Landslide boundary identification method and system fusing spatial constraints of power transmission lines, and storage medium

Abstract

The application provides a landslide boundary recognition method and system fusing spatial constraints of power transmission lines and a storage medium, aiming at overcoming the technical defects of insufficient boundary extraction accuracy caused by weak spectral texture difference between landslide boundary and background in complex mountainous environment, disconnection between recognition results and engineering application caused by lack of effective use of spatial scene information of power transmission corridor, and poor boundary continuity caused by lack of effective post-processing optimization mechanism, and relates to the technical field of natural disaster prediction. The method comprises the following steps: generating a boundary credibility field reflecting the spatial attention degree of different regions in the power transmission line scene; performing probability mapping on the enhanced boundary features to generate a landslide boundary probability response map; using the boundary credibility field, combining the spatial geometric relationship between the boundary segments to determine the connection condition, completing the broken boundary according to the connection condition, and performing contour regularization on the completed boundary to output the complete landslide boundary recognition result.

Description

Technical Field

[0001] This invention relates to the field of natural disaster prediction technology, specifically a landslide boundary identification method, system, and storage medium that integrates the spatial constraints of power transmission lines. Background Technology

[0002] In recent years, deep learning technologies, represented by Convolutional Neural Networks (CNNs) and Transformers, have developed rapidly, achieving remarkable results in remote sensing image recognition and land cover classification, laying a technological foundation for intelligent identification of geological disasters. However, when conducting landslide boundary identification in the specific scenario of power transmission line inspection, existing methods still face significant technical bottlenecks, specifically in the following three aspects.

[0003] First, the complex mountainous environment features a rich variety of land cover types, with bare soil, vegetation, rock masses, and shaded areas intertwined. The spectral and textural differences between the landslide boundary and the background are weak, and the true boundary response is easily drowned out by noise, making accurate boundary extraction difficult. While existing deep learning-based semantic segmentation methods can achieve high-level land cover classification, their ability to finely characterize boundary locations is insufficient, making it difficult to accurately locate the landslide outline under low-contrast conditions.

[0004] Secondly, existing boundary detection methods mainly rely on local image features for judgment, lacking effective utilization of spatial scene information of transmission lines. The prediction results of deep learning models depend only on the feature expression of the image pixels themselves, and do not incorporate the spatial distribution patterns of key facilities such as transmission towers and transmission lines in the transmission corridor into the decision-making process. Therefore, it is difficult to accurately assess the actual impact of the extracted boundaries on the safe operation of transmission facilities, and there is a significant disconnect between the recognition results and the needs of engineering applications.

[0005] Third, landslide boundaries often exhibit local missing parts, fractures, and discontinuous outlines. Traditional methods lack effective post-processing optimization mechanisms, making it difficult to effectively optimize and correct the continuity of boundary results. Consequently, the identification results fail to meet the needs of landslide integrity assessment in practical engineering applications.

[0006] In summary, how to fully utilize the spatial prior information of power transmission facilities in power transmission corridor scenarios to guide landslide boundary identification, while solving the problems of difficulty in weak boundary extraction and poor boundary continuity, has become a key technical problem that urgently needs to be solved in the field of intelligent landslide identification in complex mountainous power transmission corridors. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a landslide boundary identification method, system and storage medium that integrates the spatial constraints of transmission lines. It aims to overcome the technical defects of the prior art, such as insufficient boundary extraction accuracy due to the slight difference between the landslide boundary and the background spectral texture in complex mountainous environments, the disconnect between the identification results and engineering applications due to the lack of effective use of the spatial scene information of the transmission corridor, and poor boundary continuity due to the lack of an effective post-processing optimization mechanism.

[0008] The technical solution adopted by this invention to solve its technical problem is: a landslide boundary identification method integrating transmission line spatial constraints, comprising the following steps: Acquire input images of the power transmission corridor area; Based on the input image, spatial location information of each transmission tower is obtained; a spatial prior model of the transmission tower is established based on the spatial location information, and a boundary confidence field reflecting the degree of spatial attention in different areas of the transmission line scene is generated. A feature extraction network is used to extract boundary feature maps from the input image; the boundary feature maps are weighted and enhanced using the boundary confidence field as spatial constraint weights to strengthen the real landslide boundary response in the transmission line area of ​​interest and suppress the background pseudo-boundary response in the non-interest area, resulting in enhanced boundary features; the enhanced boundary features are then subjected to probability mapping to generate a landslide boundary probability response map. Boundary extraction is performed on the landslide boundary probability response map to obtain an initial boundary set. Fault analysis is then performed on the initial boundary set to extract boundary segments in a fractured state. The boundary confidence field is used to determine the connection conditions based on the spatial geometric relationship between the boundary segments. The fractured boundary is then completed according to the connection conditions, and the completed boundary is contour-normalized to output a complete landslide boundary recognition result.

[0009] Furthermore, after acquiring the input image and before acquiring the spatial location information of each transmission tower, the method further includes a preprocessing step for the input image, which includes geometric correction, radiometric correction, and image enhancement.

[0010] Furthermore, a pre-trained target detection network is used to detect transmission tower targets in the input image to obtain the spatial coordinates of each transmission tower.

