Landslide disaster-bearing body identification method fusing unmanned aerial vehicle DOM and residual Unet transfer learning algorithm

Abstract

The invention relates to a landslide disaster-bearing body identification method fusing an unmanned aerial vehicle DOM and a residual Unet transfer learning algorithm, and the method comprises the steps: obtaining landslide data, controlling an unmanned aerial vehicle to carry out the simulated flight according to the landslide data, collecting image data, processing the image data, and obtaining a landslide orthoimage map; inputting the processed landslide orthoimage into a Res-Unet model to obtain a disaster-bearing body identification result; wherein the Res-Unet model is obtained by taking a pre-trained ResNet-34 network as a backbone network to replace an encoder structure of the Unet model, and updating parameters through training iteration by utilizing a training set; in the training process of the training set, the ResNet-34 network serves as a backbone network to carry out cross-layer transmission on training samples in the training set and add the training samples with a convolution result so as to consider the Res-Unet model with more features. According to the invention, quicker and more accurate target area monitoring and data acquisition can be realized, and the method is suitable for application scenes requiring high precision and high real-time requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide hazard-prone body identification, and in particular to a landslide hazard-prone body identification method integrating unmanned aerial vehicle DOM and residual Unet transfer learning algorithm. Background Art

[0002] Among many geological disasters, landslides cause the most serious casualties and property losses. Landslides may cause large-scale movement of land, posing dangers to buildings, roads and people along the way. They may also block rivers and lakes and form barrier lakes.

[0003] Therefore, from the perspective of disaster prevention and mitigation, in addition to evaluating the susceptibility of landslides and preventing disasters from occurring, accurately and timely identifying landslide-affected bodies is of great significance for reducing the risk of landslide disasters and protecting people's lives and property. Traditional landslide-affected body identification methods mainly rely on manual interpretation of remote sensing images combined with field surveys, which have many problems. First, the image resolution may not be sufficient to capture the required detailed information, especially in complex terrain or densely vegetated areas. Second, factors such as noise, shadows, or occlusions in the image will reduce the accuracy of the interpretation. In addition, the interpretation process is affected by the subjective judgment and experience of the interpreter, which may lead to differences in results between different interpreters. In addition, for complex objects or scenes, multi-source data and multiple technologies may be combined, which increases the difficulty and complexity of interpretation. Most importantly, manual interpretation and field surveys require a lot of time and manpower costs, and there are problems of low efficiency and high cost. The development of deep learning technology in the field of computer vision provides a new solution for landslide-affected body identification. By using deep learning semantic segmentation networks, the boundary information of landslide-affected bodies can be automatically and efficiently extracted from remote sensing images, and accurate identification of landslide-affected bodies can be achieved, providing important support for disaster risk assessment, post-disaster rescue and other work. As a result, methods based on convolutional neural networks (CNNs) were quickly introduced into image segmentation. CNNs can automatically learn feature information in images, such as edges, textures, and colors, through multi-layer convolution operations. As the number of network layers increases, higher-level features are gradually abstracted to achieve a deeper understanding of the image. CNNs combine components such as convolutional layers, pooling layers, and activation functions, which can effectively reduce the number of parameters and have translation invariance, making them insensitive to changes in the target position in the image. In addition, some CNN-based deep learning models such as U-Net, SegNet, and FCN have been optimized for image segmentation tasks, retaining spatial information while efficiently processing input images of different sizes, and have achieved remarkable results in various image segmentation tasks.

[0004] However, these methods still have limitations. Generally speaking, it is difficult to achieve satisfactory performance with a small amount of training data, especially in complex tasks and fields, where generalization and learning require more data to improve performance. Training and using CNNs require a lot of computing resources, especially in network structures with high depth and complexity. This can be a challenge for resource-constrained environments. And because the ground resolution of satellite images is low, it will affect the effect of image recognition, and the update speed is slow, making it impossible to flexibly conduct real-time observation and monitoring of specific areas. In addition, the model used for landslide areas needs to have good robustness, be able to accurately identify landslide-affected bodies under different geographical environments and imaging conditions, and have a certain tolerance for interference factors such as light and shadow. Summary of the invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to propose a landslide hazard identification method that integrates drone DOM and residual Unet transfer learning algorithm, conducts fast and effective measurement based on the landslide range, and uses a small drone remote sensing system to automatically collect landslide images. Through digital image processing technology, high-precision and high-resolution remote sensing results of the slope are generated. At the same time, transfer learning is used to enhance the feature extraction of the image segmentation model, which can accelerate the training of the model and improve the model performance.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A landslide hazard identification method integrating UAV DOM and residual Unet transfer learning algorithm, including:

