Optimized U-Net model-based geological disaster monitoring and identification method and system

Abstract

The invention discloses a geological disaster detection and identification method and system based on an optimized U-Net model, and belongs to the technical field of unmanned aerial vehicle remote sensing monitoring, and the method comprises the steps: obtaining and processing multi-source remote sensing data, the multi-source remote sensing data comprising an earth surface optical image, earth surface three-dimensional point cloud data, and earth surface deformation data; and based on the processed multi-source remote sensing data and the optimized U-Net semantic segmentation model, carrying out intelligent identification analysis on the geological hazard hidden danger, and generating a geological hazard hidden danger identification report. Multi-dimensional perception of a geological environment is realized by acquiring multi-source remote sensing data, and on the basis, an optimized U-Net deep learning model is fused, and an attention mechanism and a multi-scale feature fusion and residual connection structure are introduced, so that the extraction and identification capability of key geological disaster features is remarkably enhanced, and the accuracy and the accuracy of the key geological disaster feature extraction are improved. Therefore, high-precision identification, dynamic monitoring and intelligent early warning of geological disaster hidden dangers are realized, and the operation and maintenance safety and disaster prevention and control level of the power transmission line in a complex geological environment are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) remote sensing monitoring technology, and in particular to a method and system for geological disaster detection and identification based on an optimized U-Net model. Background Technology

[0002] Transmission lines are a crucial component of the power system, and their operational safety and stability directly impact the reliability of the power grid. However, in complex terrain areas such as mountains, hills, and valleys, transmission lines are highly susceptible to geological disasters such as landslides, mudslides, collapses, and ground subsidence, which can lead to serious consequences such as line damage, tower collapses, and line breaks, affecting power supply and socio-economic development. Therefore, efficient and accurate monitoring and early warning of geological disasters along transmission lines are of great significance for ensuring power grid safety.

[0003] Traditional methods for monitoring geological hazards along power transmission lines mainly include manual inspection, monitoring with fixed equipment, and satellite remote sensing. Manual inspection is limited by factors such as terrain, weather, and environment, resulting in long cycles, high costs, low efficiency, and significant safety risks. While fixed monitoring equipment (such as tilt sensors and ground-penetrating radar) can provide real-time monitoring of localized areas, its coverage is limited, making it difficult to meet the needs for monitoring and early warning of high-frequency, large-scale geological hazards. Although satellite remote sensing possesses wide-area monitoring capabilities, it is affected by factors such as weather and temporal resolution, making it difficult to achieve high-frequency, high-resolution monitoring.

[0004] In recent years, the rapid development of drone technology has provided an efficient, flexible, and intelligent solution for monitoring geological hazards along power transmission lines. Drones can be equipped with various sensors to achieve functions such as high-resolution image acquisition, 3D terrain modeling, and hazard identification. However, existing drone inspection solutions often rely on a single sensor (such as a regular optical camera or infrared camera), making it difficult to comprehensively perceive changes in the geological environment of power transmission lines and their surroundings. This results in significant data limitations and insufficient information fusion. Furthermore, existing systems still rely on manual judgment for data processing and analysis, hindering the achievement of intelligent early warning systems.

[0005] In summary, existing methods for monitoring geological hazards along power transmission lines suffer from problems such as limited monitoring methods, limited data acquisition, insufficient data fusion, and low levels of intelligence, making it difficult to meet the safety monitoring needs in complex geological environments. Specific shortcomings are as follows:

[0006] (a) Monitoring methods are limited, and data acquisition is restricted:

[0007] Traditional power transmission line inspections mainly rely on manual inspections, fixed monitoring equipment, or single-sensor drone inspections. Manual inspections are limited by terrain and environmental factors, resulting in low efficiency and high risk; fixed monitoring equipment can only monitor specific areas and is difficult to achieve large-scale coverage; single sensors (such as ordinary optical cameras) are difficult to effectively identify complex geological hazards, thus limiting monitoring capabilities.

[0008] (ii) Insufficient data integration makes it difficult to accurately identify potential disaster risks:

[0009] Existing drone inspection solutions mostly rely on a single type of sensor, making it difficult to comprehensively perceive the geological environment along power transmission lines. For example, optical imaging alone is insufficient to identify potential deformation hazards beneath the surface, and thermal imaging alone cannot obtain terrain elevation information. This results in significant data limitations, hindering accurate identification and assessment of geological hazards.

[0010] (iii) The monitoring data has low real-time performance and cannot provide rapid early warning:

[0011] Traditional geological disaster monitoring typically relies on periodic manual inspections, lacking high-frequency, dynamic monitoring capabilities. This allows potential hazards to develop undetected until serious accidents occur. Meanwhile, some satellite remote sensing monitoring methods are limited by temporal resolution and weather conditions, failing to meet the real-time inspection and early warning needs of power transmission lines.

[0012] (iv) Low level of intelligence, relying on manual judgment:

[0013] Most monitoring systems still rely on manual identification and judgment, lacking efficient and intelligent identification algorithms, making it difficult to automatically detect and predict disaster risks, thus limiting the efficiency of early warning response.

