Group-occurring landslide disaster volume rapid extraction method based on low-altitude unmanned aerial vehicle remote sensing

Abstract

The invention relates to the technical field of mass-occurring landslide disaster analysis, in particular to a mass-occurring landslide disaster volume rapid extraction method based on low-altitude unmanned aerial vehicle remote sensing. Comprising the following steps: firstly, proposing a mass-occurring landslide recognition model to perform mass-occurring landslide recognition on a target mountain, and learning landslide features of different scales through a dynamic landslide feature attention mechanism in a landslide recognition module so as to accurately recognize landslides; thirdly, performing mass-occurring landslide volume calculation in combination with the mountain digital surface model, and designing a compactness adjuster to optimize a calculation result according to the compactness of the landslide area to improve the accuracy of volume calculation; secondly, providing a landslide hazard grade evaluation method, and comprehensively evaluating a landslide result and a landslide volume through a landslide hazard evaluation function; finally, a landslide reason analysis method is provided, related data of the landslide are analyzed through the landslide analysis function, the contribution degree of each landslide factor is obtained, and therefore landslide treatment is better carried out.

Description

Technical Field

[0001] The invention relates to the technical field of mass landslide disaster analysis, and in particular to a method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing. Background Art

[0002] Cluster landslides refer to multiple landslides that occur in a certain area within a certain period of time or under specific weather conditions. Such landslides often occur in adjacent or connected areas and may be triggered by the same cause (such as heavy rainfall, earthquakes), resulting in multiple landslides in a wide area.

[0003] Mass landslides are a very serious natural disaster that frequently occurs in mountain structures. Since the mountain itself has a complex geological structure, such as faults and cracks, coupled with heavy rainfall, vegetation loss and human activities, the probability of mass landslides in the mountain increases; when mass landslides occur in the mountain, they will cause damage to roads, bridges and water resources; especially, when mass disasters in the mountain are particularly serious, they will also pose a serious threat to the housing and life safety of people around; therefore, it is very necessary to study and analyze mass landslide disasters in the mountain.

[0004] Traditional research on mass landslide disasters in mountains mainly relies on field investigations by staff to collect relevant data for analysis; however, this method has great disadvantages; for example, during the field investigation, the mountain may slide again, making the life safety of the staff unsafe; in addition, during the data collection process, since landslide is a dynamic process, traditional field investigations cannot reflect the changes in the mountain conditions in real time, and data updates may lag behind the actual situation. With the development of science and technology, today's staff basically use low-altitude drones to monitor and analyze mass landslides in mountains; in particular, low-altitude drone remote sensing has been widely used in the rapid calculation of the volume of mass landslide disasters; low-altitude drones can take high-resolution images, capture the details of the landslide area, and more accurately evaluate the volume and morphology of landslides.

[0005] In the existing technology, researchers have proposed many excellent methods for quickly extracting the volume of mass landslide disasters based on the remote sensing technology of low-altitude unmanned aerial vehicles; however, there are still some defects in the implementation process. For example, in the process of quickly extracting the volume of mass landslide disasters, there is a lack of relevant data analysis on different landslide areas where mass landslides occur, resulting in errors in the calculation of the volume of mass landslide disasters, which makes it impossible to effectively control them later.

[0006] To this end, the present invention proposes a method for rapid volume extraction of cluster landslide disasters based on low-altitude UAV remote sensing. Summary of the invention

[0007] The purpose of the present invention is to propose a method for quickly extracting the volume of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing, which can clarify the mass landslide disasters of the target mountain for analysis; first, it is proposed to use a mass landslide recognition model to identify the landslide area, and the model obtains the remote sensing image through a backbone feature extraction module, a neck feature fusion module, a landslide recognition module and an output module to obtain the landslide location and landslide area; in the landslide recognition module, different landslide features are learned by adding landslide feature attention, thereby improving the accuracy of landslide recognition; then, the landslide volume is calculated based on the landslide recognition result combined with the mountain digital surface model; in order to accurately calculate the volume of the landslide area before and after the landslide occurs, a density adjuster is used for adjustment; the density adjuster divides the target mountain A into N areas, sets the density coefficient according to the actual conditions of the N areas, and adjusts the density coefficient according to the actual conditions of the N areas. The density coefficient is adjusted according to the data changes before and after the occurrence of the mountain digital surface model, so as to obtain the optimal adjustment coefficient; the landslide volume calculation is optimized by the density adjuster, which not only improves the calculation accuracy and model stability, but also enhances the adaptability to complex terrain and different landslide situations; secondly, a landslide hazard level assessment method is proposed, which assesses different landslide areas based on the establishment of a landslide hazard assessment function, and assigns different weights to the hazard levels of different landslide areas; the weight is obtained by analyzing the location and area of ​​the landslide; the analysis of the landslide location includes: slope and distance from the main structure; the analysis of the landslide area includes: landslide area area and geological stability; by conducting a detailed analysis of different influencing factors of the landslide (slope, distance from the main structure, landslide area area and geological stability) and assigning corresponding weights, the assessment value of the landslide hazard can be accurately calculated. This method can comprehensively consider multiple factors, making risk assessment more comprehensive and scientific, thereby providing reliable data support for disaster prevention and mitigation; finally, the landslide analysis function is used to analyze the cause of the landslide of the target mountain A. The function adjusts parameters through historical landslide data and calculates the contribution of the collected characteristic data, which improves the accuracy of the analysis.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A rapid extraction method of the volume of cluster landslide disasters based on low-altitude UAV remote sensing includes:

[0010] Use low-altitude drones to take aerial photos of the target mountain and obtain high-resolution image data;

[0011] Further, the image data is preprocessed to obtain standard image data; wherein the preprocessing includes: geometrically correcting the image data to obtain geometrically corrected image data; atmospherically correcting the geometrically corrected image data to obtain atmospherically corrected image data; removing clouds and shadows from the atmospherically corrected image data to obtain de-imaged data; mosaicking and splicing the de-imaged data to obtain complete image data; and resolution enhancing the complete image data to obtain the standard image data.

