An InSAR deformation monitoring method for a dam

Abstract

The application discloses a kind of InSAR deformation monitoring methods of dam, it is related to image processing technical field.The method includes: the multi-temporal SAR image of dam is separated and phase separation by amplitude, obtains multi-temporal amplitude image and multi-temporal phase image;Difference ratio is calculated, and the amplitude difference ratio graph and phase difference ratio graph of multi-temporal are obtained;Mark the significant point in amplitude difference ratio graph and phase difference ratio graph;The amplitude diffusion distribution graph and phase diffusion distribution graph of multi-temporal are constructed;The amplitude diffusion distribution graph and phase diffusion distribution graph of the multi-temporal are accumulated and mapped respectively, and amplitude activity graph and phase activity graph are obtained;According to the amplitude image and phase image of first and last two time phases, obtain whole time period linkage binary image;Full time period linkage binary image, amplitude activity graph and phase activity graph are processed using double-path attention fusion neural network, and dam deformation result is obtained.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to an InSAR deformation monitoring method for dams. Background Technology

[0002] Synthetic Aperture Radar Interferometry (InSAR) technology, with its advantages of all-weather, all-time, high precision, and large-scale monitoring, has been widely used in the field of deformation monitoring for major infrastructures such as dams. Traditional InSAR deformation monitoring methods for dams are usually based on techniques such as Permanent Scatterer (PS-InSAR) or Small Baseline Set (SBAS-InSAR), which extract the slow deformation field on the dam surface by analyzing the phase information of stable scatterers in long-term SAR images.

[0003] In existing technologies, a typical monitoring method involves a series of complex processes, including registration of multi-temporal SAR images, interferogram generation, phase unwrapping, and atmospheric phase correction, to ultimately obtain the deformation time series of the dam along the radar line of sight. These methods typically assume that the dam surface maintains stable scattering characteristics during the monitoring period and rely on accurate modeling and calculation of phase information.

[0004] However, existing methods generally suffer from low accuracy in deformation monitoring in practical applications. Summary of the Invention

[0005] To address the low accuracy of existing methods for monitoring dam deformation, this invention aims to provide an InSAR deformation monitoring method for dams. The specific technical solution adopted is as follows: In a first aspect, embodiments of the present invention provide an InSAR deformation monitoring method for dams, wherein the InSAR deformation monitoring method includes: performing amplitude separation and phase separation on multi-temporal SAR images of the dam to obtain multi-temporal amplitude images and multi-temporal phase images; calculating the difference ratio of amplitude images and phase images of adjacent temporal phases respectively to obtain multi-temporal amplitude difference ratio maps and phase difference ratio maps; marking salient points in the amplitude difference ratio maps and phase difference ratio maps; constructing multi-temporal amplitude diffusion distribution maps and phase diffusion distribution maps based on the number of salient points within a local range in the amplitude difference ratio maps and phase difference ratio maps of adjacent temporal phases; accumulating and mapping the multi-temporal amplitude diffusion distribution maps and phase diffusion distribution maps respectively to obtain amplitude activity maps and phase activity maps; obtaining a full-time linked binary image based on the amplitude images and phase images of the first and last two temporal phases; and processing the full-time linked binary image, amplitude activity map, and phase activity map using a dual-path attention fusion neural network to obtain dam deformation results.

[0006] Optionally, the step of calculating the difference ratios of amplitude images and phase images of adjacent time phases to obtain a multi-time phase amplitude difference ratio map and a phase difference ratio map includes: using the ratio of the amplitude difference of corresponding pixels in the amplitude images of adjacent time phases to the amplitude of the same pixel in the previous time phase as the amplitude difference ratio to obtain a multi-time phase amplitude difference ratio map; and using the ratio of the phase difference of corresponding pixels in the phase images of adjacent time phases to the phase of the same pixel in the previous time phase as the phase difference ratio to obtain a multi-time phase phase difference ratio map.

[0007] Optionally, marking significant points in the amplitude difference ratio map and the phase difference ratio map includes: selecting a pixel in the amplitude difference ratio map and calculating the absolute value of the amplitude difference ratio corresponding to the pixel; when the absolute value is greater than the amplitude difference ratio threshold, marking the selected pixel as a significant amplitude point; selecting a pixel in the phase difference ratio map and calculating the absolute value of the phase difference ratio corresponding to the pixel; when the absolute value is greater than the phase difference ratio threshold, marking the selected pixel as a significant phase point.

[0008] Optionally, the step of constructing a multi-temporal amplitude diffusion distribution map and a phase diffusion distribution map based on the number of significant points within a local range in the amplitude difference ratio map and the phase difference ratio map of adjacent temporal phases includes: performing isolated detection on significant amplitude points in the amplitude difference ratio map of each temporal phase and discarding significant amplitude points whose detection results are isolated; taking each remaining significant amplitude point as the center, counting the number of significant amplitude points within different scale windows to obtain the density of each remaining significant amplitude point at each scale; obtaining the amplitude diffusion value of each significant amplitude point based on the ratio of the density of the same significant amplitude points at the same scale in the amplitude difference ratio map of adjacent temporal phases; and setting the amplitude diffusion value of non-significant amplitude points to 0 for each significant amplitude point to obtain the amplitude diffusion distribution map at the corresponding time. For each phase difference ratio map, isolated phase salient points are detected and discarded. The number of phase salient points within different scale windows is counted using each remaining phase salient point as the center, yielding the density of each remaining phase salient point at each scale. The phase diffusion value of each phase salient point is obtained based on the ratio of the densities of the same phase salient points at the same scale in the phase difference ratio maps of adjacent phases. For each phase salient point, the phase diffusion value of non-phase salient points is set to 0, resulting in the phase diffusion distribution map for the corresponding time point.

[0009] Optionally, obtaining the amplitude diffusion value of each significant amplitude point based on the ratio of the densities of significant amplitude points at the same scale in the amplitude difference ratio diagram of adjacent time phases includes: averaging the ratios of densities at different scales and subtracting a preset ratio parameter to obtain the basic amplitude diffusion value of the significant amplitude point, wherein the preset ratio parameter is 1; setting the basic amplitude diffusion value that is less than or equal to the preset amplitude diffusion parameter to 0 to obtain the amplitude diffusion value of the significant amplitude point; wherein the preset amplitude diffusion parameter is 0. The step of obtaining the phase diffusion value of each phase salient point based on the ratio of the density of the same phase salient points at the same scale in the phase difference ratio diagram of adjacent time phases includes: averaging the ratios of the densities at different scales and subtracting a preset ratio parameter to obtain the basic phase diffusion value of the phase salient point, wherein the preset ratio parameter is 1; setting the basic phase diffusion value that is less than or equal to the preset phase diffusion parameter to 0 to obtain the phase diffusion value of the phase salient point; wherein the preset phase diffusion parameter is 0.

