A method for multi-class hidden danger identification of power channel based on multi-source remote sensing images

Abstract

The application provides a kind of power channel multi-class hidden danger identification method based on multi-source remote sensing image, the method comprises: obtaining the multi-source remote sensing image data and power line vector data of target area, and pre-processing;Using cross-modal feature alignment module, the data uniformity of the multi-source remote sensing image data after pre-processing is carried out;Based on the line vector data after pre-processing, construct a space weight graph, and embed it in the coupling of uniform semantic features to form a space-semantic joint feature;The space-semantic joint feature of the target area is input into the improved YOLOv11n model of multi-class hidden danger identification, and the multi-class hidden danger target detection result in the power channel area is output;Based on the detection result, the risk level is determined by constructing the comprehensive risk value, and the hidden danger distribution map and hidden danger list of the target area power channel are output.The application improves the recognition accuracy and positioning ability of multi-class hidden danger target, improves the detection efficiency and reduces the interference of irrelevant background target.

Description

Technical Field

[0001] This invention relates to the field of power corridor hazard identification technology, and in particular to a method for identifying multiple types of power corridor hazards based on multi-source remote sensing images. Background Technology

[0002] With the continuous expansion of the power grid, the safety of transmission line corridors has become a core aspect of ensuring the stable operation of the power grid. Transmission lines often traverse complex terrains such as mountains, farmland, and cities, and their surrounding environments present various potential hazards, including floating debris, corrugated steel roofing sheets, trees, and bamboo forests. Especially under severe weather conditions such as strong winds, these hazards can easily lead to major safety accidents such as line short circuits and power outages. Currently, the mainstream inspection methods relying on manual labor, vehicles, or drones all have significant limitations: manual inspections are inefficient and costly; vehicle inspections are constrained by terrain and road conditions; and drone inspections face challenges such as short flight time and limited operating range, making it difficult to meet the urgent need for large-scale, high-frequency, and continuous monitoring of wide-area transmission corridors.

[0003] Against this backdrop, satellite remote sensing technology, with its advantages of wide coverage and lack of terrain limitations, has become an important research direction for regional monitoring of power transmission channels. However, existing remote sensing monitoring technologies mostly rely on single optical remote sensing images, whose image quality is easily affected by changes in illumination, cloud cover, haze, and other meteorological conditions, making it difficult to guarantee the accuracy and stability of hazard identification. Meanwhile, multi-source remote sensing data, such as optical images, synthetic aperture radar images, and digital elevation models, have natural complementarity in terms of information, but existing technologies lack effective cross-modal feature alignment and collaborative characterization mechanisms. This fails to overcome the inconsistencies in feature distribution caused by differences in imaging mechanisms, spatial resolution, and physical meaning, resulting in unstable multi-source data fusion effects and failing to fully leverage their comprehensive advantages.

[0004] Furthermore, existing hazard identification methods have significant shortcomings in the target detection stage. On the one hand, these methods typically fail to establish an effective monitoring buffer zone based on the actual spatial location of the transmission line, instead performing full-domain detection directly across the entire remote sensing image. This not only generates a large amount of background target interference unrelated to line safety, significantly increasing the computational burden, but also severely reduces the efficiency and accuracy of hazard identification. On the other hand, when mainstream general target detection models are directly applied to this scenario, they struggle to effectively address the unique challenges of large differences in the scale of hazard targets in power transmission channels and complex, easily confused backgrounds. Moreover, their bounding box regression loss does not adequately constrain the geometry of the targets, affecting the positioning accuracy in complex remote sensing scenarios. More importantly, existing methods generally separate hazard identification from risk assessment, relying solely on simple distance thresholds for risk determination. They fail to comprehensively consider multi-dimensional information such as the category attributes, spatial relationships, and development trends of hazard targets, and cannot effectively distinguish between the different threat levels of sudden hazards such as easily floating objects and gradual hazards such as vegetation growth. Therefore, it is difficult to achieve refined risk assessments that meet the actual needs of power grid operation and maintenance. Summary of the Invention

[0005] To address the above issues, this invention constructs a cross-modal feature alignment and collaborative representation mechanism, a power transmission line spatial constraint model, and a spatial-semantic joint feature mechanism. Based on YOLOv11n, it introduces a small target detail preservation module, a channel spatial attention mechanism, and a constraint bounding box regression loss function. This effectively integrates multi-source remote sensing data, deeply embedding prior spatial information of transmission lines into the model learning process. This significantly improves the stability and robustness of hazard identification in complex scenarios, greatly reduces background interference, and improves detection efficiency, comprehensively enhancing the identification accuracy and positioning capability of multiple types of hazard targets.

[0006] According to embodiments of the present invention, a method for identifying multiple categories of potential hazards in power transmission channels based on multi-source remote sensing imagery is provided. The method includes: Step S01: Acquire multi-source remote sensing image data and power line vector data of the target area, and perform preprocessing; Step S02: Use the cross-modal feature alignment module to unify the preprocessed multi-source remote sensing image data; Step S03: Construct a spatial weight graph based on the preprocessed line vector data, and embed it into a unified semantic feature for coupling to form a spatial-semantic joint feature; Step S04: Input the spatial-semantic joint features of the target area into the improved YOLOv11n model for multi-category hazard identification, and output the multi-category hazard target detection results within the power corridor area; Step S05: Based on the detection results, construct a comprehensive risk value to determine the risk level, and output a distribution map and list of potential hazards along the power transmission corridor in the target area.

[0007] Furthermore, the multi-source remote sensing image data mentioned in step S01 includes: optical remote sensing images, synthetic aperture radar images, and digital elevation model data, and the power line vector data includes at least the centerline of the transmission line and the location of the towers.