[0012] Furthermore, the standard image data is used to generate a mountain digital surface model; wherein the generation process of the mountain digital surface model includes: extracting features from the standard image data to obtain land object information and mountain features respectively; generating an initial digital elevation model using the land object information and the mountain features; rasterizing the initial digital elevation model to convert it into raster data in a specified format; establishing a three-dimensional point cloud using the raster data; wherein the three-dimensional point cloud includes discrete point data on the surface, each point having spatial coordinates and corresponding elevation values; interpolating the three-dimensional point cloud to fill the space between the data to obtain continuous elevation data; and generating the mountain digital surface model based on the continuous elevation data.

[0013] Furthermore, a cluster landslide identification model is trained;

[0014] The training process of the cluster landslide identification model includes:

[0015] Acquire a remote sensing image of the target mountain based on historical landslide records; perform data enhancement on the remote sensing image to increase the number of the remote sensing images; perform landslide annotation on the remote sensing image after data enhancement, and construct a landslide data set; wherein the landslide annotation includes: landslide location and landslide area;

[0016] Input the image in the landslide data set to the backbone feature extraction module to obtain a first feature map, a second feature map and a third feature map respectively; wherein the backbone feature extraction module is composed of a convolution layer and five residual structures; each of the residual structures contains a different number of residual blocks; wherein the first residual structure contains a basic residual block; the second residual structure contains two residual blocks; the third residual structure and the fourth residual structure contain eight residual blocks, which are respectively used to obtain the first feature map and the second feature map; the fifth residual structure contains four residual blocks, which are used to obtain the third feature map;

[0017] The first feature map, the second feature map and the third feature map are input into the neck feature fusion module to obtain the first fused feature map, the second fused feature map and the third fused feature map; the neck fusion module includes: a first branch, a second branch and a third branch; wherein the first branch is used for the third feature map to pass through five convolution layers to obtain the third fused feature map; the second branch is used for the third fused feature map to pass through a convolution layer and an upsampling layer, and then to splice with the second feature map to obtain a spliced ​​feature map, and the spliced ​​feature map is passed through five convolution layers to obtain the second fused feature map; the third branch is used for the second fused feature map to be used for convolution and upsampling, and then to be spliced ​​with the first feature map, and the first fused feature map is obtained through five convolution layers;

[0018] The first fusion feature map, the second fusion feature map and the third fusion feature map are input into a landslide identification module to identify the landslide location and landslide area; wherein the landslide identification module includes: a first identification branch, a second identification branch and a third identification branch; wherein the first identification branch is used to identify the third fusion feature map to obtain a first identification result; the second identification branch is used to identify the second fusion feature map to obtain a second identification result; the third identification branch is used to identify the first fusion feature map to obtain a third identification result; the first identification branch, the second identification branch and the third recognition branch both include: a convolution layer and a dynamic landslide feature attention mechanism; wherein the dynamic landslide feature attention mechanism includes: inputting the fusion feature map into the convolution layer to obtain a convolution feature map; inputting the convolution feature map into the average pooling layer to obtain an average pooling feature map; passing the average pooling feature map through two fully connected layers in sequence to obtain a fully connected feature map; passing the fully connected feature map through a normalization layer to obtain a normalized feature map; inputting the normalized feature map into a hyperfunction to generate four learning parameters; inputting the convolution feature map and the four learning parameters into an activation function to obtain an activation feature map;

[0019] The landslide position and the landslide area are output to an output module to obtain a landslide identification result;

[0020] Calculating the identification error between the landslide identification result and the landslide annotation;

[0021] Optimizing the cluster landslide identification model according to the identification error; when the preset training condition is reached, the training ends, and the trained cluster landslide identification model is obtained;

[0022] Furthermore, the remote sensing image of the target mountain is input into the trained group landslide recognition model to obtain the group landslide disaster recognition area of ​​the target mountain;

[0023] Further, the landslide volume of the mass landslide disaster area is calculated using the mountain digital surface model to obtain the landslide volume result of the target mountain;

[0024] The calculation steps of the landslide volume result include:

[0025] Step 1: Acquire digital model data of the target mountain before and after the landslide occurs;

[0026] Step 2: Processing the digital model data before and after the landslide occurs to obtain standard digital model data;

[0027] Step 3: performing joint registration based on the group landslide hazard identification area and the standard digital model data;

[0028] Step 4: Optimizing the digital model data according to the joint registration result;

[0029] Step 5: Calculate the volume before and after the landslide according to the optimized digital model data; wherein, in the step 5, the volume calculation is adjusted by the density adjuster; wherein, the specific setting process of the density adjuster includes: dividing the mountain digital surface model into N areas; setting different density standards according to the N areas; wherein, the density standards include: low density and high density; setting the density coefficient according to the actual situation of the area; adjusting the density coefficient according to the data difference of the mountain digital surface model before and after the landslide; recalculating the volume before and after the landslide using the adjusted density coefficient; verifying whether the adjusted volumes before and after the landslide are consistent; when the results are consistent, the density coefficient is the current optimal density coefficient; when the results are inconsistent, continue to optimize the density coefficient according to the error;

[0030] Step 6: Draw a graph based on the volume before the landslide and the volume after the landslide, and determine whether an intersection point appears; when an intersection point appears, the calculation is terminated; when no intersection point appears, repeat the step 5;

[0031] Step 7: Output the calculation results.

