Deformation monitoring method based on GNSS and AI video intelligent analysis

Abstract

This invention provides a deformation monitoring method based on GNSS and AI video intelligent analysis, including S1, acquiring high-precision remote sensing monitoring images of the target area to be measured; S2, homogenizing the remote sensing monitoring images to obtain clear surface feature image data; S3, preprocessing the surface feature image data to provide comprehensive data support for deformation early warning; and S4, using a trained LSTM model to assess terrain changes and predict deformation categories on the newly collected remote sensing monitoring images, generating early warning information. During the image preprocessing stage, the invention utilizes the reflectance characteristics of visible and near-infrared channels to accurately identify and process clouds and shadows, significantly improving image quality and usability.

Description

Technical Field

[0001] The present invention relates to the technical field of mountain terrain deformation analysis, and specifically to a deformation monitoring method based on GNSS and AI video intelligent analysis. Background Art

[0002] In the field of topographic mapping, traditional manual identification methods are inefficient. However, deep learning-based automatic recognition models, by fusing multiple data sources such as remote sensing imagery, digital elevation model (DEM), geological zoning, and river systems, can automatically identify landslide features, significantly improving the efficiency of screening for potential landslide hazards across large areas. These technologies utilize deep network architectures to extract complex features from optical image data and shallow network architectures to extract features from structured data such as altitude, geological composition, and river and fault zone distribution. These features are then integrated through a feature fusion module to obtain comprehensive landslide characteristics, enabling pixel-level classification and location of landslide terrain.

[0003] With the development of advanced technologies such as pre-trained large models, spatiotemporal information large-scale models, such as the "Sky Eyes" large-scale model, the "Tianquan" visual large-scale model, and the GEOVISEarth Star Map Earth Intelligence Engine, are evolving towards intelligentization, driving the transformation and upgrading of the industry. Current intelligent surveying and mapping utilizes modern technologies such as artificial intelligence, big data, cloud computing, and the Internet of Things to upgrade and optimize traditional surveying and mapping techniques, automating, intelligentizing, and refining the surveying and mapping process, thereby improving the productivity and accuracy of surveying and mapping tasks. These technologies not only process and analyze collected data, improving the accuracy and efficiency of surveying and mapping results, but also demonstrate powerful capabilities in a variety of fields, including reconnaissance, surveying and mapping in uninhabited areas, power inspections, and logistics distribution.

[0004] However, in practice, it was found that the terrain in mountainous areas is complex and traditional surveying and mapping methods face many difficulties, especially when monitoring terrain deformation in mountainous areas. Factors such as vegetation obstruction and water reflection will cause the accuracy of remote sensing images to decrease, which in turn affects the acquisition and analysis of terrain data. Vegetation cover blocks the line of sight, making it difficult to directly observe ground features, and the reflective characteristics of water bodies may distort the terrain features in the image, making terrain identification and classification more difficult. In addition, the complex terrain in mountainous areas may also lead to unstable signal reception, further increasing positioning errors. These factors work together to limit the application of traditional surveying and mapping methods in mountainous areas.

[0005] Based on this, more advanced technologies are needed, such as methods based on GNSS and AI video intelligent analysis, to improve the accuracy and efficiency of surveying and mapping. In practice, by integrating multi-source data and using deep learning models to automatically extract and analyze terrain features, more accurate terrain monitoring and identification can be achieved in complex environments. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the present invention aims to provide a deformation monitoring method based on GNSS and AI video intelligent analysis. In order to achieve the above object, the present invention is implemented through the following technical solutions: The deformation monitoring method based on GNSS and AI video intelligent analysis includes:

[0007] S1. Ground control point measurement and remote sensing equipment deployment

[0008] Deploy GNSS receivers and use GNSS technology to measure ground control points that characterize the terrain features of the target area to be measured, thereby obtaining high-precision remote sensing monitoring images of the target area to be measured;

[0009] S2. Remote sensing monitoring image preprocessing and feature extraction

[0010] Based on the reflectivity characteristics of clouds in visible light channels and near-infrared channels, cloud images and / or vegetation and soil images in the remote sensing monitoring image are identified by setting a reflectivity threshold, and the cloud images and / or vegetation and soil images are further confirmed by using a normalized difference cloud index; the reflectivity values ​​of adjacent non-shadowed vegetation areas are used for filling to achieve homogenization processing of the remote sensing monitoring image, so as to obtain clear surface feature image data;

[0011] S3. Deformation prediction model training and verification

[0012] Preprocessing the surface feature image data, fusing the preprocessed surface feature image data with other preset sensor device data, and constructing a device-level time series data vector to provide comprehensive data support for deformation warning;

[0013] The LSTM model is trained to extract the time series features reflecting the key indicators of terrain changes from the time series data vector, and the evaluation value C of the terrain changes in the target area to be tested is obtained. t,o , and simultaneously calculate the probability distribution P of the terrain change in the target area to be measured belonging to each deformation category based on the time series characteristics t,o ;

[0014] A loss function L is defined to measure the difference between the terrain change assessment value and deformation category probability distribution predicted by the LSTM model and the actual observed value, and the LSTM model parameters are updated to improve the LSTM model's prediction accuracy for terrain change;

[0015] S4. Prediction, monitoring and early warning

[0016] The trained LSTM model is used to evaluate terrain changes and predict deformation categories on newly collected remote sensing monitoring images to generate early warning information.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. In the image preprocessing stage, the present invention uses the reflectance characteristics of the visible light channel and the near-infrared channel to achieve accurate recognition and processing of clouds and shadows. By setting the reflectance threshold and calculating the normalized difference cloud index (NDCI), it can accurately identify cloud areas, and combine terrain data and sun position information to identify shadow areas. The reflectance value of the adjacent non-shadow area is used to fill and correct the shadow area, effectively eliminating the interference of clouds and shadows on surface feature analysis, significantly improving the quality and usability of the image, and providing a more accurate data basis for subsequent terrain feature extraction and deformation monitoring. At the same time, when processing complex terrain images, the present invention faces the problem that directly processing the entire image in the existing technology will lead to low computational efficiency and difficulty in capturing small-scale features. The present invention adopts image cropping and seamless splicing technology to determine the number of segments according to the number of processor threads and image size, and set the overlap ratio for cropping. After the cropped sub-image blocks are processed, the complete terrain image is reconstructed through a precise splicing algorithm to ensure the visual and data consistency of the spliced ​​image, further improving the monitoring accuracy.

