Slope displacement monitoring method based on computer vision

Abstract

The invention relates to the technical field of side slope displacement monitoring, and discloses a side slope displacement monitoring method based on computer vision, which comprises the following steps: step 1, acquiring multi-temporal remote sensing image data of a monitoring area, constructing a time sequence based on a vegetation index, analyzing a disturbance trend of a sub-area, and generating a disturbance risk level map; 2, initial image feature points are extracted based on the slope monitoring image, the feature points are constructed into a graph structure according to the spatial proximity relation, and a graph data model with nodes connected with edges is formed; and step 3, based on the disturbance risk level graph and the graph structure, inputting the graph neural network model to carry out feature point stability modeling. The technical scheme of disturbance risk level map guidance, map neural network stability modeling and feature point stability score screening is adopted, and the technical effect that stable and traceable key feature points can be effectively recognized and screened out under the dynamic disturbance conditions of vegetation disturbance, climate change and the like is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of slope displacement monitoring, and in particular to a slope displacement monitoring method based on computer vision. Background Art

[0002] Changes in slope displacement are often a precursor to geological hazards such as landslides and collapses. By monitoring slope displacement, anomalies can be detected, warning of potential geological disasters and enabling appropriate measures to avoid or mitigate the resulting losses. Therefore, displacement monitoring is a key parameter in slope monitoring. Currently, commonly used monitoring methods include total stations, target imaging methods, and connecting pipes.

[0003] For example, the Chinese invention application with publication number CN107741211A discloses a slope displacement monitoring method, which includes the following steps: (1) cleaning the slope surface; (2) after cleaning, laying a horizontal borehole with a diameter of 80 to 150 mm in the middle and lower part of the slope, and the depth of the borehole is 8 to 20 m; (3) setting a measuring rod on the first, second, and third anchor heads; (4) successively extending the first, second, and third anchor heads of the measuring rod into the bottom, middle, and front of the borehole, injecting concrete slurry to fix the anchor heads to the borehole rock wall, and then leaving it to stand for 12 to 48 hours; (5) setting a mounting base at the borehole mouth, setting a sensor on the base, connecting one end of the sensor to one end of the measuring rod, and connecting the other end of the sensor to an external wire; (6) installing a protective cover with a wire through hole on the outside of the sensor, and leading out the wire; (7) connecting the wire to a reader, and the reader reads the initial reading of the sensor and records the reading. The present invention is accurate in measuring slope displacement and is economical and practical.

[0004] For example, a Chinese invention application with publication number CN118936323A discloses a slope displacement monitoring method, in which a monitoring starting point and a monitoring target point are set on the surface of the slope where displacement monitoring is required, and one or more relay monitoring points are set between the monitoring starting point and the monitoring target point to form a sequence of monitoring points arranged in sequence, and each monitoring point in the monitoring point sequence is spaced in the horizontal lateral direction; then, the displacement change of the adjacent subsequent monitoring points in the monitoring point sequence is monitored in sequence with the preceding monitoring point as the reference point.

[0005] The shortcomings of the above patents are:

[0006] In real-world slope environments, there is a large amount of vegetation cover, which can change frequently due to external factors such as wind sway, seasonal changes, or rain and snow. This can lead to significant differences in the position and appearance of visual feature points in images of the same area captured at different times. Traditional displacement monitoring methods based on image feature point matching often rely on the stable visibility of feature regions in time-series images. However, under such dynamic interference, feature points are prone to being unrecognized or mismatched, resulting in interruptions, misjudgments, or omissions in displacement monitoring data. This seriously affects the reliability of visual monitoring systems and limits their engineering applications in complex field environments.

[0007] Due to the complex and ever-changing natural environment of slopes, images are easily affected by weather changes, vegetation growth, and human activities, resulting in significant scene changes in a short period of time. Existing visual displacement monitoring systems mostly use fixed image processing strategies, such as static templates and preset parameters. These systems lack dynamic perception and adaptability to environmental changes, and are unable to automatically switch recognition models or optimize parameter configurations based on actual scene changes. They are prone to failure in unstructured and dynamically disturbed scenes, reducing the adaptability and generalization capabilities of the monitoring system and hindering its stable operation in long-term, continuous monitoring tasks.

[0008] To this end, the present invention proposes a slope displacement monitoring method based on computer vision to solve the above-mentioned problems. Summary of the Invention

[0009] In view of the deficiencies of the prior art, the present invention provides a slope displacement monitoring method based on computer vision to solve the problems raised in the above background technology.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: A slope displacement monitoring method based on computer vision, comprising:

[0011] Step 1: Obtain multi-temporal remote sensing image data of the monitoring area, construct a time series based on vegetation index, analyze the disturbance trend of the sub-region, and generate a disturbance risk level map;

[0012] Step 2: Extract initial image feature points based on the slope monitoring image, construct the feature points into a graph structure according to their spatial proximity, and form a graph data model with nodes and edges connected;

[0013] Step 3: Based on the disturbance risk level graph and graph structure, the graph neural network model is input to perform feature point stability modeling, and a set of feature points with stability scores higher than a preset threshold is screened out;

[0014] Step 4: Use the feature point set to perform feature matching and displacement calculation on each round of images in the image sequence, and output the slope displacement monitoring results;

[0015] Step 5: Monitor the disturbance level changes in the area where the feature points are located in real time. If the disturbance level exceeds the preset threshold or the matching fails, the failed feature points are dynamically replaced by the candidate feature points based on the latest stability score of the graph neural network.

[0016] Preferably, the vegetation index includes the Normalized Difference Vegetation Index (NDVI) or the Enhanced Vegetation Index (EVI), which is used to indicate the trend of vegetation growth changes.

[0017] Preferably, the step 1 further comprises:

[0018] Sub-step 1.1, remote sensing image preprocessing and region division:

[0019] Perform radiometric correction, atmospheric correction, and geometric registration on the acquired multi-temporal remote sensing image data to obtain a spatially aligned image sequence.

