Ground fracture detection method based on remote sensing image visual analysis

Abstract

The invention discloses a ground fracture detection method based on remote sensing image visual analysis, which comprises the following steps of: sending a remote sensing image to a trained ground fracture dynamic monitoring model, outputting a fracture path by the ground fracture dynamic monitoring model, outputting a vector element graph layer containing a detected fracture position, and identifying a section of path identified as a fracture corresponding to each element, displaying through a GIS platform; the construction and training process of the ground fracture dynamic monitoring model comprises the following steps: S1, constructing the ground fracture dynamic monitoring model, and manually marking the accuracy of a detection result of a DOM image sample; s2, improving the recognition precision of the fine-line object target in the model structure; s3, optimizing the extraction efficiency of the interference features; s4, obtaining landform parameters of a to-be-monitored area, and introducing a high-speed interpretation image sample; and S5, obtaining a trained ground fracture dynamic monitoring model. The invention belongs to the technical field of geological image processing, and significantly improves the efficiency and accuracy of ground fracture recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image processing and relates to a ground fissure detection method based on remote sensing image visual analysis. Technical Background The continued exploitation of underground coal resources can lead to a series of geological environmental problems, such as surface subsidence in mining areas and ground fissures caused by underground mining disturbances. These can cause surface buildings to sink and crack, resulting in damage to overlying farmland and landscapes, and endangering the lives and property of residents living near mining areas. Therefore, identifying ground fissures is of great theoretical and practical significance for environmental protection in mining areas, post-mining environmental restoration and management, safeguarding the lives and property of residents in mining areas, and promoting the coordinated development of coal resource development and the environment.

[0002] Due to the vast distribution of mining areas and the difficulty of ground transportation, traditional manual surveys are time-consuming, labor-intensive, and inefficient. There is an urgent need to develop a new ground fissure detection method based on remote sensing visual imagery and an improved ground fissure dynamic monitoring model to automatically identify and extract ground fissures in mining areas. Summary of the Invention

[0003] The purpose of the present invention is to provide a ground fissure detection method based on remote sensing image visual analysis, which solves the problem that in the process of constructing a dynamic ground fissure monitoring model in the prior art, a large number of image samples are required for training and the inspection results are not accurate enough.

[0004] The technical solution adopted by the present invention is a ground fissure detection method based on visual analysis of remote sensing images. The remote sensing images are sent to a trained ground fissure dynamic monitoring model. The model outputs the crack path and a vector feature layer containing the detected crack locations. Each feature corresponds to a section of the path identified as a crack, which is displayed through a GIS platform. The construction and training process of the ground fissure dynamic monitoring model is as follows: S1: Construct a dynamic monitoring model for ground fissures and calculate the accuracy of detection results by manually marking DOM image samples; S2: Improve the accuracy of thin line object recognition in the model structure; S3: Optimize the extraction efficiency of interference features; S4: Obtain the geomorphological parameters of the area to be monitored and introduce high-speed interpretation image samples; S5: Obtain the trained ground fissure dynamic monitoring model.

[0005] The beneficial effect of the present invention is that it is specifically used to automatically detect ground cracks from drone images, realizes automated analysis of drone images through artificial intelligence technology, significantly improves the efficiency and accuracy of ground crack identification, effectively replaces the traditional method of relying on manual inspections or expensive detection equipment, and helps mining area management departments to conduct wider-area, higher-frequency, and lower-cost mining area inspections, providing technical support for mining area environmental protection and restoration management. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a flowchart of the construction and training process of the ground fissure dynamic monitoring model in the method of the present invention; Figure 2 is a schematic diagram of optical remote sensing image features in the method of the present invention; Figure 3 It is a schematic diagram of marking ground fissure sample points in the method of the present invention. DETAILED DESCRIPTION

[0007] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0008] The ground fissure detection method based on remote sensing image visual analysis in the present invention sends remote sensing images to a trained ground fissure dynamic monitoring model. The model outputs the crack path and a vector feature layer containing the detected crack locations. Each feature corresponds to a section of the path identified as a crack, which is displayed through a GIS platform. The remote sensing image includes at least one of a raster image, a mosaic dataset, and an image service set, wherein the resolution of the remote sensing image is less than 2 cm / pixel.

