Ground collapse monitoring method and system based on twin neural network fusion

Abstract

The application relates to the technical field of ground collapse monitoring, in particular to a ground collapse monitoring method based on twin neural network fusion, which comprises the following steps: acquiring a monitoring range of ground collapse, and arranging control points and target points; acquiring a real-time InSAR scanning image, and first-preprocessing the same to obtain an InSAR collapse field image; acquiring a plurality of real-time 3D local scanning images, second-preprocessing the same, and acquiring a plurality of 3D local collapse field images through a twin neural network; formulating a weight rule, marking the weight of the plurality of 3D local collapse field images according to the weight rule, and splicing the same to obtain a 3D global collapse field image with weight marks; splicing the InSAR collapse field image and the 3D global collapse field image with weight marks through the twin neural network to obtain an InSAR seamless collapse field fusion image, and monitoring the same to obtain a monitoring result. The method can make up for the shortcomings of a single technical means, and realize high-precision and full-coverage monitoring of ground collapse.

Description

Technical Field

[0001] This invention relates to the field of ground subsidence monitoring technology, and more specifically, to a ground subsidence monitoring method and system based on Siamese neural network fusion. Background Technology

[0002] Ground subsidence is a common geological hazard, usually caused by natural factors such as excessive groundwater extraction, seismic activity, and karstification, but it can also be caused by human factors such as mining activities, tunnel construction, and underground engineering projects. This phenomenon causes the ground to sink, affecting the stability of buildings, damaging infrastructure, and even endangering human lives. Ground subsidence often occurs suddenly and insidiously, posing a significant challenge to monitoring and early warning systems.

[0003] Existing methods for monitoring ground subsidence mainly rely on single technologies. For example, Interferometric Synthetic Aperture Radar (InSAR) can provide large-scale surface deformation data, but may suffer from data gaps or errors; 3D laser scanning technology can provide high-precision local surface deformation data, but its monitoring range is limited; existing technologies also combine the above two technologies, but only use simple stitching techniques, which cannot achieve high accuracy and precise monitoring. Each of these methods has its own advantages and disadvantages, and none can comprehensively cover and accurately monitor ground subsidence on its own. Therefore, necessary improvements to ground subsidence monitoring methods are needed. Summary of the Invention

[0004] The present invention aims to overcome at least one of the defects (deficiencies) of the prior art and provide a ground subsidence monitoring method and system based on Siamese neural network fusion, so as to achieve high-precision and full-coverage monitoring of ground subsidence.

[0005] Obtain the monitoring range of ground subsidence, and pre-deploy control points and target points within the monitoring range;

[0006] The monitoring range is monitored in real time. The InSAR scan image is preprocessed first, and the InSAR collapse field image is obtained based on the InSAR scan image after the first preprocessing. The InSAR scan image is obtained by interferometric synthetic aperture radar (InSAR).

[0007] A number of 3D local scan images of the monitoring range in real time are acquired, and the number of 3D local scan images are subjected to a second preprocessing. The number of 3D local scan images after the second preprocessing are processed by a Siamese neural network to obtain a number of 3D local collapse field images, wherein the 3D local scan images are acquired by 3D laser scanning.

[0008] Weighting rules are formulated based on the control points, the target points, the InSAR collapse field images, and several 3D local collapse field images. After weighting and marking the several 3D local collapse field images according to the weighting rules, they are stitched together to obtain a 3D global collapse field image with weight markings.

[0009] The InSAR collapse field image and the weighted 3D global collapse field image are stitched together using the Siamese neural network to obtain a seamless InSAR collapse field fused image. The InSAR seamless collapse field fused image is then monitored to obtain the monitoring results for the monitoring range.

[0010] Understandably, current technologies primarily utilize Interferometric Synthetic Aperture Radar (InSAR) and 3D laser scanning for monitoring ground subsidence. However, while InSAR alone can detect large areas of terrain, it suffers from significant data errors in certain regions. Similarly, 3D laser scanning, while providing high-precision data, only scans small areas at a time, requiring complex and extensive stitching, and exhibiting significant superposition errors at the stitching boundaries. Therefore, this application combines InSAR and 3D laser scanning to monitor subsided ground, employing a twin neural network and weighted rules for stitching to reduce stitching errors, enabling precise monitoring of critical areas and ensuring comprehensive coverage of key subsidence area images.

[0011] Optionally, the InSAR collapse field image is an aperture InSAR collapse field image, wherein the aperture region of the aperture InSAR collapse field image is obtained by removing several error regions using a masking method, wherein the attribute data of the error regions meet a preset error threshold.

[0012] Understandably, using a masking method to remove error regions in the InSAR collapse field image to form holes ensures that error regions do not affect subsequent stitching work, making the data used in the stitching work basically reliable data, thereby improving the accuracy of the entire collapse field image.

[0013] Optionally, the step of formulating weighting rules based on the control points and the target points, the InSAR collapse field image, and several of the 3D local collapse field images specifically includes:

[0014] The control point and the target point are selected as the first selected area;

[0015] The obvious collapse regions in the InSAR collapse field image and several of the 3D local collapse field images are used as the second selected regions, wherein the attribute data of the obvious collapse regions meet the preset collapse data threshold.

[0016] Areas prone to collapse are selected as the third selected area, wherein the areas prone to collapse meet the preset collapse risk attributes;

[0017] The high-quality regions in several of the 3D local collapse field images are selected as the fourth selected regions, wherein the attribute data of the high-quality regions meet the preset data quality threshold.

[0018] Weights are assigned to the first selected area, the second selected area, the third selected area, and the fourth selected area, and four independent weight rules are formulated for the first selected area, the second selected area, the third selected area, and the fourth selected area and their corresponding weights.