[0020] The monitoring area is divided into grids according to a fixed scale, and each sub-area is denoted as R i , i is the region index, R is the sub-region set;

[0021] Sub-step 1.2, vegetation index calculation and time series construction:

[0022] Calculate each sub-region R for each remote sensing image i The vegetation index NDVI is used to construct a time series. The NDVI calculation formula is:

[0023]

[0024] in, is the sub-region R at time t i The normalized difference vegetation index,

[0025] R i The average reflectivity of the near-infrared band in the image at time t,

[0026] R i The average reflectivity of the red light band in the image at time t,

[0027] t∈{1,2,...,T}, T is the total number of time series images;

[0028] Each sub-region R i Forming NDVI time series S i :

[0029]

[0030] in, At the T time point, the i-th sub-region Ri Normalized Difference Vegetation Index value;

[0031] Sub-step 1.3, disturbance trend analysis and risk level determination:

[0032] For each NDVI time series S i To analyze the changing trend, the sliding difference disturbance index D is used. i To determine the degree of vegetation disturbance in a sub-region, the disturbance index calculation formula is:

[0033]

[0034] Among them, D i Subregion R i The mean of the disturbance change, is the sub-region R at time t-1 i Normalized Difference Vegetation Index;

[0035] According to D i The disturbance risk level is divided by the set threshold value {θ1, θ2, θ3}:

[0036] If D i <θ1, then R i Determined to be a low disturbance risk area;

[0037] If θ1≤D i <θ2, then R i Determined to be a low to medium disturbance risk area;

[0038] If θ2≤D i <θ3, then R i Determined to be a medium to high disturbance risk area;

[0039] If D i ≥θ3, then R i Determined to be a high disturbance risk area;

[0040] The final output is the disturbance risk level map RiskMap(x,y), where each pixel is mapped to its sub-region R i The perturbation level is used to assign the perturbation attribute of the feature point graph structure in step 2.

[0041] Preferably, the image feature points are extracted using scale-invariant feature transform (SIFT);

[0042] The graph structure is an undirected graph, in which each node represents a feature point, and an edge connection represents the spatial adjacency relationship between adjacent feature points, and the weight is the Euclidean distance.

[0043] Preferably, the step 2 further comprises:

[0044] Sub-step 2.1, image feature point extraction:

[0045] Perform feature detection on the slope monitoring image and use the scale-invariant feature transformation algorithm to extract the initial image feature point set;

[0046] For the image frame, the Gaussian pyramid response function is calculated at each image point (x, y), and the local extreme point is obtained as the feature point, which is recorded as:

[0047]

[0048] Among them, p j is the jth feature point, (x j ,y j ) is the coordinate position of the feature point on the image, s j is the scale information of the feature point, d j is the SIFT descriptor vector of the feature point, and N is the total number of extracted feature points;

[0049] After the feature point extraction is completed, the results are used for subsequent graph structure construction;

[0050] Sub-step 2.2: Construct a feature point graph structure based on spatial proximity:

[0051] Represent the feature point set P obtained in sub-step 2.1 as a graph structure G = (V, E, W), where V is the set of nodes in the graph, E is the set of edges in the graph, and W is the set of edge weights.

[0052] The construction of the edge is based on the Euclidean distance between feature points. The adjacency judgment threshold r is set. When the distance between feature points is less than r, an undirected edge is established:

[0053]

[0054] Among them, ||p j -p k ||2 is p j With p k The Euclidean distance between them, E is the edge set, e jk For node p j With p k A logical representation of whether there is an edge connection relationship between them or a set element of edges;

[0055] Sub-step 2.3: graph structure integration of disturbance risk level attributes:

[0056] Map the disturbance risk level map RiskMap(x,y) output in step 1 to each feature point p j The image position (x j ,y j ), assign the disturbance level label lj ∈{1,2,3,4}, and expand the feature point attribute vector:

[0057] f j =[d j ||l j ],

[0058] Among them, f j For node p j The final input attribute vector, || is the vector concatenation operation, d j is the SIFT descriptor vector, l j is the disturbance risk level label;

[0059] Pre-classify nodes according to disturbance level labels:

[0060] If l j =4, then p j It is a high disturbance point and is marked as an unstable candidate node;

[0061] If l j ≤2, then p j It is a low disturbance point and is marked as a stable candidate node.

[0062] Preferably, the step 3 further comprises:

[0063] Sub-step 3.1, graph neural network input preparation and initial feature normalization:

[0064] Before inputting the perturbation-aware graph structure G = (V, E, W) constructed in step 2 into the graph neural network model, j The attribute vector f j =[d j ||l j ] to perform normalization, unify the feature scale and eliminate the dimension effect, and use the z-score normalization method:

[0065]

[0066] in, For node p j Normalized eigenvalue in the kth dimension, μ k and σ k is the mean and standard deviation of the k-th dimension attribute in all nodes, k∈{1,2,…,129}, where the first 128 dimensions are SIFT descriptors and the last dimension is the perturbation level label;

[0067] Sub-step 3.2: stability score prediction based on the disturbance perception graph structure:

[0068] A graph convolutional neural network is used to perform node embedding and stability score calculation on the disturbance-aware graph structure. The model structure is defined as follows:

[0069] The graph convolution operation is:

[0070]

[0071] Among them, H (l+1) is the node feature matrix output of the l+1th layer, is the adjacency matrix of the graph plus the identity matrix, for The degree matrix, H (l) is the node feature matrix output of the lth layer, W (l) is the trainable weight matrix of the l-th layer graph convolution, σ(·) is the activation function;

[0072] After the last layer of graph convolution output, obtain the stability score s of each node j ∈[0,1] by the following formula:

[0073]

[0074] in, For node p j The embedding vector output by the last layer of graph convolution, w s is the scoring weight vector, b s is the scoring bias item;

[0075] Sub-step 3.3, filter the key feature point set based on the scoring threshold:

[0076] Set the stability score threshold τ s ∈(0,1), the stability score s is selected from all nodes j ≥τ s The nodes constitute the key feature point set P * ={p j ∣s j ≥τ s},

[0077] In addition, in order to improve the spatial distribution uniformity of the stability point set, the minimum spatial spacing constraint is introduced, and the minimum spacing threshold d is set min , if the selected feature point p j ,p k ∈P * satisfy:

[0078] ||p j -p k ||2 <d min ,

[0079] Keep the points with high scores and remove the other points, finally forming a set of feature points with no duplicates and balanced distribution. Feature point set P ** It will be directly used as the target point set for image feature matching and displacement calculation in step 4.