[0011] Furthermore, the step of establishing a spatial prior model of the transmission tower based on the spatial location information and generating a boundary confidence field reflecting the degree of spatial interest in different areas of the transmission line scenario includes: Calculate the spatial distance between each pixel in the input image and each transmission tower; The spatial distance is converted into a weighted spatial influence of each transmission tower on the pixel based on the attenuation function. The spatial influence weights of each transmission tower are accumulated and fused pixel by pixel to obtain a global transmission line spatial context feature map; The spatial context feature map is normalized to obtain normalized spatial context feature values; The normalized spatial context eigenvalues ​​are subjected to nonlinear mapping transformation and Gaussian smoothing to obtain the boundary confidence field.

[0012] Furthermore, the attenuation function is a Gaussian attenuation function, and the spatial influence weight is calculated in the following way: ; in, For spatial influence range control parameters; Indicates the first Transmission towers for pixels Spatial influence weight, Represents any pixel in the image To the The Euclidean distance between the center points of each transmission tower.

[0013] Furthermore, the nonlinear mapping transformation is as follows: ; Where γ is the credibility adjustment coefficient; This is the transformed confidence level value; These are the normalized spatial context feature values.

[0014] Furthermore, the feature extraction network adopts an encoder-decoder network architecture and fuses multi-scale features through a feature pyramid structure to extract the boundary feature map.

[0015] Furthermore, the step of using the boundary confidence field as spatial constraint weights to perform weighted enhancement on the boundary feature map includes: The boundary confidence field is extended along the channel dimension to the same dimension as the boundary feature map to obtain a spatial weight feature map. The boundary feature map is enhanced element-wise using the spatial weight feature map: ; in, This is the boundary feature map before enhancement; This is the expanded spatial weight feature map; This is an element-wise multiplication operation; Enhance adaptive coefficients for the boundary; This refers to the enhanced boundary features.

[0016] Further, the step of generating a landslide boundary probability response map by probabilistic mapping of the enhanced boundary features includes: The enhanced boundary features are compressed through a 1×1 convolution and normalized using a Sigmoid activation function to obtain the landslide boundary probability response map: ; in, This represents a 1×1 convolution operation; This represents the Sigmoid activation function; Represents pixels The probability response value belonging to the landslide boundary. This refers to the enhanced boundary features.

[0017] Furthermore, the step of using the boundary confidence field, combined with the spatial geometric relationship between the boundary segments, to determine the connection condition includes: For any two boundary endpoints and Calculate the boundary connection confidence weights using the following formula: ; in, This represents the Euclidean distance between two boundary endpoints; and The boundary confidence field is located at the endpoints. and The credibility value at that location; This is the distance attenuation parameter; The final output boundary connection confidence weights; When the boundary connection confidence weight is greater than the preset connection threshold When the condition is met, the corresponding boundary segment is determined to satisfy the connection condition.

[0018] A landslide boundary identification system integrating transmission line spatial constraints, used to execute the aforementioned landslide boundary identification method integrating transmission line spatial constraints, includes: The image acquisition module is used to acquire input images of the power transmission corridor area; The spatial location information acquisition module is used to acquire the spatial location information of each transmission tower based on the input image; The boundary credibility field construction module is used to establish a spatial prior model of the transmission tower based on the spatial location information, and generate a boundary credibility field that reflects the degree of spatial attention in different areas of the transmission line scenario. The boundary feature extraction and enhancement module is used to extract boundary feature maps from the input image using a feature extraction network, and to enhance the boundary feature maps by weighting the boundary confidence field as spatial constraint weights, so as to strengthen the real landslide boundary response in the transmission line area of ​​interest and suppress the background pseudo boundary response in the non-interest area, thereby obtaining the enhanced boundary features. The probability mapping module is used to generate a landslide boundary probability response map by probabilistically mapping the enhanced boundary features. The boundary optimization module is used to extract the boundaries from the landslide boundary probability response map to obtain an initial boundary set, perform fracture analysis on the initial boundary set to extract boundary segments in a fractured state, use the boundary confidence field combined with the spatial geometric relationship between the boundary segments to determine the connection conditions, complete the fractured boundary according to the connection conditions, and perform contour regularization on the completed boundary to output a complete landslide boundary recognition result.

[0019] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying landslide boundaries based on spatial constraints of power transmission lines.

[0020] The beneficial effects of this invention are as follows: Compared with the prior art, this invention addresses the core problems of strong background interference, insufficient weak boundary response, and discontinuous boundary fractures in landslide boundary identification in complex mountainous power transmission corridor scenarios. First, it transforms the knowledge of the power transmission corridor scenario into a computable boundary confidence field by constructing a spatial prior model of the transmission tower, overcoming the limitation of existing methods that rely solely on local image features and lack scene information utilization. Based on this, the boundary feature map is adaptively weighted and enhanced using the boundary confidence field as spatial constraint weights, which strengthens the weak boundary response in the areas of interest of the power transmission facility and effectively suppresses background pseudo-boundaries in non-interested areas. Furthermore, by fusing spatial geometric distance and scene prior confidence with topological connection weights, the quantitative determination and completion of fracture boundaries are performed, solving the boundary discontinuity problem and avoiding erroneous connections. This invention forms a complete end-to-end processing chain of "spatial prior construction—boundary response enhancement—fracture boundary connection correction," and the output results can directly serve the landslide disaster assessment and operation and maintenance decision-making of power transmission corridors, possessing significant engineering practical value. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] like Figure 1As shown, the landslide boundary identification method integrating transmission line spatial constraints of the present invention includes the following steps: S1. Acquire input images of the power transmission corridor area.

