Landslide crack identification method based on unmanned aerial vehicle orthoimage and pyramid scene analysis network image segmentation algorithm

Through drone tilt photography technology and pyramid scene analysis network image segmentation algorithm, the problems of high cost, time-consuming and limited accuracy of traditional landslide crack recognition technology are solved, and efficient, accurate, real-time identification and monitoring of landslide cracks are achieved.

CN119048932APending Publication Date: 2024-11-29CHINA THREE GORGES CORPORATION +2

Patent Information

Application Number
CN202410932927.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

Traditional ground observation and satellite imaging technologies have problems such as high cost, long time, limited data accuracy and inability to achieve continuous real-time monitoring in landslide crack identification, especially in complex terrain and harsh environments, which are difficult to effectively monitor the distribution and changes of landslides.

Method used

UAV tilt photography technology combined with pyramid scene analysis network image segmentation algorithm is used to obtain high-resolution images through drones, perform image preprocessing and semantic segmentation of deep learning technology, and realize automatic identification and monitoring of landslide cracks.

Benefits of technology

It realizes efficient, accurate and real-time identification of landslide cracks. Compared with traditional methods, it can better capture subtle changes on the surface, improves identification efficiency and accuracy, and is suitable for monitoring large-scale and complex terrain.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048932A_ABST
    Figure CN119048932A_ABST
Patent Text Reader

Abstract

The invention discloses a landslide crack identification method based on an unmanned aerial vehicle orthoimage and pyramid scene analysis network image segmentation algorithm. The method comprises the following steps: S1, landslide data collection; s2, the unmanned aerial vehicle flies in a simulated manner; s3, unmanned aerial vehicle image processing; s4, deep learning early-stage image processing; s5, creating a training sample; s6, model and model training; and S7, crack identification. The invention provides a landslide crack identification method based on an unmanned aerial vehicle orthoimage and pyramid scene analysis network image segmentation algorithm. The landslide crack identification method has the advantages of high efficiency, accuracy and high real-time performance. Compared with traditional ground survey or a monitoring method based on satellite images, the unmanned aerial vehicle oblique photography technology can capture subtle changes of the earth surface with higher spatial resolution, and the digital image processing and pattern recognition algorithm can automatically and quickly analyze data, so that existence of landslide cracks can be found in time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of landslide crack recognition, and in particular to a landslide crack recognition method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm. Background Art

[0002] Geological disasters occur frequently around the world, and various geological disasters have caused great harm to public safety and property. Among them, landslides are one of the most serious types of geological disasters that cause casualties and property losses. They are characterized by a large number of developments, high density, and high frequency. In the formation and development of landslides, landslide cracks are an important associated phenomenon.

[0003] Landslide cracks refer to cracks, fault zones or crack networks that appear on the surface or around the landslide body. Their formation is one of the important geological characteristics in the landslide evolution process. By observing and analyzing the number, length, width, distribution density and morphology of cracks in detail, we can obtain the following important information: First, the characteristics of cracks are often used to determine the scale of landslides. Generally speaking, the more, longer and wider the cracks are, the larger the scale of the landslide may be. Therefore, the number and morphology of cracks can be used as one of the important clues to assess the scale and potential danger of landslides. Secondly, the morphology, distribution and changes of cracks can reflect the mechanism and process of landslide formation. By analyzing the characteristics of cracks, we can understand the causes and triggering conditions of landslides, as well as the influence of geological and topographic factors on landslide evolution, so as to further explore the mechanism of landslide formation. In addition, the morphology and changes of cracks can also provide important information about the development trend of landslides. By monitoring and analyzing the evolution of cracks, we can predict the possible changes and evolution trends of landslides, providing an important basis for early warning and response to landslide disasters. Finally, the characteristics of cracks can also guide the monitoring and control of landslides. Regular monitoring of changes in cracks can timely detect abnormal changes in landslides and provide early warning of possible disasters. In addition, by analyzing the distribution and morphology of cracks, it can provide a scientific basis for the management and prevention of landslides and guide the implementation of relevant engineering measures and the evaluation of their effects.

