Automatic identification and geolocation methods, systems and electronic equipment for power transmission towers based on satellite remote sensing and deep learning

Abstract

This invention discloses a method, system, and electronic device for automatic identification and geolocation of power transmission towers based on satellite remote sensing and deep learning. The method includes: S1, collecting satellite remote sensing images containing power transmission towers and establishing a sample dataset; S2, training a deep learning object detection model based on the sample dataset; S3, automatically acquiring a satellite base map of the target area; S4, using the trained deep learning object detection model to perform batch inference and identification on the satellite base map, and employing a non-maximum suppression algorithm to remove duplicates from the identification results; S5, converting the pixel coordinates of the identified power transmission towers in the image into geographic latitude and longitude coordinates; S6, mapping the geographic latitude and longitude coordinates to a publicly available map base map for visualization. This invention achieves rapid, accurate, and automated identification and location of power transmission towers, overcoming problems such as low efficiency of traditional manual inspections, difficulties in power grid asset inventory, and a lack of visual map bases.

Description

Technical Field

[0001] This invention relates to the field of power facility inspection and geographic information technology, and in particular to a method, system and electronic equipment for automatic identification and geographic positioning of transmission towers based on satellite remote sensing and deep learning. Background Technology

[0002] With the rapid development of the national economy, the scale of power grid infrastructure is expanding year by year, covering a wide range of complex geographical environments, including cities, rural areas, mountains, and hills. Transmission towers, as the backbone support structure of the power grid system, are constantly exposed to the natural environment and face direct threats from various natural disasters such as windstorms, landslides, and floods. Their safe and stable operation is crucial. Currently, routine inspections and asset inventory of transmission towers mainly rely on manual inspections. However, this method is not only costly and inefficient, but also has a long inspection cycle and is difficult to respond quickly to equipment condition assessments after sudden disasters. More importantly, for the entire power grid system spanning a vast area, traditional manual methods are almost impossible to achieve efficient, comprehensive, large-scale rapid surveys and systemic risk assessments.

[0003] In recent years, automatic identification of ground targets using high-resolution satellite remote sensing imagery has become a promising technological approach. However, when applying this technology to the identification and location of power transmission towers, existing solutions still face several prominent problems, limiting its large-scale application. First, in the data acquisition and processing stage, the acquisition, stitching, and coordinate correction of large-scale satellite base maps remain cumbersome and have a low degree of automation, becoming a major bottleneck in the application process. Second, regarding target recognition models, publicly available, high-quality datasets specifically for power transmission towers are extremely scarce. Furthermore, satellite imagery itself is affected by various factors such as terrain type, seasonal changes, weather conditions, and imaging angle, causing deep learning models to exhibit unstable recognition accuracy when transferred from one region to another or applied from one imaging scene to another, resulting in significantly insufficient generalization ability. Third, in terms of target size, even in high-resolution satellite images, a single power transmission tower typically appears as a small target, placing extremely high demands on the sensitivity and small target detection capabilities of the detection algorithm. Finally, in terms of positioning accuracy, many existing solutions use a relatively simple conversion method from image pixel coordinates to geographic latitude and longitude coordinates, which fails to fully consider the geometric distortion of the image itself and the systematic deviation between different coordinate systems. This results in a non-negligible error between the final output latitude and longitude coordinates of the transmission tower and its actual physical location, making it difficult to use directly for engineering applications that require precise positioning.

[0004] Therefore, there is an urgent need in this field for a method, system, and corresponding equipment that can overcome the above-mentioned defects and achieve high efficiency, high precision, and high degree of automation for large-scale automatic identification and precise positioning of transmission towers. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method, system, and electronic device for automatic identification and geolocation of power transmission towers based on satellite remote sensing and deep learning. This method achieves rapid, accurate, and automated inventory and location of power transmission tower assets over a large area by constructing a high-quality sample dataset, training a robust deep learning model, and realizing automated processing and precise coordinate transformation of satellite base maps.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for automatic identification and geolocation of transmission towers based on satellite remote sensing and deep learning, comprising the following steps:

[0008] S1. Collect satellite remote sensing images containing power transmission towers and establish a sample dataset;

[0009] S2. Based on the sample dataset, train a deep learning object detection model;

[0010] S3. Automatically acquire satellite base maps of the target area;

[0011] S4. Use the trained deep learning target detection model to perform batch inference and recognition on the satellite base map, and use a non-maximum suppression algorithm to remove duplicates from the recognition results;

[0012] S5. Convert the pixel coordinates of the identified power transmission towers in the image after deduplication into geographic latitude and longitude coordinates;

[0013] S6. Map the geographic latitude and longitude coordinates to a public map base map for visualization.

