Satellite remote sensing image power tower disaster damage identification method and system and related device

Abstract

A satellite remote sensing image power tower disaster damage identification method, system and related device, the method comprises collecting multi-source power tower satellite remote sensing image and constructing a data set; input the data set into the pre-constructed rotating target detection model, output the power tower spatial position and identify the orientation angle through the rotating target detection model; according to the spatial position of the power tower and the identified orientation angle, the maximum likelihood estimation model of the angle distribution of the power tower is constructed through angle specialization processing, and the maximum likelihood estimation model parameters are solved by using the expectation maximization algorithm to determine the orientation angle distribution of the power tower in the image to be measured; according to the orientation angle distribution of the power tower in the image to be measured, abnormal detection is carried out, and the objects exceeding the threshold range are screened out, that is, the power tower damaged by disaster. The present application can avoid the problem of sample scarcity, significantly reduce the data preparation threshold, realize full automation from data preprocessing to abnormal screening in the detection process, and effectively improve the disaster damage perception efficiency and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of power emergency technology, specifically relating to a method, system and related devices for identifying damage to power poles from satellite remote sensing images. Background Technology

[0002] In the field of power emergency response, ensuring the stable operation of the power system is crucial. Power poles, as key infrastructure supporting power transmission, are directly affected by the speed of power supply recovery after disasters due to their post-disaster damage. Traditional power pole damage perception relies primarily on manual inspections. In large disaster-stricken areas, manual inspections are difficult to conduct quickly and comprehensively, and are easily limited by geographical environment and weather conditions, making it impossible to grasp the damage situation of poles in a timely and accurate manner. Compared to manual inspections, drone inspections, which have been widely used in recent years, are more efficient and faster, but are still limited by factors such as weather, terrain, and equipment battery life. Compared to traditional manual and drone inspections, satellite remote sensing technology has advantages such as large-area coverage, periodic observation, and no limitations imposed by ground conditions. In recent years, with the rapid development of remote sensing technology, large amounts of high-resolution satellite remote sensing imagery have become increasingly readily available. Satellite remote sensing technology has been widely applied in many fields such as land planning, natural disaster risk assessment, and environmental monitoring, providing a new data acquisition method and application paradigm for disaster damage perception technology in power emergency response.

[0003] Currently, there are no technologies for identifying power pole damage using satellite remote sensing imagery. For example, patent application CN116863353A proposes a "method for detecting the tilt of power poles based on a rotating target detection network," which belongs to the pole tilt identification scheme based on UAV imagery. The method includes acquiring multiple UAV images of power poles and creating training, validation, and test sets; preprocessing the images in the training, validation, and test sets; constructing a rotating target detection model; establishing a loss function for a multi-branch network and introducing a Kalman filter-based regression loss term into the loss function; inputting the preprocessed training, validation, and test sets into the rotating target detection model for training; inputting the preprocessed image to be tested into the trained rotating target detection model for inference and outputting the tilt detection result; achieving dual classification of power pole tilt and pole category, as well as accurate positioning of power pole targets, solving the problem of low accuracy in existing rotating intersection-union calculations. This scheme uses UAV imagery to capture pole tilt, which is essentially a different technical field from satellite remote sensing image processing. Compared to satellite remote sensing image recognition, using UAV imagery for pole damage identification is subject to various limitations such as terrain, equipment endurance, and weather factors, resulting in low efficiency and high time and labor consumption. For example, patent application CN118212546A proposes a "Remote Sensing Target Detection Method for Power Pole Towers Based on Large Kernel Selection Feature Fusion Network," which is a pole target identification scheme based on satellite remote sensing imagery. Its output pole detection results are vertical frames without tilt angles, making damage identification impossible. This method involves selecting YOLOv5 as the base model, designing a large kernel spatial selection attention fusion module to improve the backbone network, expand the model's receptive field, and accurately locate the power pole tower position. It also designs a multi-scale feature alignment fusion structure to improve the neck network, addressing the problem of large and inconsistent scale differences between transmission and distribution towers, improving the detection accuracy of small poles such as distribution towers, and achieving multi-scale feature fusion of power poles in complex backgrounds. The training model incorporates MPDIoU to improve CIoU, and a sliding weighted loss is designed to make the model pay more attention to negative samples during training. The optimal model obtained from training is then used to detect and identify power poles, and the model's performance is evaluated. This approach effectively improves the detection accuracy of power poles in satellite remote sensing imagery and can provide important technical support for intelligent power line inspection based on satellite remote sensing. However, pole target identification using satellite remote sensing imagery is a different task from rotating target detection. The core difference is that pole target identification does not output the pole's orientation angle, therefore it can only be used for pole location and cannot be used for disaster damage identification. Summary of the Invention

[0004] The purpose of this invention is to address the limitations of existing technologies, such as limited coverage and low efficiency in disaster damage perception, caused by various conditions including terrain, weather, equipment battery life, and road accessibility, when using manual or drone-based on-site surveys. Furthermore, the extreme scarcity of satellite remote sensing samples of damaged power poles makes it difficult to implement data-driven traditional deep learning methods. This invention provides a method, system, and related devices for identifying disaster damage to power poles using satellite remote sensing images. It circumvents the problem of sample scarcity, significantly lowers the data preparation threshold, and automates the entire detection process from data preprocessing to anomaly screening, effectively improving the efficiency and accuracy of disaster damage perception.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Firstly, a method for identifying damage to power poles using satellite remote sensing imagery is provided, including:

[0007] Collect satellite remote sensing images of power poles from multiple sources and construct a dataset;

[0008] The dataset is input into a pre-built rotating target detection model, which outputs the spatial location of the power pole and identifies its orientation angle.

[0009] Based on the spatial location of the power poles and the identified orientation angles, a maximum likelihood estimation model of the power pole angle distribution is constructed through angle specialization processing. The parameters of the maximum likelihood estimation model are iteratively solved using the expectation-maximization algorithm to determine the orientation angle distribution of the power poles in the image to be measured.

[0010] Anomaly detection is performed based on the distribution of the orientation angles of power poles in the image to be tested, and objects exceeding the threshold range are selected as power poles that have suffered damage.

[0011] As a preferred approach, in the step of acquiring multi-source satellite remote sensing images of power poles and constructing a dataset, satellite remote sensing images of power corridors covering different geographical areas, terrain types, and climatic conditions are selected as the original data source. The spatial resolution of the satellite remote sensing images is not less than 1 meter, covering a variety of power pole types and including complete pole bodies and surrounding environmental features.

[0012] As a preferred embodiment, in the step of acquiring multi-source satellite remote sensing images of power poles and constructing a dataset, the acquired satellite remote sensing images are pre-processed by slicing. The slicing pre-processing uses a sliding window method to divide the original satellite remote sensing images into sub-images of fixed size, and sets a 20%-30% overlap area between adjacent slices, so that power pole targets across slices are fully presented in at least one sub-image.

[0013] As a preferred approach, in the step of acquiring multi-source satellite remote sensing images of power poles and constructing a dataset, an image annotation tool is used to manually annotate the sliced ​​sub-images. The annotation content includes the rotated bounding box of the pole target and category information. The rotated bounding box defines the main body of the pole in pixel coordinates, and the angle parameter of the bounding box is consistent with the actual orientation of the power pole. The category information is used to annotate the pole type. The annotation does not distinguish between disaster damage status. The annotation process follows the consistency principle, and the annotation error does not exceed 5 pixels and 5 degrees.

[0014] As a preferred approach, in the step of acquiring satellite remote sensing images of multi-source power poles and constructing a dataset, the labeled dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1, and the three types of datasets after division are consistent in all dimensions.