[0008] Acquire landslide data, control the UAV to fly imitation terrain according to the landslide data to collect image data, process the image data, and acquire an orthophoto of the landslide;

[0009] Input the processed orthophoto of the landslide into the Res-Unet model to obtain the hazard-affected body identification result; wherein the Res-Unet model uses the pre-trained ResNet-34 network as the backbone network to replace the encoder structure of the Unet model, and uses the training set to iterate the training to update the parameters;

[0010] During the training of the training set, the ResNet-34 network is used as a backbone network to perform cross-layer transfer on the training samples in the training set and add the results of convolution to consider a Res-Unet model with more features.

[0011] Optionally, the landslide data include: landslide range, terrain height difference and surface characteristics.

[0012] Optionally, controlling the unmanned aerial vehicle to perform terrain-simulating flight and collect image data includes:

[0013] The flight parameters of the UAV are set according to the target environment and target characteristics; the flight parameters include: flight altitude, speed and camera parameters; the target characteristics include: terrain, height difference and vegetation;

[0014] The flight route is planned according to the target terrain and target position, and then image data is collected through terrain simulation flight; the image data includes: multiple images of the entire landslide range and POS data recording the geographical location, posture and other positioning auxiliary information of the image.

[0015] Optionally, it is characterized in that controlling the UAV to fly in a terrain-simulating manner to collect image data further comprises: manually controlling the UAV to collect image data in an area where there is no GPS signal or the signal is unstable.

[0016] Optionally, obtaining the landslide orthophoto includes:

[0017] Add control points in the target area, and select parameter configuration according to the processing requirements of the image data; the parameter configuration includes: camera parameters, image quality and image registration accuracy;

[0018] Based on the parameter configuration, feature matching and geometric calibration are performed on overlapping images by combining the image data with the POS data to generate sparse and dense three-dimensional point clouds, and a digital surface model is created to geometrically correct the image, and the geometrically corrected image is reprojected onto the surface to eliminate deformation, and an orthophoto map aligned in a geographic coordinate system is generated;

[0019] Check point errors by making an error judgment between the position of the control point and the corresponding position of the image processing result. If the accuracy requirement is not met, edit the puncture points, correct the image processing result, and obtain the orthophoto of the landslide. If the accuracy requirement is met, use the image processing result as the orthophoto of the landslide.

[0020] Optionally, it is characterized in that obtaining the Res-Unet model includes:

[0021] Performing image cropping, sample labeling, and data enhancement on the historical landslide orthophoto map to obtain the training set;

[0022] The training set is used to train an initial Res-Unet model, hyperparameters of the initial Res-Unet model are adjusted, and the Res-Unet model after the hyperparameters are adjusted is evaluated based on a cross entropy loss function and a Dice loss function to obtain the Res-Unet model.

[0023] Optionally, the method for evaluating the Res-Unet model after adjusting the hyperparameters based on the cross entropy loss function is:

[0024]

[0025] Where N is the total number of pixels; C is the number of categories; q i,c is the probability of the cth class in the true label of the i-th pixel position; p i,c The probability that the i-th pixel position belongs to the c-th class is predicted by the model.

[0026] Optionally, the method for evaluating the Res-Unet model after adjusting the hyperparameters based on the Dice loss function is:

[0027]

[0028] Among them, X is the predicted segmentation result of the model, and Y is the actual segmentation result.

[0029] Optionally, the ResNet-34 network includes: a plurality of residual blocks;

[0030] Among them, the form of each residual block is expressed as:

[0031]

[0032] x l+1 =f(y l )

[0033] Among them, x l and x l+1 is the input and output of the lth residual unit, is the residual function, f(y l ) is the activation function, h(x l ) is the identity mapping function.