[0014] As a preferred embodiment of the present invention, step S1 specifically includes: collecting approximately 1,000 high-resolution 4K satellite remote sensing images containing power transmission towers, ensuring that the images cover different landforms, seasonal phases, and imaging conditions; annotating the power transmission tower targets in the images with data, using bounding boxes to accurately mark the location of each power transmission tower, and generating corresponding annotation files; after completing the annotation, randomly dividing the entire dataset into a training set and a validation set in a ratio of approximately 8:2.

[0015] As a preferred embodiment of the present invention, step S2 specifically includes: converting the labeled dataset into a specific format required by the model; training the model using both YOLOv8 and DETR detection frameworks; for the YOLOv8 model, selecting its pre-trained weights as initial parameters, setting the input image size to 640x640 pixels, using a stochastic gradient descent optimizer, training for 100 epochs, and enabling mosaic data augmentation; for the DETR model, loading the pre-trained weights, using the AdamW optimizer, setting the initial learning rate to 1e-4, and training for 100 epochs; during training, monitoring precision, recall, F1 score, and mean precision on the validation set, and saving the model weights with the best performance.

[0016] As a preferred embodiment of the present invention, step S3 specifically includes: given the latitude and longitude range and zoom level of the target area, calling the map service interface through the program to cut the map of the specified range into multiple 256x256 pixel tiles and download them in batches; converting the tiles from the Web Mercator projection coordinate system to the WGS84 geographic coordinate system, and performing correction from the GCJ-02 coordinate system to the WGS84 coordinate system on the data within China; stitching all the tiles together with a 20% overlap to generate a complete high-resolution satellite base map, and recording its precise latitude and longitude coordinates; the zoom level is not less than level 19.

[0017] As a preferred embodiment of the present invention, step S4 specifically includes: cropping the satellite base map obtained in step S3 to a size that conforms to the model input, inputting it into the trained model for batch inference, and obtaining the recognition result including the bounding box and confidence score; using the non-maximum suppression algorithm to deduplicate all detection results, setting the confidence threshold to 0.5 and the IOU threshold to 0.45, and retaining an optimal detection result for each transmission tower.

[0018] As a preferred embodiment of the present invention, step S5 specifically includes: for each transmission tower target obtained after deduplication, taking the pixel coordinates of the center point of the bottom of its bounding box; based on the known latitude and longitude coordinates of the four corners of the satellite base map, establishing a mapping relationship from the image pixel coordinate system to the WGS84 geographic coordinate system, and accurately converting the pixel coordinates into geodetic latitude and longitude coordinates through affine transformation.

[0019] As a preferred embodiment of the present invention, step S6 specifically includes: outputting the latitude and longitude coordinates of all transmission towers into a standard format file, wherein the file is in CSV format and contains TIFF file name, latitude and longitude coordinates of transmission towers, bounding box coordinates and confidence information; importing the file into professional GIS software or loading it into an online base map service for visualization, wherein the online base map includes OpenStreetMap, Baidu Map or Gaode Map, to form a transmission tower distribution point layer.

[0020] As a preferred embodiment of the present invention, in step S2, the image input size is not less than 640×640 pixels during model training; and in step S3, the satellite base map is saved as a TIFF file containing geographic coordinate information.

[0021] Secondly, the present invention provides an automatic identification and geolocation system for transmission towers, the system comprising:

[0022] The data collection and annotation module is used to collect satellite remote sensing images containing power transmission towers and build a sample dataset;

[0023] The model training module is used to train a deep learning object detection model based on the sample dataset;

[0024] The base map acquisition and processing module is used to automatically acquire satellite base maps of the target area;

[0025] The reasoning, recognition, and deduplication module is used to perform batch reasoning and recognition on satellite base maps using a trained model, and to deduplicate the results.

[0026] The coordinate transformation module is used to convert the pixel coordinates of the identified transmission towers into geographic latitude and longitude coordinates;

[0027] The visualization output module is used to map latitude and longitude coordinates to a public map base for visualization.

[0028] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in any one of the first aspects above.

[0029] Compared with the prior art, the present invention achieves the following beneficial technical effects:

[0030] This application achieves rapid, accurate, and automated identification of transmission towers over a large area by constructing a high-quality sample dataset and an advanced deep learning model. Combined with fully automatic satellite base map processing and precise coordinate transformation technology, it effectively overcomes the problems of low efficiency, insufficient model generalization ability, and large geographic positioning error in traditional manual inspections, significantly improving the efficiency and accuracy of power facility inventory and possessing good engineering practicality. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the overall process of the automatic identification and geolocation method for transmission towers based on satellite remote sensing and deep learning, as described in this invention.