[0019] Understandably, for the monitoring range, some accurate data is needed as a benchmark for other data. Therefore, control points need to be set up. The control points are located on relatively stable terrain features to ensure their long-term stability and representativeness. The coordinates of the control points can be found in the 3D local subsidence field image, and their weights are increased to prioritize the stitching of these areas, thus providing a reliable benchmark for subsequent measurements. Corresponding to the monitoring range, target points also need to be deployed to facilitate the registration of 3D point cloud data. The coordinates of the target points can be found in the 3D local subsidence field image, and their weights are increased. Weights are assigned to prioritize the stitching of these regions to provide registration data for 3D scan images. Regions with significant collapse data in the InSAR collapse field image and the 3D local collapse field image are prioritized for stitching to reduce the impact of other data. For certain areas within the monitoring range, which may be at risk of collapse due to terrain and historical factors, areas prone to collapse within the monitoring range are recorded and given higher weights for priority stitching. This ensures the reliability and high accuracy of the data in these areas.

[0020] Optionally, the step of stitching together several 3D local collapse field images after weighting them according to the weighting rule to obtain a weighted global collapse field image specifically includes:

[0021] Traverse each of the 3D local collapse field images, and assign weight labels to the regions in each of the 3D local collapse field images according to the weight rules to obtain several 3D local collapse field images with weight labels;

[0022] Obtain all weights on each of the 3D local collapse field images with weight labels, calculate all the obtained weights, and obtain the total weight of each of the 3D local collapse field images;

[0023] The weighted 3D local collapse field images are stitched together sequentially according to their corresponding total weights to obtain a weighted global collapse field image.

[0024] Understandably, the weighting rules assigned to regions in each 3D local collapse field image provide a priority order for subsequent stitching. This is because the stitching process uses boundary data from two scanned images to integrate them according to an algorithm. The boundary data from both scanned images will affect the integrated data; if one boundary data point has an error, the stitched data will also have an error. Furthermore, the integrated data will be stitched with other boundary data, leading to the accumulation of errors. Using weighting rules to limit the stitching order allows for the priority stitching of areas of focus, ensuring that the accuracy of these areas is not affected by multiple stitching processes, thus improving the reliability of the monitoring data.

[0025] Optionally, the step of assigning weights to regions in each of the 3D local collapse field images according to the weighting rules specifically includes:

[0026] For regions of the 3D local collapse field image that satisfy one of the weighting rules, corresponding weight labels are applied.

[0027] For regions of the 3D local collapse field image that satisfy two or more of the weighting rules, weights are superimposed and marked.

[0028] For other regions of the 3D local collapse field image that do not meet the weighting rules, the weights of the other regions are all set to the initial weights.

[0029] Understandably, the process involves assigning corresponding weights to regions of 3D local collapse field images that satisfy the weighting rules, and superimposing weights on regions of 3D local collapse field images that satisfy multiple weighting rules, to ensure they are stitched together in a more prioritized order. For other regions of 3D local collapse field images that do not satisfy the weighting rules, initial weights are set, and then the Siamese neural network stitches them together sequentially based on the total weight of each 3D local collapse field image, ensuring that 3D local collapse field images with high weights are stitched together first, thus guaranteeing the data reliability of these regions.

[0030] Optionally, the initial weights are all smaller than the weights set in the weighting rules.

[0031] Understandably, the initial weight is smaller than the weight set in the weighting rule, which ensures that the regions that meet the weighting rule are spliced ​​first, and other regions are spliced ​​after the regions that meet the weighting rule are spliced.

[0032] Optionally, the step of stitching the InSAR collapse field image and the weighted 3D global collapse field image together using the Siamese neural network to obtain a seamless InSAR collapse field fused image, and monitoring the seamless InSAR collapse field fused image to obtain monitoring results for the monitoring range, specifically includes:

[0033] The phase information and amplitude information of the InSAR collapse field image are obtained respectively, and the phase information and amplitude information are encoded in different channels by the InSAR data encoder to obtain InSAR collapse field encoded data, wherein the InSAR encoder adopts a convolutional neural network structure.

[0034] The 3D global collapse field image with weighted labels is encoded by a 3D point cloud data encoder to obtain 3D global collapse field encoded data, wherein the 3D point cloud data encoder adopts a point cloud convolutional neural network structure.

[0035] The 3D global collapse field coding data and the InSAR collapse field coding data are input into the Siamese neural network. The Siamese neural network extracts information about the hole region in the InSAR collapse field image from the InSAR collapse field coding data, and uses the 3D global collapse field coding data and the information about the hole region to stitch and fill the hole region to obtain a seamless InSAR collapse field image.

[0036] The twin neural network extracts weighted regions from the 3D global collapse field coding data as selected regions, obtains the selected regions corresponding to the InSAR seamless collapse field image based on the selected regions, and uses the 3D global collapse field coding data to perform permutation and fusion on the corresponding selected regions in the InSAR seamless collapse field image to obtain the InSAR seamless collapse field fused image.

[0037] The InSAR seamless collapse field fusion images at different times are monitored to obtain the monitoring results for the monitoring range.

[0038] Understandably, using different encoders to encode the InSAR collapse field image and the 3D global collapse field image allows for corresponding processing of different data, thereby enabling more accurate capture of the features of both types of data. Using a Siamese neural network to process the InSAR collapse field image and the 3D global collapse field image can quickly and accurately extract the joint feature vector of the two sets of data. By calculating the similarity between the joint feature vectors, data changes at different times can be monitored, and the collapse area can be accurately captured.

