0053] Basic concept of the embodiments of the present invention:

[0054] Traditional offset tracking methods based on regular rectangular registration windows fail to consider the differences in motion attributes across different surface regions and image noise levels. Relying on subjective human experience, they use the similarity of rectangular registration windows of the same size to represent the similarity of central pixels, resulting in insufficient accuracy and stability of offset estimation results. In contrast, this invention constructs a robustness index for offset estimation using the ratio of the range to the mean difference of the similarity plane. Through iterative optimization, the registration window is adaptively adjusted during the offset estimation process, employing the optimal registration window size and shape based on the degree of motion heterogeneity and noise levels in different surface regions. This improves the accuracy of deformation estimation results in SAR pixel offset tracking technology.

[0055] Example 1,

[0056] A SAR pixel offset tracking deformation estimation method based on iterative adaptive window, such as Figure 1 As shown, it includes the following steps:

[0057] Step S1: Acquire two SAR images of the target area at different times, preprocess the SAR images, and define the earlier image as the master image and the later image as the slave image.

[0058] Step S11: After acquiring the raw SAR data, steps including data decompression, data format conversion, and data correction are required. Data decompression involves uncompressing the acquired data file to obtain the raw data. Data format conversion converts the raw data into a commonly used data format, such as GeoTIFF. Data correction corrects the raw data to eliminate the influence of instrument and atmospheric factors on the data.

[0059] Step S12: Sentinel-1 data (specific SAR data provided by the Sentinel-1 satellite) is typically distributed across different regions in the form of multiple files. To facilitate subsequent analysis, these data files need to be stitched and mosaicked. Data stitching merges multiple data files into a single file, while data mosaicking stitches together data from different regions to form a complete dataset of Earth's surface cover.

[0060] Step S13: SAR data is affected by speckle noise, which makes SAR image interpretation difficult and is not conducive to the similarity between master and slave images. Therefore, SAR data needs to be filtered to reduce the influence of speckle noise. In this embodiment, refined Lee filtering is used to filter and reduce noise in SAR images.

[0061] Step S2: Initial registration window generation. An elliptical registration window is constructed based on the resolution ratio of the range direction to the azimuth direction of the SAR image.

[0062] Unlike conventional pixel offset tracking methods that use a regular rectangular registration window, this invention uses an elliptical registration window as the initial registration window. The main purpose is to ensure that the registration window can expand and increase in all directions during subsequent iterations. The ratio of the major axis a to the minor axis b of the elliptical initial registration window changes from the SAR image range to the resolution P. rng and azimuth resolution P azi Decision, such as Figure 2 As shown. Assuming the major axis a is parallel to the image range direction and the minor axis b is parallel to the image azimuth direction, then a:b = P rng :P azi If the range resolution and azimuth resolution of a SAR image are close, then a:b=P rng :P azi If the ratio is approximately 1:1, then the initial registration window is circular.

[0063] Step S3: Evaluate the similarity plane of the master-slave image registration window using the maximum likelihood similarity estimation criterion;

[0064] Unlike conventional pixel offset tracking methods that use SAR intensity images to calculate the similarity of two registration windows based on the normalized cross-correlation coefficient method, this embodiment of the invention uses SAR image amplitude as input information and introduces a maximum likelihood estimation criterion to calculate the similarity of two registration windows. This effectively suppresses the influence of intensity speckle noise caused by traditional intensity normalized cross-correlation coefficient methods. The salt-and-pepper noise level caused by amplitude is significantly lower than that of intensity images. This embodiment of the invention can bring more robust offset estimation results and improve the accuracy of offset estimation. The calculation formula is as shown in equation (1):

[0065] (1)

[0066] in, This represents the maximum likelihood similarity index (ML is an abbreviation for maximum likelihood). It is the pixel amplitude value within the main image registration window. These are the pixel amplitude values ​​within the image registration window, and they possess statistical independence. The number of pixels in the registration window (the number of pixels in the registration window of the master image and the slave image is equal), i and j represent the azimuth and range numbers of the corresponding pixels in the registration window, respectively.

[0067] Removing the exponents from both sides of equation (1), it can be simplified to:

[0068] (2)

[0069] Window similarity calculation based on amplitude information can effectively suppress the influence of intensity speckle noise caused by traditional intensity-normalized cross-correlation coefficients, thereby improving the accuracy of offset estimation.