[0008] Furthermore, the preprocessing described in step S01 specifically includes: performing radiometric calibration, atmospheric correction, orthorectification, and image cropping on the optical remote sensing images in sequence; performing radiometric calibration, noise filtering, geometric correction, and backscattering normalization on the synthetic aperture radar images in sequence; removing outliers and resampling the digital elevation model data; and performing spatial registration, resolution unification, and regional cropping on the optical remote sensing images, synthetic aperture radar images, and digital elevation model data, using the coordinate system of the power line vector data as a unified spatial reference.

[0009] Furthermore, step S02 specifically includes the following steps: extracting optical remote sensing image features respectively. Synthetic Aperture Radar Image Features and characteristics of digital elevation model data ; And establish a cross-modal mapping function: ; Weighted fusion is performed to form unified semantic features: ,in, For feature extraction functions; For different modal features; It is a modal mapping function; To unify semantic space features; For modal weighting coefficients, satisfying It learns adaptively through an attention mechanism.

[0010] Furthermore, the specific steps of step S03 are as follows: For any point in the target area Establish a passage area The area within a preset range around the power line is defined as the power channel constraint area; Constructing a spatial weight function ,in, Represents any point in the target area With transmission line vector The shortest spatial distance between them The preset power channel threshold distance; This is the distance attenuation coefficient; The spatial weighting function is rasterized into a spatial weighted map of the line corresponding to the cropped sample based on the georeferenced relationships of the GeoTIFF image. And resample it according to the size of the test sample or the scale of the intermediate feature map; Embedding the spatial weight function into the unified semantic features makes the unified semantic features With spatial weight function Couple the features to establish a spatial-semantic joint feature: ,in This is the spatial weight enhancement coefficient.

[0011] Furthermore, the improved YOLOv11n model for multi-category hazard identification is based on the YOLOv11n network and improved by introducing an LCD-Conv module into the Backbone network to preserve the edge and texture details of small-scale hazard targets during feature downsampling. An LCA module is introduced into the Backbone network to replace the original self-attention module; An LCA module is introduced into the Neck network to convert the spatial weight map into a spatial attention map and enhance the target feature response within the channel region.

[0012] Furthermore, the specific contents of the LCD-Conv module are as follows: Extracting main semantic features using ordinary convolutional branches: ,in, This represents a convolution operation with a kernel size of 3×3 and a stride of 2. This represents the feature map after conventional downsampling. Set up a local detail preservation branch to perform shallow texture extraction on the input feature map: ,in, Used to maintain pixel-level channel response Used to extract local neighborhood texture and edge features Indicates local details; Local detail features Adjust the scale to align with the main semantic features. The spatial dimensions remain consistent: ,in, Indicates a downsampling operation; The main semantic features are fused with local detail features to obtain the LCD-Conv output features: ,in, This indicates a channel splicing operation. This represents the output features after processing by the small target detail-preserving convolutional module.

[0013] Furthermore, the specific content of the LCA module is as follows: Receiver Line Spatial Weighting Map Intermediate feature maps output by the backbone module ; Weighting the closing space Adjust to intermediate feature map Based on the spatial scale, a uniformly scaled spatial weighted map of the lines is obtained: , in, Indicates scaling operations. This represents a line space weight map with the same size as the current feature map. Spatial attention maps of transmission line channels are generated using convolutional mapping and the sigmoid activation function. , in, This represents a spatial attention map of a transmission line corridor. This represents the Sigmoid activation function. Through this operation, the line space weight map is transformed into an attention response suitable for network feature fusion; Spatial attention map of transmission line channels With intermediate feature map By fusing the data, the enhanced features of the line channel are obtained: , in, This represents the feature map after enhancement by the LCA module. This represents the spatial attention enhancement coefficient.

[0014] Furthermore, in step S04, the improved YOLOv11n model for multi-category hazard identification uses the LCIoU line constraint bounding box regression loss function instead of the original bounding box regression loss function in the basic YOLOv11n model. The LCIoU line constraint bounding box regression loss function is expressed as: , in, For prediction boxes With real frame The crossover ratio between them; This is the center point offset loss; This is due to the loss of spatial consistency of the line; and These are the weighting coefficients for the corresponding loss terms; For the prediction box, For the true frame, The coordinates of the center point of the prediction box. The coordinates of the center point of the true bounding box. To predict the width and height of the bounding box, For the actual frame width and height, and These represent the values ​​of the center points of the predicted bounding box and the center points of the ground truth bounding box in the spatial weighting function, respectively. The total loss of the improved YOLOv11n model for multi-category hazard identification is: , in, This indicates the loss due to multiple categories of hidden dangers. Indicates the target confidence loss. This represents the regression loss of the line constraint bounding box. , and These represent the weighting coefficients of the corresponding loss terms.

[0015] Furthermore, the detection result described in step S05 is ,in, This indicates the number of potential hazards detected. Indicates the first The categories of potential hazards; Indicates the first The pixel position of the center point of each potential target in the current detection sample coordinate system; Indicates the first The detection confidence level of each potential hazard target; Indicates the first The bounding box of a potential hazard target. Represented as: ,in, Indicates the coordinates of the center point of the bounding box. and These represent the width and height of the bounding box, respectively, both in pixels; The formula for calculating the comprehensive risk value is as follows: , in, This represents the risk function corresponding to the target category; This represents the spatial weight of the target center point in the CGCS2000 projection coordinate system. Indicates the confidence level. Represents the target scale risk function. , To preset the scale normalization parameters, This represents the area of ​​the bounding box. , , and Let be the weight coefficient, and satisfy... .