[0032] Furthermore, a landslide hazard level assessment is performed based on the landslide volume result; wherein the specific process of the landslide hazard level assessment includes:

[0033] According to the result of the mass landslide identification model and the landslide volume result, a landslide set is established; wherein the landslide set is expressed as: MSL = {(P k ,A k ,V k )|1≤k≤K}; where (Pk ,A k ,V k ) represents the information of the kth landslide area of ​​the target mountain; P k A is represented as the coordinate position of the k-th landslide area, and the coordinate position is obtained based on the spatial coordinates established by the target mountain; k Represents the regional information of the kth landslide area; V k It is represented as the landslide volume of the kth landslide area; K is represented as the total number of landslide areas of the target mountain;

[0034] Generate a landslide hazard level weight for each landslide area; use a landslide hazard assessment function to assess the target mountain to obtain a landslide hazard assessment value;

[0035] The landslide hazard assessment value is obtained in the following way: Wherein, LHAV represents the landslide hazard assessment value of the target mountain; f() represents the landslide hazard assessment function; R k It is expressed as the landslide hazard level weight of the k-th landslide area; wherein, the landslide hazard level weight is obtained according to the location analysis results and regional analysis results of the landslide area; the location analysis results include: slope influence and main structure distance, and the calculation formula is: Among them, PR k Represented as the location analysis result of the kth landslide area; NS k Denotes the normalized slope of the kth landslide area; D k It is expressed as the distance between the kth landslide area and the main structure; w p1 Expressed as the slope weight coefficient; w p2 It is expressed as the distance influence coefficient; the regional analysis results include: regional area and geological conditions, and the calculation formula is: A k =w a1 *NA k +w a2 *S k Among them, A k Represents the regional analysis result of the kth landslide area; NA k It is represented as the normalized area of ​​the kth landslide area; S k Expressed as the geological stability coefficient of the kth landslide area; w a1 Expressed as area weight coefficient; w a2 It is expressed as a geological weight coefficient; the landslide hazard level weight is calculated according to the location analysis results and the regional analysis results, and the formula is: R k =w p *PR k +w a *A k ; Among them, wp Expressed as position influence weight; w a It is expressed as regional impact weight;

[0036] Comparing the landslide hazard assessment value with an assessment threshold to obtain a landslide hazard level;

[0037] Furthermore, the cause of the landslide is analyzed for the mass landslide; wherein the specific process of the landslide cause analysis includes:

[0038] Acquire relevant data of the landslide area of ​​the target mountain; wherein the relevant data includes: landslide information, environmental information, rock information, characteristic information and meteorological data before the landslide;

[0039] Further, the relevant data are characterized to establish landslide characteristic data;

[0040] Further, obtaining historical group landslide information of the target mountain;

[0041] Further, the parameters in the landslide analysis function are adjusted according to the mass landslide information; wherein the landslide analysis function is expressed as: Wherein, S represents the probability of mass landslides occurring in the target mountain; LAF() represents the landslide analysis function; It is represented as the weight vector of the landslide analysis function; HALF is represented as the characteristic data of the historical cluster landslide information;

[0042] Furthermore, the landslide characteristic data is input into the landslide analysis function after parameter adjustment to obtain the landslide contribution; wherein the landslide contribution represents the proportion of various factors causing mass landslides in the target mountain.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention proposes a cluster landslide recognition model for cluster landslide recognition of a target mountain; the model comprises: a backbone feature fusion module, a neck feature fusion module, a landslide recognition module and an output module; wherein the backbone feature fusion module is used to extract features of remote sensing images; the neck feature fusion module is used to fuse learning features; the landslide recognition module is used to identify the location and area of ​​the landslide; the output module is used to output the recognition result; a dynamic landslide feature attention mechanism is introduced into the landslide recognition module; the dynamic landslide feature attention mechanism learns four learning parameters from a hyperfunction through operations such as convolution, pooling and full connection of input features; and the four learning parameters are input into an activation function together with the original features, thereby realizing effective learning of the original features and improving the accuracy of cluster landslide recognition; a reliable landslide area is provided for subsequent cluster landslide volume calculation, and can help effectively control cluster landslides in the target mountain.

[0045] 2. The present invention proposes a method for calculating the volume of mass landslides. The method calculates the volume of landslides based on landslide identification results combined with a digital surface model of a mountain. In order to accurately calculate the volume of a landslide area before and after a landslide occurs, a density adjuster is used for adjustment. The density adjuster divides a target mountain into multiple areas, sets density coefficients according to the actual conditions of the multiple areas, and adjusts the density coefficient according to data changes before and after the occurrence of the digital surface model of the mountain, so as to obtain an optimal adjustment coefficient. The landslide volume calculation is optimized by the density adjuster, thereby improving the accuracy of mass landslide volume calculation, which is beneficial to the management of mass landslide areas.

[0046] 3. The present invention proposes a landslide hazard level assessment method, which assesses different landslide areas based on the establishment of a landslide hazard assessment function, and assigns different weights to the hazard levels of different landslide areas; the weights are obtained by analyzing the location of the landslide and the landslide area; the analysis of the landslide location includes: slope and distance from the main structure; the analysis of the landslide area includes: landslide area area and geological stability; by conducting a detailed analysis of different influencing factors of the landslide (slope, distance from the main structure, landslide area area and geological stability) and assigning corresponding weights, the assessment value of the landslide hazard can be accurately calculated, providing a scientific basis for the later landslide control of the mountain.