[0010] Optionally, the step of accumulating and mapping the amplitude diffusion distribution map and phase diffusion distribution map of the multi-temporal phases to obtain the amplitude activity map and phase activity map includes: adding the amplitude diffusion values ​​of the same pixels in the amplitude diffusion distribution map of the multi-temporal phases to obtain the cumulative amplitude diffusion value; performing exponential normalization mapping on the cumulative amplitude diffusion value to obtain the amplitude activity map; adding the phase diffusion values ​​of the same pixels in the phase diffusion distribution map of the multi-temporal phases to obtain the cumulative phase diffusion value; and performing exponential normalization mapping on the cumulative phase diffusion value to obtain the phase activity map.

[0011] Optionally, obtaining a full-time linked binary image based on the amplitude and phase images of the first and last time phases includes: obtaining an amplitude difference ratio based on the amplitude images of the first and last time phases, and comparing the absolute value of the amplitude difference ratio with an amplitude difference ratio threshold to generate a full-time amplitude binary image; obtaining a phase difference ratio based on the phase images of the first and last time phases, and comparing the absolute value of the phase difference ratio with a phase difference ratio threshold to generate a full-time phase binary image; and performing a pixel-wise AND operation between the full-time amplitude binary image and the full-time phase binary image to obtain a full-time linked binary image.

[0012] Optionally, the dual-path attention fusion neural network includes: a first attention generation unit, a second attention generation unit, and an initial feature encoding unit arranged in parallel; a first feature enhancement unit connected to the first attention generation unit and the initial feature encoding unit respectively; a second feature enhancement unit connected to the second attention generation unit and the initial feature encoding unit respectively; a first deep encoding unit connected to the first feature enhancement unit; a second deep encoding unit connected to the second feature enhancement unit; an adder connected to the first deep encoding unit and the second deep encoding unit respectively; and a deformation decoding output unit connected to the adder.

[0013] Optionally, the first attention generation unit is used to generate a first attention weight based on the amplitude activity map; the second attention generation unit is used to generate a second attention weight based on the phase activity map; the initial feature encoding unit is used to perform global feature encoding on the full-time linked binary image; the first feature enhancement unit is used to apply the first attention weight to the global feature encoding and form a residual concatenation to obtain a first enhanced global feature; the second feature enhancement unit is used to apply the second attention weight to the global feature encoding and form a residual concatenation to obtain a second enhanced global feature; the first deep encoding unit is used to extract a first global deep feature from the first enhanced global feature; the second deep encoding unit is used to extract a second global deep feature from the second enhanced global feature; the adder is used to add the first global deep feature and the second global deep feature element-wise to obtain a global deep fusion feature; and the deformation decoding output unit is used to obtain the dam deformation result based on the global deep fusion feature.

[0014] Optionally, the expression for the first feature enhancement unit is formula (1): (1) In the above formula (1), The output features of the first feature enhancement unit, The output features of the initial feature encoding unit, For element-wise multiplication, To add elements together, The output weights of the first attention generation unit; The expression for the second feature enhancement unit is formula (2): (2) In the above formula (2), The output features of the second feature enhancement unit, The output weights for the second attention generation unit.

[0015] This invention addresses the problem of low accuracy in deformation monitoring commonly found in practical applications by proposing a joint processing method that combines image preprocessing and deep learning decision-making, ultimately yielding highly accurate dam deformation monitoring results. The specific technical means are as follows: First, the multi-temporal SAR images of the dam are preprocessed to perform amplitude separation and phase separation to obtain multi-temporal amplitude images and multi-temporal phase images; the difference ratio of the amplitude images and phase images of adjacent temporal phases is calculated to obtain multi-temporal amplitude difference ratio maps and phase difference ratio maps; significant points in the amplitude difference ratio maps and phase difference ratio maps are marked; based on the number of significant points in the local range of the amplitude difference ratio maps and phase difference ratio maps of adjacent temporal phases, multi-temporal amplitude diffusion distribution maps and phase diffusion distribution maps are constructed; the amplitude diffusion distribution maps and phase diffusion distribution maps of the multi-temporal phases are accumulated and mapped to obtain amplitude activity maps and phase activity maps; based on the amplitude images and phase images of the first and last two temporal phases, a full-time linked binary image is obtained; in the decision-making stage, a dual-path attention fusion neural network is used to process the full-time linked binary image, amplitude activity map, and phase activity map to obtain the dam deformation result. During the decision-making stage, the constructed dual-path attention fusion neural network structure, which includes attention generation unit, feature enhancement unit, deep encoding unit, adder and deformation decoding output unit, can adaptively highlight the deformation-sensitive area of ​​the dam, suppress decoherent noise and non-deformation interference, and significantly improve the accuracy of deformation monitoring.

[0016] Thus, the present invention has the following beneficial effects: This invention effectively captures the dynamic changes of local areas on the dam surface by separating the amplitude and phase of multi-temporal SAR images during the preprocessing stage, extracting the difference ratio maps of amplitude and phase respectively, and marking significant points. This reduces the phase unwrapping errors caused by changes in scattering characteristics in traditional methods.

[0017] This invention can accurately identify the local deformation region when significant non-uniform deformation occurs on the dam surface by constructing amplitude diffusion distribution map and phase diffusion distribution map, avoiding the influence of phase decoherence and decoherence error propagation that may occur during the phase unwrapping and deformation inversion process of stable scatterers.

[0018] This invention obtains amplitude activity maps and phase activity maps by accumulating amplitude diffusion distribution maps and phase diffusion distribution maps of multiple time phases. These maps are used to assess the spatial and temporal continuity of deformation over a continuous time range, accurately depicting the true deformation characteristics of the dam's local abrupt deformation area. This avoids the problem of inaccurate deformation estimation in traditional methods when facing local abrupt deformation.