[0014] Therefore, there is an urgent need for a geological disaster monitoring and identification method that combines multiple sensors such as multi-view oblique photography cameras, LiDAR, and miniSAR to achieve multi-dimensional perception of the geological environment and significantly enhance the ability to extract and identify key geological disaster features. This would enable high-precision identification, dynamic monitoring, and intelligent early warning of geological disaster hazards, effectively improving the operation and maintenance safety and disaster prevention level of transmission lines in complex geological environments. Summary of the Invention

[0015] The purpose of this invention is to overcome the shortcomings of existing technologies, such as single monitoring methods, limited data acquisition, insufficient data fusion, and low level of intelligence, and to provide a geological disaster detection and identification method and system based on an optimized U-Net model.

[0016] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0017] On the one hand, this invention provides a method for geological hazard detection and identification based on an optimized U-Net model, comprising the following steps:

[0018] S1: Acquire and process multi-source remote sensing data, including surface optical images, surface three-dimensional point cloud data, and surface deformation data;

[0019] S2: Based on the processed multi-source remote sensing data and the optimized U-Net semantic segmentation model, perform intelligent identification and analysis of geological hazard risks, and generate a geological hazard risk identification report;

[0020] The optimized U-Net semantic segmentation model introduces an attention mechanism module, a multi-scale feature fusion module, and a residual connection module on the basis of the U-Net encoder-decoder structure. The attention mechanism module is used to enhance the perception of key feature regions of disasters, the multi-scale feature fusion module extracts spatial semantic information, and the residual connection module alleviates gradient vanishing.

[0021] Preferably, the acquisition of multi-source remote sensing data in step S1 specifically includes:

[0022] The drone is equipped with a multi-view oblique photography camera to acquire optical images of the ground surface;

[0023] Acquire three-dimensional point cloud data of the earth's surface using lidar;

[0024] Data on surface deformation are obtained using synthetic aperture radar.

[0025] Preferably, the processing of multi-source remote sensing data in step S1 includes:

[0026] The surface optical imagery and surface 3D point cloud data are stitched, calibrated, and denoised, and precise alignment is achieved using control point information, ultimately generating a digital elevation model, a realistic 3D model, a digital orthophoto, and a digital surface model.

[0027] Preferably, step S2 includes the following steps:

[0028] S21: The encoder extracts disaster feature maps;

[0029] S22: The disaster feature map is processed by the attention mechanism module to obtain disaster feature maps at different scales;

[0030] S23: The disaster feature maps at different scales are fused using a multi-scale feature fusion module to obtain a fused disaster feature map;

[0031] S24: The decoder restores the spatial resolution through deconvolution and merges it with the disaster feature map extracted by the encoder to finally output a semantic segmentation map.

[0032] Preferably, the encoder performing disaster feature map extraction in step S21 includes: combining a nonlinear activation function for feature extraction expression, as shown in the following formula:

[0033] F (l) =σ(W (l) *F (l-1) +b (l) )

[0034] Wherein: F (l) The feature map extracted from layer l; W (l) With b (l) Here are the weights and biases of the convolutional kernel in layer l; * represents the convolution operation; σ is the activation function; F (l-1) The feature map extracted from layer l-1.

[0035] Preferably, the formula for the attention mechanism module described in step S22 is as follows:

[0036] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))F′=M c (F)·F

[0037] Where: F is the input feature map, F' is the weighted feature map; MLP is a multilayer perceptron; σ is the sigmoid activation function; M c (F) represents the weights of the input feature map F; AvgPool is the average pooling operation, and MaxPool is the max pooling operation.

[0038] Preferably, the formula for the multi-scale feature fusion module in step S23 is as follows:

[0039]

[0040] Wherein: F S Let S be the feature map at the s-th scale; Upsample represents the upsampling operation; S is the total number of scales; F fusion This is the fused feature map.

[0041] Preferably, the decoder described in step S24 restores the spatial resolution through deconvolution and then fuses it with the disaster feature map extracted by the encoder, as shown in the following formula:

[0042]

[0043] Among them, W (l) With b(l) σ represents the weights and biases of the convolutional kernel in layer l; * represents the convolution operation; σ represents the activation function; Upsample represents the upsampling operation; This is the feature map of the l-th layer of the decoder after concatenation and convolution. This is the feature map obtained by upsampling the features of the (l+1)th layer of the decoder. The feature map output by the encoder's symmetric layer is introduced in the decoder stage through skip connections for concatenation and fusion with the upsampled features.

[0044] Final input semantic segmentation graph P:

[0045]

[0046] in, W represents the output feature map of the last layer of the decoder. out The weights of the output layer's convolution kernels; * represents the convolution operation; b out This is the bias term for the output layer; Softmax(.) performs a softmax operation on the convolution result, transforming the output at each pixel location into the probability of belonging to each category.

[0047] Preferably, the formula for the residual connection module is as follows:

[0048] F res =F input +H(F input )

[0049] Wherein: F input The input feature map is H(·), which is the convolutional network; F is the input feature map. res This is for outputting feature maps.