[0047] 4. The landslide analysis function is used to analyze the cause of landslides in the target mountain. The function adjusts parameters through historical landslide data and calculates the contribution of the collected characteristic data, which improves the accuracy of the analysis and is conducive to the management of mass landslides. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1A flow chart of a method for rapid volume extraction of mass landslide disasters based on low-altitude UAV remote sensing provided by an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of a mountain digital surface model provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of a cluster landslide identification model provided by an embodiment of the present invention;

[0051] Figure 4 A flow chart of a dynamic landslide feature attention mechanism provided by an embodiment of the present invention;

[0052] Figure 5 A flow chart of mass landslide volume calculation provided by an embodiment of the present invention;

[0053] Figure 6 A schematic diagram of the termination conditions for calculating the volume of a mass landslide provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] 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.

[0055] Cluster landslide refers to the phenomenon that multiple landslide events occur in the same area or under the same geological conditions. This type of landslide usually manifests itself as a wide range of landslide groups, which may involve multiple landslides. Cluster landslides often occur in mountains; because mountains usually have complex geological structures, such as faults and fissures, coupled with rainfall and other reasons, the probability of cluster landslides in mountains increases. When cluster landslides occur in mountains, they will cause damage to roads, bridges, water resources, etc.; in some cases, they will also pose a threat to human life safety. Therefore, relevant researchers will obtain remote sensing images of cluster landslides in mountains through low-altitude drones to calculate the volume of landslide disasters; this can well assess the severity of the disaster so that effective governance plans can be taken; there are many excellent methods in the prior art for calculating the volume of cluster landslide disasters in mountains, but there are some defects, so the present invention proposes a method for rapid extraction of cluster landslide disaster volume based on low-altitude drone remote sensing, which is used to calculate the volume of cluster disasters in mountains; the following will be described in detail from two embodiments.

[0056] Embodiment 1:

[0057] In the embodiment of the present application, the volume of the mass landslide of the target mountain A is calculated; see Figure 1 ,exist Figure 1 The specific implementation steps are given in the following, including: S10. Use low-altitude drones to obtain image data of the target mountain; S20. Preprocess the image data; S30. Generate a digital surface model of the mountain; S40. Train a cluster landslide recognition model; S50. Identify the cluster landslide area of ​​the target mountain; S60. Calculate the landslide volume based on the recognition results; S70. Evaluate the landslide hazard level; S80. Analyze the cause of the landslide. Figure 1 To explain:

[0058] Use a low-altitude drone to take aerial photos of the target mountain A and obtain high-resolution image data, corresponding to the above step S10;

[0059] Further, the acquired image data is preprocessed to obtain standard image data, corresponding to step S20;

[0060] Among them, the process of preprocessing the image data specifically includes: performing geometric correction on the image data to obtain geometrically corrected image data; performing atmospheric correction on the geometrically corrected image data to obtain atmospherically corrected image data; removing clouds and shadows from the atmospherically corrected image data to obtain de-imaged data; performing image mosaicking and splicing on the de-imaged data to obtain complete image data; performing resolution enhancement on the complete image data to obtain the standard image data.

[0061] In the embodiment of the present application, the image data of the target mountain A taken by a low-altitude drone is preprocessed, including: geometric correction, atmospheric correction, cloud shadow removal, image mosaicking, splicing and resolution enhancement; wherein, geometric correction and atmospheric correction eliminate the influence of image distortion and climate factors, so that the image data is closer to the real terrain and the accuracy of the digital surface model is improved; cloud shadows are removed to avoid the interference of these factors on data analysis, thereby improving the clarity and reliability of the model; image mosaicking and splicing seamlessly combine multiple images to provide a continuous and high-coverage view, thereby improving the integrity and consistency of the model; resolution enhancement technology can improve image details, so that the digital surface model can more accurately reflect the subtle changes and terrain features of the mountain.

[0062] Further, a mountain digital surface model corresponding to the target mountain A is generated according to the standard image data, corresponding to the above step S30; wherein, in the embodiment of the present application, Reality Capture software is used for 3D reconstruction and digital surface model generation; the software is an advanced 3D modeling and image processing software, which is widely used to generate high-quality digital surface models (DSM) and 3D models from image data; see Figure 2 ,exist Figure 2 A schematic diagram of generating a digital surface model of a mountain is given in ;

[0063] Among them, the specific generation process of the mountain digital surface model includes: extracting features from the standard image data to obtain ground object information and mountain features respectively; generating an initial digital elevation model using the ground object information and the mountain features; rasterizing the initial digital elevation model and converting it into raster data in a specified format; establishing a three-dimensional point cloud using the raster data; wherein the three-dimensional point cloud includes discrete point data on the surface, each point having spatial coordinates and corresponding elevation values; interpolating the three-dimensional point cloud to fill the space between the data to obtain continuous elevation data; and generating the mountain digital surface model based on the continuous elevation data.

[0064] In the embodiment of the present application, professional software is used to generate a digital surface model of the target mountain A. The model establishes an initial elevation model by acquiring terrain information and mountain features, and processes the elevation data using a rasterization method to obtain three-dimensional point cloud data corresponding to the elevation data. The point cloud data is then subjected to a series of processing to obtain the final mountain data surface model. This model not only accurately calculates the volume of subsequent group landslides, but can also be used for analysis of mountain stability.

