A mine slope disaster early identification method based on multi-source remote sensing dynamic collaboration and big data fusion
By using satellite-UAV collaborative acquisition of multi-source remote sensing data and combining an improved temporal attention U-Net network and closed-loop verification mechanism, the problem of insufficient accuracy and timeliness in early identification of mine slope disasters was solved, achieving efficient and reliable identification of disaster hazards.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional ground monitoring methods and single remote sensing technologies are insufficient for early and accurate identification of large-scale mine slope disasters. Existing multi-source data fusion methods fail to fully explore the dynamic correlation of time-series data, and intelligent identification models ignore time-series dynamic changes, resulting in insufficient identification accuracy and timeliness.
By employing satellite-UAV collaborative acquisition of multiple types of remote sensing data, and through dynamic collaboration of multi-source remote sensing and big data fusion, combined with an improved temporal attention U-Net network and closed-loop verification mechanism, dynamic collaborative acquisition, fusion, and intelligent identification of multi-source data are achieved.
It enables early, accurate, and large-scale identification of mine slope hazards, improves identification efficiency and reliability, achieves an accuracy rate of over 95%, and ensures the long-term reliability of identification results.
Smart Images
Figure CN122153581A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the intersection of mine disaster monitoring, remote sensing technology, big data fusion and artificial intelligence, and is specifically applicable to the early identification and warning of mine slope disasters such as landslides and collapses, especially applicable to the dynamic monitoring of large-scale mine slopes under complex terrain conditions. Background Technology
[0002] Mine slope hazards are characterized by their suddenness and destructive power, seriously threatening mine production safety, the lives and property of workers, and the surrounding ecological environment. Early and accurate identification of hazards is crucial to reducing disaster losses. Traditional ground monitoring methods (total stations, sensors) have limited monitoring range and high costs, making it difficult to cover large areas of mine slopes. Single remote sensing technologies (optical, InSAR) have inherent limitations: optical remote sensing is susceptible to weather conditions and insufficient in identifying minute deformations, while InSAR technology performs poorly in vegetated areas and is easily affected by atmospheric delays. Existing multi-source data fusion methods mostly involve static data overlay, failing to fully explore the dynamic correlations of time-series data and lacking integration of key features such as slope structure. Intelligent identification models often focus on spatial features, ignoring temporal dynamic changes, resulting in insufficient accuracy and timeliness in early hazard identification. With the development of satellite-UAV collaborative remote sensing technology, big data processing technology, and deep learning, achieving dynamic collaborative acquisition and deep fusion of multi-source remote sensing data, combined with intelligent models to mine spatiotemporal dynamic features, has become the main direction for solving the problem of early identification of mine slope hazards. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this paper proposes an early identification method for mine slope hazards based on multi-source remote sensing dynamic collaboration and big data fusion. This method enables early, accurate, and large-scale identification of potential mine slope hazards, improving identification efficiency and reliability.
[0004] The technical solution of the present invention is as follows: An early identification method for mine slope hazards based on multi-source remote sensing dynamic collaboration and big data fusion includes the following steps: (1) Multi-source remote sensing dynamic collaborative acquisition: Multi-type remote sensing time series data of mine slope area are acquired through satellite remote sensing platform and UAV remote sensing platform. The multi-type remote sensing time series data is formed by integrating the collected multi-type remote sensing data into a time series data sequence. The collected multi-type remote sensing data includes InSAR deformation data, high-resolution optical remote sensing data, LiDAR point cloud data and thermal infrared remote sensing data. The InSAR data acquisition cycle does not exceed 12 days. UAV data covers the slope excavation area and historical disaster point area first. Other data are collected synchronously according to the corresponding cycle. (2) Data preprocessing and spatiotemporal alignment: The collected multi-type remote sensing time series data are processed by format conversion, geometric correction, radiometric correction and denoising. Spatiotemporal registration is achieved through a unified coordinate system (WGS84) and a time reference (UTC). Data missing areas are supplemented based on Kriging interpolation to generate a