[0080] Preferably, the step 4 further comprises:

[0081] Sub-step 4.1, feature point matching:

[0082] Based on the feature point set output in step 3 For any adjacent frame image I (t) with I (t+1) Perform feature point descriptor matching; suppose each feature point p j The descriptor at time t is Use Euclidean distance to identify matching pairs:

[0083]

[0084] in, In image frame I (t) and I (t+1) For feature point p j With p k Perform descriptor matching;

[0085] is the feature point p k SIFT descriptor vector in the image frame at time t+1

[0086] The ratio detection method is introduced to improve the matching robustness. Assume that the descriptor distances between the best and second-best matches are d1 and d2. If:

[0087] Then keep the match, where τ m ∈(0,1) is the matching decision ratio threshold;

[0088] The feature points that failed to match are recorded as the failure point set Subsequent steps will be replaced based on the scoring mechanism in step 3;

[0089] Sub-step 4.2, pixel-level displacement vector calculation:

[0090] The successfully matched feature point pairs Calculate its displacement vector in the image coordinate space

[0091]

[0092] in, is the feature point pj In image I (t) The pixel position in ;

[0093] For its image I (t+1) The position of the matching point;

[0094] When the displacement amplitude of any characteristic point exceeds the maximum expected displacement δ of the slope deformation max When , it is determined to be an abnormal match and the point is removed:

[0095] Between time t and t+1, all matches are successful but the displacement exceeds the preset abnormal threshold δ max A set of feature points;

[0096] Finally, the effective displacement point set is obtained:

[0097] is the set of valid feature points used for displacement estimation between time t and t+1;

[0098] Sub-step 4.3, regional slope displacement estimation and output:

[0099] Convert the pixel-level displacement vector into the actual displacement in the geographic coordinate system. Assuming that the ground distance corresponding to the unit pixel in the image space is α, the displacement of the j-th feature point in the geographic space is:

[0100]

[0101] According to the sub-region R where each feature point is located i , perform weighted average of the displacement vectors of all valid feature points falling into the sub-region to estimate the average slope displacement of the region:

[0102]

[0103] in, For sub-region R i The average displacement between time t and t+1;

[0104] is located in sub-region R at the current moment i The set of valid feature points;

[0105] like It is determined that the monitoring data of this area is insufficient, and a null value mark is output, n min is the minimum threshold of the number of valid feature points;

[0106] The displacement estimation results of all sub-areas constitute the displacement monitoring map of the entire slope.

[0107] Preferably, the step 5 further comprises:

[0108] Sub-step 5.1, joint monitoring of disturbance level and matching status:

[0109] For each feature point p in step 4 j ∈P ** Location (x j ,y j )Query the current disturbance risk level map RiskMap (t) (x,y), get the corresponding disturbance level label Set the disturbance threshold τ l , if it satisfies:

[0110] It is considered that the disturbance at this point is significant and is recorded as the high disturbance point set

[0111] At the same time, the point set that failed to match is obtained by step 4.1 Combine the high disturbance point set to get the current failure feature point set:

[0112] like Then proceed to sub-step 5.2 to perform dynamic replacement operation;

[0113] Sub-step 5.2, select replacement points based on stability scores among candidate points:

[0114] The set of stable feature points P not included in step 3 ** Construct a set of candidate feature points P from the remaining nodes cand =P\P ** , for each candidate point p k ∈P cand , extract its stability score s k with the disturbance level label Candidate points must meet the following criteria at the same time:

[0115] Stability score greater than the replacement threshold τ r , that is: s k ≥τ r ,

[0116] The current disturbance level does not exceed the maximum acceptable disturbance level τ l ,Right now:

[0117] The candidate points that meet the above conditions constitute the valid replacement point set In addition, to ensure the rationality of space, if the replacement point p k and failure point The spatial distance satisfies:

[0118] ||p k -p j ||2≤d r ,

[0119] Then establish a point pair (p j ,p k ), indicating the use of p k Replace the failure point p j , and finally form the replacement mapping:

[0120]

[0121] Among them, d r is the neighborhood radius of the space replacement, is the set of failure feature points, is a collection of replaceable feature points;

[0122] Sub-step 5.3, update the key feature point set and graph structure status:

[0123] Replace all mappings R (t) Applied to the stable feature point set P in step 3 ** , updated to a new round of feature point set:

[0124]

[0125] Synchronously update the feature point graph structure G (t) =(V (t) ,E (t) ,W (t) ),

[0126] in, V (t) is the node set, E (t) is the edge set;

[0127] Each node p j ∈V (t) The attribute vector Synchronously update the disturbance level value;

[0128] Updated feature point set and graph structure G (t) It will be used in the subsequent image sequence analysis process as input for the next round of step 4.

[0129] Preferably, the displacement calculation method in step 4 is based on an image registration and matching algorithm, and pixel-level or physical magnitude conversion is performed in combination with the coordinate changes of the feature points in consecutive image frames.

[0130] Preferably, the slope displacement monitoring method updates the disturbance level map and the input of the graph neural network in each monitoring cycle to adapt to scene changes and improve the stability of feature point matching.

[0131] The present invention provides a slope displacement monitoring method based on computer vision. It has the following beneficial effects:

[0132] 1. The present invention adopts the technical solutions of disturbance risk level map guidance, graph neural network stability modeling and feature point stability score screening to achieve the technical effect of effectively identifying and screening stable and traceable key feature points under dynamic interference conditions such as vegetation disturbance and climate change. Compared with the existing solutions that rely on static image areas or manually set templates for feature point matching, the present invention solves the problem of frequent failure or mismatch of feature points due to vegetation swing and view occlusion, and significantly improves the availability and continuity of displacement monitoring in actual field scenes.