[0009] The vector feature layer contains the crack centerline, crack boundary, and crack feature points. The crack feature points also include the crack starting and end points, crack inflection points, crack intersection points, and the widest part of the crack.

[0010] The GIS platform also includes detecting crack areas through Faster R-CNN and predicting future crack expansion trends using LSTM networks.

[0011] Reference Figure 1 ,The construction and training process of the above-mentioned ground fissure dynamic monitoring model is: S1: Construct a dynamic monitoring model for ground fissures and calculate the accuracy of detection results by manually marking DOM image samples; The DOM image samples in S1 contain crack causes, crack types, and development characteristic attributes. Crack causes include goaf subsidence, slope excavation, tectonic movement, weathering, karst collapse, engineering loading, and groundwater pumping. Crack types include tensile cracks, shear cracks, compressive cracks, and mixed cracks. Development characteristic attributes include development time, expansion rate, filling type, and surrounding environment.

[0012] S2: Improve the accuracy of thin line object recognition in the model structure; In S2, Gaussian noise, motion blur, and brightness / contrast transformation are added to the image data in the sample to simulate complex lighting. The accuracy of thin linear object recognition is improved by adding void convolution and spatial pyramid pooling.

[0013] S3: Optimize the extraction efficiency of interference features; In S3, interference features include soil cracks, shadows of high and low levels, gullies, and artificial cracks. The extraction efficiency of optimized interference features is: S31: First, a weighted penalty for difficult samples is added to the loss function to force the model to prioritize learning the features of areas prone to false detection and perform an optimization. Difficult samples include hidden cracks and cracks in complex backgrounds. S32: Construct a robustness test set, aggregate the interference features into the test set, and complete secondary optimization through manual labeling.

[0014] S4: Obtain the geomorphological parameters of the area to be monitored and introduce high-speed interpretation image samples; In S4, high-speed interpretation of image samples includes the acquisition of surface DSM images based on vehicle carriers, combined with DOM images to check for crack duplication in surface data.

[0015] S5: Obtain the trained ground fissure dynamic monitoring model.

[0016] Example 1 The present invention provides a method for detecting ground fissures based on visual analysis of remote sensing images, comprising sending remote sensing images to a trained ground fissure dynamic monitoring model, which outputs a fissure path, specifically a vector feature layer containing detected fissure locations; each feature corresponds to a path identified as a fissure, and is displayed via a GIS platform; Among them, the resolution of remote sensing images is less than 2 cm / pixel, see Figure 2 ; The vector feature layer includes the crack centerline, crack boundary, and crack feature points. The crack feature points are further divided into the crack starting point, crack end point, crack inflection point, crack intersection point, and crack widest point. The function of the GIS (Geographic Information System) platform is to detect the crack area through the Faster R-CNN network on the one hand, and predict the future expansion trend of the crack using the LSTM network on the other hand.

[0017] S1: Build a ground fissure dynamic monitoring model, use manually labeled DOM image samples as output detection results, and execute S2; Among them, DOM image samples include crack causes, crack types and development characteristic attributes; crack causes include goaf subsidence, slope excavation, tectonic movement, weathering, engineering loading, and groundwater pumping; crack types include tensile cracks, shear cracks, compressive cracks, and mixed cracks; development characteristic attributes include development time, expansion speed, filling type and surrounding environment.

[0018] The specific construction process is as follows: the original DOM image is used as a sample, and after preprocessing the original image with unified size and resolution, it is used as a training sample for the model. The DOM image sample with manual marking results is used as the output of the model. The constraint features of the model include the size and shape of the cracks, the position of the cracks, and the type of the cracks. The weights corresponding to the proportions of the above features are continuously adjusted, and multiple trainings are conducted to obtain a dynamic monitoring model for ground fissures. The specific model construction logic adopted refers to the convolutional neural network. Here, the convolutional neural network is used as a model and substituted into the training samples for training. By extracting a large number of spectral features from the samples, the results are learned and demonstrated. That is, when the result output by the model is different from the correct result of manual marking, the weight of the constraint feature, the size of the size and shape in the constraint feature, the position of the crack, and the type of the crack are modified in time until the output result is equal to the manual marking result.