[0034] The beneficial effects of the present invention are:

[0035] The present invention utilizes the learned knowledge or experience to improve the learning effect of new tasks. By migrating the model or feature knowledge learned in the original field to the target field, it can improve the model performance when data is scarce or similar but not in the same field, accelerate the model training process, improve the generalization ability, and share knowledge between different fields, thereby providing more effective solutions to new problems and significantly improving the accuracy of landslide hazard identification.

[0036] Compared with satellite-based images, the images extracted by drones in the present invention have higher resolution and flexibility, can capture richer detailed information, and are not affected by weather and clouds. They can achieve faster and more accurate target area monitoring and data collection, and are suitable for application scenarios that require high precision and real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 It is a schematic diagram of a landslide hazard-bearing body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to an embodiment of the present invention;

[0039] Figure 2 A schematic diagram showing the difference between the residual structure used in an embodiment of the present invention and a common neural unit;

[0040] Figure 3 A schematic diagram of the model structure of the combined model of an embodiment of the present invention;

[0041] Figure 4 A schematic diagram of the recognition result of the combined model according to an embodiment of the present invention;

[0042] Figure 5 Schematic diagram of the loss function and accuracy change of the combined model during the training process of an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1As shown, this embodiment discloses a method for identifying landslide hazard-prone bodies by integrating UAV DOM and residual Unet transfer learning algorithm, including: acquiring landslide data, controlling the UAV to fly imitation according to the landslide data to collect image data, processing the image data, and acquiring an orthophoto of the landslide; inputting the processed orthophoto of the landslide into the Res-Unet model to obtain a hazard-prone body identification result; wherein the Res-Unet model uses the pre-trained ResNet-34 network as the backbone network to replace the encoder structure of the Unet model, and obtains the Res-Unet model by iteratively updating parameters through training using the training set; during the training of the training set, the ResNet-34 network is used as the backbone network to perform cross-layer transfer on the training samples in the training set and add the results of the convolution to consider the Res-Unet model with more features.

[0046] Specifically, based on the scope of the landslide, rapid and effective measurements are carried out, and landslide images are automatically collected using a small UAV remote sensing system, and high-precision and high-resolution remote sensing results of the slope are generated through digital image processing technology. At the same time, the feature extraction of the image segmentation model is enhanced by transfer learning, which can accelerate the training of the model, improve the model performance, and use the learned knowledge or experience to improve the learning effect of the new task; by transferring the model or feature knowledge learned in the original field to the target field, it can improve the model performance in the case of data scarcity or similar but not the same field, accelerate the model training process, improve the generalization ability, and share knowledge between different fields, so as to provide more effective solutions to new problems; in terms of image acquisition, compared with satellite-extracted images, UAV-extracted images have higher resolution and flexibility, can capture richer details, and are not affected by weather and clouds, which can achieve faster and more accurate target area monitoring and data collection, and are suitable for application scenarios that require high precision and high real-time requirements.

[0047] Furthermore, the landslide data include: landslide range, terrain height difference and surface characteristics.

[0048] Specifically, obtain landslide data:

[0049] In the early stage of photogrammetry, it is necessary to collect data on terrain, landforms, vegetation, land use, etc., and to organize and analyze them in detail. These data are not limited to ground observations, but may also include satellite images, aerial photography, and existing map data. Through the comprehensive analysis of these data, key information such as the scope of landslides, terrain height differences, and surface features can be accurately determined, providing a reliable basis for subsequent work.

[0050] Furthermore, controlling the UAV to fly in a terrain-simulating manner to collect image data includes: setting the flight parameters of the UAV according to the target environment and target characteristics; the flight parameters include: flight altitude, speed and camera parameters; the target characteristics include: terrain, height difference and vegetation; planning the flight route according to the target terrain and target position, and then collecting image data through terrain-simulating flight; the image data includes: multiple images of the entire landslide range and POS data that records the geographical location, posture and other positioning auxiliary information of the image.

[0051] Furthermore, controlling the unmanned aerial vehicle to fly in a terrain-like manner to collect image data also includes: manually controlling the unmanned aerial vehicle to fly in a terrain-like manner to collect image data in an area where there is no GPS signal or the signal is unstable.