[0004] Existing research on landslide cracks covers a variety of technical means, including remote sensing technology, ground observation and measurement, digital image processing and pattern recognition, machine learning, and multi-source data fusion. Remote sensing technologies such as satellite images and aerial photography can quickly obtain a wide range of surface information to assist in the identification and monitoring of landslide cracks. Using digital image processing technology and pattern recognition algorithms, remote sensing images or ground photos are processed and analyzed to extract the characteristics of landslide cracks and automatically or semi-automatically identify them. This method can achieve rapid processing and analysis of large amounts of data through computer algorithms, improving recognition efficiency and accuracy. However, the use of traditional ground surveys to conduct field observations and measurements of landslide cracks requires a lot of manpower and material resources. The difficulty in obtaining data due to complex terrain or inconvenient transportation makes the cost high and time-consuming. Secondly, due to operational limitations, such as limitations of measuring equipment and the existence of human errors, data accuracy may be subject to certain restrictions, especially in areas with complex terrain. In addition, ground observation and measurement are point-to-point, and continuous and real-time data acquisition cannot be achieved, making it difficult to fully grasp the distribution and changes of landslide cracks. Moreover, ground surveys are not suitable for monitoring needs in large areas, especially in mountainous or remote areas, and observations are even more difficult in harsh environments. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a landslide crack identification method based on unmanned aerial vehicle orthophotos and pyramid scene analysis network image segmentation algorithm, which automatically identifies landslide cracks by combining unmanned aerial vehicle oblique photography technology and image segmentation algorithm. Using unmanned aerial vehicle oblique photography technology, high-resolution surface images are first obtained through aerial photography, which covers a wide range and can comprehensively record topographic information. Then, the acquired images are preprocessed using digital image processing technology, including denoising, image enhancement, etc., to optimize image quality. Next, the image segmentation algorithm is used to analyze the preprocessed images to identify possible landslide crack areas. These algorithms can identify cracks based on image features, texture, color, etc., and at the same time, combined with deep learning technology, the system has stronger automation and intelligent capabilities.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a landslide crack identification method based on drone orthophoto and pyramid scene analysis network image segmentation algorithm, comprising the following steps: S1. Landslide data collection; S2, UAV terrain imitation flight; S3, UAV image processing; S4, deep learning preliminary image processing; S5, create training samples; S6, model and model training; S7. Crack identification.

[0007] Preferably, the landslide data in step S1 includes the landslide range and height difference.

[0008] Preferably, step S2 comprises the following steps: S21. Determine the implementation plan for the emergency investigation using remote sensing by drones based on the characteristics of the cracks on site and the surrounding environment; S22, setting the flight parameters of the UAV, including flight altitude, flight speed, camera parameters (such as exposure time, interval), and ground resolution; S23. Plan the flight route based on the terrain and target of the survey area, adopt a grid-like route to ensure full coverage of the target area, and comprehensively consider the overlap between flight paths and heading changes to obtain high-quality image data.

[0009] Preferably, step S3 comprises the following steps: S31. First, import the drone data (including POS data, image data and original landslide data) into Pix4D to establish the working survey area; S32. Add control points. Control points are points with known positions, which are used to correct the accuracy of the data collected by the drone. Control points are usually recorded when collecting data on the ground. The coordinates of known points are recorded through GPS or other positioning devices, and then these coordinates are added as control points in the data processing software. Adding control points helps to improve the precision and accuracy of data processing. Before starting data processing, it is also necessary to select appropriate parameter configurations. These parameters include camera parameters, image quality, and image registration accuracy. The selection of parameters is usually adjusted according to the characteristics of the collected data and the requirements of processing. Selecting appropriate parameters can ensure the smooth progress of the data processing process and obtain high-quality processing results.

[0010] S33, checkpoint error. Checkpoint error is the difference between the actual position of a known feature or terrain and the corresponding position obtained by processing UAV data. If the checkpoint error is large, it means that the accuracy of the processed map or model may not be high enough and the checkpoints need to be edited. Editing the checkpoints can help correct erroneous or missing points caused by factors such as insufficient image overlap, sensor errors or terrain complexity to improve the accuracy and integrity of the model. Once all the operation processes are completed, the orthophoto of the landslide can be obtained.