[0039] According to a second aspect of this application, a ground subsidence monitoring system based on Siamese neural network fusion is provided, comprising:

[0040] The range acquisition module is used to acquire the monitoring range of ground subsidence, and to pre-deploy control points and target points within the monitoring range;

[0041] The InSAR data acquisition and processing module is used to acquire real-time InSAR scan images of the monitoring range, perform a first preprocessing on the InSAR scan images, and acquire an InSAR collapse field image based on the InSAR scan images after the first preprocessing, wherein the InSAR scan images are acquired by interferometric synthetic aperture radar InSAR.

[0042] A 3D scanning local data acquisition and processing module is used to acquire several 3D local scanning images in real time within the monitoring range, perform a second preprocessing on the several 3D local scanning images, and process the several 3D local scanning images that have completed the second preprocessing through a Siamese neural network to acquire several 3D local collapse field images, wherein the 3D local scanning images are acquired by 3D laser scanning.

[0043] The 3D scanning global data acquisition module is used to formulate weight rules based on the control points and target points, the InSAR collapse field image and several 3D local collapse field images, and then stitch together the several 3D local collapse field images after weight marking them according to the weight rules to obtain a 3D global collapse field image with weight markings.

[0044] The stitching monitoring module is used to stitch the InSAR collapse field image and the weighted global collapse field image together using the Siamese neural network to obtain a seamless InSAR collapse field fused image, and to monitor the seamless InSAR collapse field fused image to obtain the monitoring results of the monitoring range.

[0045] According to a third aspect of this application, an electronic device is provided, including a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the ground subsidence monitoring method based on Siamese neural network fusion described in the first aspect above.

[0046] According to a fourth aspect of this application, a computer storage medium is provided having a computer program stored thereon, which, when executed, implements a ground subsidence monitoring method based on Siamese neural network fusion as described in the first aspect above.

[0047] Based on any of the above aspects, the ground subsidence monitoring method and system based on Siamese neural network fusion provided in this application can first acquire the monitoring range of ground subsidence, then acquire InSAR subsidence field images and preprocess them; acquire several 3D local subsidence field images; stitch the several 3D local subsidence field images according to weight rules through a Siamese neural network to obtain a weighted global subsidence field image, and preprocess it; finally, stitch the InSAR subsidence field image and the weighted global subsidence field image through the Siamese neural network to obtain an InSAR seamless subsidence field fusion image, and monitor the InSAR seamless subsidence field fusion image to obtain the monitoring results of the monitoring range. This method can make up for the shortcomings of single technical means and achieve high-precision and full-coverage monitoring of ground subsidence. By combining interferometric synthetic aperture radar (InSAR) interferometry with 3D laser scanning technology, the advantages of both are fully utilized. Furthermore, by employing twin neural networks and weighting rules to process the two types of data, comprehensive monitoring of ground subsidence can be achieved, improving the accuracy and reliability of monitoring, timely detection and early warning of potential geological disasters, thereby effectively protecting human life and property. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This embodiment provides a flowchart of a ground subsidence monitoring method based on Siamese neural network fusion.

[0050] Figure 2 This is a flowchart of a method for obtaining weight rules in this embodiment.

[0051] Figure 3This is a flowchart of a method for obtaining a 3D global collapse field image according to this embodiment.

[0052] Figure 4 This is a diagram illustrating one possible weighting scheme in this embodiment.

[0053] Figure 5 This is a flowchart of a method for acquiring and monitoring a seamless InSAR collapse field fusion image according to this embodiment.

[0054] Figure 6 This embodiment provides a block diagram of a ground subsidence monitoring system based on Siamese neural network fusion.

[0055] Figure 7 This is a device structure diagram of the electronic device provided in this embodiment. Detailed Implementation

[0056] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this application. To better illustrate the following embodiments, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product; it is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0057] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0058] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0059] Ground subsidence is a common geological hazard, characterized by its suddenness and insidious nature, posing significant challenges to monitoring and early warning systems. Existing technologies commonly employ Interferometric Synthetic Aperture Radar (InSAR) or 3D laser scanning for ground monitoring; however, each of these technologies has its own advantages and disadvantages when used alone. This application combines the advantages of both technologies while mitigating their drawbacks, and utilizes a twin neural network and defined weighting rules to improve data processing accuracy, achieving high-precision and comprehensive monitoring of ground subsidence.

[0060] This embodiment provides a technical solution that can solve the above problems. The specific implementation of this application will be described in detail below with reference to the accompanying drawings.

[0061] For example, Figure 1 This document provides a flowchart of a ground subsidence monitoring method based on Siamese neural network fusion, as illustrated in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0062] S110. Obtain the monitoring range of ground subsidence, and pre-deploy control points and target points within the monitoring range;

[0063] Understandably, obtaining the monitoring range of ground subsidence delineates the scanning boundaries for subsequent scanning work, providing a clear boundary target for the area to be monitored and offering significant guidance for the stitching and registration of the two types of scanned images. Deploying control points serves as a benchmark for processing subsequent scanned images. Preferably, the control points are located in landmarks with relatively stable terrain, allowing for comparison across multiple scanned images. Understandably, accurate height data and other attribute information of the control points need to be obtained beforehand. Comparing the control point information from multiple scanned images with the accurate control point information allows for the assessment of the degree of error in the scanned images. Understandably, the target points are used for the registration of point cloud data.

[0064] S120. Obtain real-time InSAR scan images of the monitoring range, perform first preprocessing on the InSAR scan images, and obtain InSAR collapse field images based on the InSAR scan images after the first preprocessing, wherein the InSAR scan images are obtained by interferometric synthetic aperture radar InSAR.