[0052] Specifically, in this embodiment, the drone imitates the terrain to fly:

[0053] Before preparing for a drone flight, a detailed environmental assessment and investigation is required. This process includes investigating the surrounding environment of the disaster body, including topography, meteorological conditions, distribution of ground facilities, etc., and selecting suitable take-off and landing locations and determining the flight range. At the same time, it is also necessary to have a deep understanding of the characteristics of the disaster-prone body, including topography, plane distribution, elevation changes, scale and other aspects, in order to formulate an effective investigation plan. During the implementation condition investigation stage, it is necessary to evaluate the strength and stability of the GPS signal to ensure the accuracy and reliability of the drone's flight and data collection. Next, according to the characteristics of the environment and the target, the flight parameters of the drone are set, including flight altitude, speed, camera parameters, etc., to ensure the acquisition of high-quality image data. Subsequently, according to the terrain and target of the survey area, the flight route is planned. Usually, a grid route is used to ensure full coverage of the target area, and the overlap between routes and heading changes are taken into account to obtain more comprehensive and accurate data. Once the parameters are set, the drone can perform a simulated flight, execute the predetermined route and collection tasks, and provide important data support for subsequent disaster investigations.

[0054] At the same time, during field operations, it is crucial to accurately calibrate the route and flight range based on the accurate position obtained by the drone's own GPS. Ensure that the route coverage exceeds the actual distribution range of the disaster body, which can ensure that the images have sufficient overlap, thereby improving the accuracy and reliability of subsequent data processing and analysis. The setting of the starting route should take into account the characteristics of the disaster body, usually starting near the foot of the slope and gradually flying up along the terrain to near the top of the slope. Throughout the survey process, maintaining the drone's flight path from low to high helps to make full use of the drone's stability during the ascending flight and ensure the collection of high-quality image data. At the same time, ensure that the camera lens is always kept vertically downward, which can ensure the accuracy and integrity of the surface image. After planning the route, be sure to check it carefully to confirm that it is correct and import it into the flight control system to make it effective. This ensures that the drone can perform flight missions according to the predetermined route and parameters, providing reliable data support for disaster investigation and emergency response.

[0055] In practice, automatic control surveys are usually the preferred option. By utilizing the GPS information carried by the drone, the system can autonomously fly and shoot according to a precisely planned route, thereby ensuring the consistency and accuracy of the data. However, in some special cases, manual control surveys are more appropriate. For example, in deep mountain canyon areas where there is no GPS signal or the signal is unstable, or in areas where the disaster area is small and the visibility is good, manual control surveys provide a more flexible and faster way. Manually controlled aircraft can be adjusted according to the actual situation on site, making the survey more accurate and efficient. Although the flight safety and shooting quality of manual control surveys may be relatively low, its fast, maneuverable and flexible characteristics make it a suitable choice in specific situations.

[0056] Furthermore, obtaining the orthophoto of the landslide includes: adding control points in the target area, and selecting parameter configuration according to the processing requirements of the image data; the parameter configuration includes: camera parameters, image quality and image registration accuracy; based on the parameter configuration, feature matching and geometric calibration of overlapping images are performed by combining image data with POS data to generate sparse and dense three-dimensional point clouds, and a digital surface model is created to geometrically correct the image, and the geometrically corrected image is reprojected onto the surface to eliminate deformation, and an orthophoto aligned in the geographic coordinate system is generated; checking point errors, by making error judgments between the positions in the control points and the corresponding positions of the image processing results, if the accuracy requirements are not met, editing the punctum points, correcting the image processing results, and obtaining the orthophoto of the landslide, if the accuracy requirements are met, the image processing results are used as the orthophoto of the landslide.