[0019] 2. To effectively address the shortcomings of traditional surveying and mapping methods in complex mountainous environments, this invention utilizes high-precision GNSS technology for ground control point measurement, combined with AI video intelligent analysis, to obtain more accurate terrain data. This avoids the problem of reduced reflectivity caused by factors such as large terrain undulations, vegetation obstruction, and water reflection, thereby improving the accuracy and reliability of image data.

[0020] 3. By training the deformation prediction model, the present invention can evaluate terrain changes and predict deformation classification in real time, thereby generating accurate early warning information. At the same time, the LSTM model is used to extract key features from time series data and calculate terrain change assessment values ​​and deformation classification probabilities, which can provide the monitoring team with quantitative risk assessment and scientific basis. Once the early warning trigger conditions are met, the system will automatically generate early warning information, which includes areas where deformation may occur, deformation classification probabilities, and corresponding response measures. Recommendations can be made to quickly locate potential risk areas and take preventive measures, effectively improving the accuracy and real-time nature of early warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0022] Figure 1 Schematic diagram of the process of a deformation monitoring method based on GNSS and AI video intelligent analysis proposed in one embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of a remote sensing monitoring image obtained when a complex mountainous area in southwest China is selected as the target area to be measured according to an embodiment of the present invention;

[0024] Figure 3 The present invention is proposed in one embodiment of the present invention. Figure 2 The remote sensing monitoring images are preprocessed to obtain a schematic diagram of the surface feature dataset that has been corrected and fused;

[0025] Figure 4 According to the number of threads of the deformation monitoring system processor and the actual size of the surface feature image data, Figure 2 Schematic diagram of some measured slice image data obtained by cutting and splicing remote sensing monitoring images;

[0026] Figure 5 A schematic diagram of a process for generating early warning information and implementing response measures according to an embodiment of the present invention;

[0027] Figure 6 Schematic diagram of the deformation monitoring system framework proposed in one embodiment of the present invention. DETAILED DESCRIPTION

[0028] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0029] The present invention will be further described in detail below with reference to the accompanying drawings, but this does not limit the present invention.

[0030] like Figure 1 、 Figure 6 As shown, as an embodiment of the present invention, a deformation monitoring method based on GNSS and AI video intelligent analysis is proposed. In the image preprocessing stage, the reflection characteristics of the visible light channel and the near-infrared channel are used to realize the accurate identification and processing of clouds and shadows, and the reflectivity value of the adjacent non-shadow area is used to fill and correct the shadow area, effectively eliminating the interference of clouds and shadows on the surface feature analysis; in the model prediction stage, by training the deformation prediction model, it can evaluate the terrain changes and predict the deformation classification in real time, thereby generating accurate early warning information, and then use the LSTM model to extract key features from the time series data, and calculate the terrain change evaluation value and deformation classification probability, effectively improving the accuracy and real-time performance of the early warning. It includes the following specific steps:

[0031] S1. Ground control point measurement and remote sensing equipment deployment

[0032] Deploy GNSS receivers and use GNSS technology to measure the ground control points that characterize the terrain features of the target area to be measured, and calculate the precise coordinates P of each ground control point. i , preliminarily plan the flight path of the remote sensing equipment; determine an objective function C that meets the maximum coverage requirements and minimum cost conditions, and after optimizing and adjusting the flight path, obtain high-precision remote sensing monitoring images of the target area to be measured.

[0033] In one embodiment of the present invention, when executing step S1, in order to obtain the remote sensing monitoring data source of the optimal mountain terrain characteristics, it is necessary to ensure that the remote sensing equipment can effectively cover the mountain area to be measured, so as to obtain high-precision remote sensing monitoring data of the complex mountain area to be measured, as well as detailed information on the terrain characteristics. In this process, it is necessary to ensure that the ground control points can accurately reflect the key features of the terrain and provide an accurate reference for the motion trajectory planning of the remote sensing equipment. Through optimization and verification, the ground control points P i This can improve the coverage efficiency of the flight path and ensure the quality and accuracy of remote sensing monitoring data.

[0034] Based on the above technical concept, it can be understood that the preset ground control points include at least key terrain feature points used to characterize mountain tops, water body edges, and vegetation changes. The specific operations for acquiring remote sensing monitoring images are as follows:

[0035] S1-1. Construct a three-dimensional terrain model of the target area to be measured based on the measured data of ground control points, so that the deformation monitoring system can understand the subtle undulations of the terrain and the characteristics of small-scale water bodies. S1-2. Initially mark the small-scale information areas on the measured data, including at least small-scale water bodies, slight changes in vegetation cover, and subtle undulations of the terrain. S1-3. Determine the flight altitude of the remote sensing equipment based on the elevation changes of the three-dimensional terrain model to ensure that the remote sensing equipment fully covers all ground control points and small-scale information areas. And determine an objective function C that meets the maximum coverage requirements and the minimum cost conditions. Use optimization algorithms, such as genetic algorithms and simulated annealing, to adjust the remote sensing device position Q to minimize the objective function C, where d(P i ,Q) is the ground control point P i The distance from the remote sensing device position Q is expressed as a straight-line distance or a path distance calculated based on the actual flight path. λ is an adjustment factor used to balance the ground control point P i Distance and coverage C ov (Q,S j ), allowing the relative importance of these two factors to be adjusted during the optimization process, w j is the weight of the jth small-scale information region, which is used to reflect the importance of this region in deformation monitoring.j is the jth small-scale information region, C ov (Q,S j ) represents the remote sensing device position Q for the small-scale information area S j The coverage degree is in the range of [0,1], where 1 indicates complete coverage and 0 indicates no coverage. m is the total number of small-scale information areas, and n is the number of ground control points P. i S1-4, simulate the optimized flight path on the deformation monitoring system, further adjust the flight path according to the simulation results, and repeat the above steps until the coverage requirements and the objective function are met, and obtain the remote sensing monitoring image (such as Figure 2 shown).