[0050] On the other hand, the present invention provides a geological hazard detection and identification system based on an optimized U-Net model, comprising:

[0051] The data acquisition and processing module is used to acquire and process multi-source remote sensing data, including surface optical images, surface three-dimensional point cloud data, and surface deformation data.

[0052] The hazard identification module is used to intelligently identify and analyze geological hazard hazards based on processed multi-source remote sensing data and an optimized U-Net semantic segmentation model, and generate a geological hazard identification report.

[0053] The optimized U-Net semantic segmentation model introduces an attention mechanism module, a multi-scale feature fusion module, and a residual connection module on the basis of the U-Net encoder-decoder structure. The attention mechanism module is used to enhance the perception of key feature regions of disasters, the multi-scale feature fusion module extracts spatial semantic information, and the residual connection module alleviates gradient vanishing.

[0054] Compared with the prior art, the beneficial effects of this application are:

[0055] (I) Multi-source data fusion to improve monitoring accuracy

[0056] Traditional geological disaster monitoring mainly relies on a single sensor (such as an optical camera or LiDAR). Under severe weather or complex terrain conditions, data acquisition is easily limited. Integrating multiple data sources, such as surface optical images, surface 3D point cloud data, and surface deformation data, can achieve multi-source data fusion, make up for the limitations of a single data source, and improve monitoring accuracy and reliability.

[0057] (II) Intelligent analysis and automatic identification improve monitoring efficiency

[0058] Traditional geological disaster monitoring relies on manual inspections or manual interpretation of remote sensing images, which is time-consuming, labor-intensive, and easily affected by subjective factors. The optimized U-Net semantic segmentation model can intelligently identify and analyze geological disaster hazards, and automatically analyze the collected multi-source data. It can accurately identify geological disaster hazards such as landslides, debris flows, and ground subsidence, greatly improving monitoring efficiency and accuracy and reducing human intervention.

[0059] (III) High-precision 3D modeling to improve geological hazard assessment capabilities

[0060] In existing technologies, topographic information along power transmission lines is usually inferred through manual surveying or low-precision imagery, resulting in inaccurate disaster risk assessment. This invention combines LiDAR point cloud data and SAR imagery to construct a high-precision three-dimensional terrain model and identify minute deformations through time-series analysis, thereby achieving more accurate geological disaster assessment and improving disaster prevention and mitigation capabilities.

[0061] (iv) Adapt to complex environments and improve monitoring applicability

[0062] Traditional monitoring methods are difficult to operate efficiently in complex environments such as high mountains, canyons, dense forests, and severe weather, and the monitoring range is limited. The unmanned aerial vehicle system of this invention has autonomous flight capability, can adapt to various complex terrains, and can still obtain effective data under conditions such as clouds, fog, night, and extreme weather, thereby improving the applicability of monitoring.

[0063] (v) Reduce costs and improve economic efficiency

[0064] Traditional geological disaster monitoring typically requires the deployment of numerous fixed monitoring stations, which is costly and complex to maintain. This invention employs a mobile monitoring mode using unmanned aerial vehicles (UAVs), which can replace some fixed monitoring stations, reduce equipment investment, lower construction and maintenance costs, and improve economic efficiency. Attached Figure Description

[0065] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings:

[0066] Figure 1 This is a flowchart of a geological hazard detection and identification method based on an optimized U-Net model, as described in Embodiment 1 of the present invention.

[0067] Figure 2 This is a LiDAR data processing flowchart of a geological hazard detection and identification method based on an optimized U-Net model, as described in Embodiment 1 of the present invention.

[0068] Figure 3 This is a flowchart illustrating the result generation process of a geological hazard detection and identification method based on an optimized U-Net model as described in Embodiment 1 of the present invention.

[0069] Figure 4 This is a model diagram of the optimized U-Net semantic segmentation model of the geological disaster detection and identification method based on the optimized U-Net model described in Embodiment 1 of the present invention;

[0070] Figure 5 This is a technical roadmap of a geological hazard detection and identification method based on an optimized U-Net model, as described in Embodiment 1 of the present invention.

[0071] Figure 6 This is a system schematic diagram of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0073] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between components or an indirect connection via other components.

[0074] Example 1:

[0075] like Figure 1 As shown, this embodiment provides a geological hazard detection and identification method based on an optimized U-Net model, including the following steps:

[0076] S1: Acquire and process multi-source remote sensing data, including surface optical imagery, 3D surface point cloud data, and surface deformation data. Acquiring multi-source remote sensing data specifically includes:

[0077] The drone is equipped with a multi-view oblique photography camera to acquire optical images of the ground surface;

[0078] Acquire three-dimensional point cloud data of the earth's surface using lidar;

[0079] Data on surface deformation are obtained using synthetic aperture radar.