[0019] S2: Improve the accuracy of the ground fissure dynamic monitoring model in identifying thin linear objects. Gaussian noise, motion blur, and brightness / contrast transformation are added to the image data in the DOM image samples to simulate complex lighting. The accuracy of thin linear target recognition is improved by adding void convolution and spatial pyramid pooling.

[0020] S3: Optimize the extraction efficiency of interference features of the sample database in the ground fissure dynamic monitoring model, Disturbance features include soil cracks, shadows of high and low surfaces, gullies, and artificial cracks. Optimizing the extraction efficiency of interference features is specifically divided into: S31) First, a weighted penalty for difficult samples is added to the loss function to force the model to prioritize learning the features of areas prone to false detection and perform an optimization. Difficult samples include hidden cracks and cracks in complex backgrounds. S32) Construct a robustness test set, aggregate the interference features into the test set, and complete the secondary optimization through manual labeling.

[0021] S4: Obtain the geomorphological parameters of the area to be monitored and introduce high-speed interpretation image samples, wherein the high-speed interpretation image samples include the acquisition of surface DSM images based on vehicle carriers, combined with DOM image samples to check for crack duplication in surface data; S5: Obtain the trained ground fissure dynamic monitoring model; Finally, the remote sensing image of the area to be monitored is sent as input to the ground fissure dynamic monitoring model, which outputs a vector feature layer of the crack location in the area to be monitored. The vector feature layer is then used to obtain how the crack is identified, that is, the crack location path, and to show the user how to locate it from the image. Figure 3 .

[0022] Example 2 Based on Example 1, the present application further includes optimizing the model training samples to overcome the shortage of existing samples and reduce dependence on existing samples. The main content of the optimization is to learn from the existing samples to generate crack samples that are artificially synthesized with vegetation coverage and shadow superposition using image editing tools (such as Photoshop); use the generative adversarial network (GAN) to learn the sample distribution law and automatically generate virtual samples. In particular, the generative adversarial network first uses the entropy method to screen samples with high uncertainty (difficult samples), and assigns higher weights to difficult samples to reduce the training impact of simple samples (such as open ground cracks). The formula is as follows:

[0023] in, α t is the category weight, γ To focus on parameters and enhance gradient updates for difficult samples, FL is the focal loss function, which is the gradient value of the virtual sample after the weight change. P t For virtual functions.

[0024] Then, the generative adversarial network (GAN) is used to learn the distribution rules of the above-mentioned multiple virtual function samples to obtain more virtual samples.

[0025] It also includes: using the CBAM module to focus on the crack area by calculating pixel-level weights, suppressing interference from vegetation, buildings, etc., and obtaining more and higher-quality sample data.

[0026] In summary, this embodiment 2 uses a generative adversarial network, first using the entropy method to screen samples with high uncertainty (difficult samples), assigning higher weights to difficult samples, and reducing the training impact of simple samples (such as open ground cracks).

[0027] Example 3 Based on the above-mentioned embodiment 1, this embodiment 3 distinguishes the types of different scenes in the sample data as follows. After classifying the samples, secondary training is performed to obtain crack paths in multiple scenes.

[0028]

[0029] During the construction and training of the dynamic ground fissure monitoring model, preprocessed samples are classified into scene types, such as the four scenarios mentioned above. For a trained model, images of the area to be monitored are fed to the model, which automatically identifies entities, natural objects, and man-made features within the image and categorizes them into their respective types. Each type contains a large number of pre-stored fissure patterns corresponding to that scenario. This scene classification ensures a higher degree of match between the images to be monitored and the pre-stored images and fissures for individual scenes in the sample library. It can even remove entities from cracks obscured by entities, preserving the intact fissure form. For example, in vegetation-covered scenarios, using the GF-5 shortwave infrared imagery (1.6μm band has strong penetration into vegetation) combined with the DenseNet + spatial attention model, the recall rate of vegetation-covered cracks in a coal mining area in Shaanxi Province was increased from 68% to 89%.