[0015] As a preferred embodiment, the rotating target detection model is built based on the YOLOv10 model architecture and includes the following components:

[0016] The backbone network is used to extract multi-level features from the bottom to the top of the input image. It balances feature extraction capability and computational cost through lightweight convolution and attention mechanism. The backbone network adopts an improved cross-stage local network structure CSP and lightweight convolution operator to improve feature expression capability while reducing parameters and computation. It embeds attention module to aggregate features of different receptive fields through multi-scale pooling and dynamically adjusts feature weights using channel attention mechanism.

[0017] The neck network addresses the adaptation problem of target detection at different scales through bidirectional feature fusion from top to bottom and bottom to top. It takes feature maps of different scales output by the backbone network as input, corresponding to the feature representations of small, medium and large targets respectively. Each scale feature map is processed by convolutional layers and normalization layers, and then the multi-scale features are fused through bidirectional feature fusion. After feature optimization, it can retain both high-level semantic information and low-level detailed features.

[0018] The head network employs a dual-label allocation and single-output strategy. The "one-to-many" head generates abundant positive samples during the training phase, enhancing convergence stability. The "one-to-one" head outputs the final bounding box during both the training and inference phases, replacing non-maximum suppression (NMS) filtering and enabling end-to-end deployment. The detection head adopts a lightweight design, using depthwise separable convolutions, which have lightweight decoupling capabilities.

[0019] As a preferred embodiment, in the step of outputting the spatial location of power poles and identifying their orientation angles through a rotating target detection model, the rotating target detection model is trained. The loss function consists of three parts: classification loss, regression loss, and angle loss, which respectively constrain the accuracy of classification, detection boxes, and rotation angles. Data augmentation is performed using any one or more combinations of random horizontal flipping, random brightness variation, random rotation, and CutMix. The optimizer is AdamW, with an initial learning rate set to 0.001, and a cosine annealing strategy is adopted. The trained rotating target detection model is used to infer the input image. The optimal weights of the rotating target detection model are selected, and the original satellite remote sensing image is segmented using the same segmentation strategy as when constructing the dataset. After hyperparameter tuning, the output result is obtained. For a single segmented satellite remote sensing image sample, each image contains multiple detection results, and each detection result is determined by the following parameters: category, confidence, lower left corner coordinates, lower right corner coordinates, upper right corner coordinates, and upper left corner coordinates. Based on the detection results, the orientation angle of the rotating detection box is calculated and used as the sample. .

[0020] As a preferred embodiment, the step of constructing a maximum likelihood estimation model of the angle distribution of power poles through angle specialization includes:

[0021] Set sample Rotate the long side of the detection frame and The angle between the positive and negative axes, with a range of values ​​of [value missing]. By numerically specializing the angle, we can avoid classifying similar orientation angles into different categories; let the sample mean be... The angle specialization expression is:

[0022]

[0023] After angle specialization processing, the sample The range of values ​​is or ;

[0024] Assuming that the orientation angles of the power poles obtained after analyzing the post-disaster data constitute a sample set. Among them, the normal samples follow a normal distribution. Abnormal samples follow a uniform distribution ;

[0025] The orientation of power poles after a disaster is defined according to the following mixed distribution model:

[0026] Normal sample: percentage , The probability density function is:

[0027]

[0028] Abnormal samples: accounting for The probability density function is:

[0029]

[0030] This yields the joint probability density function of the mixture model, where the sample... The probability density from the mixture model is:

[0031]

[0032] In the formula, model parameters ;

[0033] For all samples The likelihood function of the mixture model is:

[0034]

[0035] Taking the logarithm, the log-likelihood function is:

[0036]

[0037] Considering maximizing In this case, because there are latent variables with "unknown sample affiliation" in the mixed distribution, the log-likelihood function is non-convex with respect to the parameter. Therefore, latent variables are introduced. The expected value maximization algorithm is used to iteratively solve the problem, where, Indicates sample These are normal power poles. Indicates sample These are power poles that have been damaged in the disaster.

[0038] As a preferred embodiment, the step of using the expectation-maximization algorithm to iteratively solve the maximum likelihood estimation model parameters to determine the orientation angle distribution of power poles in the image to be measured includes:

[0039] E-step: Calculate the posterior probability of the latent variables;

[0040] definition , indicating sample The posterior probability of belonging to the normal sample is then Indicates sample The posterior probability of belonging to an outlier sample, where, Then, according to Bayes' theorem, we have:

[0041]

[0042] M-step, maximizing the expected log-likelihood;

[0043] After executing step E, based on Construct the expected value of the log-likelihood function for all data. ,in, For the first The parameters of the next iteration maximize the expected function. And obtain new parameters ;

[0044] The log-likelihood function for all data is as follows:

[0045]

[0046] by As weights, for latent variables Expected result:

[0047]

[0048] Will about Differentiation yields:

[0049]

[0050] The solution yields:

[0051]

[0052] Solve in the same way and get:

[0053]

[0054]

[0055] Repeat steps E and M until the parameter change is less than the threshold. or log-likelihood function convergence.

[0056] As a preferred embodiment, the step of anomaly detection based on the distribution of power pole orientation angles in the image to be tested includes calculating the anomaly for each sample after model convergence. Posterior probability of belonging to an outlier Set threshold ,like Then determine the sample This is an abnormal sample.

[0057] Secondly, a satellite remote sensing imagery-based system for identifying damage to power poles is provided, including:

[0058] The data acquisition and processing module is used to acquire satellite remote sensing images of power poles from multiple sources and construct datasets.

[0059] The rotating target detection module is used to input the dataset into a pre-built rotating target detection model, and output the spatial location of the power pole and identify the orientation angle through the rotating target detection model;

[0060] The module for determining the orientation angle distribution of power poles is used to construct a maximum likelihood estimation model of the orientation angle distribution of power poles based on the spatial location of the power poles and the identified orientation angles through angle specialization processing. The module then uses the expectation-maximization algorithm to iteratively solve the parameters of the maximum likelihood estimation model to determine the orientation angle distribution of power poles in the image to be measured.

[0061] The abnormal power pole screening module is used to detect anomalies based on the distribution of the orientation angles of power poles in the image to be tested, and to screen out objects that exceed the threshold range, i.e., power poles that have been damaged.

[0062] As a preferred option, the data acquisition and processing module selects satellite remote sensing images of power corridors covering different geographical areas, terrain types, and climatic conditions as the source of raw data. The spatial resolution of the satellite remote sensing images is not less than 1 meter, covering a variety of power pole types and including the complete pole body and the surrounding environmental features.

[0063] As a preferred embodiment, the data acquisition and processing module performs slice preprocessing on the acquired satellite remote sensing images. The slice preprocessing uses a sliding window method to divide the original satellite remote sensing images into sub-images of fixed size, and sets a 20%-30% overlap area between adjacent slices. Power pole targets across slices are fully presented in at least one sub-image.

[0064] As a preferred embodiment, the data acquisition and processing module uses an image annotation tool to manually annotate the sliced ​​sub-images. The annotation content includes the rotated bounding box of the tower target and category information. The rotated bounding box defines the main body of the tower in pixel coordinates, and the angle parameters of the bounding box are consistent with the actual orientation of the power tower. The category information is used to annotate the tower type. The annotation does not distinguish between disaster damage status. The annotation process follows the consistency principle, and the annotation error does not exceed 5 pixels and 5 degrees.

[0065] As a preferred embodiment, the rotating target detection module inputs the dataset into a pre-built rotating target detection model. When the rotating target detection model outputs the spatial location of the power pole and identifies its orientation angle, the rotating target detection model is built based on the YOLOv10 model architecture and includes the following components:

[0066] The backbone network is used to extract multi-level features from the bottom to the top of the input image. It balances feature extraction capability and computational cost through lightweight convolution and attention mechanism. The backbone network adopts an improved cross-stage local network structure CSP and lightweight convolution operator to improve feature expression capability while reducing parameters and computation. It embeds attention module to aggregate features of different receptive fields through multi-scale pooling and dynamically adjusts feature weights using channel attention mechanism.