[0080] By using drones equipped with multi-view oblique photography cameras, high-resolution, multi-angle optical images of the ground surface can be acquired, and three-dimensional models of the real-world surface along the power transmission line can be quickly generated. This enables early identification and dynamic monitoring of potential hazards such as landslides and collapses, providing technical support for the prevention and control of geological disasters along power transmission lines.

[0081] High-precision three-dimensional point cloud data of the earth's surface can be acquired using lidar (LiDAR) to generate digital surface models (DSM) and digital elevation models (DEM), which can be used to map terrain undulations and identify surface deformation.

[0082] Laser point cloud data is acquired using a lidar system. The data is then segmented into uniformly sized rectangular blocks according to the flight detection area. An engineering file is created, and the point cloud data is denoised and filtered according to the block number. After removing noise points such as birds, dust, and fog, the point cloud data is further adjusted using flight paths and ground control point information to eliminate potential elevation differences between different flights and flight zones. Finally, a classification method based on echo count, echo intensity, elevation value, lowest ground point, and model key points is used to classify the regional laser point cloud into ground point cloud data, obtaining accurate surface point cloud data. An editable surface model is constructed based on the classification results, and a refined digital elevation model (DEM) is created through human-computer interaction. The resulting DEM output is then used to assist the U-Net model in improving the accuracy of terrain feature recognition.

[0083] Miniature synthetic aperture radar (miniSAR) can still acquire surface deformation data under complex weather conditions (such as cloud cover) to identify minute deformations.

[0084] Based on the interferometric image sequence acquired by miniSAR, deformation analysis was performed using the Small Baseline Set Assay (SBAS) method to extract the cumulative deformation value and annual average deformation rate of the monitored area at different time periods:

[0085]

[0086] Where: Δd is the deformation amount; Δt is the time interval; by spatiotemporally superimposing the deformation information with the recognition results of the U-Net model, more accurate risk assessment and trend judgment can be achieved.

[0087] Compared to traditional SAR, miniSAR is smaller, lighter, and consumes less power, making it suitable for high-frequency monitoring of localized areas on multi-rotor or fixed-wing UAVs. Although its resolution is relatively low, it offers advantages in rapid deployment and flexibility in scenarios such as emergency response and mountain power line inspection, and is particularly suitable for dynamic deformation monitoring in areas with potential geological hazards along power transmission lines.

[0088] The UAV used in this invention is a multi-rotor or fixed-wing UAV, which has the characteristics of long endurance, high payload capacity and stable flight.

[0089] Unmanned aerial vehicle (UAV) mission planning includes:

[0090] Task area selection: The monitoring area is automatically delineated based on areas prone to geological disasters and the route of power transmission lines;

[0091] Automatic flight path planning: Optimizes flight paths to ensure monitoring accuracy;

[0092] Takeoff and trajectory adjustment: After takeoff, the UAV autonomously adjusts its flight altitude and speed according to environmental conditions.

[0093] In optical remote sensing interpretation, visual interpretation is a commonly used technique for extracting geological disaster information. Visual interpretation, also known as visual analysis or visual interpretation, is a type of remote sensing image interpretation and is the reverse process of remote sensing imaging. It refers to the process by which professionals obtain information about specific target features from remote sensing images through direct observation or with the aid of auxiliary interpretation instruments. Optical remote sensing images can be obtained using multi-view oblique photography cameras. Then, a remote sensing interpretation technique combining 3D models and 2D images is employed. Based on the principles and characteristics of geological disaster development, ground feature identification, qualitative and spatial analysis are performed on 2D and 3D software platforms such as Google Earth, ERDAS, ENVI, MAPGIS, and ArcGIS in a human-computer interactive manner to obtain information about disasters and their geological environment. In simpler terms, it involves displaying the outdoor environment on images and identifying disaster information using geological principles.

[0094] In geological disaster monitoring, InSAR (Inductively Coupled Aperture Radar Interferometry) technology, as an important remote sensing technique, can effectively acquire information on surface deformation. By capturing minute changes in the Earth's surface, InSAR technology can accurately detect surface deformation caused by disasters such as landslides and subsidence, thus providing reliable data support for quantitative analysis and monitoring of disasters. Compared with traditional optical remote sensing technologies, InSAR technology has all-weather, all-time monitoring capabilities and can capture minute surface deformations with high precision. This makes InSAR an important tool for early warning and risk assessment of geological disasters, especially in the disaster monitoring of critical infrastructure such as power transmission lines, providing a valuable data source.

[0095] Oblique photogrammetry, employing a photogrammetric method based on structure-of-motion (SfM), can acquire high-resolution 3D models and digital orthophotos (DOMs) from aerial photography. This technology significantly reduces the workload of manual on-site surveys, provides detailed information on potential geological hazards, and generates realistic 3D models of these hazards. For monitoring geological hazards along power transmission lines, oblique photogrammetry can help accurately identify hidden hazard areas, especially in mountainous and complex terrain, providing higher-precision terrain and hazard information. After the UAV oblique photogrammetry flight is completed, the orthophotos of the hazard-prone areas are quickly stitched together to generate digital surface models (DSMs), rapidly providing 3D data of the hazard area. This provides strong technical support for the monitoring and early warning of high-altitude, hidden geological hazards along power transmission lines, particularly in mountainous, complex, and inaccessible areas, enabling efficient acquisition and analysis of hazard information.