[0057] Specifically, in this embodiment, the drone image processing:

[0058] To generate high-precision remote sensing results, first use Pix4D software for image processing, and import the drone data including POS data, image data and original landslide data into the software. After importing the data, the present embodiment establishes a working survey area and adds control points, which are points with known positions and are used to correct the accuracy of the data. Usually, these control points are recorded by GPS or other positioning devices during ground acquisition, and then they are added as control points to improve the accuracy and precision of data processing. Before starting data processing, the present embodiment needs to select appropriate parameter configurations, such as coordinate system, image overlap and texture mapping. The selection of these parameters is usually adjusted according to the characteristics of the collected data and the requirements of processing to ensure the smooth progress of the data processing process and obtain high-quality processing results. Then, by performing feature matching and geometric calibration on the overlapping images, sparse and dense three-dimensional point clouds are generated, and then a digital surface model (DSM) is created from it, and then the image is geometrically corrected and reprojected onto the surface to eliminate deformation, and finally an orthophoto map accurately aligned in the geographic coordinate system is generated. Subsequently, the present embodiment needs to check the point error to evaluate the accuracy of the processing results. Checkpoint error refers to the difference between the actual position of a known feature or terrain and the corresponding position obtained by drone data processing. If the checkpoint error is large, it is necessary to edit the puncture points to correct possible erroneous or missing points, especially when generating three-dimensional models and making maps. The purpose of editing puncture points is to improve the accuracy and integrity of the model and help correct errors caused by factors such as insufficient image overlap, sensor errors, or terrain complexity. After completing all the operation processes, this embodiment can obtain the orthophoto map of the landslide, providing a reliable data basis for further analysis and research.

[0059] Among them, POS data is generated by the drone and is in the same folder as the image data. It mainly records the geographic location, posture and other positioning auxiliary information of the image.

[0060] Obtaining the Res-Unet model includes: performing image cropping, sample labeling, and data augmentation on the historical landslide orthophoto to obtain a training set; using the training set to train the replaced Unet model, adjusting the hyperparameters of the replaced Unet model, and evaluating the replaced Unet model based on the cross entropy loss function and the Dice loss function to obtain the Res-Unet model.

[0061] Specifically, in this embodiment, deep learning early image processing:

[0062] The preliminary image processing of deep learning mainly includes three aspects: ① Image cropping: Orthophotos of landslides usually cover a large geographical area, but this embodiment may only be interested in a specific area therein. By cropping the image, the amount of data can be reduced, the processing efficiency can be improved, and only the area of ​​interest in this embodiment can be retained to facilitate subsequent analysis and application. ② Labeling sample labels: Data labeling can clearly specify the label or attribute of each sample, thereby providing an accurate supervision signal for the model. The quality of data labeling directly affects the performance of the model, affecting its precision, recall rate and overall predictive ability. During the labeling process, special attention should be paid to the accuracy of the labeling to avoid misleading the model training. At the same time, data balance is one of the key factors. It is necessary to ensure that the number of samples in each category is relatively balanced to avoid the model being biased towards a certain category and affecting its performance. ③ Data enhancement: By introducing different transformations (such as rotation, translation, blocking, etc.), the model can be made more robust and better adapt to data under different angles, scales and lighting conditions. At the same time, data enhancement can increase the diversity of training data, thereby reducing the overfitting tendency of the model and improving the generalization ability of the model. Increasing the diversity and number of data samples can improve the training effect of the model and improve the performance of the model in landslide detection or identification tasks.

[0063] Create training samples:

[0064] After the initial image processing operations, the data set has been fully expanded and more diverse training samples have been generated. These are designed to optimize the organizational structure and features of the data to improve the performance and effect of the machine learning model, so that it can better cope with various complex real-world scenarios and task requirements. In addition, the entire data set is divided into training sets and validation sets to complete different tasks.

[0065] Model and model training:

[0066] Convolutional neural networks (CNNs) play an important role in the field of deep learning. Through multiple layers of convolution and pooling operations, they gradually extract features from images, texts, or other data, thereby achieving efficient representation and processing of complex data. However, CNNs also have some disadvantages, such as the high dependence of CNNs on the amount of labeled data. Especially when dealing with high-dimensional or complex tasks, CNNs require large-scale labeled data to train the model to ensure its generalization ability and performance. This means that in some application scenarios, obtaining enough labeled data may become a challenge, especially for emerging or field-specific tasks. When the amount of training data is insufficient or the model complexity is high, overfitting problems are also prone to occur.