standardized dataset. (3) Three-level dynamic fusion of big data: Based on the Hadoop big data framework, a layered dynamic fusion strategy is adopted to process standardized datasets: ① Data layer fusion: Adaptive weighted averaging is used to eliminate redundancy in heterogeneous data from the same source, and the weights are dynamically adjusted according to the consistency of data time series and signal-to-noise ratio; WCTM weighted coherence thresholding method is introduced to optimize InSAR interferometric baseline and reduce vegetation cover interference; ② Feature layer fusion: Extract deformation features, topographic features, vegetation features, temperature features and slope structure features, and reduce the dimensionality of the features using principal component analysis to obtain a fused feature set; ③ Decision-level fusion: Based on the improved Bayesian inference model, a time-series weight factor is introduced (the weight of recent data is higher than that of distant data), and the results of feature-level fusion are subjected to multi-source decision fusion to obtain the preliminary hidden danger area; (4) Temporal attention-enhanced intelligent recognition: Constructing an improved temporal attention U-Net network: The encoding end of the U-Net network includes four convolutional blocks arranged in sequence, and each convolutional block is connected in series with a CBAM attention module. The CBAM attention module includes a channel attention sub-module and a spatial attention sub-module connected in sequence. The output end of the fourth convolutional block of the encoding end is used as the input of the temporal feature extraction branch and is connected in parallel with three GRU units arranged in parallel. The output ends of the first three convolutional blocks of the encoding end are also connected to the temporal feature extraction branch and are respectively connected to the three GRU units and the spatiotemporal fusion layer. The output ends of the three GRU units are connected to the spatiotemporal fusion layer. The spatiotemporal fusion layer is used to fuse the multi-scale spatial features of the first three convolutional blocks of the encoding end and the multi-period temporal features extracted by the three GRU units to generate spatiotemporal fusion features. The decoding end fuses the upsampled features of each layer with the spatiotemporal fusion features output by the spatiotemporal fusion layer through upsampling operation, and finally outputs the classification results of mine slope hazard levels. Using the fusion feature set obtained in step (3) ② and the time series data sequence described in step (1) as input, the unstable slope area is identified after training, and the three levels of disaster hazard level (low, medium and high) are output. (5) Closed-loop verification and dynamic update: Combine the total station measurement data and displacement sensor data measured on the ground to verify the identification results. The accuracy, recall and F1 value are calculated using the confusion matrix to correct the error. A dynamic update mechanism is established to update the fusion feature set and identification model parameters in real time according to the newly collected remote sensing data, and output the final identification results and early warning information.
[0005] Furthermore, in step (1), the multi-source remote sensing dynamic collaborative acquisition also includes data acquisition priority scheduling: the frequency of UAV acquisition in high-risk areas is 2-3 times per month.
[0006] Furthermore, in step (1), the satellite remote sensing data selected are Sentinel-1, Gaofen-6, and Landsat-8 data, and the LiDAR point cloud data point cloud density is not less than 50 points / m². 2 .
[0007] Further, in step (3), the extraction of slope structure features includes: generating a three-dimensional terrain model using LiDAR point cloud data, calculating the angle between the rock layer dip and the slope aspect, adjusting the screening threshold based on the number of free faces (single / double / three free faces), retaining slope units with an angle <30° in single free face areas, retaining slope units with an angle <60° in double free face areas, and retaining slope units with an angle <90° in three free face areas.
[0008] Further, in step (3), the deformation features include deformation rate and cumulative deformation; the terrain features include slope, aspect and elevation; the vegetation features include NDVI index; the temperature features include surface temperature anomaly; and the slope structure features include the layer dip-aspect angle and the number of free faces.
[0009] Furthermore, in step (3), the setting of the time series weight factor follows the principle that the weight of recent data is higher than that of distant data.
[0010] Furthermore, in step (4), the training iterations of the improved temporal attention U-Net network are 100-150 times, and the learning rate is dynamically adjusted using a cosine annealing strategy.
[0011] Furthermore, the update cycle of the dynamic update mechanism in step (5) is synchronized with the remote sensing data acquisition cycle. When the accuracy of high-risk area identification is less than 90%, the feature layer fusion weight and model parameters are re-optimized.