[0033] Figure 2 This is a structural block diagram of the automatic identification and geolocation system for transmission towers based on satellite remote sensing and deep learning, as described in this invention.

[0034] Figure 3 This is a schematic diagram of the training process of the deep learning model (YOLOv8 / DETR) in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram illustrating the principle of how the model in this embodiment of the invention performs reasoning, identification, and coordinate transformation on a satellite base map. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] It should be noted that, in the description of this invention, the terms “comprising,” “including,” or any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0038] Example 1

[0039] like Figure 1 As shown, this embodiment provides a method for automatic identification and geolocation of transmission towers based on satellite remote sensing and deep learning. This embodiment takes a certain area in China as the research area to demonstrate the practical application of this method.

[0040] The method includes the following steps:

[0041] S1. Collect satellite remote sensing images containing power transmission towers and establish a sample dataset.

[0042] Specifically, approximately 1000 high-resolution 4K satellite remote sensing images containing power transmission towers were collected, ensuring coverage of diverse terrains including suburbs, rural areas, and cities, as well as different seasons and imaging conditions, to enhance the generalization ability of subsequent models. Subsequently, each power transmission tower target in the images was labeled, its location precisely marked using rectangular bounding boxes, and a corresponding annotation file (such as a YOLO format .txt file or a COCO format .json file) was generated. After labeling all images, the entire dataset was randomly divided into training and validation sets in an approximately 8:2 ratio.

[0043] S2. Based on the sample dataset, train a deep learning object detection model.

[0044] Preferably, this step uses two advanced detection frameworks, YOLOv8 and DETR, for parallel model training. Those skilled in the art will know that these two object detection models can also be replaced by other detection models.

[0045] First, data preparation is performed by converting the training and validation sets, which were divided in step S1, into the specific data formats required by their respective models.

[0046] For the YOLOv8 model, the weights pre-trained on a large dataset were used as the initial parameters. The training parameters were set as follows: input image size was adjusted to 640x640 pixels, a stochastic gradient descent optimizer was used, the total training epochs were 100, and mosaic data augmentation was enabled to improve the model's robustness to changes in target scale and proportion. During training, the model automatically calculated and optimized bounding box regression loss, classification loss, etc., and monitored performance metrics such as precision, recall, F1 score, intersection-over-union ratio, and mean precision in real time on the validation set. After training, the model weight file that best performed on the entire validation set was saved.

[0047] For the DETR model, its pre-trained weights are also loaded. The training process uses the AdamW optimizer with an initial learning rate of 1e-4, coupled with a weight decay strategy. Its loss function consists of the L1 loss of the bounding box, the GIoU loss, and the cross-entropy loss of the class prediction. The model performs global computation and ensemble prediction through its built-in Transformer encoder-decoder architecture, and is trained for 100 epochs. Based on the model's overall performance on the validation set, the optimal model weights are saved.

[0048] This step will simultaneously yield two optimal detection models based on YOLOv8 and DETR, allowing for flexible selection based on different priorities regarding inference speed or detection accuracy in practical applications. The input and output images for this process are as follows: Figure 3As shown.

[0049] S3. Automatically acquire satellite base maps of the target area.

[0050] In this embodiment, the target area is a location in China. The latitude and longitude boundaries of this area are given, and the map zoom level is set to 20 (to ensure high resolution). An automated program is used to call service interfaces such as Google Maps to cut the map of the specified area into numerous standard 256×256 pixel tiles, which are then downloaded in batches.

[0051] After downloading, coordinate system transformation and correction are performed on the tiles. First, the tiles are converted from the Web Mercator projected coordinate system to the universal WGS84 geographic coordinate system. Next, for the data in China, a correction operation is performed from the GCJ-02 coordinate system to the WGS84 coordinate system to eliminate human-injected systematic biases and ensure coordinate accuracy.

[0052] Finally, all the coordinate-processed tiles are stitched together with a preset 20% overlap to generate a seamless, high-resolution satellite base map covering the area. This base map is then saved as a TIFF file containing geographic coordinate information (GeoTIFF), which records the precise latitude and longitude coordinates of the four corners of the image, as well as other necessary metadata.

[0053] S4. The trained deep learning target detection model is used to perform batch inference and recognition on the satellite base map, and the non-maximum suppression algorithm is used to remove duplicates from the recognition results.