[0065] It is understood that the InSAR scan image acquired by Interferometric Synthetic Aperture Radar (InSAR) corresponds to a single moment and includes the attribute information of the ground within the monitoring range at that current moment. Preferably, the attribute information includes phase information and amplitude information of the ground within the monitoring range, etc.; it is also understood that the InSAR collapse field image acquired from the InSAR scan image corresponds to a time period, and the InSAR collapse field image reflects the collapse data of the ground within the monitoring range.

[0066] For example, this application provides an embodiment for simulating the acquisition of InSAR collapse field images:

[0067] At time t0, the InSAR scan image at time t0 is acquired by interferometric synthetic aperture radar (InSAR).

[0068] At time t1, the InSAR scan image at time t1 is acquired by interferometric synthetic aperture radar (InSAR). Understandably, the time interval between time t0 and time t1 needs to be set in advance, preferably a few days or several months.

[0069] Understandably, to monitor ground deformation, multiple imaging acquisitions at the same orbit and angle are typically required. Preprocessing of the InSAR scan images includes, but is not limited to, one or more of existing image processing techniques such as image registration, noise removal, interferogram generation, and phase unwrapping.

[0070] Image registration: Geometrically align multiple acquired InSAR scan images to ensure they cover the same ground area;

[0071] Noise removal: Eliminates errors caused by atmospheric interference, ionospheric effects, and other noise sources through filtering;

[0072] Generate interferogram:

[0073] Interferogram calculation: An interferogram is obtained by complex multiplying two registered InSAR scan images. The interferogram contains phase information, reflecting ground height and deformation information;

[0074] Removing the flat-ground effect: By using a digital elevation model (DEM) to remove the flat-ground effect in the interferogram, the ground deformation is highlighted;

[0075] Phase unwrapping: Since the phase values ​​in the interferogram vary within the range of [-π, π], phase unwrapping is required to obtain continuous phase information; this step can obtain actual surface displacement information.

[0076] By comparing the preprocessed InSAR scan image at time t0 and the InSAR scan image at time t1, the regions where phase, amplitude, and other parameters change can be obtained. The phase difference information of the attributes is calculated to obtain the InSAR collapse data of the monitoring range in the time period t0 to t1, and an InSAR collapse field image is generated.

[0077] It is understandable that the InSAR scan image obtained by using interferometric synthetic aperture radar is a large-scale image. Due to the wide scanning range, the number of scans can be reduced, but this will also lead to large errors in some areas, resulting in large errors in some areas of the generated InSAR collapse field image.

[0078] Specifically, the InSAR collapse field image is an aperture InSAR collapse field image, wherein the aperture region of the aperture InSAR collapse field image is obtained by removing several error regions using a masking method, and the attribute data of the error regions meet a preset error threshold.

[0079] Understandably, for the identified data that meets the preset error threshold, this application uses a masking method to remove the aforementioned error data. Understandably, the InSAR collapse field image with the error data removed is a hole-filled InSAR collapse field image, where the holes represent areas lacking ground attribute information. Monitoring of these areas requires supplementation in subsequent work. The error threshold can be set using large-scale errors caused by phase unwrapping errors, noise, etc., during the interferogram processing step as reference data.

[0080] S130. Acquire several 3D local scan images of the monitoring range in real time, perform a second preprocessing on the several 3D local scan images, and process the several 3D local scan images that have completed the second preprocessing through a Siamese neural network to obtain several 3D local collapse field images, wherein the 3D local scan images are acquired by 3D laser scanning.

[0081] Understandably, 3D local scan images obtained through 3D laser scanning technology are localized, and the data possesses a certain degree of reliability and high accuracy. However, multiple scans at different angles and positions are required to acquire several 3D local scan images, generating several 3D local collapse field images. The relationship between 3D local scan images and 3D local collapse field images is the same as that between InSAR scan images and InSAR collapse field images, and will not be repeated here.

[0082] Understandably, preprocessing several 3D partial scan images requires aligning and fusing point cloud data from different scanning angles and positions using the data from the control points and target points. A classification algorithm is applied to separate the point cloud data into ground points and non-ground points, and the data is filtered. A denoising algorithm is used to remove isolated points and noise from the point cloud data, and the point cloud data of the 3D partial scan images is aligned with an existing geographic coordinate system. A coordinate transformation algorithm is then applied to transform the point cloud data to the required geographic coordinate system.

[0083] Understandably, for a local 3D scanned deformed image, the point cloud data may contain a large amount of redundant information, such as repeated points and non-critical regions. Performing the aforementioned preprocessing on all point cloud data would consume excessive computational resources and result in significant redundant calculations. Therefore, this application employs a random region selection mechanism. In the selection process of multiple local point cloud regions within a local 3D scanned deformed image, a range [n1, n2] is set, and n regions are randomly selected from all point cloud regions for preprocessing calculations, where n ∈ [n1, n2] and are random integers. This random region selection mechanism avoids redundant calculations of the global point cloud data, improves stitching efficiency, and maintains the diversity of region selection. Using randomly selected n regions for processing avoids redundancy.

[0084] Understandably, the Siamese neural network can compare the similarity of two inputs and capture subtle differences between two input samples. In this example, for the 3D local scan images of the same area at different times, theoretically, if there is no obvious collapse, the 3D local scan images at two different times will be very similar. By using the Siamese neural network to find the coordinates where the terrain data of the 3D local scan images at two different times shows subtle changes, the difference attributes of such coordinates can be calculated to generate several 3D local collapse field images. Using the Siamese neural network can free up the manual comparison process and achieve more accurate results.

