Highway slope deformation monitoring method and system based on multi-modal remote sensing

Abstract

The application discloses a highway slope deformation monitoring method based on multi-modal remote sensing, comprising: synchronously acquiring multi-modal heterogeneous original data, and generating a multi-modal heterogeneous original data set; inputting the multi-modal heterogeneous original data set into a feature extraction layer, performing feature extraction to obtain an initial feature set of each mode, inputting the initial feature set of each mode into a pre-constructed causal Transformer model, assigning a causal attention weight to the initial feature of each mode, and obtaining an effective feature set; inputting the effective feature set into an output layer of the causal Transformer model to obtain a feature correlation degree path, matching the feature correlation degree path with a preset deformation cause feature library, and judging a core cause of the current deformation; and inputting the core cause and the effective feature set into a preset risk probability calculation operator to obtain a deformation risk probability corresponding to a to-be-monitored slope. Thus, the probability of misjudgment caused by statistical bias can be greatly reduced, and the false alarm rate of deformation risk can be effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of road engineering technology, specifically to a method and system for monitoring highway slope deformation based on multimodal remote sensing. Background Technology

[0002] In mountainous and hilly areas with complex geological conditions, highway slopes are susceptible to gradual deformation or even landslide instability due to the combined effects of topography, rainfall erosion, construction disturbances, and other factors.

[0003] To effectively address highway slope disaster prevention and control, the current multimodal remote sensing method for monitoring highway slope deformation mainly involves collecting multiple heterogeneous data sources. After preprocessing each single-modal data, two fusion modes are used to output the results. The first type of output is based on feature hard fusion, which involves directly concatenating the shallow features extracted from each modality into a unified feature vector, which is then input into a general classification or regression model to obtain the deformation risk result. The second type of output is based on decision voting fusion, which involves calculating the deformation risk sub-result for each single-modal data separately, and then using rules such as majority voting and weighted averaging to obtain the final fused monitoring result. The above methods have been initially applied in the monitoring scenario of slopes in plain areas with gentle terrain and simple deformation causes.

[0004] As can be seen from the above, existing multimodal data fusion methods are limited to feature weighting or decision fusion based on statistical data features. They cannot identify the complete physical causal transmission path corresponding to slope deformation, resulting in low accuracy in identifying the core causes of deformation, high false alarm rate of deformation risk, and inability to support accurate source tracing of deformation risk. Summary of the Invention

[0005] The present invention aims to at least partially solve the technical problems in the above-mentioned technologies.

[0006] Therefore, this invention discloses a method for monitoring highway slope deformation based on multimodal remote sensing, comprising the following steps:

[0007] S1: Simultaneously acquire multimodal heterogeneous raw data including spaceborne InSAR deformation time series data, UAV visible light imagery, UAV infrared imagery, pore water pressure time series data collected by ground sensors, rainfall time series data, and surface displacement time series data, and generate a multimodal heterogeneous raw dataset.

[0008] S2: Input the multimodal heterogeneous original dataset into the feature extraction layer, extract features from each modality data to obtain the initial feature set of each modality, input the initial feature set of each modality into the pre-built causal Transformer model, assign causal attention weights to the initial features of each modality, and remove noisy features with weights lower than a preset threshold to obtain the effective feature set;

[0009] S3: Input the effective feature set into the output layer of the causal Transformer model to obtain the feature correlation path, match the feature correlation path with the preset deformation cause feature library, and determine the core cause of the current deformation. The core cause includes three categories: natural rainfall, artificial disturbance, and inherent geological defects.

[0010] S4: Input the core inducing factors and the effective feature set into a preset risk probability calculation operator to obtain the deformation risk probability of the slope to be monitored.

[0011] In addition, the highway slope deformation monitoring method based on multimodal remote sensing disclosed in this invention may also have the following additional technical features:

[0012] Furthermore, after generating the multimodal heterogeneous raw dataset in step S1, the following preprocessing steps are performed:

[0013] The multimodal heterogeneous raw dataset is input into a standardized processing operator library. Phase unwrapping and atmospheric error correction are performed on the spaceborne InSAR deformation time series data. Geometric and radiometric correction are performed on the UAV visible light image and UAV infrared image. Outlier removal and timestamp alignment are performed on the pore water pressure time series data, rainfall time series data, and surface displacement time series data to generate a standardized multimodal dataset. The standardized multimodal dataset is used as the input data source for the S2 feature extraction.