[0065] Further, the group landslide identification model is trained to identify the group landslides on the target mountain A, corresponding to step S40; see Figure 3 , Figure 3 The overall structure of the cluster landslide identification model; in FIG3 , the cluster landslide identification model includes: a backbone feature extraction module, a neck feature fusion module and an output module; the following is a specific identification process of the cluster landslide on the target mountain A:

[0066] Acquire remote sensing images based on historical landslide records of target mountain A;

[0067] Furthermore, data enhancement is performed on the remote sensing images to increase the number of the remote sensing images;

[0068] Furthermore, landslide annotation is performed on the remote sensing image after data enhancement, and a landslide data set is constructed; wherein the landslide annotation includes: landslide location and landslide area;

[0069] Furthermore, the image in the landslide dataset is input into the backbone feature extraction module to obtain a first feature map, a second feature map and a third feature map respectively; wherein the backbone feature extraction module is composed of a convolutional layer and five residual structures; each of the residual structures contains a different number of residual blocks; wherein the first residual structure contains a basic residual block; the second residual structure contains two residual blocks; the third residual structure and the fourth residual structure contain eight residual blocks, which are respectively used to obtain the first feature map and the second feature map; the fifth residual structure contains four residual blocks, which are used to obtain the third feature map;

[0070] Further, the first feature map, the second feature map and the third feature map are input into the neck feature fusion module to obtain the first fused feature map, the second fused feature map and the third fused feature map; the neck fusion module includes: a first branch, a second branch and a third branch; wherein the first branch is used for the third feature map to pass through five convolution layers to obtain the third fused feature map; the second branch is used for the third fused feature map to pass through a convolution layer and an upsampling layer, and then to splice with the second feature map to obtain a spliced ​​feature map, and the spliced ​​feature map is passed through five convolution layers to obtain the second fused feature map; the third branch is used for the second fused feature map to be used for convolution and upsampling, and then to be spliced ​​with the first feature map, and the first fused feature map is obtained through five convolution layers;

[0071] Further, the first fused feature map, the second fused feature map and the third fused feature map are input into a landslide identification module to identify the landslide location and landslide area; wherein the landslide identification module comprises: a first identification branch, a second identification branch and a third identification branch; wherein the first identification branch is used to identify the third fused feature map to obtain a first identification result; the second identification branch is used to identify the second fused feature map to obtain a second identification result; the third identification branch is used to identify the first fused feature map to obtain a third identification result; the first identification branch, the second identification branch and the third identification branch all comprise: a convolutional layer and a dynamic landslide feature attention mechanism;

[0072] In this application example, the landslide location and landslide area of ​​the target mountain A are obtained through the landslide identification module; the module includes three identification branches; according to these three identification branches, the landslide identification results on feature maps of different scales can be obtained; in this way, different details of the landslide area can be captured, from the whole to the part, thereby improving the accuracy and comprehensiveness of landslide identification.

[0073] Among them, see Figure 4The dynamic landslide feature attention mechanism includes: inputting the fusion feature map into the convolution layer to obtain the convolution feature map; inputting the convolution feature map into the average pooling layer to obtain the average pooling feature map; passing the average pooling feature map through two fully connected layers in sequence to obtain the fully connected feature map; passing the fully connected feature map through the normalization layer to obtain the normalized feature map; inputting the normalized feature map into the hyperfunction to generate four learning parameters; inputting the convolution feature map and the four learning parameters into the activation function to obtain the activation feature map;

[0074] In the dynamic landslide feature attention mechanism, the definition of the hyperfunction is θ(x); where x is expressed as the normalized feature map of the input; the role of the hyperfunction is to learn four learning parameters for the activation function in the process of processing the feature. The function compresses the feature through the input normalized feature map x through global average pooling; then, multiple element values ​​are input by two fully connected layers and a nonlinear layer; corresponding to the following formula:

[0075] in, Represented as the output vector of the first fully connected layer; is represented by the output vector of the second fully connected layer; k is represented by the kth parameter; c is represented by the cth dimension channel; α k , β k They are respectively represented as learning parameters; λa represents the learning coefficient of the first fully connected layer; λb represents the learning coefficient of the second fully connected layer; Represented as the residual value of the first fully connected layer; Represented as the residual value of the second fully connected layer;

[0076] After the above learning process, four learning parameters can be obtained: α 1 , α 2 , β 1 and β 2 .

[0077] The design of the landslide recognition module in the embodiment of the present application consists of three recognition branches; each branch includes: a convolution layer and a dynamic landslide feature attention mechanism; the dynamic landslide feature attention mechanism learns four learning parameters from a hyperfunction through operations such as convolution, pooling and full connection of input features; and inputs the four learning parameters and the original features into the activation function together, thereby realizing effective learning of the original features.

[0078] Further, the landslide position and the landslide area are output to an output module to obtain a landslide identification result;

[0079] Calculating the identification error between the landslide identification result and the landslide annotation;

[0080] Optimizing the cluster landslide identification model according to the identification error; when the preset training condition is reached, the training ends, and the trained cluster landslide identification model is obtained;

[0081] Furthermore, the current remote sensing image of the target mountain A is input into the trained group landslide recognition model to obtain the group landslide disaster recognition area of ​​the target mountain, corresponding to the above S50.

[0082] Refer to Table 1, in order to illustrate that the group landslide identification model proposed in the present invention has an effective identification effect; in this embodiment, different identification models are experimentally compared using historical data; detailed comparison results are given in Table 1;

[0083] Table 1. Model performance comparison

[0084] Model Accuracy / % Recall / % F1 Average accuracy / % Faster R-CNN 60.56 63.49 0.61 83.06 SDD 81.28 79.11 0.80 71.00 Cluster landslide identification model 90.17 89.52 0.87 90.09

[0085] In Table 1, experimental comparisons were made using Faster R-CNN, SSD and the cluster landslide recognition model. From the data in Table 1, it can be seen that the model of the present invention is higher than the above two models in precision, recall rate, FI value and average accuracy, which are 90.17%, 89.52%, 0.87% and 90.09% respectively. It can be seen that the cluster landslide recognition model has certain effectiveness in the recognition of cluster landslides.