[0019] This invention acquires a full-time linked binary image by using the amplitude and phase images of the first and last time phases. It combines a dual-path attention fusion neural network to fuse the full-time linked binary image, amplitude activity map, and phase activity map, realizing joint processing and attention weighting of amplitude change features, phase change features, and first and last time phase change features. This can adaptively highlight the deformation-sensitive area of ​​the dam, suppress phase-dissociated noise and non-deformation interference, and significantly improve the accuracy of deformation monitoring. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of protection. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 A flowchart of an InSAR deformation monitoring method for a dam, provided as an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of a dual-path attention fusion neural network structure provided in one embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the initial feature encoding unit in a dual-path attention fusion neural network structure provided in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of the first attention generation unit and the second attention generation unit in a dual-path attention fusion neural network structure provided in another embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram of the structure of the first deep coding unit and the second deep coding unit in a dual-path attention fusion neural network structure provided in another embodiment of the present invention.

[0026] Figure 6 This is a schematic diagram of the deformation decoding output unit in a dual-path attention fusion neural network structure provided in another embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. 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.

[0028] The main reason why existing technologies suffer from low deformation monitoring accuracy is that dams, as special hydraulic structures, are prone to spatiotemporal variations in their scattering characteristics due to factors such as water level changes, alternating wet and dry periods, and temperature stress at different times. When significant non-uniform deformation or surface changes occur in local areas of the dam, traditional methods based on stable scatterers for phase unwrapping and deformation inversion suffer from phase incoherence and unwrapping error propagation. This results in spatially discontinuous and temporally inconsistent deformation estimation results, making it difficult to accurately characterize the true deformation features of abrupt changes in local dam areas. Consequently, the reliability and completeness of the monitoring results are reduced, leading to low deformation monitoring accuracy.

[0029] The specific scheme of the InSAR deformation monitoring method for dams provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Please see Figure 1 The diagram illustrates a flowchart of an InSAR deformation monitoring method for dams according to an embodiment of the present invention, comprising: S1, perform amplitude and phase separation on the multi-temporal SAR images of the dam to obtain multi-temporal amplitude images and multi-temporal phase images; S2, calculate the difference ratio of the amplitude image and phase image of adjacent time phases respectively, and obtain the amplitude difference ratio map and phase difference ratio map of multiple time phases; S3, mark the significant points in the amplitude difference ratio diagram and the phase difference ratio diagram; S4. Based on the number of significant points in the local range of the amplitude difference ratio diagram and phase difference ratio diagram of adjacent time phases, construct the amplitude diffusion distribution diagram and phase diffusion distribution diagram of multiple time phases. S5, the amplitude diffusion distribution map and phase diffusion distribution map of the multi-time phase are accumulated and mapped respectively to obtain the amplitude activity map and phase activity map; S6. Based on the amplitude and phase images of the first and last time phases, a full-time linked binary image is obtained. S7 uses a dual-path attention fusion neural network to process the full-time linked binary image, amplitude activity map, and phase activity map to obtain the dam deformation results.

[0031] Specifically, for step S1, each pixel in the SAR image can be represented as: Where i is the pixel index; For the SAR image, the first The magnitude of each pixel It is a natural constant. For the SAR image, the first Phase of each pixel The unit is imaginary. This means that each pixel contains both amplitude and phase information. Extracting the amplitude information from all pixels in a SAR image creates the amplitude image for the corresponding time phase, and extracting the phase information from all pixels creates the phase image for the corresponding time phase. Following these steps, multi-temporal SAR images of the dam are processed to obtain multi-temporal amplitude and phase images. These images allow us to capture the amplitude and phase changes of the dam at different time points, providing a data foundation for subsequent deformation monitoring.

[0032] Next, the output of step S1 is preprocessed according to step S2 to obtain the amplitude difference ratio diagram and phase difference ratio diagram of multiple time phases. As an optional implementation of step S2, step S2 includes: S21, the ratio of the amplitude difference of the corresponding pixel in the amplitude image of the adjacent time phase to the amplitude of the pixel in the previous time phase is used as the amplitude difference ratio to obtain the amplitude difference ratio map of the multiple time phases. Specifically, the formula for calculating the amplitude difference ratio is:

[0033] In the above formula, t is the phase index and i is the pixel index. The amplitude difference ratio of the i-th pixel in the amplitude difference ratio diagram at time t is the amplitude difference ratio of the i-th pixel in the diagram. For the first The amplitude image of the time phase The magnitude of each pixel For the first The amplitude image of the time phase The amplitude of each pixel. The amplitude difference ratio of all pixel indices at a fixed time phase is arranged to obtain the amplitude difference ratio map of the fixed time phase. The amplitude difference ratio map of multiple time phases can be obtained by calculating it sequentially for all adjacent time phases.

[0034] S22, the ratio of the phase difference of the corresponding pixel in the phase image of the adjacent time phase to the phase of the same pixel in the previous time phase is used as the phase difference ratio to obtain the phase difference ratio map of multiple time phases.

[0035] Specifically, the formula for calculating the phase difference ratio is:

[0036] In the above formula, t is the phase index and i is the pixel index. Let be the phase difference ratio of the i-th pixel in the phase difference ratio diagram at time t. For the first The first phase image of the time phase Phase of each pixel For the first The first phase image of the time phase The phase of each pixel. The phase difference ratio of all pixel indices at a fixed time phase is arranged to obtain the phase difference ratio map of the fixed time phase. By calculating the phase difference ratio of all adjacent time phases in sequence, the phase difference ratio map of multiple time phases can be obtained.

[0037] This invention, by calculating the difference ratio of amplitude and phase, can accurately reflect the dynamic changes of the dam surface at different times, especially when abrupt deformation occurs locally, clearly identifying the deformed region from the stable region. This method effectively avoids errors caused by phase decoherence or unwrapping errors in traditional monitoring techniques, significantly improving monitoring accuracy. The construction of the difference ratio map helps to accurately identify local deformation regions when non-uniform deformation occurs on the dam surface, and effectively reduces the noise impact caused by environmental changes or non-deformation interference.

[0038] It should be noted that the SAR image acquisition strategy can be flexibly configured according to the actual monitoring needs of the dam. The acquisition strategy includes, but is not limited to, acquisition duration, acquisition interval, and the number of images acquired. For example, when monitoring short-term deformation of the dam (such as instantaneous displacement or short-term settlement), multi-temporal SAR images within one or two hours can be acquired, with the acquisition time interval between adjacent images set to several minutes to tens of minutes, and the number of images selected being 5 to 10. When monitoring long-term deformation, the acquisition duration and interval can be extended accordingly. Those skilled in the art can adjust the above parameters according to specific application scenarios, all of which fall within the scope of protection of this invention.