[0014] Furthermore, the construction steps of the causal Transformer model are as follows:

[0015] Obtain publicly available historical case datasets of highway slope deformation and a pre-defined set of slope rock mechanics and physical constraint rules. Embed the set of slope rock mechanics and physical constraint rules into the attention mechanism layer of Transformer as a causal regularization term. Use the deformation cause identification accuracy as the loss function for iterative training. Stop training when the loss function converges to a preset threshold to generate a trained causal Transformer model. Use the trained causal Transformer model as the execution subject for the S2 weight assignment.

[0016] Furthermore, the step of assigning causal attention weights to the initial features of each modality in S2 is as follows:

[0017] Retrieve pre-stored slope deformation causal chain rules, which include three types of causal transmission paths: rainfall triggering, artificial disturbance triggering, and geological defect triggering. Calculate the matching degree between the initial features of each modality and the three types of causal transmission paths. Based on the matching degree, assign causal attention weights in the range of 0 to 1 to the corresponding initial features, generate weight labels corresponding to the initial features of each modality, and use the weight labels as the basis for determining the removal of the S2 noise features.

[0018] Furthermore, the step in S3 to determine the core cause of the current deformation is as follows:

[0019] A preset deformation cause feature library is retrieved. The deformation cause feature library stores standard feature association paths corresponding to three types of causes: natural rainfall, artificial disturbance, and inherent geological defects. The cosine similarity between the feature association path and the three types of standard feature association paths is calculated. The cause type corresponding to the type with the highest cosine similarity is selected as the core cause of the current deformation. A core cause label is generated and used as the input parameter for the S4 risk probability calculation.

[0020] Furthermore, after S2 removes noise features without causal correlation, the following noise verification step is also performed:

[0021] Retrieve the pre-stored whitelist of effective features of slope deformation, match the initial feature set after removing noise with the whitelist of effective features of slope deformation, if there is a feature that is successfully matched but is judged as noise, re-include the feature in the effective feature set, generate a verified effective feature set, and use the verified effective feature set as the input data source for generating the S3 feature correlation path.

[0022] Furthermore, after obtaining the deformation risk probability in S4, the following risk classification step is also performed:

[0023] The preset risk level classification threshold is retrieved. The risk level classification threshold includes three gradients: low risk threshold, medium risk threshold, and high risk threshold. The relationship between the deformation risk probability and each gradient threshold is compared to obtain the deformation risk level corresponding to the slope to be monitored. Deformation risk results with level labels are generated and pushed to the preset highway maintenance platform interface.

[0024] Furthermore, after obtaining the deformation risk probability in step S4, the following precision verification step is also performed:

[0025] The measured deformation rate data of the preset GNSS monitoring points of the slope to be monitored are retrieved, the deformation rate calculation result corresponding to the deformation risk probability is extracted, the mean square error between the deformation rate calculation result and the measured deformation rate data is calculated, and if the mean square error is higher than the preset accuracy threshold, the re-acquisition instruction of the multimodal heterogeneous raw data is triggered, an accuracy verification qualified identifier is generated, and the accuracy verification qualified identifier is used as the pre-judgment condition for the output of the deformation risk probability.

[0026] Furthermore, after obtaining the core cause in step S3, the following source path generation step is also performed:

[0027] Extract the full-link transmission nodes of the core cause in the feature correlation path, generate a visualized deformation cause tracing path diagram, annotate the feature value and weight ratio of each node in the tracing path diagram, generate a deformation cause tracing path file, and bind the tracing path file with the deformation risk probability to a preset monitoring database.

[0028] Furthermore, this invention discloses a highway slope deformation monitoring system based on multimodal remote sensing, comprising:

[0029] The data acquisition module is used to synchronously acquire multimodal heterogeneous raw data, including spaceborne InSAR deformation time series data, UAV visible light imagery, UAV infrared imagery, pore water pressure time series data, rainfall time series data, and surface displacement time series data collected by ground sensors, and generate a multimodal heterogeneous raw dataset.