[0054] The optimal model trained in step S2 is invoked (using the YOLOv8 model as an example in this embodiment). The satellite TIFF base map of a certain location generated in step S3 is cropped or segmented into multiple sub-images according to the input size required by the model (e.g., 640×640). These sub-images are then sequentially input into the model for batch inference. The model outputs the bounding box information (usually represented by the coordinates of the top left and bottom right corners, i.e., x1, y1, x2, y2) and the corresponding confidence score for each sub-image containing suspected transmission tower targets.

[0055] Because a 20% overlap area is set during tile splicing in step S3, the same transmission tower is highly likely to be repeatedly detected in the edge areas of two or more adjacent sub-graphs. To address this issue, a non-maximum suppression algorithm is used to deduplicate all preliminary detection results. The confidence threshold for NMS is set to 0.5, and the IOU threshold is set to 0.45. This algorithm filters out redundant detection boxes pointing to the same target based on the degree of overlap and confidence level between bounding boxes, ultimately retaining the optimal detection result with the highest confidence for each real transmission tower.

[0056] S5. Convert the pixel coordinates of the identified transmission towers in the image after deduplication into geographical latitude and longitude coordinates.

[0057] For each transmission tower target retained after deduplication in step S4, the pixel coordinates of its representative point in the original satellite base map (TIFF file) are first determined. This invention preferably uses the bottom center point of its bounding box as the representative point because this point is closer to the contact point between the transmission tower base and the ground, and can more accurately reflect its actual geographical coordinates.

[0058] Then, based on the geographic metadata (especially the latitude and longitude coordinates of the four corners of the image) recorded in the source TIFF file where the transmission tower is located, a precise mapping relationship is established from the image pixel coordinate system of the entire base map to the WGS84 geographic coordinate system. Through affine transformation (a linear geometric transformation), the pixel coordinates of the transmission tower are precisely converted into the corresponding latitude and longitude coordinates (longitude, latitude).

[0059] S6. Map the geographic latitude and longitude coordinates to a public map base map for visualization.

[0060] Integrate the identification and location results of all transmission towers obtained in step S5 to generate structured data. Data fields include: source TIFF file name, geodetic latitude and longitude coordinates (longitude, latitude) of the transmission tower, bounding box coordinates, and identification confidence score. Output this data as a standard CSV file.

[0061] Table 1 Standard Format CSV Files

[0062]

[0063] Finally, import the CSV file into professional GIS software (such as QGIS), or load it programmatically into online base map services such as OpenStreetMap, Baidu Maps, and Gaode Maps. Figure 4 As shown, the latitude and longitude coordinates of the transmission towers are overlaid on the base map as a point layer, thereby generating a clear and intuitive spatial distribution map of the transmission towers in a certain area, which can be used for subsequent asset statistics, density analysis, route planning and risk assessment.

[0064] Example 2

[0065] like Figure 2 As shown, this embodiment provides an automatic identification and geolocation system for power transmission towers, used to implement the method described in Embodiment 1. This system can be implemented through software, hardware, or a combination of both, for example, deployed on a server or a distributed computing cluster.

[0066] The system includes:

[0067] The data collection and annotation module is used to collect satellite remote sensing images containing power transmission towers and establish a sample dataset; this module provides an annotation tool interface to assist users in completing the annotation work.

[0068] The model training module is used to train a deep learning object detection model based on the sample dataset. This module integrates training frameworks such as YOLOv8 and DETR, can configure training parameters, automatically execute the training process, and monitor the training status.

[0069] The base map acquisition and processing module is used to automatically acquire satellite base maps of the target area. This module can automatically complete tile downloading, coordinate transformation, correction and stitching tasks by calling the map service API according to the input latitude and longitude range and zoom level.

[0070] The reasoning, recognition, and deduplication module is used to perform batch reasoning and recognition on satellite base maps using a trained model, and to deduplicate the results. This module loads the trained model weights, automatically segments and reasons on the input large-format base map, and applies the NMS algorithm to handle duplicate detection.

[0071] The coordinate transformation module is used to convert the pixel coordinates of the identified transmission towers into geographic latitude and longitude coordinates; this module reads the geographic metadata of the base map and performs accurate affine transformation calculations.

[0072] The visualization output module is used to map latitude and longitude coordinates to a public map base map for visualization. This module generates result files in standard formats (such as CSV) and provides an interface to import the results into GIS software or publish them to online map services.

[0073] The modules mentioned above communicate and collaborate through predefined interfaces and data to jointly complete a fully automated process from data preparation to result visualization.

[0074] Example 3

[0075] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the steps of the method described in Embodiment 1.