[0039] Based on the amplitude difference ratio diagram and the phase difference ratio diagram, significant points in the corresponding diagrams can be marked according to step S3. As an optional implementation of step S3, step S3 includes: S31, select a pixel in the amplitude difference ratio map and calculate the absolute value of the amplitude difference ratio corresponding to the pixel. When the absolute value is greater than the amplitude difference ratio threshold, mark the selected pixel as a significant amplitude point. S32, select a pixel in the phase difference ratio map and calculate the absolute value of the phase difference ratio corresponding to the pixel. When the absolute value is greater than the phase difference ratio threshold, mark the selected pixel as a phase salient point.

[0040] Specifically, in S31 and S32, a threshold is used as the judgment standard, and the absolute value of the difference ratio is used as the comparison reference. After calculating the absolute value of the difference ratio, it is compared with the threshold. Pixels with an absolute value greater than the threshold are considered salient points. It should be noted that the amplitude difference ratio threshold and the phase difference ratio threshold can be flexibly set according to the actual application scenario. In this embodiment, the amplitude difference ratio threshold is set to an empirical value (e.g., 0.15), and the phase difference ratio threshold is set to an empirical value (e.g., 0.12), but the present invention is not limited to this. For example, when the image quality is good, the threshold can be appropriately lowered to retain more salient points; when the image noise is large, the threshold can be appropriately increased to reduce false detections. Those skilled in the art can adjust the threshold according to the image quality, monitoring accuracy requirements, and actual dam conditions, all of which fall within the protection scope of the present invention.

[0041] By setting a threshold and comparing it with the absolute value of the difference ratio, background noise and minor fluctuations can be effectively filtered out, and pixels that have undergone significant changes in the image can be accurately extracted, providing a reliable data foundation for the subsequent construction of diffusion distribution maps. At the same time, the threshold can be flexibly adjusted according to the actual scenario, which enhances the adaptability and robustness of the invention to different monitoring conditions (such as image quality and noise level).

[0042] The process of constructing the diffusion distribution map based on the salient point marking results is as described in step S4. As an optional implementation of step S4, step S4 includes: S41. Perform isolated detection on the amplitude saliency points in the amplitude difference ratio map of each time phase, and discard the isolated amplitude saliency points; take each remaining amplitude saliency point as the center, count the number of amplitude saliency points in the small, medium and large scale windows, and obtain the density of each remaining amplitude saliency point in the small, medium and large scale windows; according to the ratio of the density of the same amplitude saliency points in the amplitude difference ratio map of adjacent time phases at the same scale, obtain the amplitude diffusion value of each amplitude saliency point; for each amplitude saliency point, set the amplitude diffusion value of non-amplitude saliency points to 0, and obtain the amplitude diffusion distribution map of the corresponding time.

[0043] S42. For each phase difference ratio map, isolate the phase salient points and discard the isolated phase salient points. Using each remaining phase salient point as the center, count the number of phase salient points in the small, medium and large scale windows to obtain the density of each remaining phase salient point in the small, medium and large scale windows. According to the ratio of the density of the same phase salient points in the phase difference ratio map of adjacent phases at the same scale, obtain the phase diffusion value of each phase salient point. For each phase salient point, set the phase diffusion value of non-phase salient points to 0 to obtain the phase diffusion distribution map of the corresponding time.

[0044] It should be noted that the neighborhood range and multi-scale window size involved in steps S41 and S42 can be flexibly set according to the actual application scenario. In this embodiment, the neighborhood range is set to 3×3, the small-scale window is set to 5×5, the medium-scale window is set to 7×7, and the large-scale window is set to 9×9, but the present invention is not limited thereto. Those skilled in the art can adjust the window size and neighborhood range according to the image resolution, the scale of dam deformation characteristics, and the monitoring accuracy requirements, all of which fall within the protection scope of the present invention.

[0045] Specifically, when taking a significant amplitude point as the center, if there are no other significant amplitude points within a 3×3 neighborhood, the significant amplitude point at the center is an isolated point, and its marked significant amplitude point is deleted; when taking a significant phase point as the center, if there are no other significant phase points within a 3×3 neighborhood, the significant phase point at the center is an isolated point, and its marked significant phase point is deleted.

[0046] After removing isolated amplitude and phase salient points, the number of salient points within a small-scale window is counted, centered on the remaining salient points. This number is used as the density at the small scale. Similarly, the number of salient points within a medium-scale window is counted, and this number is used as the density at the medium scale. Finally, the number of salient points within a large-scale window is counted, and this number is used as the density at the large scale.

[0047] In the step of calculating the amplitude diffusion distribution map, as an optional implementation of step S41, the ratio of the density of the same amplitude salient point at the same scale in the amplitude difference ratio map of adjacent time phases is calculated to obtain the ratio at the small scale, the ratio at the medium scale, and the ratio at the large scale. The ratios at the small scale, the ratio at the medium scale, and the ratio at the large scale of the same amplitude salient point are added together, and the average value is taken. Then, 1 is subtracted to obtain the basic amplitude diffusion value of the corresponding amplitude salient point. When the basic amplitude diffusion value is greater than 0, the basic amplitude diffusion value is retained; when the basic amplitude diffusion value is less than or equal to 0, the basic amplitude diffusion value of the corresponding amplitude salient point is set to 0. The basic amplitude diffusion value of each amplitude salient point satisfies the following formula:

[0048] In the above formula, t is the phase index, j is the index of the significant amplitude point; S is the small scale, M is the medium scale, and L is the large scale; For the first The amplitude difference ratio of the time phase in the figure is shown in the figure. The base amplitude diffusion value at each significant amplitude point, For the first The amplitude difference ratio of the time phase in the figure is shown in the figure. Density at a small scale at a point of significant amplitude For the first The amplitude difference ratio of the time phase in the figure is shown in the figure. Density at a small scale at a point of significant amplitude For the first The amplitude difference ratio of the time phase in the figure is shown in the figure. Density at mesoscale at points of significant amplitude For the first The amplitude difference ratio of the time phase in the figure is shown in the figure. Density at mesoscale at points of significant amplitude For the first The amplitude difference ratio of the time phase in the figure is shown in the figure. Density at large scales at points of significant amplitude For the first The amplitude difference ratio of the time phase in the figure is shown in the figure. The density at a large scale for each significant amplitude point. Set all base amplitude diffusion values ​​less than 0 to 0 to obtain the phase diffusion value of each significant amplitude point; then fill the amplitude diffusion value of each significant amplitude point according to its pixel position, and set the positions of non-significant amplitude points to 0, thus obtaining the amplitude diffusion distribution map at the corresponding time.