[0070] Within a fixed search area in the image (the size of this area should not be less than the maximum offset of the target area, usually set subjectively based on prior information about the target area or the actual situation), the registration window can calculate the conditional probability for each pixel movement in the range and azimuth directions. , for Location similarity index ; Traversal The size of the search window can be obtained The conditional probability plane of size, also known as the similarity plane, is where the position with the highest conditional probability is the position of the point to be registered as a corresponding point in the image.

[0071] In practical applications, the maximum deformation values ​​in the range and azimuth directions of the target area are first determined. The maximum range deformation value is divided by the image range resolution to obtain the maximum range offset, and the maximum azimuth deformation value is divided by the image azimuth resolution to obtain the maximum azimuth offset. The initial values ​​are then set using these maximum range and azimuth offsets. .

[0072] Step S4: In order to evaluate the quality of offset registration and use it as an evaluation index for subsequent window iteration optimization, this embodiment of the invention uses the ratio of the range (R) of the similarity plane to the mean absolute deviation (MAD) to construct a robustness index Q for offset estimation.

[0073] Step S41: The range R of the similarity plane is In the similarity plane The difference between the maximum and minimum values, R. A larger R value indicates a more significant difference between the maximum and minimum values ​​in the similarity plane, and a more robust evaluation. The formula for calculating R is as follows:

[0074] (3)

[0075] Step S42: The mean absolute deviation (MAD) of the similarity plane is the arithmetic mean of the absolute values ​​of the deviations of the similarity values ​​at each location from their arithmetic mean. It comprehensively reflects the similarity in the plane. The smaller the mean absolute deviation (MAD), the smaller the fluctuation of the similarity plane, meaning less fluctuation caused by image noise, and the more robust the assessment. Its calculation formula is as follows:

[0076] (4)

[0077] Step S43: Finally, considering both the range R and the mean absolute deviation (MAD) of the similarity plane, a robustness index Q for offset estimation is constructed, which is the ratio of the range R to the mean absolute deviation (MAD). The calculation formula is as follows:

[0078] (5)

[0079] The constructed robustness index Q can cleverly ensure that the similarity peak is obvious while pursuing the stability of similarity estimation, thus ensuring the registration accuracy. It can effectively evaluate the stability of the offset estimation process and provide a quantitative index for the stability measurement of the offset estimation window optimization iteration.

[0080] Step S5: Determine whether the stability index Q has reached the threshold;

[0081] To evaluate the effectiveness of the registration window iterative optimization, a stability index threshold is given. Stability index threshold The higher the value, the higher the accuracy of the offset estimation, but the more easily the details of the deformation result are blurred, and vice versa. Based on the stability index Q calculated in step S43, it is determined whether it has reached a threshold. If it has reached the threshold, the iteration stops, proceeds to step S8, and the final offset estimation result is output; otherwise, the stability index threshold is not reached. If so, it is necessary to expand the size of the offset registration window, adjust the shape of the offset registration window, and proceed to step S6.

[0082] Step S6: Expand the registration window in multiple directions;

[0083] The stability index threshold was not reached under the pre-registration window conditions. In such cases, the registration window needs to be further enlarged. To adaptively increase the registration window in all directions, unlike traditional methods, this invention uses an elliptical registration window as the initial window, such as... Figure 3 As shown, the iteration step size is determined by a manually set angle. With radial iteration step size To control the growth and expansion size of the window in each iteration, theoretically, a sufficiently small angular iteration step size can be used. With radial iteration step size It allows the registration window to adaptively approximate any shape.

[0084] like Figure 3 As shown in the example, suppose there are deformed regions and stable regions in the SAR image, and their motion characteristics are significantly different. The stability index threshold is not reached in the initial registration window. In this case, let the angle iteration step size be... Then the registration window can be freely expanded in six directions with radial iteration step size. For each 8 pixels, six new registration windows are obtained. Then, proceed to step S3 to calculate the similarity between the master and slave image registration windows under each of the six window conditions. Angle iteration step size. and radial iteration step size The angle iteration step size can be set according to accuracy requirements. The smaller the size, the more directions the registration window can expand in; radial iteration step size. The smaller the value, the more refined and slower the iteration; the minimum should not be less than one pixel. To balance computational complexity and accuracy, the angle iteration step size is... to The radial iteration step size ranges from 1 to 32 pixels.