[0067] The neck network addresses the adaptation problem of target detection at different scales through bidirectional feature fusion from top to bottom and bottom to top. It takes feature maps of different scales output by the backbone network as input, corresponding to the feature representations of small, medium and large targets respectively. Each scale feature map is processed by convolutional layers and normalization layers, and then the multi-scale features are fused through bidirectional feature fusion. After feature optimization, it can retain both high-level semantic information and low-level detailed features.

[0068] The head network employs a dual-label allocation and single-output strategy. The "one-to-many" head generates abundant positive samples during the training phase, enhancing convergence stability. The "one-to-one" head outputs the final bounding box during both the training and inference phases, replacing non-maximum suppression (NMS) filtering and enabling end-to-end deployment. The detection head adopts a lightweight design, using depthwise separable convolutions, which have lightweight decoupling capabilities.

[0069] As a preferred embodiment, the rotating target detection module trains a rotating target detection model. The loss function consists of three parts: classification loss, regression loss, and angle loss, which respectively constrain the accuracy of classification, detection boxes, and rotation angle. Data augmentation is performed using any one or more combinations of random horizontal flipping, random brightness variation, random rotation, and CutMix. The optimizer is AdamW, with an initial learning rate set to 0.001, and a cosine annealing strategy is adopted. The trained rotating target detection model is used to infer the input image. The rotating target detection model uses the best weights trained beforehand. The original satellite remote sensing image is segmented using the same strategy as when constructing the dataset. After hyperparameter tuning, the output result is obtained. For a single segmented satellite remote sensing image sample, each image contains multiple detection results. Each detection result is determined by the following parameters: category, confidence score, lower left corner coordinates, lower right corner coordinates, upper right corner coordinates, and upper left corner coordinates. Based on the detection results, the orientation angle of the rotating detection box is calculated and used as the sample. .

[0070] As a preferred embodiment, when the power pole orientation angle distribution determination module constructs the maximum likelihood estimation model of the power pole angle distribution through angle specialization processing, it sets the sample... Rotate the long side of the detection frame and The angle between the positive and negative axes, with a range of values ​​of [value missing]. By numerically specializing the angle, we can avoid classifying similar orientation angles into different categories; let the sample mean be... The angle specialization expression is:

[0071]

[0072] After angle specialization processing, the sample The range of values ​​is or ;

[0073] Assuming that the orientation angles of the power poles obtained after analyzing the post-disaster data constitute a sample set. Among them, the normal samples follow a normal distribution. Abnormal samples follow a uniform distribution ;

[0074] The orientation of power poles after a disaster is defined according to the following mixed distribution model:

[0075] Normal sample: percentage , The probability density function is:

[0076]

[0077] Abnormal samples: accounting for The probability density function is:

[0078]

[0079] This yields the joint probability density function of the mixture model, where the sample... The probability density from the mixture model is:

[0080]

[0081] In the formula, model parameters ;

[0082] For all samples The likelihood function of the mixture model is:

[0083]

[0084] Taking the logarithm, the log-likelihood function is:

[0085]

[0086] Considering maximizing In this case, because there are latent variables with "unknown sample affiliation" in the mixed distribution, the log-likelihood function is non-convex with respect to the parameter. Therefore, latent variables are introduced. The expected value maximization algorithm is used to iteratively solve the problem, where, Indicates sample These are normal power poles. Indicates sample These are power poles that have been damaged in the disaster.

[0087] As a preferred embodiment, when the power pole orientation angle distribution determination module uses the expectation-maximization algorithm to iteratively solve the maximum likelihood estimation model parameters to determine the power pole orientation angle distribution in the image to be measured, it performs the following steps:

[0088] E-step: Calculate the posterior probability of the latent variables;

[0089] definition , indicating sample The posterior probability of belonging to the normal sample is then Indicates sample The posterior probability of belonging to an outlier sample, where, Then, according to Bayes' theorem, we have:

[0090]

[0091] M-step, maximizing the expected log-likelihood;

[0092] After executing step E, based on Construct the expected value of the log-likelihood function for all data. ,in, For the first The parameters of the next iteration maximize the expected function. And obtain new parameters ;

[0093] The log-likelihood function for all data is as follows:

[0094]

[0095] by As weights, for latent variables Expected result:

[0096]

[0097] Will about Differentiation yields:

[0098]

[0099] The solution yields:

[0100]

[0101] Solve in the same way and get:

[0102]

[0103]

[0104] Repeat steps E and M until the parameter change is less than the threshold. or log-likelihood function convergence.

[0105] As a preferred embodiment, the abnormal power pole screening module calculates the values ​​for each sample after the model converges. Posterior probability of belonging to an outlier Set threshold ,like Then determine the sample This is an abnormal sample.

[0106] Thirdly, an electronic device is provided, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the satellite remote sensing image power pole damage identification method.

[0107] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the satellite remote sensing image power pole damage identification method.

[0108] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects:

[0109] The satellite remote sensing imagery-based power pole damage identification method proposed in this invention enables rapid damage perception across large-scale power corridors, significantly improving emergency response efficiency. Traditional manual inspections or drone surveys are limited by terrain, flight time, and weather conditions, typically covering only a few kilometers per day, which is insufficient to meet the rapid assessment needs of large-scale power poles after a disaster. This invention leverages the full-coverage characteristics of satellite remote sensing imagery, combined with automated rotating target detection and anomaly screening processes, overcoming the spatial and temporal limitations of on-site surveys. Using sub-meter resolution satellite imagery, hundreds of square kilometers of power corridors can be covered simultaneously. Through automated model processing, tower damage screening in areas affected by disasters such as earthquakes and typhoons can be completed within the golden repair period of 24-72 hours after a disaster, significantly increasing the coverage compared to traditional methods. On the other hand, the satellite remote sensing image-based power pole damage identification method proposed in this invention overcomes the technical bottleneck of extremely scarce damage samples. Existing deep learning-based damage identification methods heavily rely on large-scale labeled damage samples, but power pole damage is infrequent and sudden, resulting in an extremely scarce number of available damage samples, making model training and convergence difficult and generalization capabilities very poor. This invention proposes a "normal sample modeling + angle anomaly screening" technical solution. It utilizes easily obtainable normal pole samples to construct an angle distribution model, identifies pole orientation through target rotation detection, and then uses the maximum likelihood estimation model (MLE) and expectation-maximization algorithm (EM) to screen angle anomaly samples to determine the damage status, fundamentally eliminating the dependence on scarce damage samples. Experimental verification shows that even without damage samples for training, the method of this invention can still effectively identify typical damaged poles such as collapsed and severely tilted poles, successfully solving the model training dilemma caused by the scarcity of damage samples and significantly improving the practical application value of the technical solution.

[0110] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0111] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0112] Figure 1 Overall flowchart of the satellite remote sensing image disaster damage identification method for power poles according to an embodiment of the present invention;

[0113] Figure 2 A schematic diagram of the rotating target detection results in satellite remote sensing imagery according to an embodiment of the present invention;

[0114] Figure 3 A schematic diagram of the screening results of damaged power poles in satellite remote sensing images according to an embodiment of the present invention;

[0115] Figure 4 A schematic diagram of the structure of the satellite remote sensing image power pole damage identification system according to an embodiment of the present invention. Detailed Implementation

[0116] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0117] Satellite remote sensing image target detection refers to the localization and classification of small, densely packed targets such as poles and towers in multispectral or panchromatic images acquired by satellites using deep convolutional networks (such as Faster R-CNN, YOLO, and RetinaNet). Key technical points include: multi-scale feature fusion (FPN, BiFPN) to mitigate scale differences caused by the "nearer objects appear larger, farther objects appear smaller" principle in satellite images; high-resolution slicing strategies, dividing large images into 512×512 or 1024×1024 pixel blocks for batch inference, balancing image memory and detail; and data augmentation (rotation, random brightness, CutMix) to compensate for the scarcity of damaged samples in satellite images.