[0036] In order to ensure that the motion trajectory of the remote sensing equipment can effectively cover the entire target area to be measured, in the above flight path optimization process, it is also necessary to use the gradient descent formula according to the objective function C For ground control point P i Real-time update is performed to ensure that the motion trajectory of the remote sensing equipment covers the entire target area to be measured. i new is the updated ground control point P i Coordinates, P i old is the current ground control point P i Coordinates, ground control point P i Coordinates are determined by the longitude, latitude and elevation of the geographic space in the three-dimensional model of the terrain. is the update amplitude, is the objective function C with respect to the ground control point P i The gradient of , indicating the direction of minimum cost.

[0037] As an embodiment of the present invention, the deformation monitoring method based on GNSS and AI video intelligent analysis further includes the following steps:

[0038] S2. Remote sensing monitoring image preprocessing and feature extraction. It should be noted that in existing remote sensing monitoring, in the visible light channel, the reflectivity of clouds is usually higher than that of the ground surface. This is because the water droplets and ice crystals in the clouds have a high scattering ability for the light in the visible light channel. In specific channels such as red, green, and blue, the high reflectivity of clouds makes them appear bright in satellite images. This characteristic is one of the key factors in identifying clouds. The reflectivity characteristics of the near-infrared channel: In the near-infrared channel, the reflectivity of clouds is usually lower than that of the ground surface. Compared with the visible light channel, the light of the near-infrared channel can penetrate the clouds more deeply because the clouds have a relatively weak scattering ability for near-infrared light. However, the water droplets and ice crystals in the clouds will still reflect some near-infrared light, but this reflection is usually lower than the reflectivity of the ground surface because the ground surface has a higher reflectivity in the near-infrared channel due to the reflective characteristics of vegetation and soil. Based on this, the reflectance characteristics of the visible light channel and the near-infrared channel can be used to distinguish surface features and clouds. That is, the visible light channel identifies clouds by detecting higher reflectivity because clouds reflect a large amount of sunlight, while the near-infrared channel further confirms clouds by detecting lower reflectivity because clouds have relatively low reflectivity in the near-infrared channel. At the same time, the higher reflectivity characteristics of the near-infrared channel are used to identify vegetation on the surface because the chlorophyll in the vegetation strongly reflects near-infrared light, thereby achieving accurate identification of the type of ground objects.

[0039] In one embodiment of the present invention, the specific operations of step S2 are as follows:

[0040] S2-1. Based on the reflectance characteristics of clouds in the visible light channel and near-infrared channel, cloud images and / or vegetation and soil images in remote sensing monitoring images are identified by setting a reflectance threshold, and the cloud images and / or vegetation and soil images are further confirmed by the Normalized Difference Cloud Index (NDCI).

[0041] It should be noted that the reflectivity threshold is obtained based on statistical analysis. Through statistical analysis, it is found that the reflectivity of clouds in the visible light channel is mostly higher than 0.5, so the threshold is set to 0.5. Among them, when judging cloud images and / or vegetation and soil images in remote sensing monitoring images based on the reflectivity threshold, when the reflectivity of the remote sensing monitoring image in the visible light channel is higher than the visible light reflectivity threshold of the cloud layer, and the reflectivity in the near-infrared channel is lower than the near-infrared reflectivity threshold of the cloud layer, then the remote sensing monitoring image area is considered to be cloud layer, and the identification result is recorded as 1; otherwise, the identification result is recorded as 0, and it is necessary to continue to further confirm the cloud image and / or vegetation and soil image through the normalized difference cloud index to optimize the accuracy of cloud layer identification in remote sensing monitoring images. The specific operations are as follows:

[0042] S2-11. Obtain the cloud reflectance data R in the visible light channel and near infrared channel respectively 可见光 ={R 可见光1 ,R可见光2 ,...,R 可见光r} and R 近红外 ={R 近红外1 ,R 近红外2 ,...,R 近红外r} as the basis for normalized difference cloud index data;

[0043] S2-12. Assign a weight α to each channel based on its reflection characteristics and contribution to cloud identification, reflecting the importance of the channel in cloud identification. i , For each channel, the NDCI value representing the difference in reflectance between the visible light channel and the near-infrared channel is calculated.

[0044] S2-13. Set the NDCI threshold th based on experience and / or statistical analysis. In specific implementation, the NDCI value in the cloud area is usually greater than 0.3, while the NDCI value in the non-cloud area is usually less than 0.3. Therefore, 0.3 is used as the NDCI threshold. If the NDCI value is greater than the NDCI threshold th, the remote sensing monitoring image is ultimately judged to be a cloud image. Otherwise, the remote sensing monitoring image is judged to be a non-cloud image and output as a vegetation and soil image.

[0045] Due to the undulating terrain or objects (such as clouds) blocking the sun's rays from directly reaching the ground, shadow areas will appear. Shadow areas will affect the accuracy and reliability of remote sensing monitoring. Specifically, shadow areas will reduce image quality, affect the accurate identification of surface features, lead to incomplete data, and interfere with the monitoring and analysis of terrain changes. In one embodiment of the present invention, in order to overcome this defect, the present invention also proposes to use shadow recognition and correction methods to improve the accuracy and reliability of remote sensing monitoring. The specific operation is as follows:

[0046] S2-14. After further confirming the cloud image and / or vegetation soil image, and before achieving homogenization processing of the remote sensing monitoring image, based on the identified cloud image and / or vegetation soil image, based on the mapping relationship between the ground control points and the terrain data and the sun position information, the shadow area that affects the accuracy of the remote sensing monitoring image is optimized and identified by iteratively normalizing the shading measurement value SOI, so as to eliminate the problem of reduced reflectivity caused by terrain undulation, shading due to changes in vegetation cover, and light reflection from small areas of water bodies. It should be noted that the relationship between terrain data and ground control points is generally approximated using a polynomial. The coefficients of the polynomial function can be determined by the known coordinates and elevation values ​​of the ground control points, thereby constructing a high-order polynomial model that can accurately describe the terrain characteristics. The relationship between the sun position information and the ground control points is described by the formula of the sun altitude angle and azimuth angle. The purpose is to use the terrain data and the sun position information, combined with the coordinates of the ground control points, to identify the shadow area caused by terrain undulation or object shading.