[0133] 2. The present invention adopts a technical solution of dynamic replacement mechanism of feature points, real-time perception of disturbance levels and update and evolution of graph structure, which can automatically perceive disturbance changes and replace invalid feature points during the monitoring process, maintain stable evolution of the feature point set structure, and support long-term continuous displacement monitoring. Compared with the processing strategy of static parameters or fixed feature point sets in the existing technology, it solves the deficiency that traditional systems cannot update recognition models or adapt to environmental disturbances in unstructured scenarios, and effectively improves the generalization ability and operation stability of the monitoring system in complex field environments such as natural slopes. BRIEF DESCRIPTION OF THE DRAWINGS

[0134] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0135] To help those skilled in the art understand the present invention, the following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0136] The present invention is described in detail below with reference to the accompanying drawings:

[0137] Example:

[0138] Please see the attached Figure 1 The embodiment of the present invention provides a slope displacement monitoring method based on computer vision, comprising:

[0139] Step 1: Obtain multi-temporal remote sensing image data of the monitoring area, construct a time series based on vegetation index, analyze the disturbance trend of the sub-region, and generate a disturbance risk level map;

[0140] Vegetation indices include the Normalized Difference Vegetation Index (NDVI) or the Enhanced Vegetation Index (EVI), which are used to indicate the trend of vegetation growth;

[0141] Sub-step 1.1, remote sensing image preprocessing and region division:

[0142] Perform radiometric correction, atmospheric correction, and geometric registration on the acquired multi-temporal remote sensing image data to obtain a spatially aligned image sequence.

[0143] The monitoring area is divided into grids according to a fixed scale, and each sub-area is denoted as R i , i is the region index, R is the sub-region set;

[0144] Sub-step 1.2, vegetation index calculation and time series construction:

[0145] Calculate each sub-region R for each remote sensing image i The vegetation index NDVI is used to construct a time series. The NDVI calculation formula is:

[0146]

[0147] in, is the sub-region R at time t i The normalized difference vegetation index,

[0148] R i The average reflectivity of the near-infrared band in the image at time t,

[0149] R i The average reflectivity of the red light band in the image at time t,

[0150] t∈{1,2,...,T}, T is the total number of time series images;

[0151] Each sub-region R i Forming NDVI time series S i :

[0152]

[0153] in, At the T time point, the i-th sub-region R i Normalized Difference Vegetation Index value;

[0154] Sub-step 1.3, disturbance trend analysis and risk level determination:

[0155] For each NDVI time series S i To analyze the changing trend, the sliding difference disturbance index D is used. i To determine the degree of vegetation disturbance in a sub-region, the disturbance index calculation formula is:

[0156]

[0157] Among them, D i Subregion R i The mean of the disturbance change, is the sub-region R at time t-1 i Normalized Difference Vegetation Index;

[0158] According to D i The disturbance risk level is divided by the set threshold value {θ1, θ2, θ3}:

[0159] If D i <θ1, then R i Determined to be a low disturbance risk area;

[0160] If θ1≤D i <θ2, then R i Determined to be a low to medium disturbance risk area;

[0161] If θ2≤D i <θ3, then R i Determined to be a medium to high disturbance risk area;

[0162] If D i ≥θ3, then R i Determined to be a high disturbance risk area;

[0163] The final output is the disturbance risk level map RiskMap(x,y), where each pixel is mapped to its sub-region R i The perturbation level is used to assign the perturbation attribute of the feature point graph structure in step 2;

[0164] Step 2: Extract initial image feature points based on the slope monitoring image, construct the feature points into a graph structure according to their spatial proximity, and form a graph data model with nodes and edges connected;

[0165] Image feature points are extracted using scale-invariant feature transform (SIFT);

[0166] The graph structure is an undirected graph. Each node in the graph represents a feature point. The edge connection represents the spatial adjacency relationship between adjacent feature points. The weight is the Euclidean distance.

[0167] Sub-step 2.1, image feature point extraction:

[0168] Perform feature detection on the slope monitoring image and use the scale-invariant feature transformation algorithm to extract the initial image feature point set;

[0169] For the image frame, the Gaussian pyramid response function is calculated at each image point (x, y), and the local extreme point is obtained as the feature point, which is recorded as:

[0170]

[0171] Among them, p j is the jth feature point, (x j ,y j ) is the coordinate position of the feature point on the image, s j is the scale information of the feature point, d j is the SIFT descriptor vector of the feature point, and N is the total number of extracted feature points;

[0172] After the feature point extraction is completed, the results are used for subsequent graph structure construction;

[0173] Sub-step 2.2: Construct a feature point graph structure based on spatial proximity:

[0174] Represent the feature point set P obtained in sub-step 2.1 as a graph structure G = (V, E, W), where V is the set of nodes in the graph, E is the set of edges in the graph, and W is the set of edge weights.

[0175] The construction of the edge is based on the Euclidean distance between feature points. The adjacency judgment threshold r is set. When the distance between feature points is less than r, an undirected edge is established:

[0176]

[0177] Among them, ||p j -p k ||2 is p j With p k The Euclidean distance between them, E is the edge set, e jk For node p j With p k A logical representation of whether there is an edge connection relationship between them or a set element of edges;

[0178] Sub-step 2.3: graph structure integration of disturbance risk level attributes:

[0179] Map the disturbance risk level map RiskMap(x,y) output in step 1 to each feature point p j The image position (x j ,y j ), assign the disturbance level label l j ∈{1,2,3,4}, and expand the feature point attribute vector:

[0180] f j =[d j ||l j ],

[0181] Among them, f j For node p j The final input attribute vector, || is the vector concatenation operation, d j is the SIFT descriptor vector, l j is the disturbance risk level label;

[0182] Pre-classify nodes according to disturbance level labels:

[0183] If l j =4, then p j It is a high disturbance point and is marked as an unstable candidate node;

[0184] If l j ≤2, then p j is a low disturbance point and is marked as a stable candidate node;