[0085] Understandably, the amount of data for the aforementioned 3D local scan images is enormous, and scanning multiple local areas will result in a lot of redundant data and increase the risk of errors. If all 3D local scan images are stitched together sequentially, it will cause the accumulation of errors. Therefore, it is necessary to prioritize stitching together the areas of interest to maintain the reliability of the data in those areas and achieve accurate monitoring.

[0086] S140. Based on the control points and target points, the InSAR collapse field image and several 3D local collapse field images, a weighting rule is formulated. After weighting and marking several 3D local collapse field images according to the weighting rule, they are stitched together to obtain a 3D global collapse field image with weight marking.

[0087] Specifically, such as Figure 2 As shown, the method for formulating the weighting rules can be further divided into the following cases:

[0088] The control point and the target point are selected as the first selected area;

[0089] The obvious collapse regions in the InSAR collapse field image and several of the 3D local collapse field images are used as the second selected regions, wherein the attribute data of the obvious collapse regions meet the preset collapse data threshold.

[0090] Areas prone to collapse are selected as the third selected area, wherein the areas prone to collapse meet the preset collapse risk attributes;

[0091] The high-quality regions in several of the 3D local collapse field images are selected as the fourth selected regions, wherein the attribute data of the high-quality regions meet the preset data quality threshold.

[0092] Weights are assigned to the first selected area, the second selected area, the third selected area, and the fourth selected area, and four independent weight rules are formulated for the first selected area, the second selected area, the third selected area, and the fourth selected area and their corresponding weights.

[0093] Understandably, stitching two scanned images requires merging and calculating their boundary data using an algorithm to obtain the stitched boundary data, thus combining the two scanned images into one. Therefore, any error in the boundary data of either of the two scanned images will affect the merged boundary data; moreover, some boundary data are stitched more than once, leading to the accumulation of errors during the stitching process.

[0094] In this application, the control points and target points are selected as the weighting rules, and the regions are spliced ​​first. This ensures that the data in the regions corresponding to the control points and target points are not affected by other data, thus guaranteeing the original accuracy and benchmark effect of the region data.

[0095] For areas in the InSAR collapse field image and 3D local collapse field image where obvious collapse data can be found, it indicates that the area may have collapse phenomenon or may be error data. Therefore, this area is prioritized for stitching to preserve the originality of the data in the area. Subsequent analysis focuses on determining whether the area is a collapse area or error data. The collapse data threshold can be set according to historical experience.

[0096] For certain areas within the monitoring range, there may be a risk of collapse due to terrain and historical factors. It is necessary to focus on monitoring the data of these areas to ensure the reliability and high accuracy of the data, and to accurately monitor the areas prone to collapse. The collapse risk attribute can be set according to the risk value predicted by experts for each terrain.

[0097] Understandably, during the stitching process, the noise level and point cloud density of each 3D local collapse field image are monitored. If a certain region has low noise and high data density, it is selected as the region by the weighting rule. This ensures that high-quality data is stitched first, improving the overall stitching accuracy and stability. The data quality threshold can be set based on historical experience.

[0098] Weights are assigned to the selected regions mentioned above, and the weighting rules are formulated. Understandably, the weight value for each region is pre-set based on expert experience.

[0099] Specifically, such as Figure 3 As shown, the process of weighting and stitching the 3D local collapse field images can be further divided into the following steps:

[0100] S141. Traverse each of the 3D local collapse field images, and assign weight labels to the regions in each of the 3D local collapse field images according to the weight rules to obtain several 3D local collapse field images with weight labels.

[0101] S142. Obtain all weights on each of the 3D local collapse field images with weight labels, calculate all the obtained weights, and obtain the total weight of each of the 3D local collapse field images;

[0102] S143. The weighted 3D local collapse field images are stitched together sequentially according to their corresponding total weight values ​​to obtain a weighted global collapse field image.

[0103] Specifically, such as Figure 4 As shown, the weighting of regions in each 3D local collapse field image according to the weighting rule can be further subdivided into the following cases:

[0104] For regions of the 3D local collapse field image that satisfy one of the weighting rules, corresponding weight labels are applied.

[0105] For regions of the 3D local collapse field image that satisfy two or more of the weighting rules, weights are superimposed and marked.

[0106] For other regions of the 3D local collapse field image that do not meet the weighting rules, the weights of the other regions are all set to the initial weights.

[0107] Specifically, the initial weights are all smaller than the weights set in the weighting rules.

[0108] Understandably, the splicing mechanism is as follows:

[0109] • Geometric Alignment: Using control point and target point data, geometric alignment is performed on randomly selected 3D local collapse field images to ensure seamless stitching between regions. Geometric alignment employs least squares error optimization and combines it with the ICP (Iterative Closest Point) algorithm to further improve alignment accuracy.

[0110] • Weight calculation: During the stitching process, the 3D local collapse field images are stitched sequentially according to their weights to ensure the stability and accuracy of the final stitching result.

[0111] Understandably, stitching together all the 3D local collapse field images after assigning weights allows for priority stitching of regions that meet the weighting rules, achieving the following effect:

[0112] By prioritizing regions that meet the weighting rules, high-precision change information in these regions can be determined earlier, reducing the accumulation of local errors in the overall stitching; ensuring the accuracy of important regions and preventing important data from being covered by low-precision point clouds.

[0113] S150. The InSAR collapse field image and the weighted global collapse field image are stitched together using the Siamese neural network to obtain an InSAR seamless collapse field fusion image. The InSAR seamless collapse field fusion image is then monitored to obtain the monitoring results for the monitoring range.