[0086] Further, the landslide volume is calculated based on the identified group landslide identification result of the target mountain A, corresponding to S60; see Figure 5 Flowchart, the specific calculation process is as follows:

[0087] Step 1: Acquire digital model data of the target mountain before and after the landslide occurs;

[0088] Step 2: Processing the digital model data before and after the landslide occurs to obtain standard digital model data;

[0089] Step 3: performing joint registration based on the group landslide hazard identification area and the standard digital model data;

[0090] Step 4: Optimizing the digital model data according to the joint registration result;

[0091] Step 5: Calculate the volume before and after the landslide according to the optimized digital model data; the calculation formula for the volume of the landslide area is:

[0092]

[0093] Where V is the cumulative volume of the landslide; a and b are the lowest and highest altitudes, respectively; S (m 2 ) is the horizontal area between a and b; h (m) is the elevation difference between the landslide digital model before and after the landslide; ΔS is the horizontal segment area; Δh i It is represented as the altitude difference of the i-th segment; n is represented as the number of segments;

[0094] Wherein, in the step five, the volume calculation is adjusted by the density adjuster; wherein, the specific setting process of the density adjuster includes: dividing the mountain digital surface model into N areas; setting different density standards according to the N areas; wherein, the density standards include: low density and high density; setting the density coefficient according to the actual situation of the area; adjusting the density coefficient according to the data difference of the mountain digital surface model before and after the landslide; recalculating the volume before and after the landslide using the adjusted density coefficient; verifying whether the adjusted volumes before and after the landslide are consistent; when the results are consistent, the density coefficient is the current optimal density coefficient; when the results are inconsistent, continuing to optimize the density coefficient according to the error;

[0095] In order to accurately calculate the volume of the landslide area before and after the landslide occurs in the embodiment of the present application, a density adjuster is used for adjustment; the density adjuster divides the target mountain A into N areas, sets the density coefficient according to the actual conditions of the N areas, and adjusts the density coefficient according to the data changes before and after the occurrence of the digital surface model of the mountain, so as to obtain the optimal adjustment coefficient; the optimization of the landslide volume calculation by the density adjuster not only improves the calculation accuracy and model stability, but also enhances the adaptability to complex terrain and different landslide scenarios.

[0096] Step 6: Draw a graph based on the volume before the landslide and the volume after the landslide, and determine whether an intersection occurs; when an intersection occurs, the calculation is terminated; when no intersection occurs, repeat the step 5; refer to Figure 6 ;

[0097] from Figure 6 It can be seen that point A0 A1 A2 is the volume change curve after the landslide; B0 B1 B2 is the landslide curve when no landslide occurs; the change in volume is due to the change in altitude during the calculation process, which causes the volume to fluctuate; Figure 6 The coordinates of each point in are expressed as: and the intersection point O(x,y); when the intersection of the two curves occurs for the first time, the termination condition is as follows: like Figure 6 As shown, the coordinates of the intersection point are solved according to the coordinates of the points given by the two curves, as shown below:

[0098] Step 7: Output the calculation results.

[0099] Furthermore, the landslide hazard level of the target mountain A is evaluated, corresponding to the above step S70; the specific process includes:

[0100] According to the result of the mass landslide identification model and the landslide volume result, a landslide set is established; wherein the landslide set is expressed as: MSL = {(P k ,A k ,V k )|1≤k≤K}; where (P k ,A k ,V k ) represents the information of the kth landslide area of ​​the target mountain; P k A is represented as the coordinate position of the k-th landslide area, and the coordinate position is obtained based on the spatial coordinates established by the target mountain; k Represents the regional information of the kth landslide area; V k It is represented as the landslide volume of the kth landslide area; K is represented as the total number of landslide areas of the target mountain;

[0101] Further, a landslide hazard level weight is generated for each landslide area; the target mountain is evaluated using a landslide hazard assessment function to obtain a landslide hazard assessment value;

[0102] The landslide hazard assessment value is obtained in the following way: Wherein, LHAV represents the landslide hazard assessment value of the target mountain; f() represents the landslide hazard assessment function; R k It is expressed as the landslide hazard level weight of the k-th landslide area; wherein, the landslide hazard level weight is obtained according to the location analysis results and regional analysis results of the landslide area; the location analysis results include: slope influence and main structure distance, and the calculation formula is: Among them, PR k Represented as the location analysis result of the kth landslide area; NS k Denotes the normalized slope of the kth landslide area; D k It is expressed as the distance between the kth landslide area and the main structure; w p1 Expressed as the slope weight coefficient; w p2 It is expressed as the distance influence coefficient; the regional analysis results include: regional area and geological conditions, and the calculation formula is: A k =w a1 *NA k +w a2 *S k Among them, A kRepresents the regional analysis result of the kth landslide area; NA k It is represented as the normalized area of ​​the kth landslide area; S k Expressed as the geological stability coefficient of the kth landslide area; w a1 Expressed as area weight coefficient; w a2 It is expressed as a geological weight coefficient; the landslide hazard level weight is calculated according to the location analysis results and the regional analysis results, and the formula is: R k =w p *PR k +w a *A k ; Among them, w p Expressed as position influence weight; w a It is expressed as regional impact weight;

[0103] In the process of calculating the landslide hazard assessment value, different weights are assigned to the hazard level of the landslide; the weight is obtained by analyzing the location and area of ​​the landslide; the analysis of the landslide location includes: slope and distance from the main structure; the analysis of the landslide area includes: landslide area and geological stability; by conducting a detailed analysis of the different influencing factors of the landslide (slope, distance from the main structure, landslide area and geological stability) and assigning corresponding weights, the assessment value of the landslide hazard can be accurately calculated. This method can comprehensively consider multiple factors, making risk assessment more comprehensive and scientific, thereby providing reliable data support for disaster prevention and mitigation.