[0024] The input images include satellite remote sensing images or UAV inspection images, which contain spatial details such as surface texture, color variation, terrain undulation and vegetation cover, to support subsequent transmission tower target detection and landslide boundary feature extraction.

[0025] The transmission corridor area refers to the strip-shaped area on the ground traversed by the transmission line, encompassing the base of the transmission towers and a certain width of area on both sides of the line. This width is determined comprehensively based on the line protection range, tower spacing, image resolution, and engineering requirements to ensure coverage of all landslide areas that may threaten the line's safety. The corridor width guides the selection of the spatial range of the input image; subsequently, a spatial prior model based on Gaussian attenuation of the transmission towers will automatically generate a continuous spatial attention distribution within this range. The acquired input image should have sufficient spatial resolution to clearly identify the landslide boundaries and the detailed features of the transmission tower targets.

[0026] Preferably, after acquiring the input image, the method further includes a preprocessing step. The preprocessing is used to eliminate distortion and noise introduced during image acquisition, improve image quality, and provide a reliable data foundation for subsequent steps. Specifically, the preprocessing includes: Geometric correction is used to eliminate image geometric distortions caused by factors such as sensor pose changes and terrain undulations, so as to establish an accurate mapping relationship between each pixel of the image and its corresponding geographic coordinates. Radiometric correction is used to eliminate the effects of atmospheric scattering, absorption, and changes in lighting conditions on the radiometric values ​​of images, and to restore the true reflectance of ground features. Image enhancement is used to highlight edges, textures, and structural information in an image by means of contrast stretching, spatial domain filtering, or frequency domain enhancement, thereby improving the visual recognizability of weak boundary areas.

[0027] Preferably, the geometric correction uses a rational polynomial model (RPC) combined with a digital elevation model (DEM) for orthorectification; the radiometric correction uses an atmospheric correction model to eliminate the effects of atmospheric scattering and absorption; the image enhancement uses adaptive histogram equalization or Gamma correction to adjust image contrast, and anisotropic filtering is used to preserve edge structure information while denoising.

[0028] The preprocessed input images have a unified georeference, balanced radiometric characteristics, and clear texture edge information, which can effectively support subsequent tasks such as transmission tower target detection and landslide boundary recognition.

[0029] S2. Construction of boundary credibility field based on spatial prior of transmission tower.

[0030] This step addresses the problem that existing landslide boundary identification methods primarily rely on local texture, color, and gradient information for boundary discrimination, lacking effective utilization of spatial scene information of transmission lines and failing to reflect the varying degrees of importance of different areas to the safe operation of transmission facilities. It proposes a landslide boundary credibility field construction method based on spatial priors of transmission towers. This method fully utilizes the spatial distribution patterns of transmission towers within transmission lines, transforming knowledge of transmission operation and maintenance scenarios into computable spatial constraint information, and constructing a landslide boundary credibility field reflecting the degree of regional concern. This provides scenario prior support for subsequent boundary response enhancement and continuity optimization.

[0031] Overall, this step includes the following sub-steps: S21, obtaining the spatial location information of each transmission tower based on the input image; S22, establishing a spatial prior model of the transmission towers based on the spatial location information, and generating a boundary confidence field reflecting the degree of spatial attention in different areas of the transmission line scene. The following provides a detailed explanation of each sub-step.

[0032] S21. Based on the input image, obtain the spatial location information of each transmission tower.

[0033] In one embodiment of the present invention, a pre-trained target detection network is used to obtain the spatial coordinates of each transmission tower in the input image. The target detection network adopts a deep learning-based target detection model, including but not limited to any one of YOLO series, Faster R-CNN or SSD. Its training process is briefly described as follows: collect remote sensing images or UAV inspection images containing transmission tower targets, annotate the bounding boxes of the transmission towers in the images, and construct a training dataset; input the annotated training set samples into the target detection network for iterative training, adopt a transfer learning strategy, use the weights pre-trained on the public dataset for parameter initialization, and then fine-tune on the self-built dataset; use the CIoU loss function for bounding box regression and the binary cross-entropy loss for classification to optimize the target localization accuracy and detection accuracy; after training, select the optimal model parameters according to the average precision (AP) and recall indicators, and use this model as a pre-trained target detection network for automatic detection of transmission tower targets in the input image.

[0034] The trained object detection network is applied to the input image to identify a discrete set of transmission towers, represented as: ; in, Indicates the first The center coordinates of each transmission tower; This indicates the total number of transmission towers detected in the image.