[0030] In summary, during the construction and training of the ground fissure dynamic monitoring model, the preprocessed samples are divided into four types of scenarios, which are specifically included. These four scenarios can be used to quickly overcome the difficulty of identifying cracks.

[0031] Example 4 On the basis of the above embodiment 1, a GIS platform is further included, and the GIS platform includes an ArcGIS platform and a MapGIS platform.

[0032] Launch ArcGIS Pro and create a new project or open an existing one. Create a new map within the project or use an existing map document. Load the ground fissure imagery data (such as remote sensing imagery, aerial photographs, etc.) and related geospatial data (such as terrain data, administrative division data, etc.) into the map. If the ground fissure imagery data is in raster format, add it directly using the "Add Data" button. Vector data (such as ground fissure vector boundaries extracted through image analysis) can also be added in the same way.

[0033] Input digital elevation model (DEM) data and an image of the area to be monitored into the dynamic ground fissure monitoring model. Drag and drop it into the "Elevation Surface" layer and adjust the vertical exaggeration factor to enhance the relief of the terrain. This helps visualize the relationship between ground fissures and the terrain. For example, ground fissures may be more common in areas with high relief or at specific geological structures. Ground fissure images can be overlaid on the terrain model to provide a more intuitive 3D perspective of their distribution across the terrain.

[0034] Open MapGIS and enter the appropriate workspace. Use the data import function to import the ground fissure image data (supports multiple formats, such as raster images like TIFF and JPEG, and vector data like SHP) into the project. Raster images can be imported directly; vector data requires following the software's import wizard to set relevant parameters, such as the coordinate system and data type.

[0035] Imported ground fissure images can be preprocessed, such as geometric correction and radiometric correction, to improve image quality and accuracy. In MapGIS, these operations can be performed using corresponding image processing tools. For example, geometric correction aligns the image with the actual geographic coordinate system to ensure the accuracy of the ground fissure locations; radiometric correction adjusts the image's brightness, contrast, and color balance to enhance the fissure features.

[0036] For collecting remote sensing images: use drone aerial surveys and satellite images (such as optical remote sensing and synthetic aperture radar InSAR) to extract the spatial distribution of cracks.

[0037] GIS modeling: Integrate multi-source data (such as terrain, geology, and monitoring data) into vector layers to perform spatial analysis (such as buffer analysis and overlay analysis).

[0038] The method of the present invention includes sending remote sensing images to a trained ground fissure dynamic monitoring model, which outputs the crack path, specifically including outputting a vector feature layer containing the detected crack location, each feature corresponding to a path identified as a crack, and displaying it through a GIS platform.

[0039] Example 5 Reference Figure 1 ,The construction and training process of the ground fissure dynamic ,monitoring model includes the following steps: S1: Build a dynamic monitoring model for ground fissures and perform S2 by calculating the accuracy of the detection results of manually labeled DOM image samples; S2: Improve the accuracy of thin line object recognition in the model structure and execute S3; S3: Optimize the extraction efficiency of interference features and execute S4; S4: Obtain the geomorphological parameters of the area to be monitored, introduce high-speed interpretation image samples, and execute S5; S5: Obtain the trained ground fissure dynamic monitoring model.

[0040] Using existing model identification, refer to Figure 2 ,The recognition accuracy of the ground fissure detection model does not meet the technical requirements. ,There is a lack of sufficient ground fissure samples to train the model. ,The detection result is very low in accuracy. ,Ground fissures are marked on some normal land. Figure 3 . Although the number of detection results in images from different years is similar, the overall recognition quality is still not ideal. The reasons for the low model accuracy may include: first, the limited number of training samples and uneven categories lead to limited model learning ability; second, the current model structure is insufficient in the recognition of thin linear targets, and it is difficult to effectively distinguish ground fissures from other linear interference information (such as roads, pipelines, ditches, sand ridges, etc.). In order to improve the subsequent detection results, it is planned to start from the following aspects: expand the high-quality annotated data set, optimize the model structure to enhance the perception of complex linear features, introduce attention mechanisms or multi-scale feature fusion methods, and consider integrating DSM data to further improve the model robustness and practical application value.