[0185] Step 3: Based on the disturbance risk level graph and graph structure, the graph neural network model is input to perform feature point stability modeling, and a set of feature points with stability scores higher than a preset threshold is screened out;

[0186] Sub-step 3.1, graph neural network input preparation and initial feature normalization:

[0187] Before inputting the perturbation-aware graph structure G = (V, E, W) constructed in step 2 into the graph neural network model, j The attribute vector f j =[d j ||l j ] to perform normalization, unify the feature scale and eliminate the dimension effect, and use the z-score normalization method:

[0188]

[0189] in, For node p j Normalized eigenvalue in the kth dimension, μ k and σ k is the mean and standard deviation of the k-th dimension attribute in all nodes, k∈{1,2,…,129}, where the first 128 dimensions are SIFT descriptors and the last dimension is the perturbation level label;

[0190] Sub-step 3.2: stability score prediction based on the disturbance perception graph structure:

[0191] A graph convolutional neural network is used to perform node embedding and stability score calculation on the disturbance-aware graph structure. The model structure is defined as follows:

[0192] The graph convolution operation is:

[0193]

[0194] Among them, H (l+1) is the node feature matrix output of the l+1th layer, is the adjacency matrix of the graph plus the identity matrix, for The degree matrix, H (l) is the node feature matrix output of the lth layer, W (l) is the trainable weight matrix of the l-th layer graph convolution, σ(·) is the activation function;

[0195] After the last layer of graph convolution output, obtain the stability score s of each node j ∈[0,1] by the following formula:

[0196]

[0197] in, For node p j The embedding vector output by the last layer of graph convolution, w s is the scoring weight vector, b s is the scoring bias item;

[0198] Sub-step 3.3, filter the key feature point set based on the scoring threshold:

[0199] Set the stability score threshold τ s ∈(0,1), the stability score s is selected from all nodes j ≥τ s The nodes constitute the key feature point set P * ={p j ∣s j ≥τ s},

[0200] In addition, in order to improve the spatial distribution uniformity of the stability point set, the minimum spatial spacing constraint is introduced, and the minimum spacing threshold d is set min , if the selected feature point p j ,p k ∈P * satisfy:

[0201] ||p j -p k ||2 <d min ,

[0202] Keep the points with high scores and remove the other points, finally forming a set of feature points with no duplicates and balanced distribution. Feature point set P ** It will be directly used as the target point set for image feature matching and displacement calculation in step 4;

[0203] Step 4: Use the feature point set to perform feature matching and displacement calculation on each round of images in the image sequence, and output the slope displacement monitoring results;

[0204] Sub-step 4.1, feature point matching:

[0205] Based on the feature point set output in step 3 For any adjacent frame image I (t) with I (t+1) Perform feature point descriptor matching; suppose each feature point p j The descriptor at time t is Use Euclidean distance to identify matching pairs:

[0206]

[0207] in, In image frame I (t) and I (t+1) For feature point p j With p k Perform descriptor matching;

[0208] is the feature point p k SIFT descriptor vector in the image frame at time t+1

[0209] The ratio detection method is introduced to improve the matching robustness. Assume that the descriptor distances between the best and second-best matches are d1 and d2. If:

[0210] Then keep the match, where τ m ∈(0,1) is the matching decision ratio threshold;

[0211] The feature points that failed to match are recorded as the failure point set Subsequent steps will be replaced based on the scoring mechanism in step 3;

[0212] Sub-step 4.2, pixel-level displacement vector calculation:

[0213] The successfully matched feature point pairs Calculate its displacement vector in the image coordinate space

[0214]

[0215] in, is the feature point p j In image I (t) The pixel position in ;

[0216] For its image I (t+1) The position of the matching point;

[0217] When the displacement amplitude of any characteristic point exceeds the maximum expected displacement δ of the slope deformation max When , it is determined to be an abnormal match and the point is removed:

[0218] Between time t and t+1, all matches are successful but the displacement exceeds the preset abnormal threshold δ max A set of feature points;

[0219] Finally, the effective displacement point set is obtained:

[0220] is the set of valid feature points used for displacement estimation between time t and t+1;

[0221] Sub-step 4.3, regional slope displacement estimation and output:

[0222] Convert the pixel-level displacement vector into the actual displacement in the geographic coordinate system. Assuming that the ground distance corresponding to the unit pixel in the image space is α, the displacement of the j-th feature point in the geographic space is:

[0223]

[0224] According to the sub-region R where each feature point is located i , perform weighted average of the displacement vectors of all valid feature points falling into the sub-region to estimate the average slope displacement of the region:

[0225]

[0226] in, For sub-region R i The average displacement between time t and t+1;

[0227] is located in sub-region R at the current moment i The set of valid feature points;

[0228] like It is determined that the monitoring data of this area is insufficient, and a null value mark is output, n min is the minimum threshold of the number of valid feature points;

[0229] The displacement estimation results of all sub-areas constitute the displacement monitoring map of the entire slope;

[0230] Step 5: Monitor the disturbance level changes in the area where the feature points are located in real time. If the disturbance level exceeds the preset threshold or the matching fails, the failed feature points are dynamically replaced by candidate feature points based on the latest stability score of the graph neural network.