[0118] Satellite remote sensing image rotating target detection: Introducing an angle parameter θ into horizontal bounding box detection aligns the detection box with the target's main direction, significantly reducing background redundancy and improving the boundary accuracy of elongated targets (such as towers and insulator strings). Mainstream frameworks include: Anchor-based Rotated Faster R-CNN: Generating rotated anchor boxes with angle θ in the RPN stage, extending ROI-Align to Rotated ROI-Align to maintain feature and geometric consistency; Anchor-free methods (Rotated FCOS, Oriented RepPoints): Avoiding the sensitivity of pre-defined anchor boxes to dense tower scenes by regressing offsets and angles pixel-by-pixel. Angle loss functions (Smooth L1-θ, GCL, KLD) address training instability caused by angle periodicity.

[0119] Tower geometry identification: In post-disaster scenarios, the geometric state of towers directly reflects their structural integrity. Existing research mainly utilizes the following two types of features: Tilt angle: The angle α between the principal axis and the vertical line is calculated by rotating the minimum bounding rectangle output by the target detection. When |α|>threshold T, it is judged as tilted; Orientation consistency: The design orientation of adjacent towers within the same span is basically consistent. If a significant deviation occurs after a disaster, it indicates foundation slippage or tower breakage. Currently, in the published literature, tilt angle detection mostly relies on close-range photogrammetry by UAVs. Satellite remote sensing, due to its high viewing angle and extremely scarce samples, has not yet formed a systematic method for screening damaged towers based on tower orientation anomalies.

[0120] This invention proposes a method for identifying power pole damage from satellite remote sensing images, providing solutions to two major pain points in assessing large-scale power pole damage after a power disaster.

[0121] First, on-site reconnaissance by humans or drones is limited by various conditions such as terrain, weather, equipment battery life, and road accessibility, resulting in limited coverage and low efficiency in disaster damage perception. After a disaster, power corridors are often difficult to reach quickly due to complex terrain and road closures. Manual inspections require checking each tower individually, with a daily coverage area of ​​only a few kilometers. While drone inspections are more efficient than manual inspections, they are still affected by terrain, battery life, airspace control, and severe weather, making it difficult to complete a comprehensive survey of a large affected area in a short time. This inefficiency directly delays power repair decisions, prolongs power outages, and causes significant losses.

[0122] Second, satellite remote sensing samples of pole damage are extremely scarce, making it difficult to implement data-driven traditional deep learning methods. Pole damage is infrequent and sudden, and satellite remote sensing imagery has a limited window of opportunity, resulting in a very small number of labeled damage samples available in real-world scenarios. Current methods, typically based on deep learning for identification and detection, require a large number of damage samples to support model training. The scarcity of samples prevents the model from fully learning the damage characteristics, leading to difficulty in model convergence, and a high risk of misclassification and underclassification, making it difficult to meet the practical needs of accurate post-disaster assessment.

[0123] Please see Figure 1 The satellite remote sensing image disaster damage identification method for power poles according to embodiments of the present invention mainly includes:

[0124] S1. Collect satellite remote sensing images of multi-source power poles and construct a dataset;

[0125] S2. Input the dataset into the pre-built rotating target detection model, and output the spatial location of the power pole and identify the orientation angle through the rotating target detection model;

[0126] S3. Based on the spatial location of the power poles and the identified orientation angles, a maximum likelihood estimation model of the power pole angle distribution is constructed through angle specialization processing. The parameters of the maximum likelihood estimation model are iteratively solved using the expectation-maximization algorithm to determine the orientation angle distribution of the power poles in the image to be measured.

[0127] S4. Based on the distribution of the orientation angles of the power poles in the image to be tested, perform anomaly detection and filter out objects that exceed the threshold range, i.e., power poles that have suffered damage.

[0128] In one possible implementation, step S1 of this embodiment of the invention selects satellite remote sensing images of power corridors covering different geographical areas, terrain types (such as plains, mountains, hills, etc.), and climatic conditions (such as sunny days, cloudy days, partly cloudy days, etc.) as the original data source during data acquisition. The spatial resolution of the satellite remote sensing images is not less than 1 meter, with sub-meter (0.3-1 meter) high-resolution satellite images preferred. Data sources include, but are not limited to, Google Earth, Gaofen series satellites (such as Gaofen-7), WorldView series satellites, etc. The acquisition range needs to cover various tower types, including but not limited to common power tower structures such as straight towers, tension towers, and angle towers. At the same time, it is ensured that the images contain complete tower bodies and surrounding environmental features, avoiding the loss of tower features due to cloud cover, shadow coverage, or motion blur.

[0129] In one possible implementation, since single-scene satellite remote sensing images are typically large (e.g., several kilometers to tens of kilometers in width), directly using them for model training would require excessive GPU memory resources. Therefore, step S1 of this embodiment of the invention performs slice preprocessing on the acquired satellite remote sensing images. This slice preprocessing uses a sliding window method to divide the original satellite remote sensing images into sub-images of fixed sizes, with the slice size set to 512×512 pixels. To avoid tower targets being sliced ​​to the boundaries of multiple sub-images or resulting in incomplete features due to boundary truncation, a 20%-30% overlap area is set between adjacent slices, with an overlap width of 100-300 pixels, ensuring that tower targets spanning multiple slices are fully presented in at least one sub-image.

[0130] In one possible implementation, step S1 of this embodiment of the invention uses industry-standard image annotation tools (such as LabelMe) to manually annotate the sliced ​​sub-images. The annotation content includes the rotated bounding box of the tower target and category information. The rotated bounding box must accurately define the tower body in pixel coordinates, and the angle parameter of the bounding box must be consistent with the actual orientation of the tower, i.e., the long side of the bounding box is parallel to the extension direction of the tower column. The category information clearly indicates the tower type (such as the aforementioned straight tower). Since the solution proposed in this embodiment of the invention does not rely on disaster damage samples, the annotation does not distinguish between disaster damage states. The annotation process follows the consistency principle, and is cross-checked by at least two annotators with expertise in power towers to ensure that the annotation error (bounding box position deviation, angle deviation) does not exceed 5 pixels and 5 degrees, respectively.

[0131] In one possible implementation, step S1 of this embodiment of the invention involves randomly dividing the labeled dataset into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for learning and optimizing model parameters, containing rich diversity of pole and tower samples covering various pole and tower features under different terrain and climatic conditions. The validation set is used for hyperparameter adjustment and overfitting monitoring during model training, ensuring the model's generalization ability on unseen data. The test set is used to evaluate the model's final performance, maintaining data distribution independence from the training and validation sets to avoid data leakage affecting evaluation accuracy. After partitioning, statistical analysis of the datasets is required to ensure consistency among the three datasets in dimensions such as pole and tower type distribution, terrain coverage ratio, and image resolution distribution, thereby guaranteeing the reliability of model training and evaluation.

[0132] In one possible implementation, the rotating target detection model described in step S2 of this embodiment of the invention is constructed based on the YOLOv10 model architecture and includes the following parts:

[0133] The backbone network is used to extract multi-level features from the input image, from the bottom layer to the top layer. It balances feature extraction capability and computational cost through lightweight convolution and attention mechanisms. The backbone network adopts an improved cross-stage partial network structure (CSP) and lightweight convolution operators to improve feature representation capability while reducing parameters and computation. It embeds an attention module to aggregate features from different receptive fields through multi-scale pooling and dynamically adjusts feature weights using channel attention mechanisms. In remote sensing scenarios, the backbone network can effectively enhance the feature response of small targets and low-contrast targets, thereby improving the accuracy of subsequent detection.