[0049] In the step of calculating the phase diffusion distribution map, as an optional implementation of step S42, the ratio of the density of the same phase salient point at the same scale in the phase difference ratio map of adjacent time phases is calculated to obtain the ratio at the small scale, the ratio at the medium scale, and the ratio at the large scale. The ratios at the small scale, the ratio at the medium scale, and the ratio at the large scale of the same phase salient point are added together, the average is taken, and then 1 is subtracted to obtain the phase diffusion value of the corresponding phase salient point. When the phase diffusion value is greater than 0, the phase diffusion value is retained; when the phase diffusion value is less than or equal to 0, the phase diffusion value of the corresponding phase salient point is set to 0. The basic phase diffusion value of each phase salient point satisfies the following formula:

[0050] In the above formula, t is the phase index, j is the phase saliency index; S is the small scale, M is the medium scale, and L is the large scale. For the first The phase difference ratio diagram of the time phase The fundamental phase diffusion value at each phase salient point For the first The phase difference ratio diagram of the time phase Density at a small scale at each phase salient point For the first The phase difference ratio diagram of the time phase Density at a small scale at each phase salient point For the first The phase difference ratio diagram of the time phase Density at mesoscale at each phase salient point For the first The phase difference ratio diagram of the time phase Density at mesoscale at each phase salient point For the first The phase difference ratio diagram of the time phase Density at large scales of phase salient points For the first The phase difference ratio diagram of the time phase The density of each phase salient point on a large scale. Set all basic phase diffusion values ​​less than 0 to 0 to obtain the phase diffusion value of each phase salient point; then fill the phase diffusion value of each phase salient point according to its pixel position, and set the positions of non-phase salient points to 0, thus obtaining the phase diffusion distribution map at the corresponding time.

[0051] This invention uses small, medium, and large-scale windows of 5×5, 7×7, and 9×9 to statistically analyze the density of remaining significant points. Combined with the density ratio at the corresponding scale in adjacent time phases, the diffusion value is obtained by averaging and positive / negative judgment. This quantifies the changing trend of the density of significant points in adjacent time phases. The setting of positive / negative judgment for amplitude diffusion value and phase diffusion value can effectively preserve the positive diffusion characteristics of significant point density and suppress the interference caused by invalid negative changes.

[0052] S41 and S42 calculate the density ratio of salient points with the same amplitude / phase at adjacent time phases and across three scales (small, medium, and large). They then average these ratios across the three scales and subtract 1 to obtain the amplitude / phase diffusion value, retaining only the positive diffusion value. This multi-scale windowing provides a comprehensive characterization of the salient point distribution across different spatial ranges (small-scale local clustering, medium-scale correlation trends, and large-scale diffusion range), and quantifies the spatiotemporal variation of the density ratio to indicate the extent of salient point diffusion. Time phase The diffusion and growth of the time phase effectively distinguished between isolated salient points caused by noise and surface perturbations and the aggregation and diffusion of salient points caused by dam deformation, eliminating invalid interference from negative diffusion.

[0053] The beneficial effects of the above steps are also reflected in the following: by statistically analyzing the density of salient points through multi-scale windows, the aggregation characteristics of salient points can be captured from different spatial scales, avoiding the information loss caused by a single scale; at the same time, the window size can be flexibly adjusted to adapt to images of different resolutions and dam deformation characteristics of different scales.

[0054] After obtaining the multi-temporal amplitude diffusion distribution map and phase diffusion distribution map based on the above steps, the amplitude activity map and phase activity map are obtained according to step S5. As an optional implementation of step S5, step S5 includes: S51. Add the amplitude diffusion values ​​of the same pixels in the amplitude diffusion distribution map of multiple time phases to obtain the cumulative amplitude diffusion value; S52. Perform exponential normalization mapping on the cumulative amplitude diffusion value to obtain the amplitude activity map; S53. Add the phase diffusion values ​​of the same pixels in the phase diffusion distribution map of multiple time phases to obtain the cumulative phase diffusion value; S54. Perform exponential normalization mapping on the cumulative phase diffusion value to obtain the phase activity map.

[0055] Specifically, the formula for calculating the cumulative amplitude diffusion value in step S51 is as follows:

[0056] In the above formula, i is the pixel index, t is the phase index, and T is the total number of phases involved in the accumulation; For the first Cumulative amplitude diffusion value of each pixel, For the first The amplitude diffusion distribution diagram of the time phase. The amplitude diffusion value of each pixel.

[0057] The formula for calculating the exponentially normalized mapping of the cumulative amplitude diffusion value in step S52 is as follows:

[0058] In the above formula, i is the pixel index. For the first Normalized amplitude diffusion cumulative value of each pixel, For the first Cumulative amplitude diffusion value of each pixel, This represents the cumulative value of the maximum diffusion. It is a natural constant.

[0059] The formula for calculating the cumulative value of phase diffusion in step S53 is as follows:

[0060] In the above formula, i is the pixel index. For the first The cumulative phase diffusion value of each pixel. For the first Phase diffusion distribution diagram of time phase Phase diffusion value of each pixel.

[0061] In step S54, the formula for exponentially normalizing the cumulative phase diffusion value is as follows:

[0062] in, For the first Normalized phase spread cumulative value of each pixel, For the first The cumulative phase diffusion value of each pixel. This is the maximum cumulative phase diffusion value. It is a natural constant.

[0063] In steps S51 and S53, the amplitude and phase diffusion values ​​of multiple time phases are added together to obtain the cumulative amplitude and phase diffusion values ​​of the corresponding pixels. In steps S52 and S54, exponential normalization mapping is used to process the cumulative values, highlighting areas of active deformation. This effectively amplifies the differences in cumulative changes between different pixels, emphasizing the distinction between deformation-sensitive and non-deformation areas of the dam. The resulting amplitude and phase activity maps accurately reflect the degree of deformation activity in different areas of the dam throughout the monitoring period.