[0011] Preferably, the deep learning preliminary image processing in step S4 adopts three image processing operations: image cropping, sample labeling and data enhancement.

[0012] Preferably, the step S6, model and model training comprises the following steps: S61. Extract feature representation of the image through convolutional neural network, and then introduce pyramid pooling module to fuse features at different scales to capture global context information. S62, using spatial pyramid pooling to perform multi-scale spatial average pooling on the feature map to obtain semantic information at different scales; S63. Use the deconvolution layer to restore the feature map to the original image size and perform pixel-level classification to achieve accurate semantic segmentation of the image.

[0013] Preferably, in step S7, the model and the orthophoto map are imported and run to obtain the crack positions of the landslide.

[0014] The present invention provides a landslide crack identification method based on drone orthophoto and pyramid scene analysis network image segmentation algorithm, which has the advantages of high efficiency, accuracy and strong real-time performance. Compared with traditional ground surveys or monitoring methods based on satellite images, drone oblique photography technology can capture subtle changes in the surface with higher spatial resolution, and digital image processing and pattern recognition algorithms can automatically and quickly analyze data, thereby timely discovering the existence of landslide cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a flow chart of the method of the present invention; FIG2 (a) is a diagram of the encoder pooling structure of the present invention; FIG2( b ) is a schematic diagram of an encoder-encoder structure in the prior art; Figure 3 It is a model structure diagram of the present invention; Figure 4 This is the recognition result diagram of the present invention. DETAILED DESCRIPTION

[0016] like Figure 1 As shown, a landslide crack identification method based on UAV orthophoto and pyramid scene parsing network image segmentation algorithm includes the following steps: S1, obtain landslide data The early stage of photogrammetry requires a lot of data collection, collation and preparation to determine the scope of the landslide, height difference, etc. Only when the scope of the landslide and other data are determined can subsequent work be carried out.

[0017] S2, drone terrain-following flight Before conducting a UAV ground-simulation flight, an environmental investigation and assessment is required. The content is divided into an investigation of the surrounding environment of the disaster body, including topography, local meteorological conditions, distribution of high-altitude and ground facilities, visible range and visibility, selection of take-off and landing sites, and determination of flight range; investigation of crack characteristics, including the topography, plane distribution characteristics, elevation change characteristics, and scale characteristics (length, width, area, etc.) of the cracks and the affected range; investigation of implementation conditions, including GPS signal strength and stability, and the layout plan of ground image control points. Its main purpose is to determine a suitable small UAV remote sensing emergency investigation implementation plan based on the characteristics of the on-site cracks and surrounding environment. Then set the flight parameters of the UAV, including flight altitude, flight speed, camera parameters (such as exposure time, interval), ground resolution, etc. Then plan the route according to the terrain and target of the survey area. Usually, a grid route is used to ensure full coverage of the target area, and the overlap and heading changes between flight paths are comprehensively considered to obtain high-quality image data. After the parameters are set, the UAV ground-simulation flight can be carried out.

[0018] It should be noted that: on site, the route and flight range must be accurately corrected according to the actual position obtained by the drone's built-in GPS; the route coverage must be larger than the actual distribution range of the disaster body to ensure that the images within the disaster body have sufficient overlap; the starting route should be set near the foot of the disaster body, and the camera lens should always be kept vertically downward, and the flight should gradually fly upward to set the ending route near the top of the disaster body. That is, the drone should always fly from low to high during the investigation. In principle, the drone's height from the ground should be kept relatively fixed to make full use of the drone's more stable characteristics during the ascending flight; the planned route must be carefully checked and confirmed to be correct before it is imported into the flight control system for implementation.

[0019] For the selection of survey schemes, the automatic survey scheme should be adopted in most cases. This scheme uses the GPS information of the drone to enable the system to fly autonomously and take pictures automatically according to the detailed planned route. However, the manual survey scheme should be adopted in the following situations: first, in areas without GPS signals or unstable signals such as deep mountain valleys, and second, in areas with extremely small disaster bodies and good visibility. The manual survey completely relies on manual control of the flight of the aircraft and the shooting of pictures. The flight safety and picture shooting quality of this scheme are generally lower than those of the automatic method, but it is faster, more maneuverable and flexible.