[0114] Specifically, such as Figure 5 As shown, the method for acquiring and monitoring seamless InSAR collapse field fusion images can be further divided into the following steps:

[0115] S151. The phase information and amplitude information of the InSAR collapse field image are obtained respectively, and the phase information and amplitude information are encoded in different channels through the InSAR data encoder to obtain InSAR collapse field encoded data, wherein the InSAR encoder adopts a convolutional neural network structure.

[0116] Understandably, InSAR collapse field image data is a two-dimensional image. Therefore, the InSAR encoder uses a convolutional neural network (CNN), which employs multiple convolutional and pooling layers to extract surface deformation features from the InSAR collapse field image data, resulting in better performance. Furthermore, the phase and amplitude information of the InSAR collapse field image are obtained separately and converted into different channel inputs to ensure that the Siamese neural network can capture complete surface displacement information. Preferably, a self-attention mechanism is also added to make the model pay more attention to areas with significant terrain changes, thereby improving the ability to capture important features.

[0117] S152. The 3D global collapse field image with weighted labels is encoded by a 3D point cloud data encoder to obtain 3D global collapse field encoded data, wherein the 3D point cloud data encoder adopts a point cloud convolutional neural network structure.

[0118] Understandably, the data of the 3D global collapse field image is characterized by irregular three-dimensional points. The Point-CNN convolutional neural network structure is a neural network structure specifically designed to process point cloud data, which can extract geometric features and local changes in the point cloud. Preferably, in order to obtain terrain change information at different scales, a multi-layer receptive field design can be adopted to extract local features and global deformation features of the point cloud. Preferably, since the point cloud data is relatively sparse, sparse convolution technology is used to reduce the amount of computation and retain the effective information in the data.

[0119] S153. The 3D global collapse field coding data and the InSAR collapse field coding data are input into the Siamese neural network. The Siamese neural network extracts information about the hole region in the InSAR collapse field image from the InSAR collapse field coding data, and uses the 3D global collapse field coding data and the information about the hole region to stitch and fill the hole region to obtain a seamless InSAR collapse field image.

[0120] Understandably, in the previous steps, error regions in the InSAR collapse field image were removed to form hole regions using a masking method. In this step, 3D global collapse field encoded data is used to fill the corresponding hole regions, thereby obtaining a seamless collapse field image.

[0121] Understandably, this application uses an interpolation algorithm—inverse distance weighted interpolation—to fill in the void regions, and performs point cloud thinning as needed to retain key points. For void regions caused by missing data or inconsistent resolution during the stitching process, inverse distance weighted interpolation is used for completion.

[0122] S154. The twin neural network extracts the weighted regions from the 3D global collapse field coding data as selected regions, obtains the corresponding selected regions in the InSAR seamless collapse field image based on the selected regions, and uses the 3D global collapse field coding data to perform permutation and fusion on the corresponding selected regions in the InSAR seamless collapse field image to obtain the InSAR seamless collapse field fused image.

[0123] S155. Monitor the InSAR seamless collapse field fusion images at different times and obtain the monitoring results of the monitoring range.

[0124] Understandably, using a Siamese neural network, the InSAR collapse field image and the 3D global collapse field image, after being processed by their respective encoders, are fused through a Siamese neural network, including:

[0125] • Feature stitching: The features of the InSAR collapse field image and the 3D global collapse field image are stitched together to form a joint feature vector.

[0126] • Shared weights: By using the shared weights of the Siamese neural network, the data from the two channels can share similar feature extraction methods in subsequent processing.

[0127] • Similarity calculation: By calculating the similarity between feature vectors, the Siamese neural network can monitor the data changes of the InSAR seamless collapse field fusion image at different times and accurately capture the collapse area.

[0128] Based on the same inventive concept, embodiments of this application also provide a ground subsidence monitoring system based on Siamese neural network fusion, such as... Figure 6 As shown, the ground subsidence monitoring system based on Siamese neural network fusion may include:

[0129] The system includes: a range acquisition module 311, an InSAR data acquisition and processing module 312, a 3D scan local data acquisition and processing module 313, a 3D scan global data acquisition module 314, and a stitching monitoring module 315. Among these:

[0130] The range acquisition module 311 is used to acquire the monitoring range of ground subsidence, and to pre-deploy control points and target points within the monitoring range;

[0131] In this embodiment, the range acquisition module 311 can be used to perform... Figure 1 For a detailed description of the range acquisition module 311 shown in step S110, please refer to the description of step S110.

[0132] InSAR data acquisition and processing module 312 is used to acquire real-time InSAR scan images of the monitoring range, perform a first preprocessing on the InSAR scan images, and acquire an InSAR collapse field image based on the InSAR scan images after the first preprocessing, wherein the InSAR scan images are acquired by interferometric synthetic aperture radar InSAR.

[0133] In this embodiment, the InSAR data acquisition-processing module 312 can be used to perform... Figure 1 For a detailed description of the InSAR data acquisition and processing module 312, see step S120 shown below.

[0134] The 3D scanning local data acquisition and processing module 313 is used to acquire several 3D local scanning images in real time within the monitoring range, perform a second preprocessing on the several 3D local scanning images, and process the several 3D local scanning images that have completed the second preprocessing through a Siamese neural network to acquire several 3D local collapse field images, wherein the 3D local scanning images are acquired by 3D laser scanning.

[0135] In this embodiment, the 3D scanning local data acquisition and processing module 313 can be used to perform... Figure 1 For a detailed description of step S130, the 3D scanning local data acquisition and processing module 313 can be found in the description of step S130.

[0136] The 3D scanning global data acquisition module 314 is used to formulate weight rules based on the control points and the target points, the InSAR collapse field image and several 3D local collapse field images, and then stitch together the several 3D local collapse field images after weight marking them according to the weight rules to obtain a 3D global collapse field image with weight markings.