[0134] The neck network addresses the adaptation problem of target detection at different scales through bidirectional feature fusion from top to bottom and bottom to top. It takes feature maps of different scales output by the backbone network as input, corresponding to the feature representations of small, medium and large targets respectively. Each scale feature map is processed by convolutional layers and normalization layers, and then the bidirectional feature fusion is used to achieve the fusion of multi-scale features. After feature optimization, it can retain both high-level semantic information and low-level detailed features, providing sufficient basis for the accurate localization and classification of rotating targets.

[0135] The head network employs a dual-label allocation and single-output strategy. The "one-to-many" head generates abundant positive samples during the training phase, enhancing convergence stability. The "one-to-one" head outputs the final bounding box during both the training and inference phases, replacing non-maximum suppression (NMS) filtering and enabling end-to-end deployment. The detection head adopts a lightweight design, using depthwise separable convolutions, which have lightweight decoupling capabilities.

[0136] Furthermore, in this embodiment of the invention, the rotating target detection model is trained. The loss function consists of three parts: classification loss, regression loss, and angle loss, which respectively constrain the accuracy of classification, detection box, and rotation angle. Data augmentation is performed using any one or more combinations of random horizontal flipping, random brightness variation, random rotation, and CutMix. The optimizer is AdamW, with an initial learning rate set to 0.001, and a cosine annealing strategy is adopted.

[0137] The trained rotating object detection model is used to infer the input image. The model selects the optimal weights from the training and applies the same segmentation strategy used when constructing the dataset to the original satellite remote sensing image. After hyperparameter tuning, the output result is obtained. For a single segmented satellite remote sensing image sample, each image contains multiple detection results. Each detection result is determined by the following parameters: category, confidence score, lower left corner coordinates, lower right corner coordinates, upper right corner coordinates, and upper left corner coordinates. Based on the detection results, the orientation angle of the rotating detection box is calculated and used as the sample. .

[0138] The orientation angle of tall structures such as power poles in satellite remote sensing imagery is primarily determined by the satellite's tilt angle. While the satellite's tilt angle is fixed within the same image, subsequent factors (topographic elevation differences, building height variations, sensor distortion, and post-processing errors) introduce additive errors, resulting in angular deviations. This causes the orientation of power poles in the final satellite imagery to deviate from the theoretical value. Considering the additive effect of these error factors, the orientation of normal power poles should follow a normal distribution. Collapsed power poles are not significantly different in appearance from normal ones, with the main difference being their orientation. Although the orientation of collapsed power poles is strongly correlated with wind direction under meteorological conditions such as typhoons, and their collapse orientation angle should also follow a normal distribution, considering broader disaster scenarios (such as heavy rainfall and earthquakes), the collapse orientation is more correlated with topographical factors, exhibiting a completely random state and following a uniform distribution. Therefore, this invention assumes that normal power poles follow a normal distribution, while collapsed power poles follow a uniform distribution, and subsequent derivations are based on this assumption.

[0139] In one possible implementation, step S3 of this embodiment of the invention, which constructs a maximum likelihood estimation model of the angle distribution of power poles through angle specialization processing, includes:

[0140] Set sample Rotate the long side of the detection frame and The angle between the positive and negative axes, with a range of values ​​of [value missing]. When screening for outliers, orientation angles of 0° and 180° are the same and should be considered as identical angles. Similarly, orientation angles... ( (smaller positive number) and orientation angle These should be considered similar angles, belonging to either normal or damaged samples. Therefore, angle specialization is needed numerically to avoid similar orientation angles being classified into different categories. Considering that the deviation angle of normal samples is small, let the sample mean be... The angle specialization processing proposed in this embodiment of the invention is based on the following mathematical expression:

[0141]

[0142] After angle specialization processing, the sample The range of values ​​is or ;

[0143] Assuming that the orientation angles of the power poles obtained after analyzing the post-disaster data constitute a sample set. Among them, the normal samples follow a normal distribution. Abnormal samples follow a uniform distribution ;

[0144] The orientation of power poles after a disaster is defined according to the following mixed distribution model:

[0145] Component 1 (normal sample): The proportion is , The probability density function is:

[0146]

[0147] Component 2 (abnormal samples): accounting for The probability density function is:

[0148]

[0149] This yields the joint probability density function of the mixture model, where the sample... The probability density from the mixture model is:

[0150]

[0151] In the formula, model parameters ;

[0152] For all samples The likelihood function of the mixture model is:

[0153]

[0154] Taking the logarithm, the log-likelihood function is:

[0155]

[0156] Considering maximizing In this case, because the mixed distribution contains latent variables where "sample affiliation is unknown" (each sample belongs to the normal / abnormal component), the log-likelihood function is non-convex with respect to the parameters, and cannot be solved analytically by simple differentiation. Therefore, latent variables are introduced. The expected value maximization algorithm is used to iteratively solve the problem, where, Indicates sample These are normal power poles. Indicates sample These are power poles that have been damaged in the disaster.

[0157] In one possible implementation, step S3 of this embodiment of the invention uses the expectation-maximization algorithm to iteratively solve the maximum likelihood estimation model parameters to determine the orientation angle distribution of power poles in the image to be measured, including:

[0158] E-step: Calculate the posterior probability of the latent variables;

[0159] definition , indicating sample The posterior probability of belonging to the normal sample is then Indicates sample The posterior probability of belonging to an outlier sample, where, Then, according to Bayes' theorem, we have:

[0160]

[0161] M-step, maximizing the expected log-likelihood;

[0162] After executing step E, based on Construct the expected value of the log-likelihood function for all data. ,in, For the first The parameters of the next iteration maximize the expected function. And obtain new parameters ;

[0163] The log-likelihood function for all data is as follows:

[0164]

[0165] by As weights, for latent variables Expected result:

[0166]

[0167] Will about Differentiation yields:

[0168]

[0169] The solution yields:

[0170]

[0171] Solve in the same way and get:

[0172]

[0173]

[0174] Repeat steps E and M until the parameter change is less than the threshold. or log-likelihood function convergence.

[0175] In one possible implementation, step S4 of this embodiment of the invention, which involves anomaly detection based on the distribution of the orientation angles of power poles in the image to be tested, includes calculating the anomaly distribution for each sample after model convergence. Posterior probability of belonging to an outlier Set threshold ,like Then determine the sample This is an abnormal sample.

[0176] According to the above description, the main process of the satellite remote sensing image power pole damage identification method of this invention includes: acquiring multiple satellite remote sensing images of power poles and creating training, validation, and test sets; preprocessing the images in the training, validation, and test sets; constructing a rotating target detection model; constructing a loss function; inputting the preprocessed training, validation, and test sets into the rotating target detection model for training and testing; inputting the preprocessed image to be tested into the trained rotating target detection model for inference and outputting the tilt detection result; performing angle specialization; constructing an MLE model; solving the model using the EM algorithm; filtering out abnormal samples and outputting the final identification result.

[0177] The core innovations of this invention are: firstly, it pioneers a technical approach of "satellite remote sensing full-area coverage + rotating target detection," which breaks free from the time and space limitations of manual / drone on-site surveys and enables rapid scanning of disaster-stricken areas at the hundred-kilometer level through sub-meter-level satellite imagery; secondly, it breaks through by adopting the logic of "normal sample modeling + angle anomaly screening," which does not rely on scarce disaster damage samples and can complete damage identification solely through a normal tower angle distribution model, fundamentally avoiding the model training difficulties caused by sample scarcity.

[0178] Please see Figure 2 and Figure 3 To verify the practical effect of the satellite remote sensing image-based power pole damage identification method of this invention, experiments were conducted on a self-built dataset. To maintain consistency between training, validation, and inference, the input images underwent the same preprocessing steps, being sliced ​​into 512*512 pixels. The final rotation detection accuracy reached 75.7%, recall 82.3%, and mAP@0.5 reached 66.3%. The detection results are as follows... Figure 2 As shown (sourced from the same satellite remote sensing image, therefore the normal tower orientations are the same). The anomaly sample screening threshold was set to 0.5, and the anomaly sample screening results are as follows. Figure 3 As shown in the yellow box.