[0096] The multi-source data obtained by UAVs carrying multi-source payloads complement each other, enabling the delineation of geological hazard boundaries along power transmission lines and providing more accurate hazard risk assessments, thus providing strong support for the early identification, monitoring, and response to geological hazards.

[0097] Processing multi-source remote sensing data includes:

[0098] The surface optical imagery and 3D point cloud data are stitched, calibrated, and denoised. Precise alignment is achieved using control point information, ultimately generating a digital elevation model, a realistic 3D model, a digital orthophoto, and a digital surface model, such as... Figure 2 As shown, Figure 2 This is a flowchart of the LiDAR data processing workflow, ultimately generating high-precision DEM, DOM, DSM, and other output data products, such as... Figure 3 As shown.

[0099] LiDAR technology utilizes lidar modules to precisely "penetrate" vegetation and acquire three-dimensional point cloud data of the earth's surface. Digital elevation models (DEMs) are generated from this point cloud data, clearly reflecting changes in topography. For power transmission lines, LiDAR technology can help monitor topographic changes along the route, especially in complex terrains such as mountains and hills, providing information on micro-topographic parameters such as slope, contour lines, aspect, and roughness. This data helps assess potential geological hazard risks, such as landslides, collapses, and subsidence, thus providing a scientific basis for disaster prevention and mitigation efforts along power transmission lines.

[0100] SBAS-InSAR (Small Baseline Subset InSAR) is a more precise monitoring method developed based on InSAR technology, specifically for complex mountainous terrain environments. By utilizing radar data with smaller baselines, SBAS-InSAR effectively improves the monitoring accuracy in mountainous areas, especially those prone to geological disasters such as landslides and collapses. SBAS-InSAR is particularly suitable for complex terrain and high-altitude regions, where conventional remote sensing techniques may be affected by factors such as weather and lighting. SBAS-InSAR, however, provides stable and reliable surface deformation monitoring data, ensuring the timeliness and accuracy of monitoring. Combining SBAS-InSAR technology with time-series deformation calculations and analysis can further reveal the dynamic characteristics of changes during disaster occurrence.

[0101] InSAR technology, carried by drones, provides a more comprehensive and detailed data acquisition method for geological disaster monitoring. SBAS-InSAR technology performs time-series deformation analysis on atlases, not only providing accurate geological disaster monitoring data but also summarizing disaster deformation characteristics, further improving monitoring accuracy and efficiency. Especially in geological disaster monitoring of critical infrastructure such as power transmission lines, it can accurately identify potential geological disaster hazards, enabling timely intervention and repair, and ensuring the safe operation of the power system.

[0102] Therefore, SBAS-InSAR technology can provide more accurate and comprehensive geological disaster monitoring data, providing strong support for risk assessment and disaster prevention of critical infrastructure such as power transmission lines.

[0103] Finally, the acquired high-precision 3D point cloud data, surface micro-deformation data, and high-resolution, multi-angle optical images, as well as the generated high-precision data elevation model, real-scene 3D model, digital orthophoto, and digital surface model, are hierarchically stored and uniformly archived into the geological disaster monitoring database.

[0104] S2: Based on the processed multi-source remote sensing data and the optimized U-Net semantic segmentation model, perform intelligent identification and analysis of geological hazard risks, and generate a geological hazard risk identification report;

[0105] The optimized U-Net semantic segmentation model, based on the U-Net encoder-decoder structure, introduces an attention mechanism module, a multi-scale feature fusion module, and a residual connection module. The attention mechanism module enhances the perception of key disaster feature regions, the multi-scale feature fusion module extracts spatial semantic information, and the residual connection module alleviates gradient vanishing. The model includes the following steps:

[0106] S21: The encoder extracts disaster feature maps;

[0107] S22: The disaster feature map is processed through the attention mechanism module to obtain disaster feature maps at different scales;

[0108] S23: A multi-scale feature fusion module is used to fuse disaster feature maps at different scales to obtain a fused disaster feature map;

[0109] S24: The decoder restores the spatial resolution through deconvolution and merges it with the disaster feature map extracted by the encoder to finally output a semantic segmentation map, namely a geological disaster risk heat map.

[0110] Specifically, in the monitoring and early warning of geological disasters along power transmission lines, traditional visual interpretation methods, relying on human experience, are inefficient and highly subjective, making them insufficient to meet the demands for high efficiency, accuracy, and real-time monitoring of large-scale, high-frequency disasters. With the development of intelligent remote sensing image recognition technology, deep learning models are gradually becoming an effective tool to replace manual interpretation.