[0041] Example 6 Building on the existing ground fissure detection model, new interpreted image samples were added for areas surrounding highways, and 600 high-quality training samples were introduced. Transfer learning optimization was performed in conjunction with the existing model, significantly improving the model's recognition capabilities. The updated model detected 4,296 ground fissures in the 2023 DOM imagery of the Dahaize area, with lengths ranging from 8.76 meters to 282.33 meters; and 4,611 cracks in the 2024 imagery, with lengths ranging from 9.39 meters to 302.3 meters. Through manual supervision and evaluation of 500 detection result points, the model's accuracy increased to 68.6%, a significant improvement compared to the pre-optimization period, validating the effectiveness of the transfer learning strategy.

[0042] It should be emphasized that the newly added interpreted image samples in the areas around highways are based on the current situation. Domestic highways have covered most areas of the country, and the vehicle's driving recorder can capture part of the terrain on both sides of the highway. The construction of highways has a large impact on the mountains, soil, vegetation, etc. on both sides of the highway. If ground cracks appear on the highway, it will greatly affect the subsequent highway maintenance, cause a lot of property losses, and have a great safety hazard. Therefore, it is particularly important to monitor ground cracks in the landforms around the highway. Therefore, highway scenes can be constructed to further improve the richness of model classification, which is also conducive to the accurate indexing of ground cracks in this type of scene.

[0043] Among them, training data: manually annotate cracks in mining area images to generate a dataset containing coordinates and categories; Model training: training deep learning models on GPU servers and optimizing loss functions (such as cross entropy loss); Inference Application: The trained model is deployed to the cloud or edge computing device to detect cracks in real-time images. A 3D point cloud model of the mining area is reconstructed from multi-view drone images to calculate surface elevation changes and identify sudden changes in terrain caused by cracks. Point cloud filtering (such as RANSAC plane fitting) is used to separate surface points from vegetation and building points, and then height difference analysis is used to locate crack areas.

[0044] (1) Continue to expand the training sample library, especially samples under different types of road conditions and terrain conditions; (2) Introducing multi-temporal images and terrain data (such as DSM) for multi-source data fusion to improve the model's sensitivity to weak or obscured cracks; (3) Establish a dynamic monitoring system for ground fissures, implement the deployment and linkage of the model in the actual inspection system, and improve the level of intelligent management of highway diseases; Through the optimization results of the above parameters, the model accuracy is greatly improved compared with before optimization, verifying the effectiveness of the transfer learning strategy.

[0045] In S3 of this embodiment 6, the interference features include soil cracks, shadows of high and low surfaces, gullies, and artificial cracks. Optimizing the extraction efficiency of interference features includes S31: first, adding a weighted penalty for difficult samples to the loss function to force the model to prioritize learning the features of areas prone to false detection, and performing an optimization, where the difficult samples include hidden cracks and cracks in complex backgrounds; S32: constructing a robustness test set, summarizing the interference features into the test set, and completing secondary optimization through manual labeling to construct a test set containing extreme noise (such as strong shadows, low-resolution images) and rare geological conditions (such as abnormal cracks after an earthquake) to evaluate the performance of the model in "unexpected scenarios."

[0046] Add Gaussian noise, motion blur, and brightness / contrast transformations to remote sensing image data to simulate complex lighting or imaging conditions. Inject random jitter and missing values ​​into time series data (such as displacement rate) to train models to adapt to data anomalies. Denoising preprocessing: Use methods such as wavelet transform and median filtering to remove salt and pepper noise from remote sensing imagery. Use sliding average (MA) and Kalman filtering to smooth out abnormal fluctuations in time series data.

[0047] In summary, the ground fissure detection model unique to this application further optimizes remote sensing sample data and processes image details (including sample data in S1, S2, S3, and S4) among the above-mentioned multiple sample processing technologies.

[0048] Example 7 Collect fracture data from different geological conditions (such as mountainous areas, plains, and mining areas), different genesis (mining, tectonic, and karst), and different development stages (initial, expansion, and stable) to avoid overfitting the model to a single scenario.