[0035] In other embodiments, the method of obtaining the spatial coordinates of transmission towers is not limited to target detection networks. For example, the spatial location of each transmission tower in the image can be obtained by reading the GIS database of transmission lines and mapping the latitude and longitude coordinates of the transmission towers to the pixel coordinate system of the input image; alternatively, a template matching algorithm can be used to search for transmission tower targets in the input image and locate the transmission tower position based on the peak value of the matching response; manual annotation can also be used, with professionals directly annotating the center position of the transmission tower on the input image. In a preferred embodiment, GIS coordinate data can be combined with the target detection network, using the GIS coordinates as a priori position to guide the detection network to perform precise positioning in a local area, thereby improving detection efficiency and accuracy.

[0036] S22. Establish a spatial prior model of the transmission tower based on the spatial location information, and generate a boundary confidence field that reflects the degree of spatial attention in different areas of the transmission line scenario.

[0037] Specifically, step S22 establishes a priori spatial model of the transmission towers based on the spatial coordinates of each transmission tower obtained in S21, and then generates a boundary confidence field, including the following sub-steps: S221. Calculate the spatial distance between each pixel in the input image and each power transmission tower. S222. Based on the attenuation function, the spatial distance is converted into the spatial influence weight of each transmission tower on the pixel. S223. The spatial influence weights of each transmission tower are accumulated and fused pixel by pixel to obtain the global transmission line spatial context feature map. S224. Normalize the spatial context feature map to obtain normalized spatial context feature values; S225. The normalized spatial context feature values ​​are subjected to nonlinear mapping transformation and Gaussian smoothing in sequence to obtain the boundary confidence field.

[0038] The following provides a detailed explanation of each sub-step of S22 above.

[0039] S221. Calculate the spatial distance between each pixel in the input image and each power transmission tower.

[0040] As a critical infrastructure in power transmission lines, transmission towers exhibit significant spatial continuity and regional correlation in their spatial distribution. In real-world scenarios, the physical threat and degree of damage posed by landslides to power transmission lines are typically closely related to the distance of the landslide body relative to the transmission tower, the spatial geometric relationship between adjacent tower locations, and the overall route of the power grid.

[0041] To characterize the spatial correlation between landslide risk areas and transmission towers—that is, areas closer to the transmission towers require higher attention for safe operation and maintenance, and correspondingly stronger landslide boundary credibility constraints—this invention uses Euclidean distance to quantitatively describe the relative spatial geometric relationship between any pixel p=(x,y) in an image and each transmission tower. The distance formula is defined as follows: ; in, Represents any pixel in the image To the The Euclidean distance between the center points of each transmission tower.

[0042] In some embodiments of the present invention, in addition to Euclidean distance, Manhattan distance or Chebyshev distance or other measurement methods can also be used to calculate the spatial distance between a pixel and a transmission tower.

[0043] S222. Based on the attenuation function, the spatial distance is converted into the spatial influence weight of each transmission tower on the pixel.

[0044] To quantify the physical characteristic that the importance of the area surrounding the transmission tower for boundary identification decreases non-linearly with distance, a spatial influence model based on a Gaussian decay function is introduced, transforming discrete physical distances into continuous spatial attention weights: ; in, This is a spatial influence range control parameter used to adjust the radius of the risk concern zone radiating outward from the transmission tower. Its value is determined based on the attenuation characteristics of the Gaussian function. Based on the "3" property of the Gaussian function... The principle of "effective scope" Generally, one-third of the expected influence radius is taken, and then converted into pixel units according to the image spatial resolution. Indicates the first Transmission towers for pixels The spatial influence weight, its numerical range is naturally constrained within... Between these parameters, the closer the distance to the tower, the higher the weight calculated.

[0045] In some embodiments of the present invention, the attenuation function of the spatial influence model is not limited to a Gaussian function, but may also be a power-law function, an inverse distance weighted function, or other function forms that can reflect the spatial distance attenuation effect.

[0046] The aforementioned spatial influence quantification rules based on Euclidean distance metric and Gaussian decay function mapping together constitute the spatial prior model of transmission towers, which transforms the spatial relationships of transmission tower locations, line routes, and multi-tower linkages into a computable pixel-by-pixel weight distribution rule system.

[0047] S223. The spatial influence weights of each transmission tower are accumulated and fused pixel by pixel to obtain the global transmission line spatial context feature map.

[0048] Because transmission lines are distributed in a continuous corridor-like pattern, the influence areas of multiple transmission towers overlap and connect, collectively constituting the spatial scope of concern for the entire line. The established spatial prior model of transmission towers is applied to all detected towers, and the spatial influence weights of each tower are accumulated and fused pixel-by-pixel to obtain a global spatial context feature map of the transmission line. ; In the formula, The spatial attention level of the transmission line corresponding to pixel p is increased accordingly. The attention level of the corridor area with multiple towers will be increased accordingly, which can intuitively reflect the strip distribution characteristics of the transmission line.

[0049] S224. Normalize the spatial context feature map to obtain normalized spatial context feature values.

[0050] To avoid excessive weighting in overlapping regions of multiple towers leading to numerical imbalance, the spatial context feature map is further normalized, with the expression being: ; in, and These represent spatial context feature maps. The maximum and minimum values ​​in; This represents the normalized spatial context feature value, whose numerical range is constrained to the interval [0,1]. This spatial context feature map is the specific numerical output of the transmission tower spatial prior model in this step.