[0020] S3, drone image processing Generate precision results by first encrypting the three-dimensional point cloud, and then generate remote sensing results such as digital surface models, digital orthophotos, and three-dimensional models based on this. This paper uses Pix4D software for image processing. First, the drone data (including POS data, image data, and original landslide data) is imported into Pix4D to establish a working survey area, and then control points are added. Control points are points with known positions, which are used to correct the accuracy of drone data collection. Control points are usually recorded when collecting on the ground, and the coordinates of known points are recorded by GPS or other positioning devices, and then these coordinates are added as control points in the data processing software. Adding control points helps to improve the precision and accuracy of data processing. Before starting data processing, it is also necessary to select appropriate parameter configurations. These parameters include camera parameters, image quality, image registration accuracy, etc. The selection of parameters is usually adjusted according to the characteristics of the collected data and the requirements of processing. Selecting appropriate parameters can ensure that the data processing process proceeds smoothly and obtains high-quality processing results. Then check the point error. Checking the point error is to check the difference between the actual position of the known object or terrain and the corresponding position obtained by drone data processing. If the checkpoint error is large, it means that the processed map or model may not be accurate enough and needs to be edited. Editing the checkpoints can help correct erroneous or missing points caused by factors such as insufficient image overlap, sensor error, or terrain complexity to improve the accuracy and integrity of the model. After all the operation processes are completed, the orthophoto of the landslide can be obtained.

[0021] S4, deep learning early image processing The early image processing of deep learning mainly includes three aspects: ① Image cropping: Image cropping plays a vital role in the early image processing of deep learning. Its main purpose is to extract the interesting parts of the image by removing irrelevant background or irrelevant areas, so that the model can focus on learning key features and improve the performance and generalization ability of the model. In addition, image cropping also helps to reduce the complexity of data, reduce the difficulty of model training, and improve the efficiency and accuracy of the model. Through image cropping, the noise and interference in the image can be minimized, so that the model can focus more on learning the target object or area, thereby improving the performance of the model in various visual tasks. ② Labeling sample labels: Sample labeling plays a vital role in the early image processing of deep learning. Its main purpose is to provide the model with accurate and reliable label information to help the model understand and learn the task content. The key points of sample labeling include accuracy and diversity. First of all, the accuracy of the label means that the label should accurately match the target content to ensure that the model learns the correct pattern and features. The diversity of the label can capture different aspects and features of the data set and provide more comprehensive training samples for the model. ③ Data enhancement: By introducing different transformations (such as rotation, translation, and blocking), these transformations can increase the richness and complexity of the data set. This helps to improve the model's robustness and generalization ability to data, reduce the risk of overfitting, enhance the universality and adaptability of the model, make the model better adapt to the complex situations in the real world, and improve the performance of the model in various application scenarios.

[0022] S5, create training samples Generating more diverse training samples not only helps improve the performance of the model, but also improves the robustness of the model. When facing various complex real-world scenarios and task requirements, the model often needs to have strong generalization and adaptability, which is exactly what is achieved through early image processing operations such as data enhancement and sample labeling.

[0023] S6, Model and Model Training The most advanced scene analysis framework is mainly based on the fully convolutional network (FCN). FCN is a deep learning network for semantic segmentation. Its core idea is to replace the traditional fully connected layer with a fully convolutional layer to achieve end-to-end pixel-level classification. FCN first uses a convolutional neural network to extract the feature representation of the image, and then restores the feature map to the original image size through upsampling operations, and finally obtains the pixel-level classification results. During the training process, FCN optimizes the network parameters through back propagation to minimize the loss between the network output and the true label. The advantage of FCN is that it can process input images of any size, and can simultaneously achieve pixel-level classification and retain position information, thus achieving better results in semantic segmentation tasks.