[0085] Step S7: Select the registration window expansion direction;

[0086] by Figure 3 Taking the example shown, after obtaining six new registration windows by freely expanding in six directions in step S6, it is necessary to select the optimal expansion direction from the six directions. Through step S3, the similarity of the master-slave image registration windows under the six window conditions can be obtained. The direction with the highest similarity is the optimal expansion direction, thus determining the registration window in this expansion direction as the new registration iteration template. Then, step S5 is entered to calculate the registration stability index and determine whether the stability index threshold has been reached. .

[0087] Repeat steps S5-S7 until the stability index threshold is reached. End the iteration. (e.g.) Figure 3As shown, based on the distribution characteristics of the light yellow deformation area, the initial registration window will gradually expand adaptively from ① to ⑧, thereby adaptively determining the size and shape of the optimal registration window, automatically fitting the real deformation shape, ensuring the detailed information of the deformation result under the premise of accurate offset estimation, and improving the accuracy and stability of offset estimation results in complex motion scenes.

[0088] Step S8: Output the offset result;

[0089] The registration window reaches the stability index threshold. After the iteration ends, the two-dimensional similarity plane of the optimal registration window can be obtained. The position of the maximum value of the two-dimensional similarity plane corresponds to the optimal matching position of two points with the same name, which is the final two-dimensional offset result.

[0090] Example 2,

[0091] A SAR pixel offset tracking deformation estimation system based on iterative adaptive window, comprising:

[0092] The SAR image preprocessing module is used to preprocess two SAR images of the target area at different times to obtain master and slave images.

[0093] The window similarity calculation module is used to calculate the similarity plane of the master-slave image registration window based on the magnitude information through maximum likelihood similarity.

[0094] The robustness index construction module is used to construct a robustness index for offset estimation by the ratio of the range to the mean absolute deviation of the similarity plane.

[0095] The adaptive expansion module is used to expand in multiple directions with an elliptical registration window as the initial window. It calculates the similarity between the master and slave images of each expanded candidate window, takes the registration window with the highest similarity as the optimal expansion direction, and determines the registration window with the optimal expansion direction as the current registration iteration template until the robustness index of the current registration window reaches the threshold, and obtains the optimal registration window.

[0096] The output module obtains the position of the maximum value of the two-dimensional similarity plane of the optimal registration window, and then outputs the two-dimensional pixel offset.

[0097] Experimental verification:

[0098] To further verify the embodiments of the present invention, the Baige landslide, which occurred near Baiyu County, Ganzi Prefecture, Sichuan Province, and Baige Village, Boluo Township, Jiangda County, Changdu City, Tibet, was selected as the verification area. This landslide caused two collapses in October and December 2018, blocking the Jinsha River. The experimental data used two ALOS-2 PALSAR-2 images from July 24, 2017, and December 27, 2017, as the main image and the secondary image, respectively. The SAR image incident angle was 36.27°, and the spatial sampling rate was 3.78 m × 4.29 m (azimuth × range). After step S1: SAR image preprocessing, the amplitude images of the main image and the secondary image were obtained as follows: Figure 4 As shown.

[0099] Because the sliding direction of the Baige landslide is almost parallel to the range direction in the PALSAR-2 image, the range deformation of the landslide is greater than the azimuth deformation. Figure 5 As shown in Figures (a) and (b), during the nearly four months of SAR image coverage, the maximum range deformation of the upper part of the Baige landslide was approximately 1.8 pixels. The SAR image spatial sampling rate was 3.78 m × 4.29 m (azimuth × range), therefore the maximum range deformation was approximately 7.5 meters, while the azimuth deformation was relatively smaller, with a maximum offset of approximately 0.8 pixels, translating to a maximum deformation of approximately 3 meters. Figure 5 The offset estimation results show that the edges between the deformed and non-deformed regions are clear, and the signal of the rapid deformation zone inside the deformed region is obvious. The embodiments of the present invention can effectively capture the detailed information of the deformed region.