[0047] The calculation formula for the standardized shading metric SOI is:

[0048]

[0049] Where, X chr and Y int are the chromaticity and intensity of the pixel points of the cloud image and / or vegetation soil image in the shadow area to be identified. This pixel point is used to describe clouds, gaps between clouds, and soil or vegetation objects in the cloud image and non-cloud image. k chr and Y k int are the chromaticity and intensity of the kth neighboring region in the cloud image and / or vegetation and soil image in the shadow area to be identified. The neighboring region is the pixels surrounding the pixel point of the current cloud image and / or vegetation and soil image. If the pixel of the current cloud image and / or vegetation and soil image is part of the cloud image, then the neighboring region is the pixels adjacent to the pixel and can represent the cloud edge or vegetation and soil characteristics under the cloud in the cloud image under the shadow area to be identified. k is the weight of the kth neighboring area, K is the number of neighboring areas in the current shadow area to be identified, and μ is a constant between 0 and 1, which is used to balance the influence of the current shadow area to be identified and the neighboring areas.

[0050] Based on the above technical concept, it is understandable that in images of clouds and non-clouds, the boundaries between clouds are not very clear, or the brightness inside the clouds may be uneven. Therefore, through the above iterative summation optimization of SOI, the influence of the pixels surrounding each pixel can be taken into account, which helps smooth the cloud boundaries, making cloud recognition more accurate, and helping to identify the outline and internal structure of the cloud, as well as the transition area between clouds and non-clouds. At the same time, since shadow areas in images of clouds and non-clouds may be caused by clouds blocking sunlight, it is necessary to set a threshold to identify areas with lower SOI values ​​in order to determine which areas are shadow areas blocked by clouds and accurately identify the location and shape of the clouds. For example, a threshold of 0.1 is set as the standard for distinguishing between shadow and non-shadow areas. When the updated SOI value of a pixel is lower than this threshold, it is identified as a shadow area, indicating that the area may be blocked by clouds, resulting in insufficient sunlight. Conversely, if the SOI value of a pixel is greater than or equal to this threshold, it is considered a non-shadow area, meaning that these areas are not blocked by clouds or the degree of blockage is not significant. So far, in this way, the shadow and non-shadow areas in the image can be clearly distinguished.

[0051] Example: Figure 3 The remote sensing monitoring image in Figure 3 In the following table, the cloud data is marked as (03 / 06 / 07 / 08 / 09), the vegetation data is marked as (06 / 07 / 08 / 09), and the soil data is marked as (06 / 07 / 08 / 09).

[0052]

[0053] In the above data table, the above method can intuitively show the impact of the shadow of the cloud area on the vegetation area below: for example, the shadow of cloud 03 causes the SOI value of vegetation 06 below to decrease. Similarly, the shadow of cloud 07 affects vegetation 07 and soil 07. This impact is reflected in the data table through the reduction of the SOI value after optimization. In addition, the SOI value of the vegetation area is also affected by the shadow of the clouds above, while the SOI value of the soil area is affected by the vegetation above. It should be noted that the above progressive logical relationship simulates the transmission effect of light and shadow in the natural environment, making the information in the data table more consistent with actual physical processes. In this way, the deformation monitoring system can more accurately analyze and understand the shadow areas in the remote sensing monitoring image and their impact on the surrounding environment.

[0054] S2-2. Use the reflectance value of the adjacent non-shadow vegetation area for filling to achieve uniform processing of the remote sensing monitoring image to obtain clear surface feature image data.

[0055] Based on the above technical concept, it should be noted that for the identified shadow area, when using the reflectance value of the adjacent non-shadow vegetation area to fill, it is necessary to select a non-shadow area with similar terrain and similar lighting conditions to the shadow area, and use its reflectance as a reference to perform the filling of the shadow area, so that the acquired remote sensing monitoring image becomes surface feature image data that accurately reflects the actual surface conditions. The specific operations for performing the filling are as follows:

[0056] S2-21, respectively determine a set of pixel points C in the cloud layer image identified based on the reflectivity threshold,

[0057] C = {(x0, y0)|(x0, y0) are pixels in the cloud image}, which are extracted directly from the cloud image and represent the location and range of the clouds and the set N of pixels in the non-cloud image. N = {(x1, y1)|(x1, y1) are pixels in the non-cloud image}, which are candidate pixels for filling the cloud and shadow areas. They provide the reflectance values ​​that the cloud and shadow areas should have. There is also a set S of pixels in the shadow area identified based on the iterative normalized shading metric (SOI). S = {(x2, y2)|(x2, y2) are pixels in the shadow area}, which need to be identified and filled to eliminate the problem of reduced reflectance caused by terrain undulation, occlusion by changes in vegetation cover, and light reflection from small areas of water.