[0067] Based on this, this embodiment improves the CNN model through the characteristics of transfer learning. Figure 2As shown in the figure, ResNet (Residual network) is a deep residual network structure proposed by Microsoft Research Asia. It is widely used in computer vision tasks such as image classification and target detection, and has achieved remarkable results in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). The reason why ResNet has attracted widespread attention and application is that it solves the problems of gradient vanishing and gradient explosion in the training process of deep neural networks, making it possible to train deeper networks and obtain better performance. The core idea of ​​ResNet is to introduce the concept of residual learning and construct deep networks through residual blocks. The structure of the residual block contains a skip connection, also called "skip connection" or "shortcut connection", which allows the input signal to bypass some convolutional layers and be added to subsequent layers, thus forming a residual learning mechanism. The general form of each residual block can be expressed as

[0068]

[0069] x l+1 =f(y l )

[0070] where x l and x l+1 is the input and output of the lth residual unit, is the residual function, f(y l ) is the activation function, h(x l ) is the identity mapping function, typically h(x l )=x l This design allows the model to more easily learn the residual, that is, learn the residual added to the identity mapping.

[0071] like Figure 3-4As shown, the CNN model used in this embodiment is the classic Unet model. Unet is a classic and widely used model in the field of image segmentation. It was originally proposed by Ronneberger et al. in 2015. Its unique U-shaped network structure and the jump connection between the encoder and the decoder make it perform well in image segmentation tasks. By replacing the encoder structure of Unet with the ResNet-34 network pre-trained on a large-scale image dataset (such as ImageNet) as the backbone network to achieve the combination of transfer learning and deep learning, this method can better capture the semantic information of the image, thereby improving the performance of semantic segmentation. Secondly, ResNet-34 uses residual connections, which helps to alleviate the gradient disappearance problem and makes it easier to propagate gradients during training. In this way, the training process is more stable and the model is more likely to converge to a good state, thereby improving the training efficiency and the performance of the model. In addition, compared with the original encoder structure of Unet, ResNet-34 has fewer parameters. This means that less computing resources are required during training and reasoning, while still being able to achieve higher performance. Finally, since ResNet-34 is pre-trained on a large-scale dataset, the pre-trained weights can be used for transfer learning. This means that even if trained on a smaller dataset, better generalization performance can be achieved, thus reducing dependence on large datasets.

[0072] like Figure 5 As shown, in pixel-level classification tasks, commonly used loss functions include cross-entropy loss function (Cross-Entropy Loss) and Dice Loss. This embodiment adopts two loss functions to evaluate the performance of the combined model.

[0073] Cross-Entropy Loss represents the distance between two distributions, so it can be used to characterize the difference between the predicted value and the label value. The formula is as follows:

[0074]

[0075] Where N is the total number of pixels; C is the number of categories; q i,c is the probability of the cth class in the true label at the i-th pixel position; p i,c is the probability that the i-th pixel position belongs to the c-th class predicted by the model.

[0076] Dice Loss is often used in image segmentation tasks, especially medical image segmentation. It measures the loss by calculating the similarity between the segmentation results predicted by the model and the actual segmentation results. The formula is as follows:

[0077]

[0078] Among them, X is the predicted segmentation result of the model, and Y is the actual segmentation result.

[0079] In addition to the selection of the loss function, hyperparameters also need to be adjusted during the model training process. Hyperparameters include learning rate, batch size, number of iterations, etc. The learning rate determines the adjustment range of the model parameters during the training process, the batch size affects the number of samples used to update the model parameters at each iteration, and the number of iterations specifies the number of training rounds. The selection of these hyperparameters needs to be adjusted through experiments and verification to obtain the best model performance. The model also needs to be verified during the training process to evaluate its performance. Cross-validation or retention of the validation set can be used to evaluate the performance of the model on unseen data, and the model parameters can be adjusted based on the verification results to improve the generalization ability of the model.

[0080] Specifically, in this embodiment, the disaster-affected body is identified as follows:

[0081] The final hazard-prone body identification only requires applying the trained model to the entire landslide image. The hazard-prone body identification results output by the model are then verified and tuned. This can be done manually or by comparing with existing geographic information. The accuracy of the identification results and error analysis can be performed to continuously improve and optimize the performance of the model.