[0064] Subsequently, a full-time linked binary image is obtained based on step S6, and this full-time linked binary image is one of the input images for subsequent networks. As an optional implementation of step S6, step S6 includes: S61. Subtract the amplitude of the same pixel position from the amplitude image of the first phase and the amplitude image of the last phase to obtain the amplitude difference of the corresponding pixel. S62. The ratio of the amplitude difference of the corresponding pixel to the amplitude of the same pixel in the amplitude image of the last time phase is taken as the amplitude difference ratio. When the absolute value of the amplitude difference ratio is greater than the amplitude difference ratio threshold, the pixel value of the corresponding pixel is set to 1 and the pixel value of other pixels is set to 0 to obtain the full-time amplitude binary image. S63. Subtract the phase of the same pixel position in the phase image of the first phase from the phase image of the last phase to obtain the phase difference of the corresponding pixel. S64. The ratio of the phase difference of the corresponding pixel to the phase of the same pixel in the phase image of the last time phase is taken as the phase difference ratio. When the absolute value of the phase difference ratio is greater than the phase difference ratio threshold, the pixel value of the corresponding pixel is set to 1 and the pixel value of other pixels is set to 0 to obtain the phase binary image of the whole time period. S65. Perform a bitwise AND operation on the pixel values ​​of the full-time amplitude binary image and the full-time phase binary image to obtain a full-time linked binary image.

[0065] This invention filters out regions with significant amplitude changes through S61 and S62 and sets them to 1, and filters out regions with significant phase changes through S63 and S64 and sets them to 1. The intersection of the two binary images is then used to obtain a full-time linked binary image, which further eliminates false interference caused by single-dimensional changes (amplitude or phase changes caused by non-deformation).

[0066] Furthermore, by performing amplitude and phase linkage calculations, it is possible to effectively eliminate interference from single-dimensional changes caused by non-deformation factors such as environmental noise and atmospheric disturbances, while retaining regions where both amplitude and phase have changed significantly. This provides more reliable spatial prior information for subsequent neural networks, reduces false detection rates, and improves the accuracy and robustness of deformation monitoring.

[0067] Finally, the network structure of the dual-path attention fusion neural network disclosed in step S7 of this application is given. For example... Figure 2 As shown, the dual-path attention fusion neural network in S7 includes: a first attention generation unit, a second attention generation unit, an initial feature encoding unit, a first feature enhancement unit, a second feature enhancement unit, a first deep encoding unit, a second deep encoding unit, an adder A1, and a deformation decoding output unit; The first attention generation unit, the second attention generation unit, and the initial feature encoding unit are configured in parallel as the input segments of the network, respectively used to receive the amplitude activity map, the full-time linked binary image, and the phase activity map. The output of the first attention generation unit is connected to the input of the first feature enhancement unit; the output of the second attention generation unit is connected to the input of the second feature enhancement unit; the output of the initial feature encoding unit is connected to the inputs of both the first and second feature enhancement units; the output of the first feature enhancement unit is connected to the input of the first deep encoding unit; the output of the second feature enhancement unit is connected to the input of the second deep encoding unit; the input of adder A1 is connected to the outputs of both the first and second deep encoding units, and its output is connected to the input of the deformation decoding output unit; the output of the deformation decoding output unit serves as the output of the dual-path attention fusion neural network.

[0068] The functions of each unit are as follows: the first attention generation unit generates the first attention weight based on the amplitude activity map; the second attention generation unit generates the second attention weight based on the phase activity map; the initial feature encoding unit performs global feature encoding on the full-time linked binary image; the first feature enhancement unit applies the first attention weight to the global feature encoding and forms a residual concatenation to obtain the first enhanced global feature; the second feature enhancement unit applies the second attention weight to the global feature encoding and forms a residual concatenation to obtain the second enhanced global feature; the first deep encoding unit obtains the first global deep feature from the first enhanced global feature; the second deep encoding unit obtains the second global deep feature from the second enhanced global feature; the adder A1 adds the first global deep feature and the second global deep feature element-wise to obtain the global deep fusion feature; and the deformation decoding output unit obtains the dam deformation result based on the global deep fusion feature.

[0069] Through extensive experimental testing and improvements by the inventors of this application, the first attention generation unit of the network structure can generate corresponding attention weights based on the input amplitude activity map, and the second attention generation unit can generate weights based on the phase activity map. The dual attention mechanism can accurately select effective features and filter invalid interference signals, thereby improving the accuracy of feature recognition. The initial feature encoding unit performs global feature encoding on the full-time linked binary image, and the first and second feature enhancement units respectively enhance and optimize the features. The first and second deep encoding units fully extract the deep correlation information of the two types of features through deep mining, and then achieve effective fusion of dual-path features through adder A1. The integrated global deep fusion features are more representative. The deformation decoding output unit combines multi-dimensional features to accurately output the dam deformation results.

[0070] In the first feature enhancement unit, the output weights of the first attention generation unit are multiplied element-wise with the output features of the initial feature encoding unit, and the result of the multiplication is added element-wise to the output features of the initial feature encoding unit to obtain the output features of the first feature enhancement unit; in the second feature enhancement unit, the output weights of the second attention generation unit are multiplied element-wise with the output features of the initial feature encoding unit, and the result of the multiplication is added element-wise to the output features of the initial feature encoding unit to obtain the output features of the second feature enhancement unit.

[0071] The expression for the first feature enhancement unit is:

[0072] in, The output features of the first feature enhancement unit, The output features of the initial feature encoding unit, For element-wise multiplication, To add elements together, The output weights of the first attention generation unit; The expression for the second feature enhancement unit is:

[0073] in, The output features of the second feature enhancement unit, The output weights for the second attention generation unit.

[0074] This approach effectively highlights key deformation features while avoiding the loss of feature information, accurately capturing the characteristics of dam deformation and enhancing the distinguishability and expressiveness of the features.

[0075] like Figure 3 As shown, the initial feature encoding unit includes a first Conv block and a second Conv block connected in sequence.

[0076] like Figure 4 As shown, the first attention generation unit and the second attention generation unit have the same structure, both including the third Conv block, the fourth Conv block and the Sigmoid layer connected in sequence.