[0111] This invention, based on multi-source remote sensing data, utilizes an optimized U-Net semantic segmentation model to achieve intelligent identification and analysis of potential geological hazards, such as... Figure 4 As shown, this optimized model, while retaining the classic U-Net encoder-decoder structure, introduces an attention module to enhance the model's ability to perceive key disaster feature regions. Furthermore, it extracts richer spatial semantic information through a multi-scale feature fusion structure, significantly improving the model's recognition accuracy in complex geological contexts. Simultaneously, residual connections are introduced to alleviate the gradient vanishing problem during deep network training, enhancing the model's stability and convergence speed.

[0112] After acquiring remote sensing data using a drone equipped with multiple payloads including optical cameras, LiDAR, and miniSAR, the optimized U-Net model is used to intelligently identify geological hazards in the area surrounding power transmission lines. This not only improves the automation and accuracy of hazard identification but also enables the processing and output of large-scale images in a very short time, providing efficient and reliable data support for hazard risk assessment and real-time early warning. A detailed flowchart is shown below. Figure 5 As shown.

[0113] In step S21, the encoder extracts the disaster feature map, which includes: using a non-linear activation function to extract and represent the features, as shown in the following formula:

[0114] F (l) =σ(W (l) *F (l-1) +b (l) )

[0115] Wherein: F (l) The feature map extracted from layer l; W (l) With b (l) Here are the weights and biases of the convolutional kernel in layer l; * represents the convolution operation; σ is the activation function; F (l-1) The feature map extracted from layer l-1.

[0116] Step S22 describes the response capability of the attention module to critical regions (such as disaster boundaries). Taking channel attention as an example, the formula for the attention module is as follows:

[0117] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0118] F′=M c (F)·F

[0119] Where: F is the input feature map, F' is the weighted feature map; MLP is a multilayer perceptron; σ is the sigmoid activation function; M c (F) represents the weights of the input feature map F; AvgPool is the average pooling operation, and MaxPool is the max pooling operation.

[0120] In step S23, multi-scale feature fusion fuses feature maps extracted at different scales to enhance the model's contextual understanding ability. The formula for the multi-scale feature fusion module is as follows:

[0121]

[0122] Wherein: F S Let S be the feature map at the s-th scale; Upsample represents the upsampling operation; S is the total number of scales; F fusion This is the fused feature map.

[0123] Residual connections are used to improve the training stability of deep networks and alleviate the vanishing gradient problem. The formula is as follows:

[0124] F res =F input +H(F input )

[0125] Wherein: F input The input feature map is H(·), which is the convolutional network; F is the input feature map. res This is for outputting feature maps.

[0126] In step S24, the decoder restores the spatial resolution through deconvolution and then fuses it with the disaster feature map extracted by the encoder via a skip-connection, as shown in the following formula:

[0127]

[0128] Among them, W (l) With b (l)σ represents the weights and biases of the convolutional kernel in layer l; * represents the convolution operation; σ represents the activation function; Upsample represents the upsampling operation; This is the feature map of the l-th layer of the decoder after concatenation and convolution. This is the feature map obtained by upsampling the features of the (l+1)th layer of the decoder. The feature map output by the encoder's symmetric layer is introduced in the decoder stage through skip connections for concatenation and fusion with the upsampled features.

[0129] Final input semantic segmentation graph P:

[0130]

[0131] in, W represents the output feature map of the last layer of the decoder, typically a high-resolution spatial feature tensor used to map the feature map from channels to the number of classes. out The weights of the output layer's convolution kernels; * represents the convolution operation; b out This is the bias term for the output layer; Softmax(.) performs a softmax operation on the convolution result, transforming the output at each pixel location into the probability of belonging to each category.

[0132] Example 2:

[0133] Based on the same inventive concept as in Embodiment 1, such as Figure 6 As shown, this embodiment provides a geological hazard detection and identification system based on an optimized U-Net model. The system applies the above method and includes:

[0134] The data acquisition and processing module is used to acquire and process multi-source remote sensing data, including surface optical images, surface three-dimensional point cloud data, and surface deformation data.

[0135] The hazard identification module is used to intelligently identify and analyze geological hazard hazards based on processed multi-source remote sensing data and an optimized U-Net semantic segmentation model, and generate a geological hazard identification report.

[0136] The optimized U-Net semantic segmentation model introduces an attention mechanism module, a multi-scale feature fusion module, and a residual connection module on the basis of the U-Net encoder-decoder structure. The attention mechanism module is used to enhance the perception of key feature regions of disasters, the multi-scale feature fusion module extracts spatial semantic information, and the residual connection module alleviates gradient vanishing.

[0137] By acquiring multi-dimensional geological information using a drone equipped with a multi-source remote sensing payload and employing advanced image segmentation and deformation analysis algorithms, the following results can be achieved:

[0138] (1) Geological Hazard Identification Report

[0139] By combining visual interpretation techniques with a semantic segmentation algorithm based on U-Net deep learning, the automatic extraction and boundary detection of potential disaster sites such as landslides, collapses, and subsidence can be achieved. The U-Net model, through its skip connection structure and multi-scale feature fusion, has strong capabilities for recognizing ground features, and is particularly suitable for pixel-level segmentation of geological disaster bodies in complex mountainous scenes.