[0051] S225. Perform nonlinear mapping transformation and Gaussian smoothing on the normalized spatial context feature values ​​to obtain the boundary confidence field.

[0052] To enhance the distinction between high-interest and low-interest regions, a nonlinear mapping function is used to transform the normalized spatial context feature map: ; in, This is a credibility adjustment coefficient used to control the spatial distribution pattern of the credibility field; when At that time, the weight of high-attention areas was further enhanced; when At this time, the confidence field distribution is more gradual. In practical applications, Based on the prior enhancement requirements, this scenario prioritizes... This is to highlight the differences in credibility between areas adjacent to transmission towers and improve the separability between key areas and non-concern areas.

[0053] Since the confidence field essentially reflects the degree of spatial attention in the transmission line scenario, the larger its value, the closer the area is to the transmission facility and the higher its importance to the safe operation of the transmission line. Therefore, it should be given higher priority for boundary preservation and response enhancement weight in the subsequent landslide boundary identification process.

[0054] To further improve the stability and transferability of the confidence field, this invention employs a Gaussian smoothing operator to perform spatial continuity processing on the confidence field: ; in, The standard deviation is expressed as Two-dimensional Gaussian kernel; This represents the convolution operation; This represents the boundary confidence field after continuity. Based on the smoothing characteristics of Gaussian filtering, the preferred value range is 1 to 5 pixels; when the image spatial resolution is high and it is necessary to preserve the local spatial differences around the transmission tower, a smaller value should be used. Value; when the image has significant noise or requires enhanced spatial continuity, a larger value is used. value.

[0055] Through the above processing, the final boundary confidence field is obtained. (Dimension 1×H×W, where H and W are the height and width of the input image, respectively) can not only effectively characterize the spatial attention of different areas in a transmission line scene, but also has good spatial continuity and regional consistency. The larger the value, the closer the area is to the transmission facility and the higher its importance to the safe operation of the transmission line. Therefore, it should be given higher priority for boundary preservation and response enhancement weight in the subsequent landslide boundary identification process. The boundary confidence field... The numerical value of each pixel is defined as the confidence value of that point. For ease of description, the "boundary confidence field" mentioned in the following steps refers to the Gaussian smoothed field. .

[0056] In some embodiments of the present invention, the spatial prior model of the transmission tower may also adopt a geographically weighted regression model or a graph neural network, and learn the weight distribution of the influence of the transmission tower on the surrounding area from the samples in a data-driven manner. All of the above alternative modeling methods are within the scope of protection covered by the present invention.

[0057] S3. Use a feature extraction network to extract boundary feature maps from the input image; use the boundary confidence field as spatial constraint weights to perform weighted enhancement on the boundary feature maps to strengthen the real landslide boundary response in the area of ​​interest of the transmission line and suppress the background pseudo-boundary response in the non-area of ​​interest, so as to obtain enhanced boundary features; use probability mapping to generate a landslide boundary probability response map.

[0058] This step addresses the issues of weak differences between landslide boundaries and background features, insufficient weak boundary response, and boundary fracturing in complex mountainous scenarios by proposing a space-prior-guided boundary response enhancement method.

[0059] (i) Use a feature extraction network to extract boundary feature maps from the input image.

[0060] The feature extraction network can employ a CNN architecture (such as U-Net, ResNet, etc.), a Transformer architecture (such as ViT, Swin Transformer, etc.), or a hybrid CNN-Transformer architecture. In this invention, to fully extract multi-scale features of the landslide boundary and preserve spatial detail information, preferably, the feature extraction network adopts an encoder-decoder network architecture and fuses multi-scale features through a feature pyramid structure to extract the boundary feature map containing rich spatial details and semantic information. Specifically, the encoder includes multiple downsampling stages connected in sequence, each stage outputting feature maps of different spatial resolutions to represent contextual information and edge detail features at different scales in the image. Lower-level feature maps have high resolution and small receptive fields, containing rich spatial details and edge location information, which is beneficial for accurate landslide boundary localization; higher-level feature maps have low resolution and large receptive fields, containing stronger semantic contextual information, which is beneficial for determining the overall orientation and regional distribution of the landslide boundary. The decoder recovers the spatial resolution of the feature maps step by step through upsampling operations, passing high-level semantic features to lower levels. It also fuses the feature maps from each stage of the encoder with the corresponding feature maps from the decoder through skip connections, thus preserving high-precision spatial location information while restoring resolution. The feature pyramid structure further performs cross-scale fusion of feature maps output from at least two different downsampling levels of the encoder. Specifically, high-level feature maps are upsampled to match the spatial resolution of lower-level features. Figure 1 Then, the feature maps are fused by adding elements one by one or by concatenating channels to obtain a fused multi-scale feature map, which is used to extract the boundary feature map.

[0061] Let the boundary feature map extracted by the feature extraction network after cross-scale fusion through the feature pyramid structure be represented as: ; in, This indicates the number of channels in the boundary feature map; and These represent the height and width of the boundary feature map, respectively. It integrates multi-scale features from the encoder's outputs at various levels, including rich spatial details and edge gradient information in low-level feature maps, as well as semantic context and region contour information in high-level feature maps, thus providing both accurate localization and semantic discrimination criteria for landslide boundaries.