[0016] This invention constructs a cross-modal feature alignment and collaborative representation mechanism, a power transmission line spatial constraint model, and a spatial-semantic joint feature mechanism. Based on YOLOv11n, it introduces a small target detail preservation module, a channel spatial attention mechanism, and a constraint bounding box regression loss function. This effectively integrates multi-source remote sensing data and deeply embeds the prior spatial information of transmission lines into the model learning process. This significantly improves the stability and robustness of hazard identification in complex scenarios, greatly reduces background interference, and improves detection efficiency. It comprehensively enhances the identification accuracy and positioning capability of multiple types of hazard targets.

[0017] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0018] The beneficial effects of this invention are: 1. By constructing a cross-modal feature alignment and collaborative representation mechanism, the semantic gap between optical, SAR and DEM data is effectively solved, and a unified semantic expression of multi-source remote sensing data is realized. This fully leverages the complementary advantages of different data sources and significantly improves the stability and robustness of hazard identification in complex scenarios. 2. By constructing a power channel spatial constraint model, the prior spatial information of transmission lines is introduced into the detection area screening and feature response process, enabling the model to show explicit attention to the channel area during the feature learning stage, thereby significantly improving detection efficiency and effectively reducing interference from irrelevant background targets; 3. By constructing a spatial-semantic joint feature mechanism, the spatial weights of transmission lines participate in the feature learning and fusion process. Based on YOLOv11n, a small target detail preservation module, a spatial attention mechanism for transmission line channels, and a line constraint bounding box regression loss function are introduced to achieve the synergistic optimization of semantic features and spatial prior information, thereby comprehensively improving the identification accuracy and positioning capability of multi-category hidden danger targets. 4. Based on the detection results, the comprehensive risk value of hidden dangers is calculated, realizing a unified model from target detection to risk level assessment, improving the interpretability and engineering application value of the detection results, and meeting the needs of refined management of power channels. Attached Figure Description

[0019] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1 A flowchart of a method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to an embodiment of the present invention is shown; Figure 2 A schematic diagram of a cross-modal feature alignment structure according to an embodiment of the present invention is shown; Figure 3 A schematic diagram of the construction of a transmission line buffer zone according to an embodiment of the present invention is shown; Figure 4 The diagram shows the structure of an improved YOLOv11n model for multi-category hazard identification according to an embodiment of the present invention. Figure 5 A structural diagram of an LCD-Conv module according to an embodiment of the present invention is shown; Figure 6 A distribution map of potential hazards in power transmission channels in a target area, according to an embodiment of the present invention, is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] According to an embodiment of the present invention, a method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery is proposed. By constructing a cross-modal feature alignment and collaborative representation mechanism, a power transmission channel spatial constraint model, and a spatial-semantic joint feature mechanism, and by introducing a small target detail preservation module, a channel spatial attention mechanism, and a constraint bounding box regression loss function based on YOLOv11n, multi-source remote sensing data is effectively integrated. The prior spatial information of transmission lines is deeply embedded into the model learning process, which significantly improves the stability and robustness of hidden danger identification in complex scenarios, greatly reduces background interference, improves detection efficiency, and comprehensively enhances the identification accuracy and localization capability of multiple types of hidden danger targets.

[0022] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0023] Figure 1 This is a schematic flowchart of a method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery, according to an embodiment of the present invention. The method includes: Step S01: Acquire multi-source remote sensing image data and power line vector data of the target area, and perform preprocessing; Step S02: Use the cross-modal feature alignment module to unify the preprocessed multi-source remote sensing image data; Step S03: Construct a spatial weight graph based on the preprocessed line vector data, and embed it into a unified semantic feature for coupling to form a spatial-semantic joint feature; Step S04: Input the spatial-semantic joint features of the target area into the improved YOLOv11n model for multi-category hazard identification, and output the multi-category hazard target detection results within the power corridor area; Step S05: Based on the detection results, construct a comprehensive risk value to determine the risk level, and output a distribution map and list of potential hazards along the power transmission corridor in the target area.

[0024] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0025] To provide a clearer explanation of the above-mentioned method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing images, a specific embodiment will be used for illustration below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.

[0026] The following example will further illustrate the method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery.

[0027] This embodiment uses a power transmission line corridor as the target area, which includes complex terrain features such as mountains, hills, farmland, roads, residential buildings, water bodies, bare land, and vegetated areas. The potential hazards to be identified along the power corridor include easily floating objects, corrugated steel roofing sheets, construction areas, trees and bamboo groves, water bodies, buildings and structures, and debris piled up under the power line. This scenario covers common, multi-background, multi-scale, and multi-category potential hazards encountered during power corridor inspections.

[0028] Step S01: Acquire multi-source remote sensing image data and power line vector data of the target area, and perform preprocessing.

[0029] Multi-source remote sensing imagery and power line vector data of the target area were acquired. The multi-source remote sensing imagery included optical remote sensing imagery, synthetic aperture radar (SAR) imagery, and digital elevation model (DEM) data. Optical remote sensing imagery reflects the spectral, color, and texture information of the target area; SAR imagery reflects the backscattering characteristics, structural contours, and all-weather imaging information of the target area; DEM data reflects the elevation, slope, aspect, and topographic relief characteristics of the target area; and power line vector data characterizes spatial information such as the centerline of the transmission lines, tower locations, line alignment, and voltage levels. The combined introduction of multi-modal data provides complementary foundational data support for identifying potential hazards in power transmission corridors.