[0030] The feature processing module is used to input the multimodal heterogeneous original dataset into the feature extraction layer, extract features from each modality data to obtain an initial feature set for each modality, input the initial feature set of each modality into a pre-built causal Transformer model, assign causal attention weights to the initial features of each modality, and remove noisy features with weights lower than a preset threshold to obtain an effective feature set.

[0031] The cause determination module inputs the effective feature set into the output layer of the causal Transformer model to obtain the feature correlation path, matches the feature correlation path with a preset deformation cause feature library, and determines the core cause of the current deformation. The core cause includes three categories: natural rainfall, artificial disturbance, and inherent geological defects.

[0032] The result output module is used to input the core inducing factors and the effective feature set into a preset risk probability calculation operator to obtain the deformation risk probability of the slope to be monitored.

[0033] The method and system for monitoring highway slope deformation based on multimodal remote sensing disclosed in this invention have at least the following beneficial effects:

[0034] (1) The causal Transformer model is used to assign causal attention weights to the features of the synchronously collected multimodal heterogeneous raw data, replacing the existing shallow feature hard splicing or decision-level voting hard fusion mode. The causal mining logic corresponding to the slope geological and physical constraints is embedded into the entire data fusion process, which can identify the complete physical causal transmission path corresponding to slope deformation, filling the technical gap of existing fusion methods without causal correlation mining logic.

[0035] (2) By assigning causal attention weights, noise features without causal relationship are actively eliminated, avoiding interference from irrelevant noise in the feature association analysis process, and effectively improving the identification accuracy of the core cause of deformation.

[0036] (3) Based on the feature correlation path matching of the causal Transformer output, the core cause of deformation is determined. Compared with the existing technology which only relies on the fitting analysis logic of data statistical features, the probability of misjudgment caused by statistical bias is greatly reduced, and the false alarm rate of deformation risk is effectively reduced.

[0037] Additional features and advantages of this invention will be set forth in the description which follows, or may be learned by practicing the invention. Attached Figure Description

[0038] The technical solution and beneficial effects of the present invention will become apparent and readily understood from the following description in conjunction with the accompanying drawings, wherein:

[0039] Figure 1 This is a flowchart of the highway slope deformation monitoring method based on multimodal remote sensing of the present invention;

[0040] Figure 2 This is a flowchart illustrating the causal Transformer model construction process for the multimodal remote sensing-based highway slope deformation monitoring method of the present invention.

[0041] Figure 3 This is a flowchart of the deformation cause identification and risk output sub-flow of the highway slope deformation monitoring method based on multimodal remote sensing of the present invention. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0043] The method for monitoring highway slope deformation based on multimodal remote sensing disclosed in this invention will now be described with reference to the accompanying drawings.

[0044] like Figure 1 , Figure 2 and Figure 3As shown, a method for monitoring highway slope deformation based on multimodal remote sensing includes the following steps:

[0045] S1: Simultaneously acquire multimodal heterogeneous raw data including spaceborne InSAR deformation time series data, UAV visible light imagery, UAV infrared imagery, pore water pressure time series data collected by ground sensors, rainfall time series data, and surface displacement time series data, and generate a multimodal heterogeneous raw dataset.

[0046] S2: Input the multimodal heterogeneous original dataset into the feature extraction layer, extract features from each modality data to obtain the initial feature set of each modality, input the initial feature set of each modality into the pre-built causal Transformer model, assign causal attention weights to the initial features of each modality, and remove noisy features with weights below a preset threshold to obtain the effective feature set;

[0047] S3: Input the effective feature set into the output layer of the causal Transformer model to obtain the feature correlation path. Match the feature correlation path with the preset deformation cause feature library to determine the core cause of the current deformation. The core causes include three categories: natural rainfall, artificial disturbance, and inherent geological defects.

[0048] S4: Input the core causes and effective feature set into the preset risk probability calculation operator to obtain the deformation risk probability of the slope to be monitored.