[0104] Furthermore, the landslide hazard assessment value is compared with an assessment threshold to obtain a landslide hazard level; wherein the setting of the assessment threshold is obtained based on the historical landslide data analysis of the target mountain A, and the statistical distribution of the assessment threshold and the conditions for landslide occurrence are calculated.

[0105] In the embodiment of the present application, a landslide hazard level assessment method is used to assess the occurrence of mass landslides on the target mountain A; the method assesses different landslide areas based on the establishment of a landslide hazard assessment function, and makes a comprehensive judgment based on the assessment value; in the calculation process of the assessment value, the location, area and volume of the landslide are considered as assessment factors, and explained by assigning different weight information; the function can provide a comprehensive and accurate risk assessment. This method not only improves the scientificity and accuracy of landslide hazard assessment, but also effectively supports risk management and decision-making processes, and improves the efficiency of disaster prevention and emergency response.

[0106] Further, the cause of the landslide of the target mountain A is analyzed, corresponding to step S80; the specific process includes:

[0107] Obtain relevant data of the landslide area of ​​the target mountain A; wherein the relevant data includes: landslide information, environmental information, rock information, characteristic information and meteorological data before the landslide;

[0108] Further, the relevant data are characterized to establish landslide characteristic data;

[0109] Furthermore, historical group landslide information of target mountain A is obtained;

[0110] Further, the parameters in the landslide analysis function are adjusted according to the mass landslide information; wherein the landslide analysis function is expressed as: Wherein, S represents the probability of mass landslides occurring in the target mountain; LAF() represents the landslide analysis function; It is represented as the weight vector of the landslide analysis function; HALF is represented as the characteristic data of the historical cluster landslide information;

[0111] Furthermore, the landslide characteristic data is input into the landslide analysis function after parameter adjustment to obtain the landslide contribution; wherein the landslide contribution represents the proportion of various factors causing mass landslides in the target mountain.

[0112] In the embodiment of the present application, relevant data of the target mountain A is collected and characterized for later analysis; during the analysis process, the landslide analysis function is mainly used to obtain the contribution of each cause of the landslide; the function improves the accuracy of the analysis by adjusting parameters through historical landslide data.

[0113] In the application embodiment, the method for rapid extraction of the volume of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing realizes the calculation of the landslide volume when a mass landslide occurs in the target mountain A; in this process, the present invention includes multiple aspects: first, it is proposed to use a mass landslide recognition model to identify the landslide area. The model obtains the remote sensing image through a backbone feature extraction module, a neck feature fusion module, a landslide recognition module and an output module to obtain the landslide location and landslide area; in the landslide recognition module, different landslide features are learned by adding landslide feature attention, thereby improving the accuracy of landslide recognition; then, the landslide volume is calculated based on the landslide recognition result combined with the mountain digital surface model; in order to accurately calculate the volume of the landslide area before and after the landslide occurs, a density adjuster is used for adjustment; the density adjuster divides the target mountain A into N areas, and sets the density according to the actual conditions of these N areas. The density coefficient is calculated, and the density coefficient is adjusted according to the data changes before and after the occurrence of the mountain digital surface model, so as to obtain the optimal adjustment coefficient; the optimization of landslide volume calculation through the density adjuster not only improves the calculation accuracy and model stability, but also enhances the adaptability to complex terrain and different landslide situations; secondly, a landslide hazard level assessment method is proposed, which evaluates different landslide areas based on the establishment of a landslide hazard assessment function, and assigns different weights to the hazard levels of different landslide areas; the weight is obtained by analyzing the location and area of ​​the landslide; the analysis of the landslide location includes: slope and distance from the main structure; the analysis of the landslide area includes: landslide area area and geological stability; by conducting a detailed analysis of different influencing factors of the landslide (slope, distance from the main structure, landslide area area and geological stability) and assigning corresponding weights, the assessment value of the landslide hazard can be accurately calculated. This method can comprehensively consider multiple factors, making risk assessment more comprehensive and scientific, thereby providing reliable data support for disaster prevention and mitigation; finally, the landslide analysis function is used to analyze the cause of the landslide of the target mountain A. The function adjusts parameters through historical landslide data and calculates the contribution of the collected characteristic data, which improves the accuracy of the analysis.

[0114] Embodiment 2:

[0115] In the above-mentioned embodiment 1, the scheme of the present invention is used to verify the effect, and the feasibility of the present invention can be seen from the final result. In order to further illustrate the effect of the present invention, the volume of mass landslide disasters of mountain B is calculated in the embodiment of the present application. The specific process is basically the same as that of the embodiment 1, and the specific process includes:

[0116] Use low-altitude drones to take aerial photos of the target mountain and obtain high-resolution image data;

[0117] Further, the image data is preprocessed to obtain standard image data; wherein the preprocessing includes: geometrically correcting the image data to obtain geometrically corrected image data; atmospherically correcting the geometrically corrected image data to obtain atmospherically corrected image data; removing clouds and shadows from the atmospherically corrected image data to obtain de-imaged data; mosaicking and splicing the de-imaged data to obtain complete image data; and resolution enhancing the complete image data to obtain the standard image data.