[0058] S2-22. For each shadow area pixel (x2, y2), find the nearest non-cloud image pixel (x1, y1) and calculate the Euclidean distance d between them. And select the nearest non-cloud image pixel (x1, y1), Assign its reflectivity value I(x1,y1) to the corresponding shadow area pixel point (x2,y2) for filling:

[0059]

[0060] In one embodiment of the present invention, since high-resolution image processing has high requirements for computing resources, directly processing the entire image will lead to low efficiency. Therefore, the present invention also proposes the following S2-3 steps: based on the number of threads of the deformation monitoring system processor and the actual size of the surface feature image data, the surface feature image data is segmented to determine the maximum number of segments that does not exceed the number of processor threads; then, the starting point of each segment is determined avoiding the small-scale information area, and the overlap ratio is set between the sub-surface feature image data obtained after segmentation, and a new starting point is selected to perform cropping and splicing of the sub-surface feature image data to ensure that the data processing efficiency is improved without losing resolution accuracy, and the complexity of subsequent data processing is reduced, thereby effectively balancing the computing load and improving processing efficiency. In this way, it can be ensured that the number of image segments does not exceed the number of processor threads, so that each thread can process an image block in parallel, thereby speeding up the overall processing speed. At the same time, by reasonably setting the size and overlap ratio of the segments, the continuity and accuracy of the image during subsequent splicing can be guaranteed, ensuring that the quality of the final spliced ​​image meets the requirements of terrain feature extraction and analysis, improving the efficiency of image processing, retaining key terrain features, and obtaining complete and high-quality complex mountain terrain images to be tested, providing an accurate data basis for subsequent terrain feature extraction and analysis.

[0061] The specific operations of step S2-3 are as follows:

[0062] S2-31. Determine the segmentation parameters and cropping starting point: Based on the number of processor threads T and the actual image size I, calculate the maximum number of segmentation blocks n that does not exceed the number of processor threads, ensuring that n ≤ T and n × B ≥ I, where B is the target area of ​​each segmentation block and represents the size. At the same time, based on the obtained pre-processed (cloud removal, shadow identification, and radiation correction) surface feature dataset and the marked small-scale information area, determine the starting point of each segmentation block. These starting points should avoid small-scale information areas as much as possible to ensure that these key areas will not be segmented in subsequent processing, thereby completely preserving their feature information and providing accurate reference points for subsequent image stitching. It should be noted that the starting point of each segmentation block should meet the following requirements: for the starting point X in the X direction start :X start =max(X min ,X prev +BA overlap )(8), where X prev is the x coordinate of the end point of the previous segmentation block, A overlap The size of the overlapping area between the segments, for the starting point Ystart in the Y direction: Where Y min is the minimum y coordinate of the surface feature dataset (image), n Xis the number of blocks that the surface feature dataset (image) can be segmented in the X direction, and k is the index of the current segmentation block in the Y direction.

[0063] S2-32. Perform image cropping: Use the determined segmentation parameters and starting point to crop the surface feature dataset (image) preprocessed in step S2-31. During the cropping process, it is necessary to ensure that there is a certain overlap ratio between each segmented block. It can be understood that the setting of this overlapping area is based on the distribution of the identified key terrain features and small-scale information areas, as well as the texture and color characteristics of the image. In specific implementation, the size of the overlapping area should be sufficient to cover the existing terrain feature changes and the edge parts of the small-scale information, so that these features can be accurately aligned and fused during subsequent splicing to avoid feature information loss or splicing errors due to cropping.

[0064] S2-33. Perform image stitching: Stitch the cropped sub-image blocks together, ensuring visual and data consistency. Ultimately, a complete and high-quality image of the complex mountainous terrain to be measured is obtained, providing an accurate data foundation for subsequent terrain feature extraction and analysis. This improves stitching quality.

[0065] Based on the above technical concept, Figure 3 - Figure 4 As shown in the figure, the mountainous image of the experimental area contains multiple small streams and complex ridge lines. Through the cropping steps above, the image blocks of each stream and ridge line can be processed separately, which helps the model to identify these features more accurately. Then, through stitching, these processed image blocks are recombined into a complete terrain view, which enables the subsequent prediction model to make more accurate deformation predictions based on complete terrain information. Without the cropping step, directly processing the entire large image may lead to low computational efficiency and difficulty in accurately capturing key terrain features at small scales. The model will only be able to make predictions based on fragmented image blocks, which leads to inaccurate or incomplete prediction results.

[0066] As an embodiment of the present invention, the deformation monitoring method based on GNSS and AI video intelligent analysis further includes step S3, deformation prediction model training and verification. In one embodiment of the present invention, the specific operations of step S3 are as follows:

[0067] S3-1. Preprocess the surface feature image data, fuse the preprocessed surface feature image data with the data of other sensor devices pre-deployed to collect data required for subsequent terrain deformation monitoring, and construct a device-level time series data vector to provide comprehensive data support for deformation early warning.

[0068] S3-2. Train the LSTM model to extract the time series features reflecting the key indicators of terrain changes from the time series data vector, and obtain the evaluation value C of the terrain change in the target area to be tested. t,o , as a quantitative indicator of terrain stability. And based on the time series characteristics, the probability distribution P of the terrain change in the target area to be tested belonging to each registered deformation category (such as landslide, subsidence, etc.) is calculated. t,o .

[0069] Based on the above technical concept, the evaluation value C of the terrain change in the target area to be measured is obtained. t,o The specific steps include:

[0070] The first step is to train the deformation prediction model to extract time series features. The specific operations are as follows:

[0071] First, the surface feature image data is normalized and scaled to eliminate the dimension difference. It can be understood that the normalization interval is [0, 1] or [-1, 1] to simplify the calculation;

[0072] Secondly, the pre-processed surface feature image data is fused with the data of other sensor devices that are pre-arranged to collect data required for subsequent terrain deformation monitoring to construct the device-level time series data vector f t,o It should be noted that in the actual target monitoring area, a series of sensors need to be pre-deployed to collect the data required for subsequent terrain deformation monitoring. These sensors need to include at least GNSS receivers, displacement meters, and strain meters to provide accurate measurement data on terrain changes. The data collected by these sensors are then fused with the high-resolution image data obtained through the image processing system. Through this data fusion, a more comprehensive device-level time series dataset is constructed. It not only contains image information, but also integrates the terrain change data monitored by the sensors, providing key information for the deformation prediction model, enabling it to more accurately analyze and predict possible terrain deformation, thereby improving the reliability and effectiveness of the early warning system.