[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. Employing dynamic collaborative data acquisition via satellite and drones (as shown in Figure 2), this approach balances large-scale monitoring with precise detection in key areas. By prioritizing data scheduling, it improves acquisition efficiency and overcomes the limitations of a single platform. 2. An innovative three-level dynamic fusion strategy (corresponding to Figure 3) is adopted to integrate temporal features and slope structure features, introduce the WCTM method to optimize the interference baseline, dynamically adjust the fusion weight, fully explore the complementarity of data, and improve data utilization and reliability. 3. Improved Temporal Attention U-Net Network (corresponding to)Figure 4 By integrating spatial and temporal features, strengthening the extraction of disaster-sensitive features, the accuracy rate of high-risk area identification is ≥95%, and improving the automation level of early hazard identification; 4. Construct a closed-loop verification and dynamic update mechanism (corresponding to Figure 1), combine ground measurement data to correct errors, update model parameters in real time, ensure the long-term reliability of identification results, and provide dynamic technical support for mine disaster prevention and control. Attached Figure Description
[0013] Figure 1 is an overall flowchart of the method of the present invention; Figure 2 is a schematic diagram of the platform and data types for multi-source remote sensing dynamic collaborative acquisition according to the present invention; Figure 3 is a schematic diagram illustrating the principle of the three-level dynamic fusion of big data in this invention; Figure 4 is a schematic diagram of the improved temporal attention U-Net network structure of the present invention. Detailed Implementation
[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0015] This embodiment uses a slope of an open-pit coal mine as the study area (area 50 km²). 2 The method of this invention is used for early identification of slope hazards. The specific steps are as follows: 1. Dynamic collaborative acquisition of data from multiple remote sensing sources (corresponding to Figure 2) As shown in Figure 2, multi-source remote sensing data is acquired collaboratively through a dual-platform approach of satellite and UAV, forming a data collection mode of "wide coverage + focus on key areas": (1) Satellite data acquisition: Six months of InSAR deformation data were acquired by Sentinel-1 satellite (spatial resolution 10m, acquisition period 12 days), three clear and cloudless high-resolution optical images were acquired by Gaofen-6 satellite (spatial resolution 2m), and corresponding thermal infrared data were acquired by Landsat-8 satellite (spatial resolution 30m). (2) Drone data: A DJI M300 drone equipped with a LiDAR device and a full-frame camera was used to collect point cloud data (density 50 points / m²) three times in key areas such as the slope excavation area and historical landslide points. 2 (and local high-definition optical images, the acquisition range covers) Figure 2 "Drone-focused areas" in China.
[0016] (3) Data priority scheduling: The collection frequency is dynamically adjusted according to the regional risk level, and the collection is intensified in high-risk areas to form a multi-source remote sensing time series data sequence.
[0017] 2. Data preprocessing and spatiotemporal alignment (1) Optical and thermal infrared data: The format conversion (original format to TIFF), geometric correction (WGS84) and radiometric correction were completed using ENVI software to eliminate the effects of atmospheric and surface reflections; (2) InSAR data: Goldstein filtering was performed using GAMMA software to remove atmospheric delay interference and extract deformation rate and cumulative deformation. (3) LiDAR data: Outlier removal and ground point filtering were performed using CloudCompare software to generate a digital elevation model (DEM). (4) Spatiotemporal registration: By using ArcGIS software to unify the coordinate system and time reference, Kriging interpolation is used to supplement data for some cloud-obscured areas to generate a standardized dataset.