[0049] Specifically, the highway slope deformation monitoring method based on multimodal remote sensing disclosed in this application firstly imposes unified constraints on the entire processing benchmark. The time benchmark of all multimodal data is unified to Coordinated Universal Time (UTC+8h), and the sampling time alignment granularity is preset to 1h. At the feature dimension level, the initial features output by each modality are standardized and adapted, with the spaceborne InSAR deformation time series feature dimension unified to 128 dimensions, the UAV visible light and infrared image visual feature dimension unified to 256 dimensions, and the ground sensor time series feature dimension unified to 64 dimensions. The confidence benchmark of the entire data processing is set to 95%, the default rejection threshold for causal attention weight is θ=0.2, and the final output deformation risk probability retains 4 significant digits.

[0050] After S1 generates the multimodal heterogeneous raw dataset, the following preprocessing steps are performed:

[0051] The multimodal heterogeneous raw dataset is input into a standardized processing operator library. Phase unwrapping and atmospheric error correction are performed on the spaceborne InSAR deformation time series data. Geometric and radiometric corrections are performed on the UAV visible light image and UAV infrared image. Outlier removal and timestamp alignment are performed on the pore water pressure time series data, rainfall time series data, and surface displacement time series data to generate a standardized multimodal dataset. The standardized multimodal dataset is used as the input data source for S2 feature extraction.

[0052] Specifically, the standardized processing operator library has built-in processing logic specific to each modality, and the execution process is as follows:

[0053] First, the spaceborne InSAR deformation time series data is processed. The phase unwrapping adopts the minimum cost flow algorithm, and the unwrapping error threshold is controlled to not exceed 0.1 rad.

[0054] Atmospheric error correction uses ERA5 atmospheric reanalysis data and follows the correction formula. Complete the calibration, among which, The original interference phase, The atmospheric delay factor is taken as a fixed value of 2.28e-3. This represents the ground air pressure at the corresponding location. Near-surface temperature, This refers to the slope elevation. The atmospheric elevation is taken as 8km;

[0055] Afterwards, the visible light and infrared images of the UAV are processed, and the geometric correction is performed by registering the RPC model to the CGCS2000 coordinate system, controlling the registration accuracy to be no more than 0.5 pixels.

[0056] Radiation correction uses the apparent reflectance conversion formula. Complete the conversion, among which, For image radiance, The distance between the Earth and the Sun. This represents the solar irradiance for the corresponding wavelength band. The solar altitude angle;

[0057] Finally, the ground sensor time-series data is processed, and outlier removal is performed using... The criterion is to classify values ​​falling outside the interval as outliers and remove them;

[0058] Timestamp alignment uses linear interpolation to ensure that the time deviation between the timestamps of each sensor data and the InSAR and UAV data does not exceed 10 minutes after interpolation. Once completed, a standardized multimodal dataset is output.

[0059] The steps for constructing a causal Transformer model are as follows:

[0060] Obtain publicly available historical case datasets of highway slope deformation and a pre-defined set of slope rock mechanics and physical constraint rules. Embed the set of slope rock mechanics and physical constraint rules into the attention mechanism layer of Transformer as a causal regularization term. Use the deformation cause identification accuracy as the loss function for iterative training. Stop training when the loss function converges to a preset threshold to generate a trained causal Transformer model. Use the trained causal Transformer model as the execution subject for S2 weight assignment.

[0061] Specifically, the construction process of the causal Transformer model is adapted to the physical constraint logic of slope deformation:

[0062] First, a training set is constructed by selecting no less than 1,000 labeled deformation history cases from the publicly available highway slope disaster dataset. Each sample is matched with corresponding multimodal features, deformation cause labels, and rock mechanics parameters as training data.

[0063] Next, physical constraint embedding is completed, transforming the Mohr-Coulomb shear strength criterion into a causal regularization term. The formula for calculating the regularization term is as follows: ,in, For the first Calculated shear stress values ​​for each sample. To correspond to normal stress, The internal friction angle of the rock mass. To represent the cohesion of the rock mass, this regularization term is embedded in the attention mechanism layer of the Transformer, constraining the attention allocation of the model to conform to the slope mechanics transmission logic;

[0064] During the training phase, the total loss function is set to be the cross-entropy classification loss plus a regularization term, i.e. ,in, The regularization coefficient is set to 0.15, the training batch size is set to 16, the initial learning rate is 1e-4, and the AdamW optimizer is used for iterative training. The maximum number of iterations is 100. When the loss value decreases by less than 1e-5 for 5 consecutive iterations, the model is considered to have converged, and training is stopped to obtain a causal Transformer model that can be directly called.