[0073] f t,o =[Loc o ,Dis t,o ,Hum t,o ,Veg t,o ,Den t,o ,Slip t,o ,Crk t,o ,GWL t,o ,Temp t,o ,Str t,o ,Slo t,o ](10), where

[0074] Loco For the location information of other sensor devices, Dis t,o ,Hum t,o ,Veg t,o ,Den t,o ,Slip t,o ,Crk t,o ,GWL t,o ,Temp t,o ,Str t,o ,Slo t,o They are terrain displacement data, soil moisture data, vegetation coverage data, soil density data, rock slip velocity data in the target test area, terrain crack width data, terrain groundwater level data, terrain surface temperature data, terrain surface stress data, and terrain slope data obtained based on other sensor equipment;

[0075] Again, the input time series data vector f t,o As the input of the LSTM model, through the forward propagation process of the LSTM model, the time series data vector f is processed by using its internal gating mechanism and loop structure. t,o Processing is performed to extract the feature vector F that can reflect the characteristics of the time series t,o , F t,o =LSTM(f t,o ,F t-1,o ), where f t,o is the device-level time series data vector at time step T, F t-1,o is the feature vector at time step t-1, F t,o is the feature vector at time step T. Through steps S3-1 to S3-2, high-quality information is extracted from the original high-resolution image data and fused with other sensor data to form device-level time series data for subsequent deformation warning analysis to predict and locate possible deformation areas. The preprocessed data is integrated into device-level time series data, which includes sensor readings and position information of each device at different time points.

[0076] The second step is to use the eigenvector F t,o and other sensor device location information Loc o , through the LSTM model parameter W C and b c Calculate the evaluation value C of terrain change t,o :C t,o =W C ×F t,o +b c (11), where W C Input gate weight matrix for LSTM model, b cThe gate bias vector is input to the LSTM model. This step allows for a quantitative assessment of terrain changes based on data from remote sensing equipment, providing a basis for predicting areas of potential deformation. For example, if a terrain change assessment reveals a decrease in stability in a particular area, this could be a precursor to an impending landslide or other terrain disaster, triggering further analysis and response from the early warning system.

[0077] Calculate the probability distribution P based on time series characteristics t,o The formula is as follows: t,o =softmax(W p ×F t,o +b p )(12), where W p is a weight matrix of the LSTM model, which is used to convert time series features into probability distribution of deformation categories. During the LSTM model training process, W p Adjust according to the training parameters, b p Output gate bias vector for the LSTM model.

[0078] S3-3. Define the loss function L to measure the difference between the terrain change evaluation value and deformation category probability distribution predicted by the LSTM model and the actual observed value. Use an optimization algorithm (such as Adam) to update the model parameters, minimize the loss function, and update the LSTM model parameters to improve the LSTM model's prediction accuracy for terrain changes.

[0079] The calculation formula of the loss function L is as follows: Where T is the length of the time series, indicating the total number of time steps; O is the total number of terrain feature points, indicating the total number of terrain change assessment values; J is the total number of deformation categories, indicating the total number of deformation classification labels; C t,o is the evaluation value of the terrain change of the o-th terrain feature point at the t-th time step predicted by the LSTM model, is the estimated value of the actual observed terrain change, P t,o,j is the probability of the j-th deformation category at the t-th time step predicted by the LSTM model, is the probability of the deformation category actually observed.

[0080] As an embodiment of the present invention, Figure 5 As shown, the proposed deformation monitoring method based on GNSS and AI video intelligent analysis also includes step: S4, prediction monitoring and early warning. In one embodiment of the present invention, the specific operations of step S4 are as follows:

[0081] The same preprocessing and time series feature extraction as in S2 to S3 are performed on the newly collected remote sensing monitoring images. The trained LSTM model is used to evaluate the terrain change and predict the deformation category of the newly collected remote sensing monitoring images. The terrain change evaluation value C after prediction is used. t,o And the probability distribution P of deformation category t,o , generate early warning information.

[0082] It should be noted that the warning information at least includes the area where deformation may occur, the probability distribution P of deformation category, and the probability distribution P of deformation category. t,o And the corresponding response measures are recommended. If the terrain change assessment value C obtained in the target area to be tested is t,o And the probability distribution P of deformation category t,o Satisfaction: trigger warning = (C t,o >TC)V(P t,o >TP), the deformation monitoring system generates early warning information, where TC and TP are the preset evaluation threshold and classification threshold respectively. The early warning information is shown as follows:

[0083]

[0084]

[0085] If it is below the preset threshold, the system will not trigger an early warning, but it will need to continuously monitor the data in the area, record the current terrain change assessment and historical data for future analysis and model training, and take steps such as further analysis, regular evaluation, report generation, notification of relevant personnel, and preventive measures as needed.

[0086] As a second aspect of the present invention, Figure 6 As shown in the figure, a deformation monitoring system based on GNSS and AI video intelligent analysis is proposed. The specific implementation includes:

[0087] Data acquisition module: collects high-precision location information of ground control points and remote sensing monitoring images of the target area to be measured through the Global Navigation Satellite System (GNSS) receiver; AI video analysis and processing module: performs intelligent analysis of data, identifies and tracks changes in terrain features; image processing module: pre-processes remote sensing monitoring images, including cloud removal, shadow removal, radiation correction, and image segmentation and stitching to improve image quality and analysis accuracy; deformation prediction model module: analyzes and predicts the collected deformation data; early warning and response module: generates early warning information based on the output of the deformation prediction model, and proposes corresponding response measures to prevent or mitigate possible terrain disasters; user interface module: provides an intuitive user interface that enables operators to easily access monitoring data, analysis results and early warning information, and perform corresponding operations.