[0062] (ii) Using the boundary confidence field as the spatial constraint weight, the boundary feature map is weighted and enhanced to strengthen the real landslide boundary response in the area of ​​interest of the transmission line and suppress the background pseudo-boundary response in the non-area of ​​interest, so as to obtain the enhanced boundary features.

[0063] It should be noted that the downsampling operation in the feature extraction network will affect the boundary feature map. The spatial resolution is lower than that of the original input image, while the boundary confidence field Generated based on the original input image resolution. To ensure the correctness of subsequent element-wise weighted operations, the boundary confidence field is first upsampled to the boundary feature map using bilinear interpolation before weighted enhancement. With the same spatial resolution H×W, the scale-aligned boundary confidence field is obtained. .

[0064] Due to the scale-aligned boundary confidence field Single-channel spatial weighting graph (dimension: To achieve pixel-level matrix operations, the feature map is first expanded along the channel dimension to the same dimension as the boundary feature map by channel copying, thus obtaining the spatial weight feature map: ; Then, spatial weighted feature maps were used. Boundary features The expression for element-wise adaptive weighted enhancement is: ; in, This represents element-wise multiplication. Enhance adaptive coefficients for the boundary; This represents the enhanced boundary features. This is used to adjust the intensity of the influence of prior spatial information of the transmission tower on visual boundary features. Its value is set according to the enhancement requirements and is typically between 0.5 and 2. This embodiment preferably uses... This approach aims to balance the enhancement of weak boundaries with the suppression of spurious background responses.

[0065] As can be seen from the above formula, when the pixel is located in a high-confidence area near the transmission tower, Larger values ​​result in higher gains for the corresponding boundary features; conversely, lower gains are observed in low-confidence regions far from the transmission towers, effectively suppressing background pseudo-boundary responses. This residual-connected enhancement design (adding additional gains to the original features) ensures that the original features are preserved even when the confidence level is 0, avoiding the gradient vanishing problem caused by the direct "erasure" of features in low-confidence regions.

[0066] (iii) The enhanced boundary features are used to generate a landslide boundary probability response map through probability mapping.

[0067] Enhanced boundary features After probability mapping, a landslide boundary probability response map is generated. This probability mapping can employ conventional methods such as fully connected layer classifiers, Softmax normalization, or Conditional Random Fields (CRF). In this invention, to transform the enhanced boundary features into pixel-level probability responses with clear physical meaning, while balancing computational efficiency and output stability, preferably, the probability mapping uses a boundary prediction layer for channel compression and is normalized using the Sigmoid function. ; in, This represents a 1×1 convolution operation; This represents the Sigmoid activation function, used to normalize numerical values; Represents pixels The probability response value belonging to the landslide boundary.

[0068] Through the above process, the spatial prior of the transmission tower is adaptively enhanced to the boundary features, enabling the model to focus on the real landslide boundaries in the area surrounding the transmission line, thereby improving the boundary response strength and boundary extraction accuracy.

[0069] S4. Extract the boundaries from the landslide boundary probability response map to obtain an initial boundary set. Perform fracture analysis on the initial boundary set to extract boundary segments in a fractured state. Utilize the boundary confidence field and the spatial geometric relationship between the boundary segments to determine the connection conditions. Complete the fractured boundaries according to the connection conditions and perform contour regularization on the completed boundaries to output a complete landslide boundary recognition result.

[0070] To address the issues of boundary breaks and contour loss caused by vegetation occlusion, shadow coverage, and gradient regions, this step constructs a continuity optimization method under spatial prior constraints. First, a boundary extraction threshold is set. Boundary probability response diagram Perform binarization segmentation and extract the initial boundary set; ; in, This is used to determine the boundary threshold between the effective boundary and the background response. In practical applications, in accordance with The probability distribution is adaptively determined, and Otsu's method is preferred for automatic thresholding. The optimal value is calculated based on the principle of maximizing inter-class variance to adapt to different image scenarios.

[0071] Subsequently, a fracture analysis is performed on the initial boundary set to extract boundary segments in a fractured state. When analyzing fractured boundary regions using conventional edge tracking or morphological processing methods, false positives or false negatives are easily caused by noise interference, and it is difficult to accurately define the attribution relationships of fractured boundary segments. Therefore, this invention determines the spatial connectivity of each pixel in the initial boundary set, merging boundary pixels belonging to the same connected component into independent boundary segments, and extracting the endpoint pixels at both ends of each boundary segment as targets to be connected. Compared to pixel-by-pixel tracking or endpoint detection based on local windows, connected component analysis has advantages such as high computational efficiency, good noise resistance, and easy preservation of boundary segment integrity, providing accurate endpoint geometric positions and attribution information for subsequent calculation of boundary topology connection weights. Preferably, the fracture analysis includes performing connected component analysis on the initial boundary set to extract isolated boundary endpoints in a fractured state.