[0065] The steps for assigning causal attention weights to the initial features of each modality in S2 are as follows:

[0066] Retrieve the pre-stored slope deformation causal chain rules, which include three types of causal transmission paths: rainfall triggering, artificial disturbance triggering, and geological defect triggering. Calculate the matching degree between the initial features of each mode and the three types of causal transmission paths. Based on the matching degree, assign causal attention weights in the range of 0 to 1 to the corresponding initial features, generate weight labels corresponding to the initial features of each mode, and use the weight labels as the basis for removing S2 noise features.

[0067] Specifically, the causal attention weight assignment process is executed based on predefined deformable causal chain logic:

[0068] The pre-stored slope deformation causal chain rules include three types of standard causal transmission paths: rainfall-triggered, human disturbance-triggered, and geological defect-triggered. The standard feature vector corresponding to each type of path has a dimension of 512.

[0069] When assigning values, first calculate the matching degree between each initial feature and the three types of causal transmission paths. The matching degree is calculated using the cosine similarity formula. We obtained, among which, For the first 1 initial feature vector For the first Standard feature vectors of causal transmission paths These correspond to three types of conduction pathways;

[0070] Then, the maximum value of the three-class matching degree corresponding to each feature is taken as the causal attention weight of that feature. ,Right now , The value range is [0,1], and the final value will be... Features with a value less than 0.2 are considered noise features with no causal relationship and are therefore removed.

[0071] The steps for determining the core cause of the current deformation in S3 are as follows:

[0072] The system retrieves a pre-defined feature library of deformation causes, which stores standard feature association paths corresponding to three types of causes: natural rainfall, artificial disturbance, and inherent geological defects. It calculates the cosine similarity between the feature association path and the three types of standard feature association paths, selects the type of cause with the highest cosine similarity as the core cause of the current deformation, generates a core cause label, and uses the core cause label as the input parameter for S4 risk probability calculation.

[0073] Specifically, the matching and judgment process of the core trigger is performed based on a pre-set feature library of deformation triggers:

[0074] The deformation induced feature library pre-stores standard feature association paths corresponding to three types of induced factors: natural rainfall, artificial disturbance, and inherent geological defects. The standard vector dimension of each type of path is 512-dimensional and is bound to the corresponding induced factor type label.

[0075] During matching, the feature correlation path vector output by the causal Transformer is first calculated. Path vectors associated with three types of standard features The cosine similarity is calculated using the following formula: These correspond to three types of triggers;

[0076] The type of cause with the highest similarity is taken as the core cause of the current deformation. If the highest similarity is lower than 0.7, feature re-extraction and verification are triggered. If the similarity after re-extraction is still lower than the threshold, a multimodal data re-collection instruction is triggered to ensure the reliability of cause judgment.

[0077] After S2 removes noise features without causal correlation, the following noise verification steps are also performed:

[0078] Retrieve the pre-stored whitelist of effective features of slope deformation, match the initial feature set after removing noise with the whitelist of effective features of slope deformation, and if there is a feature that is successfully matched but is judged as noise, re-include the feature in the effective feature set to generate a verified effective feature set. Use the verified effective feature set as the input data source for generating the S3 feature correlation path.

[0079] Specifically, the noise verification process is performed based on a pre-defined whitelist of valid features to prevent core related features from being mistakenly removed.

[0080] The pre-stored whitelist of valid slope deformation features includes 12 mandatory core features: InSAR average deformation rate, InSAR deformation rate increment, crack ratio in visible light images, infrared image temperature anomalies, daily increment of pore water pressure, cumulative rainfall, daily increment of surface displacement, rock mass cohesion, internal friction angle, slope, slope height, and lithological type. During verification, a unique feature ID is used for precise matching. If a feature's ID falls within the whitelist, it is re-included in the valid feature set regardless of whether its causal attention weight is below 0.2. If the dimension of the valid feature set is less than 64 after verification, it is considered insufficient, triggering a multimodal data re-acquisition command.