[0088] The technical scope of the present invention is not limited to the contents of the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A deformation monitoring method based on GNSS and AI video intelligent analysis, characterized by: Including steps: S1. Ground control point measurement and remote sensing equipment deployment Deploy GNSS receivers and use GNSS technology to measure ground control points that characterize the terrain features of the target area to be measured, thereby obtaining high-precision remote sensing monitoring images of the target area to be measured; S2. Remote sensing monitoring image preprocessing and feature extraction Based on the reflectivity characteristics of clouds in visible light channels and near-infrared channels, cloud images and / or vegetation and soil images in the remote sensing monitoring image are identified by setting a reflectivity threshold, and the cloud images and / or vegetation and soil images are further confirmed by using a normalized difference cloud index; the reflectivity values ​​of adjacent non-shadowed vegetation areas are used for filling to achieve homogenization processing of the remote sensing monitoring image, so as to obtain clear surface feature image data; S3. Deformation prediction model training and verification Preprocessing the surface feature image data, fusing the preprocessed surface feature image data with other preset sensor device data, and constructing a device-level time series data vector to provide comprehensive data support for deformation warning; The LSTM model is trained to extract the time series features reflecting the key indicators of terrain changes from the time series data vector, and the evaluation value C of the terrain changes in the target area to be tested is obtained. t,o , and simultaneously calculate the probability distribution P of the terrain change in the target area to be measured belonging to each deformation category based on the time series characteristics t,o ; A loss function L is defined to measure the difference between the terrain change assessment value and deformation category probability distribution predicted by the LSTM model and the actual observed value, and the LSTM model parameters are updated to improve the LSTM model's prediction accuracy for terrain change; S4. Prediction, monitoring and early warning The trained LSTM model is used to evaluate terrain changes and predict deformation categories on newly collected remote sensing monitoring images to generate early warning information.

2. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 1 is characterized in that: The reflectivity threshold is obtained based on statistical analysis and is set to 0.5, where: When determining whether a cloud image and / or a vegetation and soil image in the remote sensing monitoring image is cloud based on the reflectivity threshold, if the reflectivity of the remote sensing monitoring image in the visible light channel is higher than the visible light reflectivity threshold of the cloud layer, and the reflectivity in the near-infrared channel is lower than the near-infrared reflectivity threshold of the cloud layer, the remote sensing monitoring image area is considered to be cloud layer, and the identification result is recorded as 1; Otherwise, the identification result is recorded as 0. It is necessary to further confirm the cloud image and / or vegetation and soil image through the normalized difference cloud index to optimize the accuracy of cloud identification in remote sensing monitoring images. The specific operation is as follows: First, obtain the cloud reflection data R in the visible light channel and near infrared channel respectively. 可见光 ={R 可见光1 ,R 可见光2 ,...,R 可见光r } and R 近红外 ={R 近红外1 ,R 近红外2 ,...,R 近红外r } as the basis for normalized difference cloud index data; Secondly, a weight α is assigned to each channel based on its reflection characteristics and contribution to cloud recognition, which reflects the importance of the channel in cloud recognition. i , For each channel, the NDCI value representing the difference in reflectance between the visible light channel and the near-infrared channel is calculated. Finally, the NDCI threshold th is set based on experience and / or statistical analysis. If the NDCI value is greater than the NDCI threshold th, the remote sensing monitoring image is ultimately identified as a cloud image. Otherwise, the remote sensing monitoring image is identified as a non-cloud image and output as a vegetation and soil image.

3. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 1 or 2, characterized in that: After further confirming the cloud image and / or vegetation soil image, and before achieving the homogenization processing of the remote sensing monitoring image, it is necessary to optimize and identify the shadow area that affects the accuracy of the remote sensing monitoring image by iteratively normalizing the light shading index value SOI based on the identified cloud image and / or vegetation soil image, so as to eliminate the reflectivity reduction problem caused by terrain undulation, vegetation cover change, and light reflection from small water bodies. The calculation formula of the normalized light shading index SOI is: Where, X chr and Y int are the chromaticity and intensity of the pixel points of the cloud image and / or vegetation soil image in the shadow area to be identified. This pixel point is used to describe clouds, gaps between clouds, and soil or vegetation objects in the cloud image and non-cloud image. k chr and Y k int are the chromaticity and intensity of the kth neighboring region in the cloud image and / or vegetation and soil image in the shadow area to be identified. The neighboring region is the pixels surrounding the pixel point of the current cloud image and / or vegetation and soil image. If the pixel of the current cloud image and / or vegetation and soil image is part of the cloud image, the neighboring region is the pixels adjacent to the pixel and can represent the cloud edge or vegetation and soil characteristics under the cloud in the cloud image under the shadow area to be identified. k is the weight of the kth neighboring area, K is the number of neighboring areas in the current shadow area to be identified, and μ is a constant between 0 and 1, which is used to balance the influence of the current shadow area to be identified and the neighboring areas.

4. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 3 is characterized by: When filling the identified shadow area using the reflectance value of the adjacent non-shadow vegetation area, it is necessary to select a non-shadow area with similar terrain and lighting conditions to the shadow area, and use its reflectance as a reference to fill the shadow area, so that the acquired remote sensing monitoring image becomes surface feature image data that accurately reflects the actual surface conditions. The specific operations for performing the filling are as follows: First, determine the pixel set C in the cloud image identified based on the reflectivity threshold, C = {(x0, y0)|(x0, y0) is a pixel in the cloud image} and the pixel set N in the non-cloud image, N = {(x1, y1)|(x1, y1) is a pixel in the non-cloud image}, as well as the pixel set S in the shadow area identified based on the iterative normalized shading index value SOI, S = {(x2, y2)|(x2, y2) is a pixel in the shadow area}; Secondly, for each shadow area pixel (x2, y2), find the nearest non-cloud image pixel (x1, y1) and calculate the Euclidean distance between them. And select the nearest non-cloud image pixel (x1, y1), Assign its reflectivity value I(x1,y1) to the corresponding shadow area pixel point (x2,y2) for filling:

5. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 1 is characterized in that: Before training the deformation prediction model, it is also necessary to determine the maximum number of segmentation blocks that does not exceed the number of processor threads based on the number of threads of the deformation monitoring system processor and the actual size of the surface feature image data for surface feature image data segmentation. Then, the starting point of each segmentation block is determined by avoiding the small-scale information area, and the overlap ratio is set between the sub-surface feature image data obtained after segmentation. The new starting point is selected for cropping and splicing of the sub-surface feature image data to ensure that the efficiency of data processing is improved without losing resolution accuracy and reduce the complexity of subsequent data processing.

6. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 1 is characterized in that: The preset ground control points include at least key terrain feature points for characterizing mountain tops, water body edges, and vegetation changes. The specific operations for acquiring remote sensing monitoring images are as follows: First, a three-dimensional terrain model of the target area to be measured is constructed based on the measured data of the ground control points. Second, small-scale information areas, including at least small water bodies, changes in vegetation cover, and subtle undulations of the terrain, are initially marked on the measured data. Third, the flight altitude of the remote sensing device is determined based on the elevation changes of the three-dimensional terrain model, and an objective function C is determined that meets the maximum coverage requirement and the minimum cost condition. Use the optimization algorithm to adjust the remote sensing device position Q to minimize the objective function C, where d(P i ,Q) is the ground control point P i The distance from the remote sensing device position Q, λ is the adjustment factor used to balance the ground control point P i Distance and coverage C ov (Q,S j ), the weight between j is the weight of the jth small-scale information region, which is used to reflect the importance of this region in deformation monitoring. j is the jth small-scale information region, C ov (Q,S j ) represents the remote sensing device position Q for the small-scale information area S j The coverage degree is in the range of [0,1], where 1 indicates complete coverage and 0 indicates no coverage. m is the total number of small-scale information areas, and n is the number of ground control points P. i total; Finally, the optimized flight path is simulated on the deformation monitoring system, and the flight path is further adjusted based on the simulation results. The above steps are repeated until the coverage requirements and the objective function are minimized, and the remote sensing monitoring image is obtained.

7. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 6 is characterized in that: In the process of flight path optimization, it is also necessary to use the gradient descent formula according to the objective function C For the ground control point P i Real-time updates are performed to ensure that the motion trajectory of the remote sensing equipment covers the entire target area to be measured, where: is the updated ground control point P i coordinate, is the current ground control point P i Coordinates, ground control point P i The coordinates are determined based on the longitude, latitude and elevation of the geographic space in the three-dimensional terrain model. is the update amplitude, is the objective function C with respect to the ground control point P i The gradient of , indicating the direction of minimum cost.

8. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 1 is characterized in that: Get the evaluation value C of the terrain change in the target area to be tested t,o The specific steps include: The first step is to train the deformation prediction model to extract time series features. The specific operations are as follows: First, normalization interval scaling is performed on the surface feature image data to eliminate dimension differences; Secondly, the pre-processed surface feature image data is fused with the data of other sensor devices that are pre-arranged to collect data required for subsequent terrain deformation monitoring to construct the device-level time series data vector f t,o , f t,o =[Loc o ,Dis t,o ,Hum t,o ,Veg t,o ,Den t,o ,Slip t,o ,Crk t,o ,GWL t,o ,Temp t,o ,Str t,o ,Slo t,o ], where Loc o For the location information of other sensor devices, Dis t,o ,Hum t,o ,Veg t,o ,Den t,o ,Slip t,o ,Crk t,o ,GWL t,o ,Temp t,o ,Str t,o ,Slo t,o They are terrain displacement data, soil moisture data, vegetation coverage data, soil density data, rock slip velocity data in the target test area, terrain crack width data, terrain groundwater level data, terrain surface temperature data, terrain surface stress data, and terrain slope data obtained based on other sensor equipment; Again, the input time series data vector f t,o As the input of the LSTM model, through the forward propagation process of the LSTM model, the time series data vector f is processed by using its internal gating mechanism and loop structure. t,o Processing is performed to extract the feature vector F that can reflect the characteristics of the time series t,o , F t,o =LSTM(f t,o ,F t-1,o ), where f t,o is the device-level time series data vector at time step T, F t-1,o is the feature vector at time step t-1, F t,o is the feature vector at time step T; The second step is to use the eigenvector F t,o and other sensor device location information Loc o , through the LSTM model parameter W C and b c Calculate the evaluation value C of terrain change t,o :C t,o =W C ×F t,o +b c , where W C Input gate weight matrix for LSTM model, b c Input gate bias vector for LSTM model.

9. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 1 or 8, characterized in that: Calculate the probability distribution P based on the time series characteristics t,o The formula is as follows: t,o =softmax(W p ×F t,o +b p ), where W p is a weight matrix of the LSTM model, which is used to convert time series features into probability distribution of deformation categories. During the LSTM model training process, W p Adjust according to the training parameters, b p Output gate bias vector for LSTM model; The calculation formula of the loss function L is as follows: Where T is the length of the time series, indicating the total number of time steps; O is the total number of terrain feature points, indicating the total number of terrain change assessment values; J is the total number of deformation categories, indicating the total number of deformation classification labels; C t,o is the evaluation value of the terrain change of the o-th terrain feature point at the t-th time step predicted by the LSTM model, is the estimated value of the actual observed terrain change, P t,o,j is the probability of the j-th deformation category at the t-th time step predicted by the LSTM model, is the probability of the deformation category actually observed.

10. The deformation monitoring method based on GNSS and AI video intelligent analysis according to claim 1, characterized in that: The specific steps of prediction, monitoring and early warning include: In the first step, the same preprocessing and time series feature extraction as in S2 to S3 are performed on the newly collected remote sensing monitoring images, and the trained LSTM model is used to evaluate terrain changes and predict deformation categories for the newly collected remote sensing monitoring images. The second step is to evaluate the value C based on the predicted terrain change. t,o And the probability distribution P of deformation category t,o , generate warning information, where if the terrain change assessment value C obtained in the target area to be tested is t,o And the probability distribution P of deformation category t,o Satisfaction: trigger warning = (C t,o >TC)V(P t,o >TP), the deformation monitoring system generates warning information, where TC and TP are the preset evaluation threshold and classification threshold respectively. The generated warning information at least includes: indicating the area where deformation may occur, providing the deformation category probability P t,o , provide suggestions for response measures to early warnings; In the third step, further analysis, regular evaluation, report generation, notification of relevant personnel, and preventive measures are taken as needed.

Images