[0077] and Figure 3 and Figure 4 The corresponding parameters are set as follows: the kernel size of the first Conv block, the second Conv block, the third Conv block, and the fourth Conv block are all 3×3, the stride is 1, and the padding is 1.

[0078] It should be noted that, Figure 3 and Figure 4 The kernel size, stride, and padding parameters are merely examples. Those skilled in the art can make adaptive adjustments based on factors such as network depth, feature map size, and computing resources, all of which fall within the scope of protection of this invention.

[0079] like Figure 5 As shown, the first deep coding unit and the second deep coding unit have the same structure, both including the following connected in sequence: the fifth Conv block, the sixth Conv block, the downsampling layer and the seventh Conv block.

[0080] and Figure 5 The corresponding parameters are set as follows: Fifth Conv block: 3×3 convolution kernel, stride 1, padding 1; Sixth Conv block: 3×3 convolution kernel, stride 1, padding 1; Downsampling layer: 2×2 max pooling, stride 2; Seventh Conv block: 5×5 convolution kernel, stride 1, padding 2.

[0081] It should also be noted that Figure 5 The kernel size, stride, padding, and pooling parameters in the examples are merely illustrative. Those skilled in the art can make adaptive adjustments according to actual needs, and all of these adjustments fall within the scope of protection of this invention.

[0082] like Figure 6 As shown, the deformation decoding output unit includes: an eighth Conv block, a ninth Conv block, an adder A1, an upsampling layer, a tenth Conv block, an eleventh Conv block, a twelfth Conv block, a Concat layer, a thirteenth Conv block, and a classification layer; the eighth Conv block, the ninth Conv block, and the adder A1 constitute the residual structure, the upsampling layer is used to restore the scale, the tenth Conv block, the eleventh Conv block, and the twelfth Conv block are used to form multi-scale feature extraction, and the Concat layer is used to concatenate the outputs of the tenth Conv block, the eleventh Conv block, and the twelfth Conv block.

[0083] All Conv blocks include: convolutional layers, BN layers, and ReLU layers.

[0084] Eighth Conv block: 3×3 kernel, stride 1, padding 1; Ninth Conv block: 3×3 kernel, stride 1, padding 1. The upsampling layer is scaled by 2 to restore the feature map.

[0085] 10th Conv block: 3×3 kernel, stride 1, padding 1; 11th Conv block: 5×5 kernel, stride 1, padding 2; 12th Conv block: 7×7 kernel, stride 1, padding 3; 13th Conv block: 1×1 kernel, stride 1, padding 1.

[0086] It needs to be stated again that, Figure 6 The kernel size, stride, padding, and scaling parameters in the examples are merely illustrative. Those skilled in the art can make adaptive adjustments according to actual needs, all of which fall within the protection scope of this invention.

[0087] This classification layer employs a fully connected layer and outputs a sigmoid activation function, providing a probability value for each pixel indicating whether it belongs to a deformed region. Specifically, the sigmoid activation function maps input values ​​to probability values ​​between 0 and 1, representing the probability that a pixel belongs to a deformed region. When the probability value is greater than or equal to 0.5, the pixel is classified as "deformed"; while when the probability value is less than 0.5, the pixel is classified as "not significantly deformed." In this way, the classification layer can effectively distinguish between deformed and non-deformed regions, achieving accurate identification and classification of deformation on the dam surface.

[0088] In summary, this invention addresses the low accuracy of dam deformation monitoring in existing technologies by proposing a joint processing method combining image preprocessing and deep learning decision-making. Through innovative experimental research, high-accuracy dam deformation monitoring was ultimately achieved. The beneficial effects of this method are as follows: By separating amplitude and phase in multi-temporal SAR images during the preprocessing stage, extracting the difference ratio maps of amplitude and phase respectively, and marking significant points, this invention effectively captures the dynamic changes in local areas of the dam surface, reducing phase unwrapping errors caused by changes in scattering characteristics in traditional methods.

[0089] This invention can accurately identify the local deformation region when significant non-uniform deformation occurs on the dam surface by constructing amplitude diffusion distribution map and phase diffusion distribution map, avoiding the influence of phase decoherence and decoherence error propagation that may occur during the phase unwrapping and deformation inversion process of stable scatterers.

[0090] This invention obtains amplitude activity maps and phase activity maps by accumulating amplitude diffusion distribution maps and phase diffusion distribution maps of multiple time phases. These maps are used to assess the spatial and temporal continuity of deformation over a continuous time range, accurately depicting the true deformation characteristics of the dam's local abrupt deformation area. This avoids the problem of inaccurate deformation estimation in traditional methods when facing local abrupt deformation.

[0091] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An InSAR deformation monitoring method for dams, characterized in that, The InSAR deformation monitoring method includes: The multi-temporal SAR images of the dam are subjected to amplitude separation and phase separation to obtain multi-temporal amplitude images and multi-temporal phase images. The difference ratios of amplitude and phase images of adjacent time phases are calculated separately to obtain the amplitude difference ratio map and phase difference ratio map of multiple time phases; Mark the significant points in the amplitude difference ratio plot and the phase difference ratio plot; Based on the number of significant points within the local range of the amplitude difference ratio diagram and phase difference ratio diagram of adjacent time phases, construct the amplitude diffusion distribution diagram and phase diffusion distribution diagram of multiple time phases; The amplitude diffusion distribution map and phase diffusion distribution map of the multi-temporal phases are accumulated and mapped respectively to obtain the amplitude activity map and phase activity map; Based on the amplitude and phase images of the first and last time phases, a full-time linked binary image is obtained; A dual-path attention fusion neural network was used to process the full-time linked binary image, amplitude activity map, and phase activity map to obtain the dam deformation results.

2. The InSAR deformation monitoring method according to claim 1, characterized in that, The step of calculating the difference ratios of amplitude and phase images of adjacent time phases to obtain multi-time phase amplitude difference ratio maps and phase difference ratio maps includes: The amplitude difference ratio is obtained by taking the ratio of the amplitude difference of corresponding pixels in the amplitude image of adjacent time phases to the amplitude of the same pixel in the previous time phase as the amplitude difference ratio. The phase difference ratio is obtained by taking the ratio of the phase difference of corresponding pixels in the phase image of adjacent time phases to the phase of the same pixel in the previous time phase as the phase difference ratio.