[0137] In this embodiment, the 3D scanning global data acquisition module 314 can be used to perform... Figure 1 For a detailed description of the 3D scanning global data acquisition module 314 shown in step S140, please refer to the description of step S140.

[0138] The stitching monitoring module 315 is used to stitch the InSAR collapse field image and the weighted global collapse field image through the Siamese neural network to obtain an InSAR seamless collapse field fusion image, and to monitor the InSAR seamless collapse field fusion image to obtain the monitoring results of the monitoring range.

[0139] In this embodiment, the splicing monitoring module 315 can be used to perform... Figure 1For a detailed description of the splicing monitoring module 315 shown in step S150, please refer to the description of step S150.

[0140] Based on the same inventive concept, this embodiment also provides an electronic device. Figure 7 The diagram shows the structure of the electronic device of this embodiment, including a memory 21 and a processor 22. The memory 21 stores computer-readable instructions, and the processor 22 executes the computer-readable instructions to implement the ground subsidence monitoring method based on Siamese neural network fusion of this embodiment.

[0141] Preferably, the electronic device further includes a bus 23 and a communication interface 24, and the processor 22, the communication interface 24 and the memory 21 are connected through the bus 23.

[0142] The memory 21 may include high-speed random access memory (RAM), and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 24 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 23 may be an ISA bus, PCI bus, or EISA bus, etc. The bus 23 can be divided into address bus, data bus, control bus, etc. (not fully shown in the figure).

[0143] The processor 22 can be an integrated circuit chip with signal processing capabilities. In specific implementations, the steps in the embodiments of the above methods can be completed by the integrated logic circuits in the hardware of the processor 22 or by instructions in the form of software. The processor 22 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, which can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor, or the processor 22 can be any conventional processor 22, etc. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 21, and processor 22 reads information from memory 21 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0144] This invention also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are invoked and executed by the processor 22, they cause the processor 22 to implement the aforementioned ground subsidence monitoring method based on Siamese neural network fusion. For specific implementation details, please refer to the embodiments described above, which will not be repeated here.

[0145] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0146] Obviously, the above embodiments of this application are merely examples for clearly illustrating the technical solution of this application, and are not intended to limit the specific implementation of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of this application should be included within the protection scope of the claims of this application.

Claims

1. A ground subsidence monitoring method based on Siamese neural network fusion, characterized in that, Specifically, it includes: Obtain the monitoring range of ground subsidence, and pre-deploy control points and target points within the monitoring range; The monitoring range is monitored in real time. The InSAR scan image is preprocessed first, and the InSAR collapse field image is obtained based on the InSAR scan image after the first preprocessing. The InSAR scan image is obtained by interferometric synthetic aperture radar (InSAR). A number of 3D local scan images of the monitoring range in real time are acquired, and the number of 3D local scan images are subjected to a second preprocessing. The number of 3D local scan images after the second preprocessing are processed by a Siamese neural network to obtain a number of 3D local collapse field images, wherein the 3D local scan images are acquired by 3D laser scanning. The control point and the target point are selected as the first selected region; the obvious collapse areas in the InSAR collapse field image and several 3D local collapse field images are selected as the second selected region, wherein the attribute data of the obvious collapse areas meets a preset collapse data threshold; the collapse-prone areas are selected as the third selected region, wherein the collapse-prone areas meet a preset collapse risk attribute; the high-quality areas in several 3D local collapse field images are selected as the fourth selected region, wherein the attribute data of the high-quality areas meets a preset data quality threshold; weights are assigned to the first selected region, the second selected region, the third selected region, and the fourth selected region, and the first selected region is... Four independent weighting rules are established for the central region, the second selected region, the third selected region, and the fourth selected region, and their corresponding weights. Each 3D local collapse field image is traversed, and regions in each 3D local collapse field image are weighted according to the weighting rules to obtain several 3D local collapse field images with weighting labels. All weights on each 3D local collapse field image with weighting labels are obtained, and all weights are calculated to obtain the total weight of each 3D local collapse field image. The several 3D local collapse field images with weighting labels are then stitched together sequentially according to their corresponding total weights to obtain a 3D global collapse field image with weighting labels. The InSAR collapse field image and the weighted 3D global collapse field image are stitched together using the Siamese neural network to obtain a seamless InSAR collapse field fused image. The InSAR seamless collapse field fused image is then monitored to obtain the monitoring results for the monitoring range.

2. The ground subsidence monitoring method based on Siamese neural network fusion according to claim 1, characterized in that, The InSAR collapse field image is an aperture InSAR collapse field image, wherein the aperture region of the aperture InSAR collapse field image is obtained by removing several error regions using a masking method, wherein the attribute data of the error regions meet a preset error threshold.

3. The ground subsidence monitoring method based on Siamese neural network fusion according to claim 1, characterized in that, The step of assigning weights to regions in each of the 3D local collapse field images according to the weighting rules specifically includes: For regions of the 3D local collapse field image that satisfy one of the weighting rules, corresponding weight labels are applied. For regions of the 3D local collapse field image that satisfy two or more of the weighting rules, weights are superimposed and marked. For other regions of the 3D local collapse field image that do not meet the weighting rules, the weights of the other regions are all set to the initial weights.

4. The ground subsidence monitoring method based on Siamese neural network fusion according to claim 3, characterized in that, The initial weights are all smaller than the weights set in the weighting rules.