[0100] Figure 6 Figures (a)-(f) show the offset estimation results of the traditional method with registration window sizes (azimuth × range) of 32×32, 64×64, and 128×128, respectively. It can be observed that under the small registration window condition (32×32), the offset estimation results contain a large number of patch errors, and even regions that should be stable show numerous incorrect estimation results. Figure 6 Under a 64×64 registration window, registration errors were significantly reduced, and the two rapidly deformed areas in the landslide region were reconciled; however, some noise still remained in the surrounding stable areas. As the registration window was further increased, in... Figure 6 Under the 128×128 registration window condition, error patches are now rarely seen, but the boundaries of deformed regions are blurred, and two rapidly deformed regions in the slope area are smoothed. Compared with the normalized cross-correlation coefficient based on the traditional regular rectangular registration window, Figure 5The method of the present invention, as demonstrated, can adaptively select the optimal registration window of appropriate size and shape. The offset estimation results are clear and stable with virtually no noise from incorrect matching. At the same time, the deformation details and contours of the landslide deformation area are clear. This shows that compared with the traditional method that uses a matching window of fixed shape and size, the present invention can adaptively select the optimal size and shape of the registration window, which can not only improve the accuracy of the deformation estimation of the target area, but also improve the ability to capture deformation detail information.

[0101] The input to this invention is SAR imagery. Due to the influence of speckle noise in SAR images, conventional offset tracking methods estimate the statistical similarity of two matching windows instead of the similarity of two corresponding center pixels. This matching strategy is feasible in homogeneous motion regions (i.e., the local motion is basically consistent), but in heterogeneous motion regions (i.e., there are large differences in local motion), pixels within the matching window often have different motion characteristics (stable or moving, fast or slow). Typical heterogeneous motion regions include landslide boundary areas, local collapse areas of landslide bodies, and other areas with large gradient changes. A larger matching window (i.e., more statistical information) is needed to increase the accuracy of similarity matching. However, while increasing the matching window increases matching accuracy, it also blurs the details of deformation results and reduces the resolution. As the matching window increases, the possibility of heterogeneous motion characteristics of pixels within the matching window also increases accordingly. Large matching windows can obtain high accuracy in homogeneous motion regions, but in heterogeneous motion regions, they can lead to unstable or overly smoothed estimation results. This is why offset tracking methods have a trade-off between estimation accuracy and resolution when choosing the matching window size.

[0102] Conventional offset tracking methods use fixed rectangular windows, inevitably introducing pixels with different motion attributes within the window. Existing technique 1 (An Adaptive Offset Tracking Method with SAR Images for Landslide Displacement Monitoring, Jiehua Cai, Changcheng Wang, Xiaokang Mao and Qijie Wang, 2017) first calculates the approximate deformation range using a traditional regular window, and then uses a fixed-size but shape-adaptive matching window based on this deformation range. This approach relies on traditional methods to provide the deformation range, yet attempts to improve upon the inaccuracy of deformation range estimation in traditional methods, creating a chicken-and-egg dilemma. The similarity calculation method used is still based on the normalized cross-correlation coefficient method using intensity information.

[0103] This invention employs an elliptical window that can adaptively expand in multiple directions, ensuring proportional growth in each direction. The optimal expansion direction is the registration window with the highest similarity, thus avoiding the introduction of pixels with different motion attributes within the window. This ensures that the similarity of the window is equivalent to the similarity of the central pixel, improving the accuracy of deformation estimation. A stability index is also designed. Using an iterative approach, the window size is increased from small to large, stopping at a threshold. It does not rely on prior deformation range information but iterates itself during offset calculation, adaptively generating not only the optimal window shape but also the optimal window size. It can adaptively select appropriate window sizes in different regions, ensuring accurate registration while using the smallest window to meet the needs of different scenes and images. This adaptive selection of optimal window size and shape improves deformation estimation accuracy while preserving deformation details. The maximum likelihood estimation method based on amplitude information effectively suppresses the influence of intensity speckle noise caused by traditional intensity-normalized cross-correlation coefficients, further improving the accuracy of offset estimation.

[0104] Figure 7 Figures (a) and (b) show the estimated offset of the Baige landslide obtained by prior art 1, compared with... Figure 5 A comparison of the offset estimation results obtained in the embodiments of the present invention reveals that, although prior art 1 obtains a clear landslide boundary in the range offset estimation results, the rapidly deforming region in the middle of the landslide is smoothed and blurred. This is because prior art 1 only considers the adaptation of the matching window shape, but does not consider the optimal selection of the matching window size. In the azimuth offset estimation results, the deformed region in the prior art 1 estimation results is completely smoothed, which is due to the erroneous estimation caused by not considering the optimal selection of the matching window size. In addition, in the estimation results of prior art 1, there are still a few speckled and patchy erroneous noise results in the stable region. This is partly because the ANCC method uses the normalized cross-correlation coefficient of intensity information as the similarity estimation criterion, which is greatly affected by SAR image noise. It also lacks a stability index for offset estimation as an auxiliary evaluation of the accuracy of the results, thus potentially leading to localized errors.