[0030] Optical remote sensing images, SAR images, and DEM data are saved in GeoTIFF format containing georeferenced information and unified to the CGCS2000 coordinate reference system; power line vector data includes at least the centerline of the transmission line and tower locations, which are used for subsequent spatial registration, channel constraint modeling, and coordinate recovery of detection results.

[0031] In a preferred embodiment, the optical remote sensing imagery may use data with a spatial resolution better than 2m and the cloud cover is preferably less than 10%; the SAR imagery may use single-polarized or dual-polarized data with the same spatial coverage as the optical imagery; the DEM data may use data with a spatial resolution of not less than 12.5m; and the power line vector data may include at least the line centerline and tower locations.

[0032] The optical remote sensing images were sequentially subjected to radiometric calibration, atmospheric correction, orthorectification, and image cropping to reduce the impact of differences in illumination, atmosphere, and imaging angle on target identification. The SAR images were sequentially subjected to radiometric calibration, noise filtering, geometric correction, and backscattering normalization to reduce the interference of speckle noise on the feature extraction of potential targets. The DEM data were subjected to outlier removal and resampling to obtain topographic features such as elevation, slope, and aspect.

[0033] Using the coordinate system of the power line vector data as a unified spatial reference, spatial registration, resolution unification, and regional cropping are performed on optical remote sensing images, SAR images, and DEM data to ensure that different modal data maintain consistent position and scale under the same coordinate system.

[0034] Let the original image be After preprocessing, a unified image is obtained: , in, This represents the preprocessing mapping function, which includes the radiation correction function, atmospheric correction function, and geometric correction function.

[0035] After preprocessing, the GeoTIFF raster data under a unified spatial reference is cropped into fixed-size (1024×1024 pixels or other sizes suitable for target detection network input) detection samples according to the transmission line corridor range. Simultaneously, the cropping window position, pixel resolution, affine transformation parameters, and coordinate reference system information of each cropped sample relative to the original GeoTIFF image are recorded, forming a one-to-one spatial index relationship between the detection samples and the original georeferenced image. Known potential hazards are labeled and converted into YOLO format labels. The label content includes the target category number, target center point coordinates, bounding box width, and bounding box height, and is divided into training and validation sets.

[0036] This embodiment adopts a unified target category system, including color steel, trees, water, floating debris, construction areas, and buildings. For ease of visualization, the detection results are labeled with English abbreviations, but their meanings correspond one-to-one with the above standard category system.

[0037] Step S02: Use the cross-modal feature alignment module to unify the preprocessed multi-source remote sensing image data.

[0038] The preprocessed optical remote sensing images, SAR images, and DEM data are respectively denoted as... , , Because optical remote sensing images, SAR images, and DEM data differ in imaging mechanisms, grayscale ranges, spatial resolutions, and physical meanings, directly inputting them into target detection models can easily lead to inconsistent feature distributions across different modalities. Therefore, we construct a cross-modal feature alignment module to provide a unified representation of data from different modalities, such as... Figure 2 As shown.

[0039] Extracting features from optical remote sensing images Synthetic Aperture Radar Image Features and characteristics of digital elevation model data .

[0040] And establish a cross-modal mapping function to achieve feature alignment: .

[0041] Then, weighted fusion is performed to map the features of different modalities to a unified semantic feature space, forming unified semantic features: This reduces the feature distribution shift between modalities and improves the collaborative representation capability of multi-source remote sensing information. For feature extraction functions; For different modal features; It is a modal mapping function; To unify semantic space features; For modal weighting coefficients, satisfying Weights can be adaptively learned through an attention mechanism.

[0042] Preferably, the weighting coefficients are adaptively learned through an attention mechanism, enabling the model to automatically adjust the contributions of different modalities based on the land cover type, imaging quality, and hazard characteristics of the target area. For example, in areas with significant cloud and fog obscuring the surface, the model can increase the weight of SAR image features; in mountainous areas with significant topographic relief, the model can increase the weight of DEM topographic features; and in plains areas with significant differences in texture and spectral density, the model can increase the weight of optical image features. Through the aforementioned cross-modal feature alignment and collaborative representation, the semantic gap between multi-source remote sensing data can be reduced, and the stable identification capability of hazard targets can be improved.

[0043] Step S03: Construct a spatial weighted graph based on the preprocessed line vector data, and embed it into a unified semantic feature for coupling to form a spatial-semantic joint feature.

[0044] In existing technologies, target recognition and detection are typically performed directly on the entire remote sensing image, which easily introduces a large number of targets unrelated to power transmission lines, affecting recognition efficiency and accuracy. This invention uses power line vector data to spatially model the area surrounding the power transmission lines, such as... Figure 3 As shown, for any point in the target area Establish a passage area The area within a preset range around the power line is defined as the power channel constraint region, and a spatial weighting function is constructed. This allows the hazard identification process to have a clear spatial focus, among which, Represents any point in the target area With transmission line vector The shortest spatial distance between them The preset power channel threshold distance; This refers to the distance attenuation factor. The power channel threshold distance is set according to the line voltage level and management requirements. For example, it can be set to 30m or 50m according to the actual scenario, or dynamically set according to the voltage level. In this embodiment... Set to 50m.

[0045] The spatial weighting function is calculated based on transmission line vector data in a unified projected coordinate system, and then rasterized into a spatial weighted map of the transmission lines corresponding to the cropped samples according to the georeferenced relationships of GeoTIFF imagery. Before inputting the target detection model, the line spatial weight map is resampled according to the detection sample size or intermediate feature map scale so that it can correspond to the image sample or intermediate feature map in spatial location.