[0231] Sub-step 5.1, joint monitoring of disturbance level and matching status:

[0232] For each feature point p in step 4 j ∈P ** Location (x j ,y j )Query the current disturbance risk level map RiskMap (t) (x,y), get the corresponding disturbance level label Set the disturbance threshold τ l , if it satisfies:

[0233] It is considered that the disturbance at this point is significant and is recorded as the high disturbance point set

[0234] At the same time, the point set that failed to match is obtained by step 4.1 Combine the high disturbance point set to get the current failure feature point set:

[0235] like Then proceed to sub-step 5.2 to perform dynamic replacement operation;

[0236] Sub-step 5.2, select replacement points based on stability scores among candidate points:

[0237] The set of stable feature points P not included in step 3 ** Construct a set of candidate feature points P from the remaining nodes cand =P\P ** , for each candidate point p k ∈P cand , extract its stability score s k with the disturbance level label Candidate points must meet the following criteria at the same time:

[0238] Stability score greater than the replacement threshold τ r , that is: s k ≥τ r ,

[0239] The current disturbance level does not exceed the maximum acceptable disturbance level τ l ,Right now:

[0240] The candidate points that meet the above conditions constitute the valid replacement point set In addition, to ensure the rationality of space, if the replacement point p k and failure point The spatial distance satisfies:

[0241] ||p k -p j ||2≤d r ,

[0242] Then establish a point pair (p j ,p k ), indicating the use of p k Replace the failure point p j , and finally form the replacement mapping:

[0243]

[0244] Among them, d r is the neighborhood radius of the space replacement, is the set of failure feature points, is a collection of replaceable feature points;

[0245] Sub-step 5.3, update the key feature point set and graph structure status:

[0246] Replace all mappings R (t) Applied to the stable feature point set P in step 3 ** , updated to a new round of feature point set:

[0247]

[0248] Synchronously update the feature point graph structure G (t) =(V (t) ,E (t) ,W (t) ),

[0249] in, V (t) is the node set, E (t) is the edge set;

[0250] Each node p j ∈V (t) The attribute vector Synchronously update the disturbance level value;

[0251] Updated feature point set and graph structure G (t) It will be used in the subsequent image sequence analysis process as input for the next round of step 4.

[0252] The advantage of step 1 is that by combining remote sensing imagery with vegetation indices, we can systematically analyze vegetation growth trends within the region. Time series analysis can identify potential disturbances and dynamically assess slope risks. The disturbance risk level map generated using this time series provides reliable basic data for subsequent monitoring and early warning. By dividing the region into subregions, we can conduct risk assessments at scale, ensuring highly accurate monitoring results.

[0253] The benefits of sub-step 1.1 include ensuring the consistency and accuracy of image data through radiometric correction, atmospheric correction, and geometric registration, addressing data errors caused by factors such as shooting angle and weather. Regionalization makes the processing of monitoring areas systematic and operational, providing a spatial structure for subsequent vegetation index calculations.

[0254] Benefits of Substep 1.2: Calculating vegetation indices can reflect vegetation growth and change trends. By constructing a time series, we can monitor vegetation changes at different points in time, providing historical data support for disturbance analysis and ensuring accurate disturbance trend analysis.

[0255] The benefits of substep 1.3 are that the sliding differential disturbance index objectively assesses vegetation disturbance and further refines the risk classification. The analytical method can sensitively capture subtle changes in slope conditions, helping to identify potential hazards early and develop appropriate preventive measures.

[0256] Benefits of Step 2: Feature point extraction is a classic technique in computer vision, enabling efficient identification and location of key points in complex environments. Graph structure modeling clearly reflects the spatial relationships between feature points in an image, providing data structure support for subsequent displacement calculations.

[0257] Benefits of Substep 2.1: Scale-invariant feature transformation is a classic and powerful feature extraction method with strong robustness and invariance. In slope monitoring, it ensures that feature points maintain good matching under different viewing angles and in different environments, reducing the influence of external factors and ensuring the stability of monitoring results.

[0258] Benefits of Sub-step 2.2: Constructing a feature point graph effectively reflects the spatial correlation between feature points in the image. Constructing adjacency relationships using Euclidean distance ensures that the relative relationships between adjacent feature points are accurately modeled.

[0259] The benefit of sub-step 2.3 is that by integrating the disturbance risk level with the node attributes in the graph structure, the attribute information of the feature points is enhanced, so that each feature point has spatial location information and important attributes such as the disturbance level.

[0260] Benefits of Step 3: Graph neural networks, an advanced technology for processing graph-structured data, can automatically learn the relationships and features of nodes through graph convolution operations, accurately predicting the stability of feature points. This method can identify potential unstable points, improve the accuracy of slope monitoring predictions, and flexibly handle complex spatial relationships between nodes.

[0261] Benefits of Substep 3.1: Normalization helps eliminate scale differences between features, ensuring that the graph neural network model can fairly process different types of features, improving model training efficiency and stability. Furthermore, normalization avoids numerical instabilities during model training and enhances the credibility of the final prediction results.

[0262] Benefits of Substep 3.2: The graph convolutional neural network can predict stability scores based on the disturbance-aware graph structure, automatically identifying feature points in areas at risk of slope instability. This method's advantage lies in its ability to perform global optimization based on the relationships between nodes in the graph structure, avoiding the local analysis limitations of traditional methods and enhancing the globality and accuracy of monitoring results.

[0263] The benefit of sub-step 3.3 is that by combining the stability score and the spatial distribution uniformity, it ensures that the selected feature point set has high stability and can maintain a uniform distribution in space, avoiding the situation where feature points are concentrated in a certain local area.

[0264] Benefits of Step 4: Feature point matching and displacement calculation are core components of slope displacement monitoring. Improving matching robustness through ratio detection can reduce matching errors caused by image noise, occlusion, or perspective changes, thereby improving monitoring accuracy. Furthermore, pixel-level displacement vector calculation enables precise estimation of slope deformation.

[0265] Benefits of Substep 4.1: Feature point matching improves the accuracy of image-to-image matching and eliminates false matches through ratio detection, effectively improving monitoring stability and accuracy. This advantage is that by comparing the ratio of the best to the suboptimal matches, low-quality matches can be filtered out, improving the robustness of the overall monitoring system.

[0266] The benefit of sub-step 4.2 is that by accurately calculating the displacement vectors of feature points, slope deformation information can be obtained at the image level. The advantage lies in providing displacement data, which can monitor small changes in the slope in real time, providing a scientific basis for decision makers.

[0267] Benefits of Sub-step 4.3: Regional-level displacement estimation enables a holistic assessment of slope displacement at a large spatial scale by weighted averaging of characteristic points. This effectively integrates local displacement data to provide regional displacement trends, helping to understand slope stability.