[0140] An improved U-Net architecture is adopted to further enhance the model's accuracy in identifying small-target disaster hazards, addressing the weakness of traditional methods in identifying fine-grained targets. The output is a visualization map fused with high-resolution remote sensing imagery and disaster hazard areas, used to assist in manual verification and disaster level determination, forming a standardized geological hazard annotation atlas.

[0141] (2) Surface Deformation Monitoring and Analysis Report

[0142] SBAS-InSAR technology was used to perform time-series analysis on surface deformation, and the annual average deformation rate and cumulative deformation of the monitored area were calculated.

[0143] By combining LiDAR point cloud data, a three-dimensional analysis of the surface change area is conducted to assess the trend and risk level of geological disaster development.

[0144] (3) Three-dimensional terrain and surface model

[0145] Generate high-precision DEM (Digital Elevation Model) and DSM (Digital Surface Model) for analyzing terrain undulation and disaster evolution trends.

[0146] It outputs a high-precision 3D point cloud model, which can be used for geological disaster simulation and assessment along power transmission lines.

[0147] (4) Stability assessment of transmission lines

[0148] By integrating U-Net image recognition results, SBAS-InSAR deformation information, and 3D point cloud structure data, a comprehensive stability index system for transmission line tower foundations and pole areas is constructed.

[0149] By combining structural risk weights and geological deformation indicators, an automatic interpretation model is formed to conduct early diagnosis of tower deformation risks and foundation collapse trends, and output early warning factor maps and reinforcement suggestions.

[0150] (5) Automated disaster early warning and risk assessment

[0151] By combining historical data and monitoring results, a geological disaster risk assessment model is constructed to provide early warning support for the operation and maintenance of power transmission lines.

[0152] Based on geological hazard features extracted by U-Net and InSAR deformation data, a multi-factor risk assessment model for geological hazards is constructed. Long-term monitoring is conducted, and training samples are accumulated and labels are expanded using the long-term monitoring data. The weights of the U-Net image segmentation model are continuously optimized to improve the model's ability to identify typical and atypical geological hazards. The deformation rate threshold and risk warning level are adjusted using cumulative InSAR deformation data feedback.

[0153] Using the above-mentioned technical solutions, transmission lines often traverse complex terrain areas such as mountains, hills, and river valleys, making them susceptible to geological disasters such as landslides, collapses, and ground subsidence. Geological disasters may lead to instability in the transmission tower foundations, damage to the lines, and even large-scale power outages, affecting the safety and stability of power supply.

[0154] Based on the UAV platform, it has advantages such as high mobility, flexible deployment, and high efficiency in aerial operations. Combined with multi-source remote sensing payloads such as multi-view oblique photography, LiDAR, and miniSAR, it can quickly acquire and accurately cover geological environment information along the power transmission line, and build a three-dimensional perception system.

[0155] In the identification of geological hazard risks, an optimized U-Net deep convolutional neural network semantic segmentation model is introduced. Through end-to-end pixel-level recognition, it automatically extracts potential hazard areas such as landslides, collapses, and subsidence from remote sensing images. The U-Net model structure has been optimized, incorporating an attention mechanism and a multi-scale feature fusion module, significantly improving the accuracy and generalization ability of geological hazard identification in complex terrain. Compared to traditional visual interpretation methods, the algorithm not only improves interpretation efficiency but also reduces human subjective error.

[0156] In terms of surface deformation monitoring, the system integrates SBAS-InSAR (Small Baseline Set Differential Interferometric Synthetic Aperture Radar) technology. By constructing a deformation time series, it extracts minute (millimeter-level) vertical or horizontal deformations of the surface. Combining radar image registration, phase unwrapping, and least squares estimation calculations, it ultimately outputs the annual average deformation rate (V0). avg ) and cumulative deformation (D cum This enables quantitative analysis of the evolution trend of geological disasters.

[0157] Further integrating multi-source remote sensing intelligent fusion algorithms, a real three-dimensional surface model is reconstructed based on LiDAR point clouds. By calculating derived factors such as slope, elevation difference, and roughness, the results of U-Net and SBAS-InSAR are assisted, realizing closed-loop identification of disaster bodies from two-dimensional images to three-dimensional structures, thus overcoming the limitations of a single data source.

[0158] By integrating key technologies such as rapid deployment of UAVs, multi-source remote sensing acquisition, U-Net intelligent identification, and SBAS-InSAR deformation inversion, this system can achieve rapid detection, accurate identification, and dynamic early warning of geological disaster hazards in large-scale, complex terrain areas, greatly improving the perception capability of the power transmission line operating environment and the level of disaster prevention and control.

[0159] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0160] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0161] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for geological hazard detection and identification based on an optimized U-Net model, characterized in that, Includes the following steps: S1: Acquire and process multi-source remote sensing data, including surface optical images, surface three-dimensional point cloud data, and surface deformation data; S2: Based on the processed multi-source remote sensing data and the optimized U-Net semantic segmentation model, perform intelligent identification and analysis of geological hazard risks, and generate a geological hazard risk identification report; The optimized U-Net semantic segmentation model introduces an attention mechanism module, a multi-scale feature fusion module, and a residual connection module on the basis of the U-Net encoder-decoder structure. The attention mechanism module is used to enhance the perception of key feature regions of disasters, the multi-scale feature fusion module extracts spatial semantic information, and the residual connection module alleviates gradient vanishing.