[0179] It should be noted that the basic YOLOv10 model architecture is used for rotating target detection in this embodiment of the invention. Other rotating target detection models or improved versions of the YOLO series can also achieve the effect of outputting the orientation angle.

[0180] Furthermore, the embodiments of the present invention use the maximum likelihood estimation model (MLE) to construct a mathematical model and the expectation-maximization algorithm (EM) to screen outlier samples. Other outlier screening methods, such as clustering algorithms, can also achieve similar results.

[0181] The core of the satellite remote sensing image-based power pole damage identification method in this invention is to bypass the limitation of the extreme lack of damage samples, which prevents traditional deep learning methods from being implemented, by combining "rotating target detection" with "abnormal sample screening". This enables large-scale power pole damage identification based on satellite remote sensing imagery.

[0182] Please see Figure 4 Another embodiment of the present invention proposes a satellite remote sensing image power pole damage identification system, comprising:

[0183] Data acquisition and processing module 410 is used to acquire satellite remote sensing images of multi-source power poles and construct datasets;

[0184] The rotating target detection module 420 is used to input the dataset into a pre-built rotating target detection model, and output the spatial position of the power pole and identify the orientation angle through the rotating target detection model;

[0185] The power pole orientation angle distribution determination module 430 is used to construct a maximum likelihood estimation model of the power pole angle distribution based on the spatial location of the power pole and the identified orientation angle, and to use the expectation-maximization algorithm to iteratively solve the parameters of the maximum likelihood estimation model to determine the power pole orientation angle distribution in the image to be measured.

[0186] The abnormal power pole screening module 440 is used to detect anomalies based on the distribution of the orientation angles of power poles in the image to be tested, and to screen out objects that exceed the threshold range, i.e., power poles that have suffered damage.

[0187] In one possible implementation, the data acquisition and processing module 410 selects satellite remote sensing images of power corridors covering different geographical areas, terrain types, and climatic conditions as the source of raw data. The spatial resolution of the satellite remote sensing images is not less than 1 meter, covering a variety of power pole types and including the complete pole body and the surrounding environmental features.

[0188] In one possible implementation, the data acquisition and processing module 410 performs slice preprocessing on the acquired satellite remote sensing images. The slice preprocessing uses a sliding window method to divide the original satellite remote sensing images into sub-images of fixed size, and sets a 20%-30% overlap area between adjacent slices. Power pole targets across slices are fully presented in at least one sub-image.

[0189] In one possible implementation, the data acquisition and processing module 410 uses an image annotation tool to manually annotate the sliced ​​sub-images. The annotation content includes the rotated bounding box of the tower target and category information. The rotated bounding box defines the main body of the tower in pixel coordinates, and the angle parameters of the bounding box are consistent with the actual orientation of the power tower. The category information is used to annotate the tower type. The annotation does not distinguish between disaster damage status. The annotation process follows the consistency principle, and the annotation error does not exceed 5 pixels and 5 degrees.

[0190] In one possible implementation, the rotating target detection module 420 inputs the dataset into a pre-built rotating target detection model, and outputs the spatial position of the power pole and identifies its orientation angle through the rotating target detection model. This rotating target detection model is built based on the YOLOv10 model architecture and includes the following components:

[0191] The backbone network is used to extract multi-level features from the bottom to the top of the input image. It balances feature extraction capability and computational cost through lightweight convolution and attention mechanism. The backbone network adopts an improved cross-stage local network structure CSP and lightweight convolution operator to improve feature expression capability while reducing parameters and computation. It embeds attention module to aggregate features of different receptive fields through multi-scale pooling and dynamically adjusts feature weights using channel attention mechanism.

[0192] The neck network addresses the adaptation problem of target detection at different scales through bidirectional feature fusion from top to bottom and bottom to top. It takes feature maps of different scales output by the backbone network as input, corresponding to the feature representations of small, medium and large targets respectively. Each scale feature map is processed by convolutional layers and normalization layers, and then the multi-scale features are fused through bidirectional feature fusion. After feature optimization, it can retain both high-level semantic information and low-level detailed features.

[0193] The head network employs a dual-label allocation and single-output strategy. The "one-to-many" head generates abundant positive samples during the training phase, enhancing convergence stability. The "one-to-one" head outputs the final bounding box during both the training and inference phases, replacing non-maximum suppression (NMS) filtering and enabling end-to-end deployment. The detection head adopts a lightweight design, using depthwise separable convolutions, which have lightweight decoupling capabilities.

[0194] In one possible implementation, the rotating target detection module 420 trains the rotating target detection model. The loss function consists of three parts: classification loss, regression loss, and angle loss, which respectively constrain the accuracy of classification, detection boxes, and rotation angle. Data augmentation is performed using any one or more combinations of random horizontal flipping, random brightness variation, random rotation, and CutMix. The optimizer is AdamW, with an initial learning rate set to 0.001, and a cosine annealing strategy is adopted. The trained rotating target detection model is used to infer the input image. The rotating target detection model uses the best weights trained beforehand. The original satellite remote sensing image is segmented using the same strategy as when constructing the dataset. After hyperparameter tuning, the output result is obtained. For a single segmented satellite remote sensing image sample, each image contains multiple detection results. Each detection result is determined by the following parameters: category, confidence, lower left corner coordinates, lower right corner coordinates, upper right corner coordinates, and upper left corner coordinates. Based on the detection results, the orientation angle of the rotating detection box is calculated and used as the sample. .

[0195] In one possible implementation, when the power pole orientation angle distribution determination module 430 constructs a maximum likelihood estimation model of the power pole angle distribution through angle specialization processing, it sets a sample... Rotate the long side of the detection frame and The angle between the positive and negative axes, with a range of values ​​of [value missing]. By numerically specializing the angle, we can avoid classifying similar orientation angles into different categories; let the sample mean be... The angle specialization expression is:

[0196]

[0197] After angle specialization processing, the sample The range of values ​​is or ;

[0198] Assuming that the orientation angles of the power poles obtained after analyzing the post-disaster data constitute a sample set. Among them, the normal samples follow a normal distribution. Abnormal samples follow a uniform distribution ;

[0199] The orientation of power poles after a disaster is defined according to the following mixed distribution model:

[0200] Normal sample: percentage , The probability density function is:

[0201]

[0202] Abnormal samples: accounting for The probability density function is:

[0203]

[0204] This yields the joint probability density function of the mixture model, where the sample... The probability density from the mixture model is:

[0205]

[0206] In the formula, model parameters ;

[0207] For all samples The likelihood function of the mixture model is:

[0208]

[0209] Taking the logarithm, the log-likelihood function is:

[0210]

[0211] Considering maximizing In this case, because there are latent variables with "unknown sample affiliation" in the mixed distribution, the log-likelihood function is non-convex with respect to the parameter. Therefore, latent variables are introduced. The expected value maximization algorithm is used to iteratively solve the problem, where, Indicates sample These are normal power poles. Indicates sample These are power poles that have been damaged in the disaster.

[0212] In one possible implementation, when the power pole orientation angle distribution determination module 430 uses the expectation-maximization algorithm to iteratively solve the maximum likelihood estimation model parameters to determine the power pole orientation angle distribution in the image to be measured, it performs the following steps:

[0213] E-step: Calculate the posterior probability of the latent variables;

[0214] definition , indicating sample The posterior probability of belonging to the normal sample is then Indicates sample The posterior probability of belonging to an outlier sample, where, Then, according to Bayes' theorem, we have:

[0215]

[0216] M-step, maximizing the expected log-likelihood;

[0217] After executing step E, based on Construct the expected value of the log-likelihood function for all data. ,in, For the first The parameters of the next iteration maximize the expected function. And obtain new parameters ;

[0218] The log-likelihood function for all data is as follows:

[0219]

[0220] by As weights, for latent variables Expected result:

[0221]

[0222] Will about Differentiation yields:

[0223]

[0224] The solution yields:

[0225]

[0226] Solve in the same way and get:

[0227]

[0228]

[0229] Repeat steps E and M until the parameter change is less than the threshold. or log-likelihood function convergence.