3. The InSAR deformation monitoring method according to claim 1, characterized in that, The significant points in the marked amplitude difference ratio diagram and phase difference ratio diagram include: Select a pixel in the amplitude difference ratio map and calculate the absolute value of the amplitude difference ratio corresponding to the pixel. When the absolute value is greater than the amplitude difference ratio threshold, mark the selected pixel as a significant amplitude point. Select a pixel in the phase difference ratio map and calculate the absolute value of the phase difference ratio corresponding to the pixel. When the absolute value is greater than the phase difference ratio threshold, mark the selected pixel as a phase salient point.

4. The InSAR deformation monitoring method according to claim 1, characterized in that, The step of constructing multi-temporal amplitude diffusion distribution maps and phase diffusion distribution maps based on the number of significant points within local ranges in the amplitude difference ratio maps and phase difference ratio maps of adjacent time phases includes: For each time phase, significant amplitude points in the amplitude difference ratio graph are isolated and discarded. Centered on each remaining significant amplitude point, the number of significant amplitude points within different scale windows is counted to obtain the density of each remaining significant amplitude point at each scale. The amplitude diffusion value of each significant amplitude point is obtained by comparing the density of significant points with the same amplitude at the same scale in the amplitude difference ratio diagram of adjacent time phases. For each significant amplitude point, the amplitude diffusion value of non-significant amplitude points is set to 0, thus obtaining the amplitude diffusion distribution map at the corresponding time. For each phase, the phase salient points in the phase difference ratio diagram are isolated and the isolated phase salient points are discarded. Centered on each remaining phase salient point, the number of phase salient points within different scale windows is counted to obtain the density of each remaining phase salient point at each scale. Based on the ratio of the density of significant points of the same phase at the same scale in the phase difference ratio diagram of adjacent time phases, the phase diffusion value of each significant point of the phase is obtained; For each phase salient point, the phase diffusion value of the non-phase salient points is set to 0, thus obtaining the phase diffusion distribution map at the corresponding time.

5. The InSAR deformation monitoring method according to claim 4, characterized in that, The step of obtaining the amplitude diffusion value of each significant amplitude point based on the ratio of the density of significant points with the same amplitude at the same scale in the amplitude difference ratio diagram of adjacent time phases includes: The baseline amplitude diffusion value of the significant point of amplitude is obtained by averaging the ratios of densities at different scales and subtracting the preset ratio parameter. The preset ratio parameter is 1. Set the base amplitude diffusion value, which is less than or equal to the preset amplitude diffusion parameter, to 0 to obtain the amplitude diffusion value at the significant amplitude point; where the preset amplitude diffusion parameter is 0; The step of obtaining the phase diffusion value of each significant point in a phase based on the ratio of the densities of significant points of the same phase at the same scale in the phase difference ratio diagram of adjacent time phases includes: The basic phase diffusion value of the phase salient point is obtained by averaging the ratios of densities at different scales and subtracting the preset ratio parameter. The preset ratio parameter is 1. Set the base phase diffusion value, which is less than or equal to the preset phase diffusion parameter, to 0 to obtain the phase diffusion value of the phase salient point; where the preset phase diffusion parameter is 0.

6. The InSAR deformation monitoring method according to claim 1, characterized in that, The step of accumulating and mapping the amplitude diffusion distribution maps and phase diffusion distribution maps of the multiple time phases to obtain the amplitude activity map and phase activity map includes: The amplitude diffusion values ​​of the same pixels in the multi-temporal amplitude diffusion distribution map are added together to obtain the cumulative amplitude diffusion value; The amplitude diffusion cumulative value is subjected to exponential normalization mapping to obtain the amplitude activity map; The phase diffusion values ​​of the same pixels in the phase diffusion distribution map of multiple time phases are added together to obtain the cumulative phase diffusion value; The phase diffusion cumulative value is subjected to exponential normalization mapping to obtain the phase activity map.

7. The InSAR deformation monitoring method according to claim 1, characterized in that, The process of obtaining a full-time linked binary image based on the amplitude and phase images of the first and last time phases includes: The amplitude difference ratio is obtained from the amplitude images of the first and last time phases, and the absolute value of the amplitude difference ratio is compared with the amplitude difference ratio threshold to generate a full-time amplitude binary map. The phase difference ratio is obtained from the phase images of the first and last time phases, and the absolute value of the phase difference ratio is compared with the phase difference ratio threshold to generate a phase binary image for the entire time period. The full-time amplitude binary image and the full-time phase binary image are ANDed by pixels to obtain a full-time linked binary image.

8. The InSAR deformation monitoring method according to claim 1, characterized in that, The dual-path attention fusion neural network includes: The first attention generation unit, the second attention generation unit, and the initial feature encoding unit are set up in parallel. A first feature enhancement unit that connects the first attention generation unit and the initial feature encoding unit respectively; A second feature enhancement unit that connects the second attention generation unit and the initial feature encoding unit respectively; A first deep coding unit connected to the first feature enhancement unit; A second deep coding unit connected to the second feature enhancement unit; Adders that are respectively connected to the first deep coding unit and the second deep coding unit; A deformation decoding output unit connected to the adder.

9. The InSAR deformation monitoring method according to claim 8, characterized in that, The first attention generation unit is used to generate a first attention weight based on the amplitude activity map; The second attention generation unit is used to generate second attention weights based on the phase activity map; The initial feature encoding unit is used to perform global feature encoding on the full-time linked binary image; The first feature enhancement unit is used to apply the first attention weight to the global feature encoding and form a residual concatenation to obtain the first enhanced global feature; The second feature enhancement unit is used to apply the second attention weight to the global feature encoding and form a residual concatenation to obtain the second enhanced global feature; The first deep coding unit is used to extract the first global deep feature from the first enhanced global feature; The second deep coding unit is used to extract the second global deep feature from the second enhanced global feature; The adder is used to add the first global deep feature and the second global deep feature element by element to obtain the global deep fusion feature; The deformation decoding output unit is used to obtain the dam deformation result based on the global deep fusion features.

10. The InSAR deformation monitoring method according to claim 8 or 9, characterized in that, The expression for the first feature enhancement unit is formula (1): （1） In the above formula (1), The output features of the first feature enhancement unit, The output features of the initial feature encoding unit, For element-wise multiplication, To add elements together, The output weights of the first attention generation unit; The expression for the second feature enhancement unit is formula (2): （2） In the above formula (2), The output features of the second feature enhancement unit, The output weights for the second attention generation unit.

Images