[0046] By embedding the spatial weight function into the unified semantic feature, the unified semantic feature... With spatial weight function Couple the features to establish a spatial-semantic joint feature: ,in This is the spatial weight enhancement coefficient. (And...) and Compared to the simple weighting method, this weighting method can enhance features in areas close to the line channel while preserving the original semantic features, avoiding the complete suppression of features in areas far from the line, which would lead to loss of detection information. In this embodiment... Setting it to 0.6 means that when a pixel is close to a power transmission line, When the pixel is larger, the semantic features of the corresponding region are enhanced; when the pixel is far away from the power transmission line... The size is smaller, and semantic features are still preserved, but the enhancement is reduced.

[0047] The proposed spatial-semantic joint feature does not directly eliminate areas outside the power transmission line corridor, nor does it apply a fixed threshold filter to all pixels. Instead, it enhances the regions near the transmission line corridor with soft weights while preserving the original remote sensing semantic features. This approach balances the identification of key potential hazards near the line with the retention of targets that may have associated risks but are far from the line. Thus, the hazard identification model considers not only the image appearance, spectral, structural, and topographical attributes of the target, but also its spatial position relative to the transmission line, achieving collaborative modeling of semantic and spatial information.

[0048] Step S04: Input the spatial-semantic joint features of the target area into the improved YOLOv11n model for multi-category hazard identification, and output the multi-category hazard target detection results within the power channel area.

[0049] Using the YOLOv11n network as the base model, an improved YOLOv11n model for multi-category hazard identification is constructed: The Conv modules in layers 1, 3, 4, and 5 of the Backbone network are replaced with LCD-Conv modules, and the C2PSA module in layer 10 of the Backbone network is replaced with LCA modules. In the Neck network, LCA modules are added after the C3k2 modules in layers 16, 19, and 22, as follows: Figure 4 As shown.

[0050] The original YOLOv11n model has advantages such as fast detection speed, small number of parameters, and convenient deployment. However, it is mainly designed for general image target detection tasks, and has the following shortcomings when directly used for remote sensing hazard identification of power transmission line channels: 1. The scale of potential hazards along power transmission line corridors varies greatly. Targets such as easily floating objects, local corrugated steel roof tiles, and piles of materials under the line are small in area in remote sensing images. The original network downsampling process can easily cause the loss of details of small targets. 2. The background of the power transmission line is complex, and features such as farmland, roads, buildings, water bodies, and vegetation can easily be confused with potential hazards. 3. The original YOLOv11n model lacks prior spatial information about transmission lines and cannot distinguish whether the target is located in an area that may affect the safety of the line; 4. The original bounding box regression loss does not adequately constrain the geometric orientation and shape of the target, affecting the positioning accuracy of potential targets in complex remote sensing scenarios.

[0051] To address the issue of detail loss in downsampling of small-scale hazardous targets in remote sensing images, this invention introduces the LCD-Conv small-target detail-preserving convolution module into the shallow feature extraction stage of the YOLOv11n backbone network. The LCD-Conv module stands for Line-Channel Detail Convolution. This module adds a local detail-preserving branch to the existing convolutional downsampling structure to enhance the edge and texture information of small-scale hazardous targets such as easily floating objects, local corrugated steel sheets, piles of materials under lines, and small-scale construction areas. The specific processing steps are as follows: Let the input feature map be: ,in, and This represents the height and width of the feature map. Indicates the number of channels; like Figure 5 As shown, the LCD-Conv module first extracts the main semantic features through a regular convolutional branch: ,in, This represents a convolution operation with a kernel size of 3×3 and a stride of 2. This represents the feature map after conventional downsampling. To avoid excessive attenuation of small target details during downsampling, the LCD-Conv module sets a local detail preservation branch to perform shallow texture extraction on the input feature map: ,in, Used to maintain pixel-level channel response Used to extract local neighborhood texture and edge features Indicates local details; Local detail features Adjust the scale to align with the main semantic features. The spatial dimensions remain consistent: ,in, Indicates a downsampling operation; The main semantic features are fused with local detail features to obtain the LCD-Conv output features: ,in, This indicates a channel splicing operation. This represents the output features after processing by the small target detail-preserving convolutional module. By introducing the LCD-Conv small target detail-preserving convolutional module, the model can retain the edge, texture, and local structural information of small-scale potential hazards while performing feature downsampling. This reduces the loss of small target features caused by traditional downsampling processes and improves the detection capability of potential hazards such as easily floating objects, corrugated steel sheets, underground piles, and small construction areas.

[0052] To differentiate the model from general object detection models, this invention introduces an LCA module in the feature fusion stage of YOLOv11n. LCA stands for Line Corridor Attention, which is spatial attention for transmission line corridors. This module is designed specifically for the scenario of identifying potential hazards along transmission lines. Its purpose is not simply to multiply the spatial weight map with the feature map, but rather to generate a spatial attention response using the spatial weight map of the transmission line. This allows the model to automatically enhance the features of potential hazards within the transmission line corridor during the feature fusion stage and reduce interference from similar features located far from the line. The specific processing steps are as follows.

[0053] Spatial weighting diagram of receiving transmission lines Intermediate feature map of YOLOv11n ; Spatial weight map Adjust to intermediate feature map Based on the spatial scale, a uniformly scaled line weight map is obtained: , in, Indicates scaling operations. This represents a line space weight map with the same size as the current feature map. Spatial attention maps of transmission line channels are generated using convolutional mapping and the sigmoid activation function. , in, This represents a spatial attention map of a transmission line corridor. This represents the Sigmoid activation function. Through this operation, the spatial weight map is transformed into an attention response suitable for network feature fusion; Spatial attention map With intermediate feature map By fusing the data, the enhanced features of the line channel are obtained: , in, This represents the feature map after enhancement by the LCA module. This represents the spatial attention enhancement coefficient.