[0049] A GIS platform for surface crack detection is a specialized system that integrates remote sensing monitoring, spatial analysis, data management, and visualization. It is primarily used for surface crack identification, location, dynamic monitoring, and risk assessment. This type of platform, with GIS technology at its core, integrates multi-source geospatial data with crack detection algorithms to enable full-process management from data collection to decision support. The following is a detailed description of the GIS platform's core functions, technical architecture, and typical applications: 1. Multi-source data integration and management.

[0050] Data types include spatial data and attribute data, which are described as follows: Spatial data: remote sensing images (satellite / UAV RGB, multispectral, SAR data), DEM (digital elevation model), geological layers (stratum distribution, fault zone data), and spatial locations of monitoring equipment (GNSS base stations and crack meter locations).

[0051] Attribute data: fracture geometry (length, width, direction), development time, genesis (mining, tectonic activity, karst), and monitoring time series data (displacement rate, deformation).

[0052] Management capabilities are described as follows: Supports the storage, query, and update of vector (point, line, and surface) and raster data, such as the spatial topological relationship management of crack vector layers; integrated databases (such as PostGIS and ArcSDE) enable efficient retrieval of massive monitoring data.

[0053] 2. Crack detection and spatial analysis.

[0054] Automatic / semi-automatic detection: Built-in machine learning models (such as YOLO and U-Net) or traditional image processing algorithms (edge ​​detection and threshold segmentation) identify cracks in remote sensing imagery and output crack vector boundaries. Combined with InSAR deformation rate data, spatial overlay analysis (such as buffer analysis and overlay analysis) locates areas of abnormal deformation and assists in identifying the cause of cracks.

[0055] Spatial analysis tools: Trend analysis: Modeling of crack propagation trajectories based on time series to predict development trends (such as Kriging interpolation and spatiotemporal autocorrelation analysis).

[0056] Risk assessment: Overlay data on population distribution, infrastructure (oil and gas pipelines, roads, buildings), and other data to generate a fracture hazard risk level map (high / medium / low risk areas).

[0057] 3D visualization: Use DEM data to construct a 3D surface model to intuitively display the relationship between the spatial distribution of cracks and terrain undulations.

[0058] 3. Dynamic monitoring and early warning.

[0059] Time series comparison: Automatically compare crack vector layers from different time periods, identify new cracks or the expansion / closure of existing cracks, and generate change detection reports.

[0060] Threshold warning: Set warning thresholds for crack width, displacement rate and other parameters (such as width exceeding 5cm, monthly displacement >10mm), trigger warning messages (SMS, email) and highlight risk areas on the map.

[0061] Real-time data access: Connect to IoT devices (such as deployed crack meters and inclinometers) to obtain micro-deformation data in real time and dynamically update GIS layers.

[0062] 4. Visualization and interactive display.

[0063] Multi-scale map service: provides basic base maps (such as Google Maps and Tiandi Map) and professional thematic maps (crack distribution thematic maps, deformation heat maps), and supports interactive operations such as zooming, querying, and annotation.

[0064] Dynamic charts: Display the changing trends of crack parameters (such as crack length growth statistics in different quarters) in the form of time series curves, spatial statistical histograms, etc.

[0065] Report generation: Automatically output PDF reports containing maps, charts, and analysis conclusions to assist decision makers in formulating prevention and control plans.

[0066] 5. The technical architecture also includes: 1) Data layer.

[0067] Data sources: Satellite remote sensing (such as Sentinel-2 and Landsat), drone aerial surveys, ground monitoring equipment, and historical geological archive data.

[0068] Data preprocessing: Remote sensing image preprocessing: radiometric calibration, geometric correction, atmospheric correction, image mosaicking / cropping.

[0069] Monitoring data cleaning: remove outliers and fill in missing values ​​(such as using Kalman filtering to repair GNSS displacement data).

[0070] 2) Algorithm layer.

[0071] Crack detection algorithm: Traditional methods: Canny edge detection, morphological operations (dilation / erosion), threshold segmentation (Otsu algorithm).

[0072] Deep learning: CNN-based semantic segmentation models (such as DeepLab and Mask R-CNN) are used for pixel-level crack extraction; object detection models (such as Faster R-CNN) are used for identifying blocky crack areas.