[0081] After obtaining the deformation risk probability, S4 performs the following risk classification steps:

[0082] The system retrieves preset risk level classification thresholds, which include three gradients: low risk, medium risk, and high risk. It compares the deformation risk probability with the magnitude of each gradient threshold to obtain the deformation risk level corresponding to the slope to be monitored, generates deformation risk results with level labels, and pushes the deformation risk results to the preset highway maintenance platform interface.

[0083] Specifically, the risk classification and notification process is adapted to the routine handling needs of highway maintenance:

[0084] The preset risk level classification thresholds are divided into three tiers: a low-risk threshold of 0.3, a medium-risk threshold of 0.7, meaning that a deformation risk probability below 0.3 is considered low-risk, between 0.3 (inclusive) and 0.7 is considered medium-risk, and above 0.7 is considered high-risk. Risk result pushes use an HTTP RESTful interface with JSON transmission format. The transmitted fields include four core fields: monitoring timestamp, deformation risk probability, risk level label, and core cause type. The interface response delay is controlled to not exceed 2 seconds. If a high-risk result is determined, an additional SMS alarm command is triggered and automatically pushed to the preset maintenance personnel's contact list interface to ensure efficient response to high-risk events.

[0085] After obtaining the deformation risk probability, S4 also performs the following precision verification step:

[0086] The measured deformation rate data of the preset GNSS monitoring points of the slope to be monitored are retrieved, the deformation rate calculation result corresponding to the deformation risk probability is extracted, the mean square error between the deformation rate calculation result and the measured deformation rate data is calculated, and if the mean square error is higher than the preset accuracy threshold, a re-acquisition command for multimodal heterogeneous raw data is triggered, an accuracy verification qualified mark is generated, and the accuracy verification qualified mark is used as a prerequisite for the output of deformation risk probability.

[0087] Specifically, the accuracy verification process uses GNSS measured data as a benchmark, while simultaneously implementing dynamic iterative optimization of the model:

[0088] The sampling frequency of the measured deformation rate at the GNSS monitoring point is 1Hz, and the daily average value is taken as the measured reference value, with the unit being mm / year;

[0089] During verification, the calculated deformation rate corresponding to the deformation risk probability is first extracted, and the mean square error is calculated by comparing it with the GNSS measured value. The calculation formula is as follows: ,in, For the first Calculated deformation rate values ​​for each GNSS point. The GNSS measured deformation rate at the corresponding point. This represents the total number of GNSS monitoring points within the monitoring area.

[0090] The preset accuracy threshold is 2.5 mm / year. If the MSE is higher than this threshold, multimodal data will be re-acquired. At the same time, the data pairs will be stored as fine-tuning samples in the model training library. Every 100 sets of fine-tuning samples will trigger a model fine-tuning, so as to achieve dynamic iterative improvement of model accuracy.

[0091] After obtaining the core cause, S3 also performs the following source path generation steps:

[0092] Extract the full-link transmission nodes of the core causes in the feature correlation path, generate a visualized deformation cause tracing path diagram, annotate the feature value and weight ratio of each node in the tracing path diagram, generate a deformation cause tracing path file, and bind the tracing path file with the deformation risk probability to the preset monitoring database.

[0093] Specifically, the process of generating the deformation cause tracing path takes into account both the needs of visualization and structured storage:

[0094] First, feature nodes with a weight ratio of no less than 5% in the feature correlation path are extracted as path nodes. The directed edges between nodes correspond to causal transmission logic, and the width of the edge is positively correlated with the weight ratio.

[0095] The source path graph is generated in the form of a directed acyclic graph (DAG), and the node colors are distinguished by modality:

[0096] Spaceborne InSAR feature nodes are set to blue, UAV imagery nodes to green, ground sensor nodes to orange, and geological attribute nodes to gray. The corresponding feature values ​​and weight percentages are also labeled next to each node.

[0097] The final generated traceability path file contains a visualization in SVG format and structured data in JSON format. The JSON fields contain four types of information for each node: feature ID, feature value, weight percentage, and modality.

[0098] A highway slope deformation monitoring system based on multimodal remote sensing includes:

[0099] The data acquisition module is used to synchronously acquire multimodal heterogeneous raw data, including spaceborne InSAR deformation time series data, UAV visible light imagery, UAV infrared imagery, pore water pressure time series data, rainfall time series data, and surface displacement time series data collected by ground sensors, and generate a multimodal heterogeneous raw dataset.