[0082] This embodiment discloses a landslide hazard-bearing body identification method integrating UAV DOM and residual Unet transfer learning algorithm, including:

[0083] 1. Obtain landslide data:

[0084] The study area of ​​this example is located on the south bank of the Yangtze River, about 56 km east of the Three Gorges Dam. Based on detailed survey data such as engineering geological mapping, drilling, trenching, sampling and analysis, it is believed that the study area is an ancient landslide that failed on different scales about 1.1 million years ago and 200,000 to 300,000 years ago. The total landslide area is about 2.2 million m 2 , 40 to 140 m thick, with a total volume of about 1.25×10 8 m 3 The landslide is a bedrock landslide with a rear edge elevation of 450m, forming a spectacular flat and smooth landslide wall, and a front edge elevation of 75m. The rock mass is strongly bent and deformed to form a reverse slope. Due to the obvious reverse tilt characteristics of the front edge shear zone, the front edge is wrinkled and curved. The study area includes two relatively independent secondary landslide bodies.

[0085] 2. Drone imitation flight:

[0086] The aerial photography area is from 110°29′20″ to 110°30′50″ east longitude and 31°1′30″ to 32°2′15″ north latitude, with an area of ​​about 2.2 million m 2The aerial photography time is from July 16 to July 18, 2023, and the weather is sunny. The drone used is the DJI Phantom 4RTK, which is a small multi-rotor high-precision aerial survey drone launched by DJI for low-altitude photogrammetry applications. It is equipped with a high-definition aerial survey camera and a GNSS positioning system with centimeter-level positioning accuracy, providing accurate aerial survey image output and supporting operations and post-processing requirements in various mission scenarios. The lens focal length is 24mm, the image effective pixels are 20 million, and the collected images include three bands: red, green, and blue.

[0087] 3. UAV image processing:

[0088] This embodiment uses Pix4D to process drone images. To use Pix4D to generate orthophotos, first import the collected aerial image data, then select the projection method and coordinate system, perform image correction and registration, then set the output parameters and resolution, run the processing flow to generate orthophotos, and finally check the results and make necessary adjustments and optimizations to realize the application of object extraction.

[0089] 4. Image processing before deep learning:

[0090] This embodiment uses a large number of image processing operations before deep learning, mainly image cropping, sample labeling and data enhancement. The landslide image data used before creating the training sample is cropped from the orthophoto of the landslide, and the water part and the blurred part of the boundary image are cropped, aiming to focus on the area used in the study, save memory and time, and thus perform deep learning recognition faster and more effectively. In terms of sample labeling, the present invention labels different groups of disaster-prone objects and disaster-prone objects of different colors, and focuses on accurately dividing the boundaries of the disaster-prone objects. At the same time, this study also enhances the image, mainly including 90° rotation and 256*256 block processing.

[0091] 5. Create training samples:

[0092] The training data finally exported for deep learning is divided into blocks from the cropped image data. After data enhancement techniques such as translation and rotation, a total of 1249 images with a pixel size of 256*256 are formed, generating a feature set of a total of 1458 sample features. In order to enhance the robustness of the model, 10% of the training data is also selected as a validation set.

[0093] 6. Model and model training:

[0094] The concept of transfer learning is introduced into the model. The combined Res-Unet model can more effectively extract features, perform transfer learning, accelerate training, and reduce the risk of overfitting. The Res-Unet model can fully utilize the advantages of the individual models, thereby improving the performance and generalization ability of the network in various tasks. The batch size of the model training is 64 images / time, and the total iteration is 20 rounds.

[0095] 7. Disaster-prone body identification:

[0096] The data set is input, and the trained model is applied to the entire image to obtain the final classification map. The f1 score of the model reaches 0.9531, and the IOU index reaches 0.9104. From the classification map, it can be verified that the combined model of this embodiment has a high accuracy in identifying landslide hazard-bearing bodies.

[0097] To further illustrate the advantages of this embodiment, this embodiment is also compared with Deeplabv3 and PSP 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 combined model recognition results. The indicator calculation formula is as follows:

[0098]

[0099] Among them, 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.

[0100] Table 1 Evaluation indicators of three models

[0101]

[0102] It can be seen from the comparative study of this embodiment with Deeplabv3 and PSP models that each prediction evaluation index shows that the recognition accuracy of the migration model of this embodiment is higher than that of other single models.