[0230] In one possible implementation, the abnormal power pole screening module 440 calculates each sample after the model converges. Posterior probability of belonging to an outlier Set threshold ,like Then determine the sample This is an abnormal sample.

[0231] Another embodiment of the present invention also proposes an electronic device, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the satellite remote sensing image power pole damage identification method.

[0232] Another embodiment of the present invention provides a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the satellite remote sensing image power pole damage identification method.

[0233] The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals. For ease of explanation, the above content only shows the parts related to the embodiments of the present invention; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. This computer-readable storage medium is non-transitory and can be stored in storage devices formed by various electronic devices, enabling the execution process described in the method of the embodiments of the present invention.

[0234] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0235] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0236] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0237] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0238] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying damage to power poles from satellite remote sensing images, characterized in that, include: Collect satellite remote sensing images of power poles from multiple sources and construct a dataset; The dataset is input into a pre-built rotating target detection model, which outputs the spatial location of the power pole and identifies its orientation angle. Based on the spatial location of the power poles and the identified orientation angles, a maximum likelihood estimation model of the power pole angle distribution is constructed through angle specialization processing. The parameters of the maximum likelihood estimation model are iteratively solved using the expectation-maximization algorithm to determine the orientation angle distribution of the power poles in the image to be measured. Anomaly detection is performed based on the distribution of the orientation angles of power poles in the image to be tested, and objects exceeding the threshold range are selected, i.e., power poles that have suffered damage. The rotating target detection model is built based on the YOLOv10 model architecture, including a backbone network, a neck network, and a head network. In the step of outputting the spatial location of power poles and identifying their orientation angles through a rotating target detection model, the rotating target detection model is trained. The loss function consists of three parts: classification loss, regression loss, and angle loss, which respectively constrain the accuracy of classification, detection boxes, and rotation angles. Data augmentation is performed using any one or more combinations of random horizontal flipping, random brightness variation, random rotation, and CutMix. The optimizer is AdamW, with an initial learning rate set to 0.001, and a cosine annealing strategy is adopted. The trained rotating target detection model is used to infer the input image. The optimal weights of the rotating target detection model are selected, and the original satellite remote sensing image is segmented using the same segmentation strategy as when constructing the dataset. After hyperparameter tuning, the output result is obtained. For a single segmented satellite remote sensing image sample, each image contains multiple detection results. Each detection result is determined by the following parameters: category, confidence, lower left corner coordinates, lower right corner coordinates, upper right corner coordinates, and upper left corner coordinates. Based on the detection results, the orientation angle of the rotating detection box is calculated and used as the sample. ; The steps for constructing the maximum likelihood estimation model of the angle distribution of power poles through angle specialization processing include: Set sample Rotate the long side of the detection frame and The angle between the positive and negative axes, with a range of values ​​of [value missing]. By numerically specializing the angle, we can avoid classifying similar orientation angles into different categories; let the sample mean be... The angle specialization expression is: After angle specialization processing, the sample The range of values ​​is or ; Assuming that the orientation angles of the power poles obtained after analyzing the post-disaster data constitute a sample set. Among them, the normal samples follow a normal distribution. Abnormal samples follow a uniform distribution ; The orientation of power poles after a disaster is defined according to the following mixed distribution model: Normal sample: percentage , The probability density function is: Abnormal samples: accounting for The probability density function is: This yields the joint probability density function of the mixture model, where the sample... The probability density from the mixture model is: In the formula, model parameters ; For all samples The likelihood function of the mixture model is: Taking the logarithm, the log-likelihood function is: Considering maximizing In this case, because there are latent variables in the mixed distribution where the sample attribution is unknown, the log-likelihood function is non-convex with respect to the parameter. Therefore, latent variables are introduced. The expected value maximization algorithm is used to iteratively solve the problem, where, Indicates sample These are normal power poles. Indicates sample For power poles that have been damaged in the disaster; The step of using the expectation-maximization algorithm to iteratively solve the maximum likelihood estimation model parameters to determine the orientation angle distribution of power poles in the image to be measured includes: E-step: Calculate the posterior probability of the latent variables; definition , indicating sample The posterior probability of belonging to the normal sample is then Indicates sample The posterior probability of belonging to an outlier sample, where, Then, according to Bayes' theorem, we have: M-step, maximizing the expected log-likelihood; After executing step E, based on Construct the expected value of the log-likelihood function for all data. ,in, For the first The parameters of the next iteration maximize the expected function. And obtain new parameters ; The log-likelihood function for all data is as follows: by As weights, for latent variables Expected result: Will about Differentiation yields: The solution yields: Solve in the same way and get: Repeat steps E and M until the parameter change is less than the threshold. or log-likelihood function convergence.

2. The method for identifying damage to power poles using satellite remote sensing imagery according to claim 1, characterized in that, In the step of acquiring multi-source satellite remote sensing images of power poles and constructing a dataset, satellite remote sensing images of power corridors covering different geographical regions, terrain types, and climatic conditions are selected as the original data source. The spatial resolution of the satellite remote sensing images is not less than 1 meter, covering a variety of power pole types and including the complete pole body and the surrounding environmental features.

3. The method for identifying damage to power poles using satellite remote sensing imagery according to claim 1, characterized in that, In the step of acquiring multi-source satellite remote sensing images of power poles and constructing a dataset, the acquired satellite remote sensing images are pre-processed by slicing. The slicing pre-processing uses a sliding window method to divide the original satellite remote sensing images into sub-images of fixed size, and sets a 20%-30% overlap area between adjacent slices. Power pole targets across slices are fully presented in at least one sub-image.

4. The method for identifying damage to power poles using satellite remote sensing imagery according to claim 3, characterized in that, In the step of acquiring multi-source satellite remote sensing images of power poles and constructing a dataset, an image annotation tool is used to manually annotate the sliced ​​sub-images. The annotation content includes the rotated bounding box of the pole target and category information. The rotated bounding box defines the main body of the pole in pixel coordinates, and the angle parameter of the bounding box is consistent with the actual orientation of the power pole. The category information is used to annotate the pole type. The annotation does not distinguish between disaster damage status. The annotation process follows the consistency principle, and the annotation error does not exceed 5 pixels and 5 degrees.

5. The method for identifying damage to power poles using satellite remote sensing imagery according to claim 4, characterized in that, In the step of collecting satellite remote sensing images of multi-source power poles and constructing a dataset, the labeled dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:

1. The three types of datasets after division are consistent in all dimensions.

6. The method for identifying damage to power poles using satellite remote sensing imagery according to claim 1, characterized in that: The backbone network is used to extract multi-level features from the bottom to the top of the input image. It balances feature extraction capability and computational cost through lightweight convolution and attention mechanism. The backbone network adopts an improved cross-stage local network structure CSP and lightweight convolution operator to improve feature expression capability while reducing parameters and computation. It embeds attention module to aggregate features of different receptive fields through multi-scale pooling and dynamically adjusts feature weights using channel attention mechanism. The neck network addresses the adaptation problem of target detection at different scales through bidirectional feature fusion from top to bottom and bottom to top. It takes feature maps of different scales output by the backbone network as input, which correspond to the feature representations of small, medium and large targets respectively. Each scale feature map is processed by convolutional layers and normalization layers, and then the fusion of multi-scale features is achieved through bidirectional feature fusion. Through feature optimization, it is possible to retain both high-level semantic information and low-level detailed features; The head network employs a dual-label allocation and single-output strategy. The "one-to-many" head generates abundant positive samples during the training phase, enhancing convergence stability. The "one-to-one" head outputs the final bounding box during both the training and inference phases, replacing non-maximum suppression (NMS) filtering and enabling end-to-end deployment. The detection head adopts a lightweight design, using depthwise separable convolutions, which have lightweight decoupling capabilities.