2. The geological hazard detection and identification method based on the optimized U-Net model according to claim 1, wherein step S1 of acquiring multi-source remote sensing data specifically includes: The drone is equipped with a multi-view oblique photography camera to acquire optical images of the ground surface; Acquire three-dimensional point cloud data of the earth's surface using lidar; Data on surface deformation are obtained using synthetic aperture radar.

3. The geological hazard detection and identification method based on the optimized U-Net model according to claim 1, characterized in that, Step S1 involves processing the multi-source remote sensing data, including: The surface optical imagery and surface 3D point cloud data are stitched, calibrated, and denoised, and precise alignment is achieved using control point information, ultimately generating a digital elevation model, a realistic 3D model, a digital orthophoto, and a digital surface model.

4. The geological hazard detection and identification method based on the optimized U-Net model according to claim 1, characterized in that, Step S2 includes the following steps: S21: The encoder extracts disaster feature maps; S22: The disaster feature map is processed by the attention mechanism module to obtain disaster feature maps at different scales; S23: The disaster feature maps at different scales are fused using a multi-scale feature fusion module to obtain a fused disaster feature map; S24: The decoder restores the spatial resolution through deconvolution and merges it with the disaster feature map extracted by the encoder to finally output a semantic segmentation map.

5. The geological hazard detection and identification method based on the optimized U-Net model according to claim 4, characterized in that, Step S21, which involves the encoder extracting disaster feature maps, includes: combining a nonlinear activation function to extract and represent features, as shown in the following formula: F (l) =σ(W (l) *F (l-1) +b (l) ) Wherein: F (l) The feature map extracted from layer l; W (l) With b (l) Here are the weights and biases of the convolutional kernel in layer l; * represents the convolution operation; σ is the activation function; F (l-1) The feature map extracted from layer l-1.

6. The geological hazard detection and identification method based on the optimized U-Net model according to claim 4, characterized in that, The formula for the attention mechanism module described in step S22 is as follows: M c (F)<σ(MLP(AvgPool(F))+MLP(MaxPool(F))) F′=M c (F)·F Where: F is the input feature map, F' is the weighted feature map; MLP is a multilayer perceptron; σ is the sigmoid activation function; M c (F) represents the weights of the input feature map F; AvgPool is the average pooling operation, and MaxPool is the max pooling operation.

7. The geological hazard detection and identification method based on the optimized U-Net model according to claim 4, characterized in that, The formula for the multi-scale feature fusion module described in step S23 is as follows: Wherein: F S Let S be the feature map at the s-th scale; Upsample represents the upsampling operation; S is the total number of scales; F fusion This is the fused feature map.

8. The geological hazard detection and identification method based on the optimized U-Net model according to claim 4, characterized in that, The decoder described in step S24 restores the spatial resolution through deconvolution and then fuses it with the disaster feature map extracted by the encoder, as shown in the following formula: Among them, W (l) With b (l) σ represents the weights and biases of the convolutional kernel in layer l; * represents the convolution operation; σ represents the activation function; Upsample represents the upsampling operation; This is the feature map of the l-th layer of the decoder after concatenation and convolution. This is the feature map obtained by upsampling the features of the (l+1)th layer of the decoder. The feature map output by the encoder's symmetric layer is introduced in the decoder stage through skip connections for concatenation and fusion with the upsampled features. Final input semantic segmentation graph P: in, W represents the output feature map of the last layer of the decoder. out The weights of the output layer's convolution kernels; * represents the convolution operation; b out This is the bias term for the output layer; Softmax(.) performs a softmax operation on the convolution result, transforming the output at each pixel location into the probability of belonging to each category.

9. The geological hazard detection and identification method based on the optimized U-Net model according to claim 1, characterized in that, The formula for the residual connection module is as follows: F res =F input +H(F input ) Wherein: F input The input feature map is H(·), which is the convolutional network; F is the input feature map. res This is for outputting feature maps.

10. A geological hazard detection and identification system based on an optimized U-Net model, characterized in that, include: The data acquisition and processing module is used to acquire and process multi-source remote sensing data, including surface optical images, surface three-dimensional point cloud data, and surface deformation data. The hazard identification module is used to intelligently identify and analyze geological hazard hazards based on processed multi-source remote sensing data and an optimized U-Net semantic segmentation model, and generate a geological hazard identification report. The optimized U-Net semantic segmentation model introduces an attention mechanism module, a multi-scale feature fusion module, and a residual connection module on the basis of the U-Net encoder-decoder structure. The attention mechanism module is used to enhance the perception of key feature regions of disasters, the multi-scale feature fusion module extracts spatial semantic information, and the residual connection module alleviates gradient vanishing.

Images