[0268] Benefits of Step 5: The real-time monitoring mechanism enables the slope monitoring system to quickly respond to changes and promptly update the feature point set graph structure, ensuring the system's dynamic adaptability. By combining the disturbance level and matching status, failed feature points can be effectively identified and dynamically replaced, ensuring the accuracy and reliability of monitoring data.

[0269] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A slope displacement monitoring method based on computer vision, characterized in that: include: Step 1: Obtain multi-temporal remote sensing image data of the monitoring area, construct a time series based on vegetation index, analyze the disturbance trend of the sub-region, and generate a disturbance risk level map; Step 2: Extract initial image feature points based on the slope monitoring image, construct the feature points into a graph structure according to their spatial proximity, and form a graph data model with nodes and edges connected; Step 3: Based on the disturbance risk level graph and graph structure, the graph neural network model is input to perform feature point stability modeling, and a set of feature points with stability scores higher than a preset threshold is screened out; Step 4: Use the feature point set to perform feature matching and displacement calculation on each round of images in the image sequence, and output the slope displacement monitoring results; Step 5: Monitor the disturbance level changes in the area where the feature points are located in real time. If the disturbance level exceeds the preset threshold or the matching fails, the failed feature points are dynamically replaced by the candidate feature points based on the latest stability score of the graph neural network.

2. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The vegetation index includes the Normalized Difference Vegetation Index (NDVI) or the Enhanced Vegetation Index (EVI), which is used to indicate the trend of vegetation growth changes.

3. The computer vision-based slope displacement monitoring method according to claim 2, characterized in that: The step 1 further comprises: Sub-step 1.1, remote sensing image preprocessing and region division: Perform radiometric correction, atmospheric correction, and geometric registration on the acquired multi-temporal remote sensing image data to obtain a spatially aligned image sequence. The monitoring area is divided into grids according to a fixed scale, and each sub-area is denoted as R i , i is the region index, R is the sub-region set; Sub-step 1.2, vegetation index calculation and time series construction: Calculate each sub-region R for each remote sensing image i The vegetation index NDVI is used to construct a time series. The NDVI calculation formula is: in, is the sub-region R at time t i The normalized difference vegetation index, R i The average reflectivity of the near-infrared band in the image at time t, R i The average reflectivity of the red light band in the image at time t, t∈{1,2,...,T}, T is the total number of time series images; Each sub-region R i Forming NDVI time series S i : in, At the T time point, the i-th sub-region R i Normalized Difference Vegetation Index value; Sub-step 1.3, disturbance trend analysis and risk level determination: For each NDVI time series S i To analyze the changing trend, the sliding difference disturbance index D is used. i To determine the degree of vegetation disturbance in a sub-region, the disturbance index calculation formula is: Among them, D i Subregion R i The mean of the disturbance change, is the sub-region R at time t-1 i Normalized Difference Vegetation Index; According to D i The disturbance risk level is divided by the set threshold value {θ1, θ2, θ3}: If D i <θ1, then R i Determined to be a low disturbance risk area; If θ1≤D i <θ2, then R i Determined to be a low to medium disturbance risk area; If θ2≤D i <θ3, then R i Determined to be a medium to high disturbance risk area; If D i ≥θ3, then R i Determined to be a high disturbance risk area; The final output is the disturbance risk level map RiskMap(x,y), where each pixel is mapped to its sub-region R i The perturbation level is used to assign the perturbation attribute of the feature point graph structure in step 2.

4. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The image feature points are extracted using scale-invariant feature transform (SIFT); The graph structure is an undirected graph, in which each node represents a feature point, and an edge connection represents the spatial adjacency relationship between adjacent feature points, and the weight is the Euclidean distance.

5. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The step 2 further includes: Sub-step 2.1, image feature point extraction: Perform feature detection on the slope monitoring image and use the scale-invariant feature transformation algorithm to extract the initial image feature point set; For the image frame, the Gaussian pyramid response function is calculated at each image point (x, y), and the local extreme point is obtained as the feature point, which is recorded as: Among them, p j is the jth feature point, (x j ,y j ) is the coordinate position of the feature point on the image, s j is the scale information of the feature point, d j is the SIFT descriptor vector of the feature point, and N is the total number of extracted feature points; After the feature point extraction is completed, the results are used for subsequent graph structure construction; Sub-step 2.2: Construct a feature point graph structure based on spatial proximity: Represent the feature point set P obtained in sub-step 2.1 as a graph structure G = (V, E, W), where V is the set of nodes in the graph, E is the set of edges in the graph, and W is the set of edge weights. The construction of the edge is based on the Euclidean distance between feature points. The adjacency judgment threshold r is set. When the distance between feature points is less than r, an undirected edge is established: Among them, ||p j -p k ||2 is p j With p k The Euclidean distance between them, E is the edge set, e jk For node p j With p k A logical representation of whether there is an edge connection relationship between them or a set element of edges; Sub-step 2.3: graph structure integration of disturbance risk level attributes: Map the disturbance risk level map RiskMap(x,y) output in step 1 to each feature point p j The image position (x j ,y j ), assign the disturbance level label l j ∈{1,2,3,4}, and expand the feature point attribute vector: f j =[d j ||l j ], Among them, f j For node p j The final input attribute vector, || is the vector concatenation operation, d j is the SIFT descriptor vector, l j is the disturbance risk level label; Pre-classify nodes according to disturbance level labels: If l j =4, then p j It is a high disturbance point and is marked as an unstable candidate node; If l j ≤2, then p j It is a low disturbance point and is marked as a stable candidate node.

6. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The step 3 further includes: Sub-step 3.1, graph neural network input preparation and initial feature normalization: Before inputting the perturbation-aware graph structure G = (V, E, W) constructed in step 2 into the graph neural network model, j The attribute vector f j =[d j ||l j ] to perform normalization, unify the feature scale and eliminate the dimension effect, and use the z-score normalization method: in, For node p j Normalized eigenvalue in the kth dimension, μ k and σ k is the mean and standard deviation of the k-th dimension attribute in all nodes, k∈{1,2,…,129}, where the first 128 dimensions are SIFT descriptors and the last dimension is the perturbation level label; Sub-step 3.2: stability score prediction based on the disturbance perception graph structure: A graph convolutional neural network is used to perform node embedding and stability score calculation on the disturbance-aware graph structure. The model structure is defined as follows: The graph convolution operation is: Among them, H (l+1) is the node feature matrix output of the l+1th layer, is the adjacency matrix of the graph plus the identity matrix, for The degree matrix, H (l) is the node feature matrix output of the lth layer, W (l) is the trainable weight matrix of the l-th layer graph convolution, σ(·) is the activation function; After the last layer of graph convolution output, obtain the stability score s of each node j ∈[0,1] by the following formula: in, For node p j The embedding vector output by the last layer of graph convolution, w s is the scoring weight vector, b s is the scoring bias item; Sub-step 3.3, filter the key feature point set based on the scoring threshold: Set the stability score threshold τ s ∈(0,1), the stability score s is selected from all nodes j ≥τ s The nodes constitute the key feature point set P * ={p j ∣s j ≥τ s }, In addition, in order to improve the spatial distribution uniformity of the stability point set, the minimum spatial spacing constraint is introduced, and the minimum spacing threshold d is set min , if the selected feature point p j ,p k ∈P * satisfy: ||p j -p k ||2<d min , Keep the points with high scores and remove the other points, finally forming a set of feature points with no duplicates and balanced distribution. Feature point set P ** It will be directly used as the target point set for image feature matching and displacement calculation in step 4.

7. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The step 4 further comprises: Sub-step 4.1, feature point matching: Based on the feature point set output in step 3 For any adjacent frame image I (t) with I (t+1) Perform feature point descriptor matching; suppose each feature point p j The descriptor at time t is Use Euclidean distance to identify matching pairs: in, In image frame I (t) and I (t+1) For feature point p j With p k Perform descriptor matching; is the feature point p k SIFT descriptor vector in the image frame at time t+1 The ratio detection method is introduced to improve the matching robustness. Assume that the descriptor distances between the best and second-best matches are d1 and d2. If: Then keep the match, where τ m ∈(0,1) is the matching decision ratio threshold; The feature points that failed to match are recorded as the failure point set Subsequent steps will be replaced based on the scoring mechanism in step 3; Sub-step 4.2, pixel-level displacement vector calculation: The feature point pairs that are successfully matched Calculate its displacement vector in the image coordinate space in, is the feature point p j In image I (t) The pixel position in ; For its image I (t+1) The position of the matching point; When the displacement amplitude of any characteristic point exceeds the maximum expected displacement δ of the slope deformation max When , it is determined to be an abnormal match and the point is removed: Between time t and t+1, all matches are successful but the displacement exceeds the preset abnormal threshold δ max A set of feature points; Finally, the effective displacement point set is obtained: is the set of valid feature points used for displacement estimation between time t and t+1; Sub-step 4.3, regional slope displacement estimation and output: Convert the pixel-level displacement vector into the actual displacement in the geographic coordinate system. Assuming that the ground distance corresponding to the unit pixel in the image space is α, the displacement of the j-th feature point in the geographic space is: According to the sub-region R where each feature point is located i , perform weighted average of the displacement vectors of all valid feature points falling into the sub-region to estimate the average slope displacement of the region: in, For sub-region R i The average displacement between time t and t+1; is located in sub-region R at the current moment i The set of valid feature points; If |P i (t) |<n min , it is determined that the monitoring data of this area is insufficient, and a null value mark is output, n min is the minimum threshold of the number of valid feature points; The displacement estimation results of all sub-areas constitute the displacement monitoring map of the entire slope.

8. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The step 5 further comprises: Sub-step 5.1, joint monitoring of disturbance level and matching status: For each feature point p in step 4 j ∈P ** Location (x j ,y j )Query the current disturbance risk level map RiskMap (t) (x,y), get the corresponding disturbance level label Set the disturbance threshold τ l , if it satisfies: It is considered that the disturbance at this point is significant and is recorded as a high disturbance point set. At the same time, the point set that failed to match is obtained by step 4.1 Combine the high disturbance point set to get the current failure feature point set: like Then proceed to sub-step 5.2 to perform dynamic replacement operation; Sub-step 5.2, select replacement points based on stability scores among candidate points: The set of stable feature points P not included in step 3 ** Construct a set of candidate feature points P from the remaining nodes cand =P\P ** , for each candidate point p k ∈P cand , extract its stability score s k with the disturbance level label Candidate points must meet the following criteria at the same time: Stability score greater than the replacement threshold τ r , that is: s k ≥τ r , The current disturbance level does not exceed the maximum acceptable disturbance level τ l ,Right now: The candidate points that meet the above conditions constitute the valid replacement point set In addition, to ensure the rationality of space, if the replacement point p k and failure point The spatial distance satisfies: ||p k -p j ||2≤d r , Then establish a point pair (p j ,p k ), indicating the use of p k Replace the failure point p j , and finally form the replacement mapping: Among them, d r is the neighborhood radius of the space replacement, is the set of failure feature points, is a collection of replaceable feature points; Sub-step 5.3, update the key feature point set and graph structure status: Replace all mappings R (t) Applied to the stable feature point set P in step 3 ** , updated to a new round of feature point set: Synchronously update the feature point graph structure G (t) =(V (t) ,E (t) ,W (t) ), in, V (t) is the node set, E (t) is the edge set; Each node p j ∈V (t) The attribute vector Synchronously update the disturbance level value; Updated feature point set and graph structure G (t) It will be used in the subsequent image sequence analysis process as input for the next round of step 4.

9. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The displacement calculation method in step 4 is based on an image registration and matching algorithm, and pixel-level or physical magnitude conversion is performed in combination with the coordinate changes of feature points in consecutive image frames.

10. The computer vision-based slope displacement monitoring method according to claim 1, characterized in that: The slope displacement monitoring method updates the disturbance level map and the input of the graph neural network in each monitoring cycle to adapt to scene changes and improve the stability of feature point matching.

Images