7. The method for identifying damage to power poles using satellite remote sensing imagery according to claim 1, characterized in that, The step of anomaly detection based on the orientation angle distribution of power poles in the image under test includes calculating each sample after model convergence. Posterior probability of belonging to an outlier Set threshold ,like Then determine the sample This is an abnormal sample.

8. A satellite remote sensing imagery-based system for identifying damage to power poles, characterized in that, include: The data acquisition and processing module is used to acquire satellite remote sensing images of power poles from multiple sources and construct datasets. The rotating target detection module is used to input the dataset into a pre-built rotating target detection model, and output the spatial location of the power pole and identify the orientation angle through the rotating target detection model; The module for determining the orientation angle distribution of power poles is used to construct a maximum likelihood estimation model of the orientation angle distribution of power poles based on the spatial location of the power poles and the identified orientation angles through angle specialization processing. The module then uses the expectation-maximization algorithm to iteratively solve the parameters of the maximum likelihood estimation model to determine the orientation angle distribution of power poles in the image to be measured. The abnormal power pole screening module is used to detect anomalies based on the distribution of the orientation angles of power poles in the image to be tested, and to screen out objects that exceed the threshold range, i.e., power poles that have been damaged. The rotating target detection model is built based on the YOLOv10 model architecture, including a backbone network, a neck network, and a head network. The rotating target detection module trains the rotating target detection model. The loss function consists of three parts: classification loss, regression loss, and angle loss, which respectively constrain the accuracy of classification, detection boxes, and rotation angle. Data augmentation is performed using any combination of random horizontal flipping, random brightness variation, random rotation, and CutMix. The optimizer is AdamW, with an initial learning rate set to 0.001 and a cosine annealing strategy. The trained rotating target detection model is used to infer the input image. The rotating target detection model uses the best weights trained beforehand. The original satellite remote sensing image is segmented using the same strategy as when constructing the dataset. After hyperparameter tuning, the output result is obtained. For a single segmented satellite remote sensing image sample, each image contains multiple detection results. Each detection result is determined by the following parameters: category, confidence, lower left corner coordinates, lower right corner coordinates, upper right corner coordinates, and upper left corner coordinates. Based on the detection results, the orientation angle of the rotating detection box is calculated and used as the sample. ; When the power pole orientation angle distribution determination module constructs the maximum likelihood estimation model of the power pole angle distribution through angle specialization processing, it sets the sample... Rotate the long side of the detection frame and The angle between the positive and negative axes, with a range of values ​​of [value missing]. ; By numerically specializing the angle, we can avoid classifying similar orientation angles into different categories; let the sample mean be... The angle specialization expression is: After angle specialization processing, the sample The range of values ​​is or ; Assuming that the orientation angles of the power poles obtained after analyzing the post-disaster data constitute a sample set. Among them, the normal samples follow a normal distribution. Abnormal samples follow a uniform distribution ; The orientation of power poles after a disaster is defined according to the following mixed distribution model: Normal sample: percentage , The probability density function is: Abnormal samples: accounting for The probability density function is: This yields the joint probability density function of the mixture model, where the sample... The probability density from the mixture model is: In the formula, model parameters ; For all samples The likelihood function of the mixture model is: Taking the logarithm, the log-likelihood function is: Considering maximizing In this case, because there are latent variables in the mixed distribution where the sample attribution is unknown, the log-likelihood function is non-convex with respect to the parameter. Therefore, latent variables are introduced. The expected value maximization algorithm is used to iteratively solve the problem, where, Indicates sample These are normal power poles. Indicates sample For power poles that have been damaged in the disaster; When the power pole orientation angle distribution determination module uses the expectation-maximization algorithm to iteratively solve the maximum likelihood estimation model parameters to determine the power pole orientation angle distribution in the image to be measured, it performs the following steps: E-step: Calculate the posterior probability of the latent variables; definition , indicating sample The posterior probability of belonging to the normal sample is then Indicates sample The posterior probability of belonging to an outlier sample, where, Then, according to Bayes' theorem, we have: M-step, maximizing the expected log-likelihood; After executing step E, based on Construct the expected value of the log-likelihood function for all data. ,in, For the first The parameters of the next iteration maximize the expected function. And obtain new parameters ; The log-likelihood function for all data is as follows: by As weights, for latent variables Expected result: Will about Differentiation yields: The solution yields: Solve in the same way and get: Repeat steps E and M until the parameter change is less than the threshold. or log-likelihood function convergence.

9. The satellite remote sensing image power pole damage identification system according to claim 8, characterized in that, The data acquisition and processing module selects satellite remote sensing images of power corridors covering different geographical regions, terrain types, and climate conditions as the source of raw data. The spatial resolution of the satellite remote sensing images is no less than 1 meter, covering a variety of power pole types and including the complete pole body and the surrounding environmental features.

10. The satellite remote sensing image power pole damage identification system according to claim 8, characterized in that, The data acquisition and processing module performs slice preprocessing on the acquired satellite remote sensing images. The slice preprocessing uses a sliding window method to divide the original satellite remote sensing images into sub-images of fixed size, and sets a 20%-30% overlap area between adjacent slices. Power pole targets that cross slices are fully presented in at least one sub-image.

11. The satellite remote sensing image power pole damage identification system according to claim 10, characterized in that, The data acquisition and processing module uses an image annotation tool to manually annotate the sliced ​​sub-images. The annotation content includes the rotated bounding box of the tower target and category information. The rotated bounding box defines the main body of the tower in pixel coordinates, and the angle parameters of the bounding box are consistent with the actual orientation of the power tower. The category information is used to annotate the tower type. The annotation does not distinguish between disaster damage status. The annotation process follows the consistency principle, and the annotation error does not exceed 5 pixels and 5 degrees.

12. The satellite remote sensing image power pole damage identification system according to claim 8, characterized in that, The rotating target detection module inputs the dataset into a pre-built rotating target detection model, and outputs the spatial position of the power pole and identifies its orientation angle through the rotating target detection model: The backbone network is used to extract multi-level features from the bottom to the top of the input image. It balances feature extraction capability and computational cost through lightweight convolution and attention mechanism. The backbone network adopts an improved cross-stage local network structure CSP and lightweight convolution operator to improve feature expression capability while reducing parameters and computation. It embeds attention module to aggregate features of different receptive fields through multi-scale pooling and dynamically adjusts feature weights using channel attention mechanism. The neck network addresses the adaptation problem of target detection at different scales through bidirectional feature fusion from top to bottom and bottom to top. It takes feature maps of different scales output by the backbone network as input, which correspond to the feature representations of small, medium and large targets respectively. Each scale feature map is processed by convolutional layers and normalization layers, and then the fusion of multi-scale features is achieved through bidirectional feature fusion. Through feature optimization, it is possible to retain both high-level semantic information and low-level detailed features; The head network employs a dual-label allocation and single-output strategy. The "one-to-many" head generates abundant positive samples during the training phase, enhancing convergence stability. The "one-to-one" head outputs the final bounding box during both the training and inference phases, replacing non-maximum suppression (NMS) filtering and enabling end-to-end deployment. The detection head adopts a lightweight design, using depthwise separable convolutions, which have lightweight decoupling capabilities.

13. The satellite remote sensing image power pole damage identification system according to claim 8, characterized in that, The abnormal power pole screening module calculates the values ​​for each sample after the model converges. Posterior probability of belonging to an outlier Set threshold ,like Then determine the sample This is an abnormal sample.

14. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the satellite remote sensing image power pole damage identification method as described in any one of claims 1 to 7.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the satellite remote sensing image power pole damage identification method as described in any one of claims 1 to 7.

Images