[0018] 3. Dynamic integration of big data at three levels (corresponding to Figure 3) As shown in Figure 3, based on the Hadoop big data framework, a three-level dynamic fusion of "data layer - feature layer - decision layer" is implemented: (1) Data layer fusion: The optical data of Gaofen-6 and UAV (from the same source but different origins) are fused using an adaptive weighted average method. The weights are dynamically adjusted according to the data signal-to-noise ratio (0.65 for UAV data and 0.35 for satellite data), improving the spatial resolution to 1.5m. The WCTM method is used to optimize the InSAR interferometric baseline. The interferogram is segmented according to the vegetation cover level and assigned coherence coefficient weights to reduce the impact of vegetation cover on data quality, thereby improving the coherence of the vegetation cover area by 20%. (2) Feature layer fusion: 12 feature parameters were extracted (deformation rate and cumulative deformation of InSAR data, slope, aspect and elevation of LiDAR data, NDVI index of optical data, surface temperature anomaly value of thermal infrared data (more than 2℃ higher than the surrounding area), rock layer-slope aspect angle, etc.). After dimensionality reduction of the feature parameters by principal component analysis, 5 principal components were retained (cumulative contribution rate of 93%). (3) Decision-making level fusion: Based on the improved Bayesian inference model, a time-series weight factor is introduced (the weight of data in the last 3 months is 0.7, and the weight of data in the early period is 0.3). The fusion feature set is input to calculate the posterior probability of each region as a potential disaster area, and three potential disaster areas with a posterior probability > 0.6 are obtained.
[0019] 4. Temporal attention-enhanced intelligent recognition (corresponding to Figure 4) like Figure 4 As shown, an improved temporal attention U-Net network is constructed to enhance the spatiotemporal feature fusion capability: (1) Network structure: The encoder contains 4 convolutional blocks, embeds a CBAM attention module to enhance the extraction of sensitive features, and adds a time-series feature extraction branch (GRU unit) to capture the dynamic changes of data in multiple periods; the decoder fuses the features of the encoder with the upsampling to output the classification results of the hidden danger level. (2) Model training: 20 historical disaster sites and 60 stable areas were selected as samples to train the improved temporal attention U-Net network (e.g., Figure 4 As shown, the model incorporates a CBAM module and a GRU branch, with an initial learning rate of 0.001. After 120 iterations, the model achieved a test set accuracy of 94%. (3) Intelligent identification: Input the fusion feature set and time series data sequence to identify 1 high-risk area, 2 medium-risk areas and 3 low-risk areas, and clarify the spatial coordinates and range of each area.
[0020] 5. Closed-loop verification and dynamic update (corresponding to Figure 1) As shown in Figure 1, a closed-loop process of "identification-verification-update" is formed: (1) Results Verification: Total station monitoring points and displacement sensors were deployed in the study area to obtain ground deformation data. The data were compared with the identification results, and the accuracy, recall, and F1 score were calculated to correct the identification error. Ground measurement data showed that the overlap between high-risk areas and areas with large measured deformation reached 96%, and the overlap between medium- and low-risk areas reached 91%. The identification accuracy was significantly higher than that of the single InSAR technology (82%). (2) Dynamic Update: Based on the newly acquired remote sensing time-series data, the remote sensing data is updated monthly, and the fused feature set and model parameters are updated synchronously to maintain the timeliness of the recognition results. The update process follows Figure 1 The "dynamic update mechanism" in China; (3) Early warning output: Generate early identification report of mine slope disaster, mark high-risk areas (coordinates: 112°35′-112°37′ east longitude, 37°18′-37°20′ north latitude, hazard level), and recommend strengthening real-time monitoring and protection.