[0072] Using the boundary confidence field constructed in step S2, the connection conditions are determined in conjunction with the spatial geometric relationships between the boundary segments. For any two boundary endpoints to be evaluated... and By integrating spatial geometric span and scene priors, a boundary topology connection weight formula is constructed: ; in, This represents the Euclidean distance between two boundary endpoints; and The boundary confidence field is located at the endpoints. and The credibility value at that location; The distance attenuation parameter is determined based on the spatial attenuation characteristics of the Gaussian function, and is usually taken as... This is the effective connection radius. Beyond this distance, the weight approaches zero, and it is considered that there is no effective connection. This refers to the boundary connection confidence weights in the final output.

[0073] As can be seen from the above formula, when two boundary endpoints are close to each other and are located in high confidence regions, their connection weight is larger and they are more likely to be identified as part of the same landslide boundary; while boundary segments located in low confidence regions are difficult to be incorrectly connected, thereby reducing the risk of false boundary reconstruction.

[0074] When the boundary connection weight is greater than a preset threshold ,Right now If the condition is met, the corresponding boundary segment is determined to satisfy the connection condition, and curve interpolation is used to complete the broken boundary, generating a fused contour set. Preferably, the curve interpolation employs cubic spline interpolation or Catmull-Rom splines to ensure that the interpolated curve smoothly passes through the endpoints and maintains the geometric continuity of the boundaries. Because... Determined by both spatial distance and prior confidence, its range is naturally constrained to [0,1]. The preferred value is 0.5 to 0.8. In this embodiment, we take... This means that a connection is only established when the connection reliability reaches 70% or higher, in order to ensure the reliability of continuous optimization.

[0075] To eliminate the jagged effect caused by digital interpolation and restore the geometric smoothness of the boundaries, a boundary smoothing operator is used on the connected contour set. After regularization, the final refined identification result of the landslide boundary is obtained: ; in, Indicates the boundary smoothing operator; This represents the final landslide boundary.

[0076] Through the aforementioned boundary continuity optimization process, adaptive connection and contour reconstruction of the fracture boundary are achieved, enabling the landslide boundary to achieve better continuity and integrity while maintaining geometric accuracy. The final output not only accurately characterizes the actual contour range of the landslide but also fully integrates prior spatial information of the transmission tower, providing reliable data support for landslide disaster risk assessment and engineering operation and maintenance decisions in transmission corridors.

[0077] This invention also provides a landslide boundary identification system that integrates transmission line spatial constraints, used to execute the above-described landslide boundary identification method that integrates transmission line spatial constraints, characterized in that it includes: The image acquisition module is used to acquire input images of the power transmission corridor area; The spatial location information acquisition module is used to acquire the spatial location information of each transmission tower based on the input image; The boundary credibility field construction module is used to establish a spatial prior model of the transmission tower based on the spatial location information, and generate a boundary credibility field that reflects the degree of spatial attention in different areas of the transmission line scenario. The boundary feature extraction and enhancement module is used to extract boundary feature maps from the input image using a feature extraction network, and to enhance the boundary feature maps by weighting the boundary confidence field as spatial constraint weights, so as to strengthen the real landslide boundary response in the transmission line area of ​​interest and suppress the background pseudo boundary response in the non-interest area, thereby obtaining the enhanced boundary features. The probability mapping module is used to generate a landslide boundary probability response map by probabilistically mapping the enhanced boundary features. The boundary optimization module is used to extract the boundaries from the landslide boundary probability response map to obtain an initial boundary set, perform fracture analysis on the initial boundary set to extract boundary segments in a fractured state, use the boundary confidence field combined with the spatial geometric relationship between the boundary segments to determine the connection conditions, complete the fractured boundary according to the connection conditions, and perform contour regularization on the completed boundary to output a complete landslide boundary recognition result.

[0078] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the landslide boundary identification method for integrating transmission line spatial constraints as described above.

Claims

1. A method for landslide boundary identification that fuses spatial constraints of power transmission lines, characterized by, Includes the following steps: Acquire input images of the power transmission corridor area; Based on the input image, spatial location information of each transmission tower is obtained; a spatial prior model of the transmission tower is established based on the spatial location information, and a boundary confidence field reflecting the degree of spatial attention in different areas of the transmission line scene is generated. A feature extraction network is used to extract boundary feature maps from the input image; the boundary feature maps are then weighted and enhanced using the boundary confidence field as spatial constraint weights to strengthen the real landslide boundary response within the transmission line area of ​​interest and suppress the background pseudo-boundary response in the non-interest area, resulting in enhanced boundary features. The enhanced boundary features are then subjected to probability mapping to generate a landslide boundary probability response map; Boundary extraction is performed on the landslide boundary probability response map to obtain an initial boundary set. Fault analysis is then performed on the initial boundary set to extract boundary segments in a fractured state. The boundary confidence field is used to determine the connection conditions based on the spatial geometric relationship between the boundary segments. The fractured boundary is then completed according to the connection conditions, and the completed boundary is contour-normalized to output a complete landslide boundary recognition result.

2. The method of claim 1, wherein the method comprises: After acquiring the input image and before acquiring the spatial location information of each transmission tower, the process also includes a preprocessing step for the input image, which includes geometric correction, radiometric correction, and image enhancement.