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A SAR pixel offset tracking deformation estimation method based on iterative adaptive window, characterized in that, Includes the following steps: Step S1: Acquire two SAR images of the target area at different times and preprocess the SAR images; Step S2: Construct an elliptical registration window based on the resolution ratio of the range direction to the azimuth direction of the SAR image; Step S3: Calculate the similarity plane of the master-slave image registration window based on the amplitude information using maximum likelihood similarity; Step S4: Construct a robustness index for offset estimation by the ratio of the range to the mean absolute deviation of the similarity plane; Step S5: Determine whether the stability index has reached the threshold; if the stability index threshold has not been reached, use the elliptical registration window as the initial window to expand in multiple directions, calculate the similarity between the master and slave images of each expanded candidate window, take the registration window with the highest similarity as the optimal expansion direction, determine the registration window with the optimal expansion direction as the current registration iteration template, until the robustness index of the current registration window reaches the threshold; then obtain the maximum value position of the two-dimensional similarity plane of the optimal registration window, thereby determining the two-dimensional pixel offset.

2. The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 1, characterized in that, Step S1 includes the following steps: Step S11: Decompress the acquired raw SAR data, convert the data format, and correct the data. Step S12: The SAR data files are stitched and mosaicked to form a complete dataset of Earth's surface cover. Step S13: SAR data is filtered to reduce the impact of speckle noise.

3. The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 1, characterized in that, In step S2, the major axis a of the elliptical registration window is parallel to the image distance direction, and the minor axis b is parallel to the image azimuth direction, where a:b=P rng :P azi P rng This represents the range resolution of the SAR image. This represents the azimuth resolution of the SAR image.

4. The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 1, characterized in that, Step S3 includes the following steps: Within a fixed search area in the image, the registration window moves by one pixel in both the range and azimuth directions. Based on the amplitude information, the similarity index between the master and slave images at that location is calculated using the maximum likelihood estimation criterion. This process is repeated throughout the entire search window to obtain a similarity plane.

5. The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 4, characterized in that, The method for calculating the similarity index between the master and slave images is as follows: （1） in, This represents the maximum likelihood similarity index. It is the pixel amplitude value within the main image registration window. It is the pixel amplitude value within the image registration window. The number of pixels in the registration window; i and j represent the azimuth and range numbers of the corresponding pixels in the registration window, respectively.

6. The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 1, characterized in that, Step S4 includes the following steps: Step S41: Calculation of the range R of the similarity plane: （3） Step S42: Calculation of the mean absolute deviation of the similarity plane: （4） Step S43: Construct a robustness metric for offset estimation Q: （5） in, for Location similarity index .

7. The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 6, characterized in that, The Method for determining: First, determine the maximum range and azimuth deformation values ​​of the target area. Divide the maximum range deformation value by the image range resolution to obtain the maximum range offset, and divide the maximum azimuth deformation value by the image azimuth resolution to obtain the maximum azimuth offset. Set the initial values ​​using these maximum range and azimuth offsets. .

8. The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 1, characterized in that, In step S5, the elliptical registration window is used as the initial window, and the angular iteration step size for multi-directional expansion is... to The radial iteration step size ranges from 1 to 32 pixels.

9. A SAR pixel offset tracking deformation estimation system based on iterative adaptive window, characterized in that, The SAR pixel offset tracking deformation estimation method based on iterative adaptive window as described in claim 1 includes: The SAR image preprocessing module is used to preprocess two SAR images of the target area at different times to obtain master and slave images. The window similarity calculation module is used to calculate the similarity plane of the master-slave image registration window based on the magnitude information through maximum likelihood similarity. The robustness index construction module is used to construct a robustness index for offset estimation by the ratio of the range to the mean absolute deviation of the similarity plane. The adaptive expansion module is used to expand in multiple directions with an elliptical registration window as the initial window. It calculates the similarity between the master and slave images of each expanded candidate window, takes the registration window with the highest similarity as the optimal expansion direction, and determines the registration window with the optimal expansion direction as the current registration iteration template until the robustness index of the current registration window reaches the threshold, and obtains the optimal registration window. The output module obtains the position of the maximum value of the two-dimensional similarity plane of the optimal registration window, and then outputs the two-dimensional pixel offset.

Images