5. The ground subsidence monitoring method based on Siamese neural network fusion according to claim 1, characterized in that, The process involves stitching the InSAR collapse field image and the weighted 3D global collapse field image together using the Siamese neural network to obtain a seamless InSAR collapse field fused image. Monitoring this seamless InSAR collapse field fused image and obtaining monitoring results for the monitored area specifically includes: The phase information and amplitude information of the InSAR collapse field image are obtained respectively, and the phase information and amplitude information are encoded in different channels by the InSAR data encoder to obtain InSAR collapse field encoded data, wherein the InSAR encoder adopts a convolutional neural network structure. The 3D global collapse field image with weighted labels is encoded by a 3D point cloud data encoder to obtain 3D global collapse field encoded data, wherein the 3D point cloud data encoder adopts a point cloud convolutional neural network structure. The 3D global collapse field coding data and the InSAR collapse field coding data are input into the Siamese neural network. The Siamese neural network extracts information about the hole region in the InSAR collapse field image from the InSAR collapse field coding data, and uses the 3D global collapse field coding data and the information about the hole region to stitch and fill the hole region to obtain a seamless InSAR collapse field image. The twin neural network extracts weighted regions from the 3D global collapse field coding data as selected regions, obtains the corresponding selected regions in the InSAR seamless collapse field image based on the selected regions, and uses the 3D global collapse field coding data to perform permutation and fusion on the corresponding selected regions in the InSAR seamless collapse field image to obtain the InSAR seamless collapse field fused image. The InSAR seamless collapse field fusion images at different times are monitored to obtain the monitoring results for the monitoring range.

6. A ground subsidence monitoring system based on Siamese neural network fusion, characterized in that, include: The range acquisition module is used to acquire the monitoring range of ground subsidence, and to pre-deploy control points and target points within the monitoring range; The InSAR data acquisition and processing module is used to acquire real-time InSAR scan images of the monitoring range, perform a first preprocessing on the InSAR scan images, and acquire an InSAR collapse field image based on the InSAR scan images after the first preprocessing, wherein the InSAR scan images are acquired by interferometric synthetic aperture radar InSAR. A 3D scanning local data acquisition and processing module is used to acquire several 3D local scanning images in real time within the monitoring range, perform a second preprocessing on the several 3D local scanning images, and process the several 3D local scanning images that have completed the second preprocessing through a Siamese neural network to acquire several 3D local collapse field images, wherein the 3D local scanning images are acquired by 3D laser scanning. The 3D scanning global data acquisition module is used to: select the control point and the target point as the first selected region; select obvious collapse areas in the InSAR collapse field image and several 3D local collapse field images as the second selected region, wherein the attribute data of the obvious collapse areas meets a preset collapse data threshold; select areas prone to collapse as the third selected region, wherein the areas prone to collapse meet a preset collapse risk attribute; select high-quality areas in several 3D local collapse field images as the fourth selected region, wherein the attribute data of the high-quality areas meets a preset data quality threshold; and assign weights to the first selected region, the second selected region, the third selected region, and the fourth selected region, respectively. Four independent weighting rules are established for the first selected region, the second selected region, the third selected region, and the fourth selected region, and their corresponding weights. Each 3D local collapse field image is traversed, and regions in each 3D local collapse field image are weighted according to the weighting rules to obtain several 3D local collapse field images with weighting labels. All weights on each 3D local collapse field image with weighting labels are obtained, and all weights are calculated to obtain the total weight of each 3D local collapse field image. The several 3D local collapse field images with weighting labels are then stitched together sequentially according to their corresponding total weights to obtain a 3D global collapse field image with weighting labels. The stitching monitoring module is used to stitch the InSAR collapse field image and the weighted global collapse field image together using the Siamese neural network to obtain a seamless InSAR collapse field fused image, and to monitor the seamless InSAR collapse field fused image to obtain the monitoring results of the monitoring range.

7. A ground subsidence monitoring system based on Siamese neural network fusion according to claim 6, characterized in that, The 3D scanning global data acquisition module is also used for: For regions of the 3D local collapse field image that satisfy one of the weighting rules, corresponding weight labels are applied. For regions of the 3D local collapse field image that satisfy two or more of the weighting rules, weights are superimposed and marked. For other regions of the 3D local collapse field image that do not meet the weighting rules, the weights of the other regions are all set to the initial weights.

8. A ground subsidence monitoring system based on Siamese neural network fusion according to claim 6, characterized in that, The splicing monitoring module is also used for: The phase information and amplitude information of the InSAR collapse field image are obtained respectively, and the phase information and amplitude information are encoded in different channels by the InSAR data encoder to obtain InSAR collapse field encoded data, wherein the InSAR encoder adopts a convolutional neural network structure. The 3D global collapse field image with weighted labels is encoded by a 3D point cloud data encoder to obtain 3D global collapse field encoded data, wherein the 3D point cloud data encoder adopts a point cloud convolutional neural network structure. The 3D global collapse field coding data and the InSAR collapse field coding data are input into the Siamese neural network. The Siamese neural network extracts information about the hole region in the InSAR collapse field image from the InSAR collapse field coding data, and uses the 3D global collapse field coding data and the information about the hole region to stitch and fill the hole region to obtain a seamless InSAR collapse field image. The twin neural network extracts weighted regions from the 3D global collapse field coding data as selected regions, obtains the corresponding selected regions in the InSAR seamless collapse field image based on the selected regions, and uses the 3D global collapse field coding data to perform permutation and fusion on the corresponding selected regions in the InSAR seamless collapse field image to obtain the InSAR seamless collapse field fused image. The InSAR seamless collapse field fusion images at different times are monitored to obtain the monitoring results for the monitoring range.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the ground subsidence monitoring method based on Siamese neural network fusion as described in any one of claims 1-5.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements a ground subsidence monitoring method based on Siamese neural network fusion as described in any one of claims 1-5.

Images