[0076] The electronic device may be a server, workstation, personal computer, or dedicated image processing equipment. The memory includes a non-volatile storage medium for storing the computer program and data generated during the execution of the method, such as trained model weights, satellite base maps, intermediate recognition results, and final coordinate files. The processor may be a central processing unit, graphics processing unit, or tensor processing unit, for performing computationally intensive model training and inference tasks. The electronic device may also include a communication interface for acquiring satellite imagery data from a network or interacting with map services.

[0077] 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic identification and geolocation of power transmission towers based on satellite remote sensing and deep learning, characterized in that, Includes the following steps: S1. Collect satellite remote sensing images containing power transmission towers and establish a sample dataset; S2. Based on the sample dataset, train a deep learning object detection model; S3. Automatically acquire satellite base maps of the target area; S4. Use the trained deep learning target detection model to perform batch reasoning and recognition on the satellite base map, and use a non-maximum suppression algorithm to remove duplicates from the recognition results; S5. Convert the pixel coordinates of the identified power transmission towers in the image after deduplication into geographical latitude and longitude coordinates; S6. Map the geographic latitude and longitude coordinates to a publicly available map base map for visualization; Step S3 specifically includes: Given the latitude and longitude range and zoom level of the target area, the program calls the map service interface to cut the map of the specified range into multiple tiles and download them in batches; The downloaded tiles are converted from the Web Mercator projected coordinate system to the WGS84 geographic coordinate system, and the data within China is corrected from the GCJ-02 coordinate system to the WGS84 coordinate system. All tiles are stitched together according to a preset overlap to generate a complete high-resolution satellite base map, and its precise latitude and longitude coordinates are recorded. The scaling level is no less than 19 levels; Step S4 specifically includes: The acquired satellite base map is cropped to fit the size of the model input and then input into the trained model for batch inference to obtain recognition results including bounding boxes and confidence scores. A non-maximum suppression algorithm is used to remove duplicates from all detection results, and an optimal detection result is retained for each transmission tower. The confidence threshold for the nonmaximum suppression algorithm is set to 0.5, and the IOU threshold is set to 0.

45.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: Collect satellite remote sensing images containing power transmission towers; annotate the power transmission tower targets in the images, accurately mark the location of each power transmission tower using bounding boxes, and generate corresponding annotation files; randomly divide the entire dataset into training and validation sets according to a preset ratio.

3. The method according to claim 2, characterized in that, Step S2 specifically includes: The labeled dataset was converted into the specific format required by the model; the model was trained using YOLOv8 and DETR frameworks respectively. For the YOLOv8 model, pre-trained weights were selected as initial parameters, the input image size was set, the stochastic gradient descent optimizer was used for training, and mosaic data augmentation was enabled. For the DETR model, pre-trained weights are loaded, and the AdamW optimizer is used for training. Monitor model performance metrics on the validation set and save the weights of the best-performing model.

4. The method according to claim 1, characterized in that, Step S5 specifically includes: For each transmission tower target obtained after deduplication, take the pixel coordinates of the center point of the bottom of its bounding box; Based on the known latitude and longitude coordinates of the four corners of the satellite base map, a mapping relationship from the image pixel coordinate system to the WGS84 geographic coordinate system is established, and the pixel coordinates are accurately converted into geodetic latitude and longitude coordinates through affine transformation.

5. The method according to claim 4, characterized in that, In step S2, during model training, the image input size is no less than 640×640 pixels; and in step S3, the satellite base map is saved as a TIFF file containing geographic coordinate information.

6. The method according to claim 4, characterized in that, In step S6, the base map used for visualization includes OpenStreetMap, Baidu Maps, or Gaode Maps; and the output standard format file is CSV format, which includes TIFF file name, transmission tower latitude and longitude coordinates, bounding box coordinates, and confidence information.

7. A transmission tower automatic identification and geolocation system for implementing the automatic identification and geolocation method for transmission towers based on satellite remote sensing and deep learning as described in claim 1, characterized in that, The system includes: The data collection and annotation module is used to collect satellite remote sensing images containing power transmission towers and build a sample dataset; The model training module is used to train a deep learning object detection model based on the sample dataset; The base map acquisition and processing module is used to automatically acquire satellite base maps of the target area; The reasoning, recognition, and deduplication module is used to perform batch reasoning and recognition on satellite base maps using a trained model, and to deduplicate the results. The coordinate transformation module is used to convert the pixel coordinates of the identified transmission towers into geographic latitude and longitude coordinates; The visualization output module is used to map latitude and longitude coordinates to a public map base for visualization.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-6.

Images