[0054] Compared to directly using a simple spatial weighting method In comparison, the LCA module of this invention first... By performing scale adaptation and convolutional mapping, the spatial prior of transmission lines can participate in network feature fusion in the form of attention maps, rather than being directly applied to the feature maps as fixed weights. This avoids the complete suppression of potentially relevant but distant hazards due to overly strong spatial priors, while simultaneously enhancing the target response within the transmission line corridor.

[0055] By introducing the LCA (Limited Area Communication) transmission line corridor spatial attention enhancement module, the model can focus on targets within the transmission line corridor during detection, reducing the response to similar features such as farmland, roads, buildings, water bodies, and vegetation that are far from the transmission line. This improvement is the key difference between this invention and general image detection models. General YOLOv11n mainly relies on target appearance features for detection, while the improved model of this invention utilizes both target appearance features and the spatial relationship of the transmission line, making the detection results more consistent with the requirements of power corridor hazard identification tasks.

[0056] To address the challenges of irregular shapes, significant scale variations, high accuracy requirements for bounding box positioning of remotely sensed potential hazards, and the close spatial relationship between target locations and transmission lines, this invention designs the LCIoU (Line-Constrained IoU) bounding box regression loss function to replace the original bounding box regression loss function in the basic YOLOv11n model. LCIoU stands for Line-Constrained IoU, i.e., the line-constrained intersection-exchange ratio loss. This loss function, based on the intersection-exchange ratio loss, introduces predicted box center point offset constraints and line spatial consistency constraints, ensuring that the predicted box not only has a high degree of overlap with the ground truth box, but also that its center position maintains consistency with the positional relationship of the ground truth target in the spatial weights of the transmission line.

[0057] Let the prediction box be: The true bounding box is: ,in, Indicates the coordinates of the center point of the prediction box. This represents the coordinates of the center point of the true bounding box. This indicates the width and height of the prediction box. This represents the actual width and height of the bounding box.

[0058] The LCIoU loss function is expressed as: ,in, This represents the intersection-over-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. This represents the center point offset loss. This indicates the loss of spatial consistency of the line. and These are the weighting coefficients for the corresponding loss terms.

[0059] The center point offset loss is expressed as: ,in, To prevent extremely small constants with a denominator of 0, this term is used to constrain the center point of the predicted bounding box to be closer to the center point of the actual bounding box, thereby improving the accuracy of the bounding box positioning of potential hazards.

[0060] The spatial consistency loss of the line is expressed as: ,in, Indicates the center point of the prediction box. Indicates the center point of the true bounding box. and These represent the values ​​of the predicted bounding box center point and the ground truth bounding box center point in the spatial weight function, respectively.

[0061] When the predicted bounding box overlaps with the actual bounding box to some extent, but the center point of the predicted bounding box deviates from the relevant area of ​​the transmission line corridor, This will increase, thereby guiding the model to learn prediction box positions that better match the spatial distribution of transmission line corridors during training.

[0062] Combining classification loss, target confidence loss, and LCIoU loss, the total loss function of the model is expressed as: , in, This indicates the loss due to multiple categories of hidden dangers. Indicates the target confidence loss. This represents the regression loss of the line constraint bounding box. , and These represent the weighting coefficients of the corresponding loss terms.

[0063] By adopting the LCIoU line-constrained bounding box regression loss function, the model can not only improve the geometric matching degree between the predicted box and the real box, but also enhance the consistency between the predicted box and the spatial positional relationship of the transmission line, thereby improving the positioning accuracy and detection reliability of hidden danger targets in complex remote sensing scenarios.

[0064] In this embodiment, the model training uses the training set and validation set generated in step S01, and the training input is the spatial-semantic joint features. The labels are YOLO format hazard target annotations. The training epochs are set to 200, the batch size to 16, and the initial learning rate to 0.01, adjusted using cosine annealing. The input size is 1024×1024 pixels. During training, enhancement methods such as random flipping, brightness perturbation, scale scaling, and slight rotation are used, but these do not change the target category or bounding box semantics.

[0065] After training is complete, the spatial-semantic joint features of the target region are... Input the improved YOLOv11n model and adjust its number of channels to the number required by the YOLOv11n backbone network through feature-adaptive convolution. Output the detection results of multiple categories of hidden danger targets in the power corridor area: ,in, This indicates the number of potential hazards detected. : No. The categories of potential hazards include floating objects, corrugated steel sheets, construction work areas, trees and bamboo forests, water bodies, buildings and structures, and objects piled up below ground, among others. : No. The pixel position of the center point of each potential target in the current detection sample coordinate system; : No. The detection confidence level of each potential hazard target; : No. The bounding box of a potential hazard target. It can be represented as: ,in, Indicates the coordinates of the center point of the bounding box. and These represent the width and height of the bounding box, respectively, both in pixels.

[0066] Step S05: Based on the detection results, construct a comprehensive risk value to determine the risk level, and output a distribution map and list of potential hazards along the power transmission corridor in the target area.

[0067] Based on the target detection results, the spatial index relationship is read, and the first... The pixel position of each potential hazard target in the coordinate system of the detection sample. The original GeoTIFF image pixel coordinate system is restored, and then converted to projected coordinates based on the GeoTIFF affine transformation parameters and CGCS2000 coordinate reference information. The target point is then calculated in the unified projected coordinate system. With transmission line vector Shortest spatial distance between .