3. The method of claim 1, wherein the method comprises: The input image is used to detect transmission tower targets using a pre-trained target detection network to obtain the spatial coordinates of each transmission tower.

4. The method of claim 1, wherein the method is characterized by: The step of establishing a spatial prior model of the transmission tower based on the spatial location information and generating a boundary confidence field reflecting the degree of spatial interest in different areas of the transmission line scenario includes: Calculate the spatial distance between each pixel in the input image and each transmission tower; The spatial distance is converted into a weighted spatial influence of each transmission tower on the pixel based on the attenuation function. The spatial influence weights of each transmission tower are accumulated and fused pixel by pixel to obtain a global transmission line spatial context feature map; The spatial context feature map is normalized to obtain normalized spatial context feature values; The normalized spatial context eigenvalues ​​are subjected to nonlinear mapping transformation and Gaussian smoothing to obtain the boundary confidence field.

5. The method of claim 4, wherein the step of identifying the landslide boundary comprises the steps of: determining a landslide boundary based on the identified landslide area and the identified landslide area's spatial constraints. The attenuation function is a Gaussian attenuation function, and the spatial influence weight is calculated in the following way: ； wherein, a spatial influence range control parameter; represents the spatial influence weight of the pixel point to the transmission tower, represents the Euclidean distance from the arbitrary pixel point to the center point of the transmission tower .​ 6. The method of identifying the landslide boundary constraining the transmission line route in space according to claim 5, wherein, The nonlinear mapping transformation is as follows: ； wherein γ is a confidence adjustment coefficient; is the transformed confidence value; is the normalized spatial context feature value.

7. The landslide boundary identification method integrating transmission line spatial constraints as described in claim 1, characterized in that, The feature extraction network adopts an encoder-decoder network architecture and fuses multi-scale features through a feature pyramid structure to extract the boundary feature map.

8. The landslide boundary identification method integrating transmission line spatial constraints as described in claim 1, characterized in that, The step of using the boundary confidence field as spatial constraint weights to perform weighted enhancement on the boundary feature map includes: The boundary confidence field is extended along the channel dimension to the same dimension as the boundary feature map to obtain a spatial weight feature map. The boundary feature map is enhanced element-wise using the spatial weight feature map: ； in, This is the boundary feature map before enhancement; This is the expanded spatial weight feature map; This is an element-wise multiplication operation; Enhance adaptive coefficients for the boundary; This refers to the enhanced boundary features.

9. The landslide boundary identification method integrating transmission line spatial constraints as described in claim 1, characterized in that, The step of generating a landslide boundary probability response map by probabilistic mapping of the enhanced boundary features includes: The enhanced boundary features are compressed through a 1×1 convolution and normalized using a Sigmoid activation function to obtain the landslide boundary probability response map: ； in, This represents a 1×1 convolution operation; This represents the Sigmoid activation function; Represents pixels The probability response value belonging to the landslide boundary. This refers to the enhanced boundary features.

10. The landslide boundary identification method integrating transmission line spatial constraints as described in claim 1, characterized in that, The step of determining the connection conditions using the boundary confidence field and the spatial geometric relationship between the boundary segments includes: For any two boundary endpoints and Calculate the boundary connection confidence weights using the following formula: ； in, This represents the Euclidean distance between two boundary endpoints; and The boundary confidence field is located at the endpoints. and The credibility value at that location; This is the distance attenuation parameter; The final output boundary connection confidence weights; When the boundary connection confidence weight is greater than the preset connection threshold When the condition is met, the corresponding boundary segment is determined to satisfy the connection condition.

11. A landslide boundary identification system integrating transmission line spatial constraints, used to execute the landslide boundary identification method integrating transmission line spatial constraints as described in any one of claims 1 to 10, characterized in that, include: The image acquisition module is used to acquire input images of the power transmission corridor area; The spatial location information acquisition module is used to acquire the spatial location information of each transmission tower based on the input image; The boundary credibility field construction module is used to establish a spatial prior model of the transmission tower based on the spatial location information, and generate a boundary credibility field that reflects the degree of spatial attention in different areas of the transmission line scenario. The boundary feature extraction and enhancement module is used to extract boundary feature maps from the input image using a feature extraction network, and to enhance the boundary feature maps by weighting the boundary confidence field as spatial constraint weights, so as to strengthen the real landslide boundary response in the transmission line area of ​​interest and suppress the background pseudo boundary response in the non-interest area, thereby obtaining the enhanced boundary features. The probability mapping module is used to generate a landslide boundary probability response map by probabilistically mapping the enhanced boundary features. The boundary optimization module is used to extract the boundaries from the landslide boundary probability response map to obtain an initial boundary set, perform fracture analysis on the initial boundary set to extract boundary segments in a fractured state, use the boundary confidence field combined with the spatial geometric relationship between the boundary segments to determine the connection conditions, complete the fractured boundary according to the connection conditions, and perform contour regularization on the completed boundary to output a complete landslide boundary recognition result.

12. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the landslide boundary identification method for fusion transmission line spatial constraints as described in any one of claims 1 to 10.

Images