[0103] The above example results show that a landslide hazard identification method that integrates UAV DOM and residual Unet transfer learning algorithm has higher accuracy and relatively high generalization ability compared with a single image segmentation model.

[0104] The above embodiments are only descriptions of the preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.

Claims

1. A landslide hazard identification method integrating UAV DOM and residual Unet transfer learning algorithm, characterized in that: include: Acquire landslide data, control the UAV to fly imitation terrain according to the landslide data to collect image data, process the image data, and acquire an orthophoto of the landslide; Input the processed orthophoto of the landslide into the Res-Unet model to obtain the hazard-affected body identification result; wherein the Res-Unet model uses the pre-trained ResNet-34 network as the backbone network to replace the encoder structure of the Unet model, and uses the training set to iterate the training to update the parameters; During the training of the training set, the ResNet-34 network is used as a backbone network to perform cross-layer transfer on the training samples in the training set and add the results of convolution to consider a Res-Unet model with more features.

2. The landslide hazard-prone body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 1 is characterized in that: The landslide data include: landslide range, terrain height difference and surface characteristics.

3. The landslide hazard-prone body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 1 is characterized in that: Controlling the UAV to fly in a terrain-like manner and collect image data includes: The flight parameters of the UAV are set according to the target environment and target characteristics; the flight parameters include: flight altitude, speed and camera parameters; the target characteristics include: terrain, height difference and vegetation; The flight route is planned according to the target terrain and target position, and then image data is collected through terrain simulation flight; the image data includes: multiple images of the entire landslide range and POS data recording the geographical location, posture and other positioning auxiliary information of the image.

4. The landslide hazard-prone body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 1 is characterized in that: Controlling the UAV to fly in a terrestrial manner to collect image data also includes: manually controlling the UAV to collect image data in an area where there is no GPS signal or the signal is unstable.

5. The landslide hazard-prone body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 3 is characterized in that: Acquiring the orthophoto of the landslide includes: Add control points in the target area, and select parameter configuration according to the processing requirements of the image data; the parameter configuration includes: camera parameters, image quality and image registration accuracy; Based on the parameter configuration, feature matching and geometric calibration are performed on overlapping images by combining the image data with the POS data to generate sparse and dense three-dimensional point clouds, and a digital surface model is created to geometrically correct the image, and the geometrically corrected image is reprojected onto the surface to eliminate deformation, and an orthophoto map aligned in a geographic coordinate system is generated; Check point errors by making an error judgment between the position of the control point and the corresponding position of the image processing result. If the accuracy requirement is not met, edit the puncture points, correct the image processing result, and obtain the orthophoto of the landslide. If the accuracy requirement is met, use the image processing result as the orthophoto of the landslide.

6. The landslide hazard-prone body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 1 is characterized in that: Obtaining the Res-Unet model includes: Performing image cropping, sample labeling, and data enhancement on the historical landslide orthophoto map to obtain the training set; The training set is used to train an initial Res-Unet model, hyperparameters of the initial Res-Unet model are adjusted, and the Res-Unet model after the hyperparameters are adjusted is evaluated based on a cross entropy loss function and a Dice loss function to obtain the Res-Unet model.

7. The landslide hazard-prone body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 6 is characterized in that: The method for evaluating the Res-Unet model after adjusting the hyperparameters based on the cross entropy loss function is: Where N is the total number of pixels; C is the number of categories; q i,c is the probability of the cth class in the true label of the i-th pixel position; p i,c The probability that the i-th pixel position belongs to the c-th class is predicted by the model.

8. The landslide hazard-bearing body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 6 is characterized in that: The method for evaluating the Res-Unet model after adjusting the hyperparameters based on the Dice loss function is: Among them, X is the predicted segmentation result of the model, and Y is the actual segmentation result.

9. The landslide hazard-prone body identification method integrating UAV DOM and residual Unet transfer learning algorithm according to claim 1 is characterized in that: The ResNet-34 network includes: a number of residual blocks; Among them, the form of each residual block is expressed as: x l+1 =f(y l ) Among them, x l and x l+1 is the input and output of the lth residual unit, is the residual function, f(y l ) is the activation function, h(x l ) is the identity mapping function.

Images