[0068] Detection results of the improved YOLOv11n model based on multi-class hazard identification: Taking into account the target category, target spatial distance, detection confidence level, and target bounding box, a comprehensive risk value is constructed: , in, : Represents the risk function corresponding to the target category; : Represents the spatial weight of the target center point in the CGCS2000 projection coordinate system. , , and Let be the weight coefficient, and satisfy... ; Confidence level; The target-scale risk function, calculated based on the area of ​​the target bounding box, reflects the impact of the target size on the safety of transmission lines. Meanwhile, , , and All need to be normalized to the [0,1] interval.

[0069] In this embodiment, , , and The values ​​are set to 0.35, 0.30, 0.20, and 0.15 respectively. The category risk function is pre-normalized based on operation and maintenance experience. For example, the category risk is higher for easily floating objects and construction work areas, followed by the category risk for offline piles, trees, and corrugated steel sheets, while the risk contribution of water bodies and buildings is determined according to their specific location and scale.

[0070] Target scale risk function The area of ​​the bounding box is used to calculate the impact of the target size on the safety of the transmission line. For example, large areas of corrugated steel sheets, large construction areas, and large areas of trees and bamboo forests pose a higher risk than scattered small targets.

[0071] Let the first The bounding box of each potential hazard target is: Then its bounding box area is: ,in, and They represent the first The width and height of the boundary frame of each potential hazard target.

[0072] The target scale risk function can be expressed as: ,in, The preset scale normalization parameter, in pixels squared, is used to characterize the reference area scale of potential hazards. It is determined based on the statistical results of the bounding box areas of real potential hazards in the training set; that is, the bounding box areas of all real potential hazards in the training set are statistically analyzed, and the median of the bounding box areas is taken as the reference area. The larger the target area, The larger the value, the greater the potential impact area of ​​the target; the smaller the target area, the smaller the potential impact area. The smaller the size, the smaller the potential impact area of ​​the target. For example, large areas of corrugated steel sheets, large construction areas, and large areas of trees and bamboo forests pose a higher risk than scattered small targets.

[0073] Finally, based on the risk value The size determines the risk level, such as Defined as extremely high risk, Defined as high risk, Defined as medium risk, Defined as low risk. Specifically, the risk level threshold can be adjusted based on historical inspection sample statistics, expert experience, and power system operation and maintenance standards.

[0074] The final output includes the category, bounding box, spatial distance from the transmission line, risk value, risk level, and location of the corresponding hazard in the image for each detected hidden danger. This risk inference model maps the target identification results to risk quantification assessment, thereby elevating hazard analysis from a single detection result to a multi-factor comprehensive decision-making result. Through this risk inference model, a unified quantitative evaluation can be performed on hidden danger targets of different categories and locations, and further classified into extremely high risk, high risk, medium risk, and low risk levels.

[0075] The system ultimately outputs a map showing the distribution of potential hazards along the power transmission corridor in the target area and a list of hazards. The hazard list includes at least the hazard target number, hazard category, pixel coordinates within the detected sample, shortest distance to the centerline of the transmission line, detection confidence level, bounding box, risk value, and risk level. In practical engineering applications, the CGCS2000 projected coordinates can be further converted to latitude and longitude coordinates to facilitate on-site verification and handling by maintenance personnel.

[0076] Distribution map of potential hazards in the target area of ​​the power corridor detected by the system, as shown below. Figure 6 As shown in Table 1, the list of potential hazards detected is as follows: Table 1

[0077] The floating object, number 64, is located 6.8 meters from the power transmission line, with a detection confidence level of 0.90. Given its high risk of sudden occurrence, its proximity to the transmission line corridor, and the fact that its bounding box area meets the preset risk level, the risk inference model calculates a comprehensive risk value of 0.92, classifying it as an extremely high-risk hazard based on the preset threshold. Maintenance personnel can prioritize on-site verification and handling based on the hazard distribution map and hazard list.

[0078] This embodiment unifies multi-source remote sensing data and power line vector data under the same spatial reference, establishing a spatial mapping closed loop from geographic reference imagery, model detection samples, and detection box pixel coordinates to CGCS2000 projection coordinates. It integrates spectral texture, scattering structure, terrain undulation, and prior spatial information of the power line through cross-modal feature alignment and spatial-semantic joint features. LCD-Conv, LCA, and LCIoU are used to enhance the spatial consistency of small-scale hazards, line corridor-related targets, and predicted boxes. A risk inference model converts the detection results into interpretable hazard levels. Therefore, this embodiment can achieve integrated processing of power corridor hazard identification, location, distance calculation, and risk classification, providing an implementable technical solution for refined monitoring and operation and maintenance decisions regarding power transmission line corridor hazards.

[0079] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0080] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery, characterized in that, The method includes: Step S01: Acquire multi-source remote sensing image data and power line vector data of the target area, and perform preprocessing; Step S02: Use the cross-modal feature alignment module to unify the preprocessed multi-source remote sensing image data; Step S03: Construct a spatial weight graph based on the preprocessed line vector data, and embed it into a unified semantic feature for coupling to form a spatial-semantic joint feature; Step S04: Input the spatial-semantic joint features of the target area into the improved YOLOv11n model for multi-category hazard identification, and output the multi-category hazard target detection results within the power corridor area; Step S05: Based on the detection results, construct a comprehensive risk value to determine the risk level, and output a distribution map and list of potential hazards along the power transmission corridor in the target area.

2. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 1, characterized in that, The multi-source remote sensing image data mentioned in step S01 includes: optical remote sensing images, synthetic aperture radar images, and digital elevation model data. The power line vector data includes at least the centerline of the transmission line and the location of the towers.

3. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 2, characterized in that, The preprocessing described in step S01 specifically includes: performing radiometric calibration, atmospheric correction, orthorectification, and image cropping on the optical remote sensing images in sequence; performing radiometric calibration, noise filtering, geometric correction, and backscattering normalization on the synthetic aperture radar images in sequence; removing outliers and resampling the digital elevation model data; and performing spatial registration, resolution unification, and regional cropping on the optical remote sensing images, synthetic aperture radar images, and digital elevation model data, using the coordinate system of the power line vector data as a unified spatial reference.

4. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 1, characterized in that, Step S02 specifically includes the following steps: extracting optical remote sensing image features respectively. Synthetic Aperture Radar Image Features and characteristics of digital elevation model data ; And establish a cross-modal mapping function: ; Weighted fusion is performed to form unified semantic features: ,in, For feature extraction functions; For different modal features; It is a modal mapping function; To unify semantic space features; For modal weighting coefficients, satisfying It learns adaptively through an attention mechanism.

5. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 1, characterized in that, The specific steps of step S03 are as follows: For any point in the target area Establish a passage area The area within a preset range around the power line is defined as the power channel constraint area; Constructing a spatial weight function ,in, Represents any point in the target area With transmission line vector The shortest spatial distance between them The preset power channel threshold distance; This is the distance attenuation coefficient; The spatial weighting function is rasterized into a spatial weighted map of the line corresponding to the cropped sample based on the georeferenced relationships of the GeoTIFF image. And resample it according to the size of the test sample or the scale of the intermediate feature map; Embedding the spatial weight function into the unified semantic features makes the unified semantic features With spatial weight function Couple the features to establish a spatial-semantic joint feature: ,in This is the spatial weight enhancement coefficient.

6. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 1, characterized in that, The improved YOLOv11n model for multi-category hazard identification is based on the YOLOv11n network and improved by introducing an LCD-Conv module into the Backbone network to preserve the edge and texture details of small-scale hazard targets during feature downsampling. An LCA module is introduced into the Backbone network to replace the original self-attention module; An LCA module is introduced into the Neck network to convert the spatial weight map into a spatial attention map and enhance the target feature response within the channel region.

7. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 6, characterized in that, The specific contents of the LCD-Conv module are as follows: Extracting main semantic features using ordinary convolutional branches: ,in, This represents a convolution operation with a kernel size of 3×3 and a stride of 2. This represents the feature map after conventional downsampling. Set up a local detail preservation branch to perform shallow texture extraction on the input feature map: ,in, Used to maintain pixel-level channel response Used to extract local neighborhood texture and edge features Indicates local details; Local detail features Adjust the scale to align with the main semantic features. The spatial dimensions remain consistent: ,in, Indicates a downsampling operation; The main semantic features are fused with local detail features to obtain the LCD-Conv output features: ,in, This indicates a channel splicing operation. This represents the output features after processing by the small target detail-preserving convolutional module.

8. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 6, characterized in that, The specific content of the LCA module is as follows: Receiver Line Spatial Weighting Map Intermediate feature maps output by the backbone module ; Weighting the closing space Adjust to intermediate feature map Based on the spatial scale, a uniformly scaled spatial weighted map of the lines is obtained: ， in, Indicates scaling operations. This represents a line space weight map with the same size as the current feature map. Spatial attention maps of transmission line channels are generated using convolutional mapping and the sigmoid activation function. ， in, This represents a spatial attention map of a transmission line corridor. This represents the Sigmoid activation function. Through this operation, the line space weight map is transformed into an attention response suitable for network feature fusion; Spatial attention map of transmission line channels With intermediate feature map By fusing the data, the enhanced features of the line channel are obtained: ， in, This represents the feature map after enhancement by the LCA module. This represents the spatial attention enhancement coefficient.

9. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 1, characterized in that, The improved YOLOv11n model for multi-category hazard identification described in step S04 uses the LCIoU line constraint bounding box regression loss function to replace the original bounding box regression loss function in the basic YOLOv11n model. The LCIoU line constraint bounding box regression loss function is expressed as: ， in, For prediction boxes With real frame The crossover ratio between them; This is the center point offset loss; This is due to the loss of spatial consistency of the line; and These are the weighting coefficients for the corresponding loss terms; For the prediction box, For the true frame, The coordinates of the center point of the prediction box. The coordinates of the center point of the true bounding box. To predict the width and height of the bounding box, For the actual frame width and height, and These represent the values ​​of the center points of the predicted bounding box and the center points of the ground truth bounding box in the spatial weighting function, respectively. The total loss of the improved YOLOv11n model for multi-category hazard identification is: , in, This indicates the loss due to multiple categories of hidden dangers. Indicates the target confidence loss. This represents the regression loss of the line constraint bounding box. , and These represent the weighting coefficients of the corresponding loss terms.

10. The method for identifying multiple types of hidden dangers in power transmission channels based on multi-source remote sensing imagery according to claim 1, characterized in that, The detection result mentioned in step S05 is ,in, This indicates the number of potential hazards detected. Indicates the first The categories of potential hazards; Indicates the first The pixel position of the center point of each potential target in the current detection sample coordinate system; Indicates the first The detection confidence level of each potential hazard target; Indicates the first The bounding box of a potential hazard target. Represented as: ,in, Indicates the coordinates of the center point of the bounding box. and These represent the width and height of the bounding box, respectively, both in pixels; The formula for calculating the comprehensive risk value is as follows: ， in, This represents the risk function corresponding to the target category; This represents the spatial weight of the target center point in the CGCS2000 projection coordinate system. Indicates the confidence level. Represents the target scale risk function. , To preset the scale normalization parameters, This represents the area of ​​the bounding box. , , and Let be the weight coefficient, and satisfy... .

Images