[0024] As shown in Figure 2 (a) and 2 (b), PSPNet is a semantic segmentation network based on FCN. Compared with the previous semantic segmentation model, it introduces the pyramid pooling module and the spatial pyramid pooling layer, which is a significant difference from the traditional model. The pyramid pooling module enables the network to capture global context information at different scales, so as to understand the semantic content of the image more comprehensively. This acquisition of global context information is not available in traditional models. At the same time, the spatial pyramid pooling layer realizes multi-scale feature extraction, enabling the network to adapt to target objects of different sizes and scales, and improving the robustness and generalization ability of the network. PSPNet first extracts the feature representation of the image through a convolutional neural network, and then introduces the pyramid pooling module to fuse features at different scales to capture global context information. Then, spatial pyramid pooling (SPP) is used to perform multi-scale spatial average pooling on the feature map to obtain semantic information at different scales. Finally, the deconvolution layer is used to restore the feature map to the original image size, and pixel-level classification is performed to achieve accurate semantic segmentation of the image. PSPNet can effectively solve the problems of scale inconsistency and insufficient contextual information in semantic segmentation through pyramid-style information fusion and multi-scale feature extraction, and improve the accuracy and robustness of semantic segmentation.

[0025] In addition, PSPNet also integrates ResNet (Residual network) into its structure to take advantage of ResNet's advantages in image feature extraction. ResNet is a deep residual network that solves the gradient vanishing and gradient exploding problems in the training process of deep neural networks by introducing residual connections, enabling the network to learn image features more deeply, thereby improving the performance and generalization ability of the model. In PSPNet, ResNet is used as a feature extractor to extract high-level semantic features from input images. By using a pre-trained ResNet model, PSPNet can extract rich feature representations from the input image, which contain the semantic information of the image and provide strong support for subsequent semantic segmentation tasks.

[0026] Specifically, PSPNet usually uses ResNet as the backbone network of its feature extractor. In the process of feature extraction, the feature map extracted by ResNet is used, and then the pyramid pooling module and spatial pyramid pooling layer of PSPNet are used to perform multi-scale feature fusion and capture contextual information, thereby obtaining a feature representation with more semantic information for subsequent pixel-level classification tasks.

[0027] In pixel-level classification tasks, commonly used loss functions include Cross-Entropy Loss and Dice Loss. The present invention adopts two loss functions to evaluate the performance of the combined model.

[0028] The cross entropy loss function is often used in classification tasks, especially multi-category classification. It measures the difference between the probability distribution of the model output and the true label. The formula is as follows: ; in, is the total number of pixels; is the number of categories; It is The true label of the pixel position Probability of class; The model predicts The pixel position belongs to The probability of the class.

[0029] The Dice Loss coefficient is a criterion for judging the segmentation effect. Its formula is equivalent to the intersection-over-union ratio of the predicted result area and the true area. Because Dice Loss directly uses the segmentation effect evaluation index as the loss to supervise the network, and ignores a large number of background pixels when calculating the intersection-over-union ratio, it solves the problem of imbalance between positive and negative samples, so it converges quickly. The formula is as follows: ; in, is the predicted segmentation result of the model. is the real segmentation result.

[0030] In addition to selecting a suitable loss function, the division of the data set is also crucial. Dividing the entire data set into a training set and a validation set can effectively evaluate the performance and generalization ability of the model. By training the model on the training set and then verifying the performance of the model on the validation set, the overfitting and underfitting problems of the model can be discovered in time, so that the model structure and hyperparameters can be adjusted in time to optimize the loss function until the loss is minimized.

[0031] S7, crack identification like Figure 3 As shown in the figure, the final crack identification only requires importing the model and the orthophoto map, and running it to obtain the crack location of the landslide. Finally, careful verification is carried out, and the model is optimized through comprehensive evaluation of the evaluation indicators and the identification results. Figure 4 Shown is the recognition result diagram of the present invention.

[0032] Below, taking the Hemenkou landslide as an example, the embodiment is further described, and the implementation steps are as follows: 1. Obtain landslide data The study area of ​​this embodiment is the Hemenkou landslide, which is located in Ma'anping Village, Xin'an Dai Township, Huili County, Sichuan Province. It is located on the slope of the left bank exit section of Chen River, a first-level tributary of Jinsha River. The geographical location is: 102°7′20.41″ east longitude and 26°7′14.09″ north latitude. The Hemenkou landslide is nearly east-west distributed in the left bank exit section of Chen River. It is irregular and long in plane. The rear edge is located on the bedrock wall of the slope at an elevation of about 1210m, and the front edge has extended into the riverbed of Chen River at an elevation of about 920m. The front and back height difference is nearly 300m. The overall slope shape is steep in front and gentle in the back, and the local terrain is undulating. A platform with a width of about 50m is developed at an elevation of 1065-1072m in the middle. The terrain slope above the platform is about 18°~25°, and the slope above 1175m is about 10°. The terrain is relatively gentle. The north and south sides of the landslide are bounded by gullies, which are V-shaped and generally 15 to 30 meters deep. The landslide is about 830 meters long from east to west and 250 to 350 meters wide from north to south, with a plane area of ​​25×10 4 m 2 , with a volume of about 6 million m 3 , the main sliding direction is 282°.