[0021] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for early identification of mine slope hazards based on multi-source remote sensing dynamic collaboration and big data fusion, characterized in that, Includes the following steps: (1) Multi-source remote sensing dynamic collaborative acquisition: Multi-type remote sensing time series data of mine slope area are acquired through satellite remote sensing platform and UAV remote sensing platform. The multi-type remote sensing time series data is formed by integrating the collected multi-type remote sensing data into a time series data sequence. The collected multi-type remote sensing data includes InSAR deformation data, high-resolution optical remote sensing data, LiDAR point cloud data and thermal infrared remote sensing data. The InSAR data acquisition cycle does not exceed 12 days. UAV data covers the slope excavation area and historical disaster point area first. Other data are collected synchronously according to the corresponding cycle. (2) Data preprocessing and spatiotemporal alignment: The collected multi-type remote sensing time series data are processed by format conversion, geometric correction, radiometric correction and noise reduction. Spatiotemporal registration is achieved by unifying the coordinate system and time reference. Data missing areas are supplemented by Kriging interpolation to generate a standardized dataset. (3) Three-level dynamic fusion of big data: Based on the Hadoop big data framework, a layered dynamic fusion strategy is adopted to process standardized datasets: ① Data layer fusion: Adaptive weighted averaging is used to eliminate redundancy in heterogeneous data from the same source, and the weights are dynamically adjusted according to the consistency of data time series and signal-to-noise ratio; WCTM weighted coherence thresholding method is introduced to optimize InSAR interferometric baseline and reduce vegetation cover interference; ② Feature layer fusion: Extract deformation features, topographic features, vegetation features, temperature features and slope structure features, and reduce the dimensionality of the features using principal component analysis to obtain a fused feature set; ③ Decision-level fusion: Based on the improved Bayesian inference model, a time-series weight factor is introduced (the weight of recent data is higher than that of distant data), and the results of feature-level fusion are subjected to multi-source decision fusion to obtain the preliminary hidden danger area; (4) Temporal Attention Enhanced Intelligent Recognition: Constructing an improved temporal attention U-Net network: The encoding end of the U-Net network includes four convolutional blocks arranged sequentially, each convolutional block being connected in series with a CBAM attention module. The CBAM attention module includes a channel attention sub-module and a spatial attention sub-module connected in sequence. The output of the fourth convolutional block of the encoding end serves as the input of the temporal feature extraction branch and is connected in parallel to three GRU units arranged in parallel. The outputs of the first three convolutional blocks of the encoding end are also connected to the temporal feature extraction branch and are respectively connected to the three GRU units and the spatiotemporal fusion layer. The outputs of the three GRU units are connected to the spatiotemporal fusion layer, which is used to fuse the multi-scale spatial features of the first three convolutional blocks of the encoding end and the three GRU units. The multi-period time series features extracted by the unit are used to generate spatiotemporal fusion features; the decoding end performs upsampling operation to fuse the upsampled features of each layer with the spatiotemporal fusion features output by the spatiotemporal fusion layer, and finally outputs the classification results of mine slope hazard levels; the fusion feature set obtained in step (3) ② and the time series data sequence described in step (1) are used as input, and after training, unstable slope areas are identified and low, medium and high three-level disaster hazard levels are output; (5) Closed-loop verification and dynamic update: Combine the total station measurement data and displacement sensor data measured on the ground to verify the identification results. The accuracy, recall and F1 value are calculated using the confusion matrix to correct the error. A dynamic update mechanism is established to update the fusion feature set and identification model parameters in real time according to the newly collected remote sensing data, and output the final identification results and early warning information.
2. The method according to claim 1, characterized in that, In step (1), the multi-source remote sensing dynamic collaborative acquisition also includes data acquisition priority scheduling: the frequency of UAV acquisition in high-risk areas is 2-3 times per month.
3. The method according to claim 1, characterized in that, In step (1), the satellite remote sensing data selected are Sentinel-1, Gaofen-6, and Landsat-8 data, and the LiDAR point cloud data point cloud density is not less than 50 points / m². 2 .
4. The method according to claim 1, characterized in that, In step (3), the extraction of slope structure features includes: generating a three-dimensional terrain model using LiDAR point cloud data, calculating the angle between the rock layer dip and the slope aspect, adjusting the screening threshold based on the number of free faces, retaining slope units with an angle <30° in single-face free areas, retaining slope units with an angle <60° in double-face free areas, and retaining slope units with an angle <90° in three-face free areas.
5. The method according to claim 1, characterized in that, In step (3), the deformation features include deformation rate and cumulative deformation; the terrain features include slope, aspect and elevation; the vegetation features include NDVI index; the temperature features include surface temperature anomaly; and the slope structure features include the layer dip-aspect angle and the number of free faces.
6. The method according to claim 1, characterized in that, In step (3), the setting of the time series weight factor follows the principle that the weight of recent data is higher than that of distant data.
7. The method according to claim 1, characterized in that, The improved temporal attention U-Net network in step (4) is trained 100-150 times, and the learning rate is dynamically adjusted using a cosine annealing strategy.
8. The method according to claim 1, characterized in that, In step (5), the update cycle of the dynamic update mechanism is synchronized with the remote sensing data acquisition cycle. When the accuracy of high-risk area identification is less than 90%, the feature layer fusion weight and model parameters are re-optimized.