[0073] Spatial analysis engine: Integrates basic GIS algorithms (such as Thiessen polygons and spatial interpolation) with professional models (such as the geological hazard susceptibility index model).

[0074] 3) Platform layer (GIS basic functions).

[0075] Development Framework: Desktop: Based on ArcGIS Pro and QGIS secondary development, data editing and advanced analysis are enabled. Web: Using open source libraries such as Leaflet and OpenLayers to build a Web GIS platform, supporting browser-based access and interaction.

[0076] Key technologies: Vector Tile technology: Optimizes front-end rendering efficiency for large-scale crack vector data. Spatiotemporal databases (such as TimescaleDB): Store and manage time-series monitoring data (such as InSAR deformation rate time series).

[0077] In summary, the above description of this technology is based on technical architecture, algorithms and data integration. Through the joint application of multiple technologies, a complete technical solution is obtained.

Claims

1. A ground fissure detection method based on remote sensing image visual analysis is characterized by: The remote sensing image is sent to the trained ground fissure dynamic monitoring model, which outputs the crack path. Specifically, it outputs a vector feature layer containing the detected crack locations. Each feature corresponds to a path identified as a crack, which is displayed through the GIS platform. The construction and training process of the ground fissure dynamic monitoring model is as follows: S1: Construct a dynamic monitoring model for ground fissures and calculate the accuracy of detection results by manually marking DOM image samples; S2: Improve the accuracy of thin line object recognition in the model structure; S3: Optimize the extraction efficiency of interference features; S4: Obtain the geomorphological parameters of the area to be monitored and introduce the interpreted image samples; S5: Obtain the trained ground fissure dynamic monitoring model.

2. The ground fissure detection method based on remote sensing image visual analysis according to claim 1, characterized in that: The remote sensing image includes at least one of a raster image, a mosaic dataset, and an image service set, wherein the resolution of the remote sensing image is less than 2 cm / pixel.

3. The ground fissure detection method based on remote sensing image visual analysis according to claim 2, characterized in that: The vector element layer includes a crack centerline, a crack boundary, and crack feature points. The crack feature points also include a crack starting point and end point, a crack inflection point, a crack intersection point, and the widest point of the crack.

4. The ground fissure detection method based on remote sensing image visual analysis according to claim 1, characterized in that: In S1, the DOM image sample contains the crack cause, crack type and development characteristics. Among them, the causes of cracks include goaf subsidence, slope excavation, tectonic movement, weathering, karst collapse, engineering loading, and groundwater pumping; Crack types include tensile cracks, shear cracks, compressive cracks, and mixed cracks; Developmental characteristics attributes include development time, expansion rate, filler type, and surrounding environment.

5. The ground fissure detection method based on remote sensing image visual analysis according to claim 1, characterized in that: In S2, Gaussian noise, motion blur, and brightness / contrast transformation are added to the image data in the sample to simulate complex lighting. The accuracy of thin linear object recognition is improved by adding void convolution and spatial pyramid pooling.

6. The ground fissure detection method based on remote sensing image visual analysis as claimed in claim 1, characterized in that: In S3, interference features include soil cracks, shadows of high and low surfaces, gullies, and man-made cracks. The process of optimizing the extraction efficiency of interference features is as follows: S31: First, a weighted penalty for difficult samples is added to the loss function to force the model to prioritize learning the features of areas prone to false detection and perform an optimization. Difficult samples include hidden cracks and cracks in complex backgrounds. S32: Construct a robustness test set, aggregate the interference features into the test set, and complete secondary optimization through manual labeling.

7. The ground fissure detection method based on remote sensing image visual analysis as claimed in claim 1, characterized in that: In S4, high-speed interpretation of image samples includes the acquisition of surface DSM images based on vehicle carriers, combined with DOM images to check for crack duplication in surface data.

8. The ground fissure detection method based on remote sensing image visual analysis according to claim 1, characterized in that: The GIS platform uses Faster R-CNN to detect crack areas and uses LSTM network to predict the future expansion trend of cracks.

Images