[0033] 2. Drone Terrain Flight The aerial photography area is from 102°7′3″ to 102°7′30″ east longitude and 26°7′5″ to 26°7′35″ north latitude, with an area of ​​about 2 km 2The aerial photography was taken on January 9, 2024, and the weather was sunny. The drone used was the DJI Matrice 350 RTK, a professional-grade six-rotor aircraft that uses an advanced RTK positioning system. It uses satellite signals and ground base station data to achieve high-precision aerial positioning, with an accuracy of up to centimeters, ensuring the stability and accuracy of the aircraft in complex environments. Equipped with a Zenmuse L2 lens, which uses advanced optical technology and high-definition sensors to capture clear, detailed, and high-resolution images and videos, it is also equipped with GNSS-assisted positioning functions and ranging and infrared sensors, which can achieve high-precision distance and height measurements of ground targets. Supports operations and post-processing requirements in a variety of mission scenarios. The focal length of the lens is 24mm, the effective pixel of the image is 20 million, and the size of the captured image is 5280 × 3956, including three bands: red, green, and blue.

[0034] 3. UAV image processing The present invention processes drone images through Pix4D and imports the image data taken by the drone into it. Pix4D will go through a series of processing steps, including image initialization processing, registration, point cloud and texture, and finally generate a 3D point cloud model. This point cloud model can be used to generate digital surface models, digital elevation models and texture maps, as well as other models based on three-dimensional information. These products can be applied to map making, terrain analysis, object extraction and other fields.

[0035] 4. Image processing before deep learning The present invention uses three image processing operations, namely image cropping, sample labeling and data enhancement, before deep learning. The landslide image data used before creating the training sample is cropped from the orthophoto of the landslide, cutting out the water area and a large area outside the landslide, aiming to focus on the area used in the study, saving memory and time, so as to conduct deep learning recognition faster and more effectively. In terms of sample labeling, the present invention appropriately labels the landslide cracks, and focuses on labeling some small cracks to prevent missed recognition. At the same time, this study also rotated the image 90° and processed it in 228*228 blocks. After the enhancement processing, the number of data sets was expanded, and the model could be trained more fully.

[0036] 5. Create training samples The early image processing operation makes the landslide images better for model input training, generating a total of 275 images with a pixel size of 228 × 228. These images constitute a feature set containing 2418 sample features. On this basis, 10% of the training samples were selected as the validation set to ensure the robustness of the model.

[0037] 6. Model and model training The pyramid pooling module of PSPNet enables the network to capture global context information at different scales, thereby more comprehensively understanding the semantic content of the image. At the same time, combined with the feature extraction capability of ResNet, it can effectively learn richer and more abstract feature representations in the image, improving the model's ability to extract landslide features and recognition accuracy. The training samples generated in the previous step were imported. After 20 iterations, the model was continuously optimized and gradually stabilized.

[0038] 7. Crack identification The trained PSPNet was applied to the entire landslide orthophoto map to obtain the final classification map, whose f1 score reached 0.648 and IOU index reached 0.48. From the classification map, it can be verified that the model of the present invention has a high accuracy in landslide crack identification.

[0039] To further illustrate the advantages of the present invention, the present invention is also compared with Deeplabv3 and Unet models which are widely used in the field of image segmentation. Precision, recall, accuracy (ACC), f1 score, and intersection over union (IOU) are selected as performance evaluation indicators of the recognition results of the combined model. The indicator calculation formula is as follows: ; ; ; ; ; TP, TN, FP, and FN refer to True Positives, True Negatives, False Positives, and False Negatives, respectively. The evaluation indicators of the three models are shown in Table 1.