[0100] The feature processing module is used to input the multimodal heterogeneous raw dataset into the feature extraction layer, extract features from each modality data to obtain the initial feature set of each modality, input the initial feature set of each modality into the pre-built causal Transformer model, assign causal attention weights to the initial features of each modality, and remove noisy features with weights below a preset threshold to obtain the effective feature set;

[0101] The cause judgment module inputs the effective feature set into the output layer of the causal Transformer model to obtain the feature correlation path. It then matches the feature correlation path with the preset deformation cause feature library to determine the core cause of the current deformation. The core causes include three categories: natural rainfall, human disturbance, and inherent geological defects.

[0102] The results output module is used to input the core causes and effective feature sets into the preset risk probability calculation operator to obtain the deformation risk probability of the slope to be monitored.

[0103] In summary, the highway slope deformation monitoring method and system based on multimodal remote sensing disclosed in this invention has at least the following beneficial effects:

[0104] (1) The causal Transformer model is used to assign causal attention weights to the features of the synchronously collected multimodal heterogeneous raw data, replacing the existing shallow feature hard splicing or decision-level voting hard fusion mode. The causal mining logic corresponding to the slope geological and physical constraints is embedded into the entire data fusion process, which can identify the complete physical causal transmission path corresponding to slope deformation, filling the technical gap of existing fusion methods without causal correlation mining logic.

[0105] (2) By assigning causal attention weights, noise features without causal relationship are actively eliminated, avoiding interference from irrelevant noise in the feature association analysis process, and effectively improving the identification accuracy of the core cause of deformation.

[0106] (3) Based on the feature correlation path matching of the causal Transformer output, the core cause of deformation is determined. Compared with the existing technology which only relies on the fitting analysis logic of data statistical features, the probability of misjudgment caused by statistical bias is greatly reduced, and the false alarm rate of deformation risk is effectively reduced.

[0107] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A highway slope deformation monitoring method based on multi-modal remote sensing, characterized in that, Includes the following steps: S1: Simultaneously acquire multimodal heterogeneous raw data including spaceborne InSAR deformation time series data, UAV visible light imagery, UAV infrared imagery, pore water pressure time series data collected by ground sensors, rainfall time series data, and surface displacement time series data, and generate a multimodal heterogeneous raw dataset. S2: Input the multimodal heterogeneous original dataset into the feature extraction layer, extract features from each modality data to obtain the initial feature set of each modality, input the initial feature set of each modality into the pre-built causal Transformer model, assign causal attention weights to the initial features of each modality, and remove noisy features with weights lower than a preset threshold to obtain the effective feature set; S3: Input the effective feature set into the output layer of the causal Transformer model to obtain the feature correlation path, match the feature correlation path with the preset deformation cause feature library, and determine the core cause of the current deformation. The core cause includes three categories: natural rainfall, artificial disturbance, and inherent geological defects. S4: Input the core inducing factors and the effective feature set into a preset risk probability calculation operator to obtain the deformation risk probability of the slope to be monitored.

2. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, After generating the multimodal heterogeneous raw dataset in step S1, the following preprocessing steps are performed: The multimodal heterogeneous raw dataset is input into a standardized processing operator library. Phase unwrapping and atmospheric error correction are performed on the spaceborne InSAR deformation time series data. Geometric and radiometric correction are performed on the UAV visible light image and UAV infrared image. Outlier removal and timestamp alignment are performed on the pore water pressure time series data, rainfall time series data, and surface displacement time series data to generate a standardized multimodal dataset. The standardized multimodal dataset is used as the input data source for the S2 feature extraction.

3. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, The steps for constructing the causal Transformer model are as follows: Obtain publicly available historical case datasets of highway slope deformation and a pre-defined set of slope rock mechanics and physical constraint rules. Embed the set of slope rock mechanics and physical constraint rules into the attention mechanism layer of Transformer as a causal regularization term. Use the deformation cause identification accuracy as the loss function for iterative training. Stop training when the loss function converges to a preset threshold to generate a trained causal Transformer model. Use the trained causal Transformer model as the execution subject for the S2 weight assignment.

4. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, The step in S2 of assigning causal attention weights to the initial features of each modality is as follows: Retrieve pre-stored slope deformation causal chain rules, which include three types of causal transmission paths: rainfall triggering, artificial disturbance triggering, and geological defect triggering. Calculate the matching degree between the initial features of each modality and the three types of causal transmission paths. Based on the matching degree, assign causal attention weights in the range of 0 to 1 to the corresponding initial features, generate weight labels corresponding to the initial features of each modality, and use the weight labels as the basis for determining the removal of the S2 noise features.

5. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, The step in S3 to determine the core cause of the current deformation is as follows: A preset deformation cause feature library is retrieved. The deformation cause feature library stores standard feature association paths corresponding to three types of causes: natural rainfall, artificial disturbance, and inherent geological defects. The cosine similarity between the feature association path and the three types of standard feature association paths is calculated. The cause type corresponding to the type with the highest cosine similarity is selected as the core cause of the current deformation. A core cause label is generated and used as the input parameter for the S4 risk probability calculation.

6. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, After S2 removes noise features without causal correlation, the following noise verification steps are also performed: Retrieve the pre-stored whitelist of effective features of slope deformation, match the initial feature set after removing noise with the whitelist of effective features of slope deformation, if there is a feature that is successfully matched but is judged as noise, re-include the feature in the effective feature set, generate a verified effective feature set, and use the verified effective feature set as the input data source for generating the S3 feature correlation path.

7. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, After obtaining the deformation risk probability in step S4, the following risk classification steps are also performed: The preset risk level classification threshold is retrieved. The risk level classification threshold includes three gradients: low risk threshold, medium risk threshold, and high risk threshold. The relationship between the deformation risk probability and each gradient threshold is compared to obtain the deformation risk level corresponding to the slope to be monitored. Deformation risk results with level labels are generated and pushed to the preset highway maintenance platform interface.

8. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, After obtaining the deformation risk probability in step S4, the following precision verification step is also performed: The measured deformation rate data of the preset GNSS monitoring points of the slope to be monitored are retrieved, the deformation rate calculation result corresponding to the deformation risk probability is extracted, the mean square error between the deformation rate calculation result and the measured deformation rate data is calculated, and if the mean square error is higher than the preset accuracy threshold, the re-acquisition instruction of the multimodal heterogeneous raw data is triggered, an accuracy verification qualified identifier is generated, and the accuracy verification qualified identifier is used as the pre-judgment condition for the output of the deformation risk probability.

9. The method for monitoring highway slope deformation based on multimodal remote sensing according to claim 1, characterized in that, After obtaining the core cause, S3 also performs the following source path generation step: Extract the full-link transmission nodes of the core cause in the feature correlation path, generate a visualized deformation cause tracing path diagram, annotate the feature value and weight ratio of each node in the tracing path diagram, generate a deformation cause tracing path file, and bind the tracing path file with the deformation risk probability to a preset monitoring database.

10. A highway slope deformation monitoring system based on multimodal remote sensing, used to execute the highway slope deformation monitoring method based on multimodal remote sensing as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to synchronously acquire multimodal heterogeneous raw data, including spaceborne InSAR deformation time series data, UAV visible light imagery, UAV infrared imagery, pore water pressure time series data, rainfall time series data, and surface displacement time series data collected by ground sensors, and generate a multimodal heterogeneous raw dataset. The feature processing module is used to input the multimodal heterogeneous original dataset into the feature extraction layer, extract features from each modality data to obtain an initial feature set for each modality, input the initial feature set of each modality into a pre-built causal Transformer model, assign causal attention weights to the initial features of each modality, and remove noisy features with weights lower than a preset threshold to obtain an effective feature set. The cause determination module inputs the effective feature set into the output layer of the causal Transformer model to obtain the feature correlation path, matches the feature correlation path with a preset deformation cause feature library, and determines the core cause of the current deformation. The core cause includes three categories: natural rainfall, artificial disturbance, and inherent geological defects. The result output module is used to input the core inducing factors and the effective feature set into a preset risk probability calculation operator to obtain the deformation risk probability of the slope to be monitored.

Images