[0040] Table 1 Evaluation indicators of three models

[0041] It can be seen from the comparative study of the present invention with Deeplabv3 and Unet models that each prediction evaluation index shows that the recognition accuracy of the model of the present invention is higher than that of other models.

[0042] The above example results show that PSPNet, a landslide crack identification method based on drone orthophotos and pyramid scene parsing network image segmentation algorithm, has higher accuracy than other image segmentation models.

[0043] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A landslide crack identification method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm, characterized in that: The following steps are involved: S1. Landslide data collection; S2, UAV terrain imitation flight; S3, UAV image processing; S4, deep learning preliminary image processing; S5, create training samples; S6, model and model training; S7. Crack identification.

2. A landslide crack identification method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm according to claim 1, characterized in that: The landslide data in step S1 include the landslide range and height difference.

3. A landslide crack identification method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm according to claim 1, characterized in that: The step S2 comprises the following steps: S21. Determine the implementation plan for the emergency investigation using remote sensing by drones based on the characteristics of the cracks on site and the surrounding environment; S22, setting the flight parameters of the UAV, including flight altitude, flight speed, camera parameters, and ground resolution; S23. Plan the flight route based on the terrain and target of the survey area, adopt a grid-like route to ensure full coverage of the target area, and comprehensively consider the overlap between flight paths and heading changes to obtain high-quality image data.

4. A landslide crack identification method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm according to claim 1, characterized in that: The step S3 comprises the following steps: S31. First, import the drone data into Pix4D and establish the working survey area; S32, add control points, which are points with known positions and are used to correct the accuracy of the data collected by the drone. The control points are usually recorded when collecting data on the ground, by using GPS or other positioning devices to record the coordinates of the known points, and then adding these coordinates as control points in the data processing software. Adding control points helps to improve the precision and accuracy of data processing. Before starting data processing, it is also necessary to select appropriate parameter configurations. These parameters include camera parameters, image quality, and image registration accuracy. S33, checkpoint error. Checkpoint error is the difference between the actual position of a known feature or terrain and the corresponding position obtained by processing UAV data. If the checkpoint error is large, it means that the accuracy of the processed map or model may not be high enough and the checkpoints need to be edited. Editing the checkpoints can help correct erroneous or missing points caused by factors such as insufficient image overlap, sensor errors or terrain complexity to improve the accuracy and integrity of the model. Once all the operation processes are completed, the orthophoto of the landslide can be obtained.

5. A landslide crack identification method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm according to claim 1, characterized in that: The step S4 deep learning preliminary image processing uses three image processing operations: image cropping, sample labeling, and data enhancement.

6. A landslide crack identification method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm according to claim 1, characterized in that: The step S6, model and model training comprises the following steps: S61. Extract feature representation of the image through convolutional neural network, and then introduce pyramid pooling module to fuse features at different scales to capture global context information. S62, using spatial pyramid pooling to perform multi-scale spatial average pooling on the feature map to obtain semantic information at different scales; S63. Use the deconvolution layer to restore the feature map to the original image size and perform pixel-level classification to achieve accurate semantic segmentation of the image.

7. A landslide crack identification method based on unmanned aerial vehicle orthophoto and pyramid scene analysis network image segmentation algorithm according to claim 1, characterized in that: In step S7, the model and the orthophoto map are imported and run to obtain the crack positions of the landslide.

Citation Information

Patent Citations

  • Surface mining-induced crack unmanned aerial vehicle image recognition method based on deep learning algorithm

    CN116343071A

  • Method for identifying potential landslide hazard of reservoir bank based on rock mass degradation feature

    US20220383475A1

Cited By

  • Digital twin terrain rapid construction method based on oblique photography

    CN119339021A

  • Loess landslide crack evolution trend map construction method and device

    CN120106195A

  • Ragweed identification method and device based on unmanned aerial vehicle remote sensing technology and deep learning

    CN120356127A

  • Crack quantitative identification method and system fusing photogrammetry and semantic key points

    CN121883571A

  • Crack quantitative identification method and system fusing photogrammetry and semantic key points

    CN121883571B