[0118] Furthermore, the standard image data is used to generate a mountain digital surface model; wherein the generation process of the mountain digital surface model includes: extracting features from the standard image data to obtain land object information and mountain features respectively; generating an initial digital elevation model using the land object information and the mountain features; rasterizing the initial digital elevation model to convert it into raster data in a specified format; establishing a three-dimensional point cloud using the raster data; wherein the three-dimensional point cloud includes discrete point data on the surface, each point having spatial coordinates and corresponding elevation values; interpolating the three-dimensional point cloud to fill the space between the data to obtain continuous elevation data; and generating the mountain digital surface model based on the continuous elevation data.

[0119] Furthermore, a cluster landslide identification model is trained;

[0120] The training process of the cluster landslide identification model includes:

[0121] Acquire a remote sensing image of the target mountain based on historical landslide records; perform data enhancement on the remote sensing image to increase the number of the remote sensing images; perform landslide annotation on the remote sensing image after data enhancement, and construct a landslide data set; wherein the landslide annotation includes: landslide location and landslide area;

[0122] Input the image in the landslide data set to the backbone feature extraction module to obtain a first feature map, a second feature map and a third feature map respectively; wherein the backbone feature extraction module is composed of a convolution layer and five residual structures; each of the residual structures contains a different number of residual blocks; wherein the first residual structure contains a basic residual block; the second residual structure contains two residual blocks; the third residual structure and the fourth residual structure contain eight residual blocks, which are respectively used to obtain the first feature map and the second feature map; the fifth residual structure contains four residual blocks, which are used to obtain the third feature map;

[0123] The first feature map, the second feature map and the third feature map are input into the neck feature fusion module to obtain the first fused feature map, the second fused feature map and the third fused feature map; the neck fusion module includes: a first branch, a second branch and a third branch; wherein the first branch is used for the third feature map to pass through five convolution layers to obtain the third fused feature map; the second branch is used for the third fused feature map to pass through a convolution layer and an upsampling layer, and then to splice with the second feature map to obtain a spliced ​​feature map, and the spliced ​​feature map is passed through five convolution layers to obtain the second fused feature map; the third branch is used for the second fused feature map to be used for convolution and upsampling, and then to be spliced ​​with the first feature map, and the first fused feature map is obtained through five convolution layers;

[0124] The first fusion feature map, the second fusion feature map and the third fusion feature map are input into a landslide identification module to identify the landslide location and landslide area; wherein the landslide identification module includes: a first identification branch, a second identification branch and a third identification branch; wherein the first identification branch is used to identify the third fusion feature map to obtain a first identification result; the second identification branch is used to identify the second fusion feature map to obtain a second identification result; the third identification branch is used to identify the first fusion feature map to obtain a third identification result; the first identification branch, the second identification branch and the third recognition branch both include: a convolution layer and a dynamic landslide feature attention mechanism; wherein the dynamic landslide feature attention mechanism includes: inputting the fusion feature map into the convolution layer to obtain a convolution feature map; inputting the convolution feature map into the average pooling layer to obtain an average pooling feature map; passing the average pooling feature map through two fully connected layers in sequence to obtain a fully connected feature map; passing the fully connected feature map through a normalization layer to obtain a normalized feature map; inputting the normalized feature map into a hyperfunction to generate four learning parameters; inputting the convolution feature map and the four learning parameters into an activation function to obtain an activation feature map;

[0125] The landslide position and the landslide area are output to an output module to obtain a landslide identification result;

[0126] Calculating the identification error between the landslide identification result and the landslide annotation;

[0127] Optimizing the cluster landslide identification model according to the identification error; when the preset training condition is reached, the training ends, and the trained cluster landslide identification model is obtained;

[0128] Furthermore, the remote sensing image of the target mountain is input into the trained group landslide recognition model to obtain the group landslide disaster recognition area of ​​the target mountain;

[0129] Further, the landslide volume of the mass landslide disaster area is calculated using the mountain digital surface model to obtain the landslide volume result of the target mountain;

[0130] The calculation steps of the landslide volume result include:

[0131] Step 1: Acquire digital model data of the target mountain before and after the landslide occurs;

[0132] Step 2: Processing the digital model data before and after the landslide occurs to obtain standard digital model data;

[0133] Step 3: performing joint registration based on the group landslide hazard identification area and the standard digital model data;

[0134] Step 4: Optimizing the digital model data according to the joint registration result;

[0135] Step 5: Calculate the volume before and after the landslide according to the optimized digital model data; wherein, in the step 5, the volume calculation is adjusted by the density adjuster; the specific setting process of the density adjuster includes: dividing the mountain digital surface model into N areas; setting different density standards according to the N areas; wherein, the density standards include: low density and high density; setting the density coefficient according to the actual situation of the area; adjusting the density coefficient according to the data difference of the mountain digital surface model before and after the landslide; recalculating the volume before and after the landslide using the adjusted density coefficient; verifying whether the adjusted volumes before and after the landslide are consistent; when the results are consistent, the density coefficient is the current optimal density coefficient; when the results are inconsistent, continue to optimize the density coefficient according to the error;

[0136] Step 6: Draw a graph based on the volume before the landslide and the volume after the landslide, and determine whether an intersection point appears; when an intersection point appears, the calculation is terminated; when no intersection point appears, repeat the step 5;

[0137] Step 7: Output the calculation results.

[0138] Furthermore, a landslide hazard level assessment is performed based on the landslide volume result; wherein the specific process of the landslide hazard level assessment includes:

[0139] According to the result of the mass landslide identification model and the landslide volume result, a landslide set is established; wherein the landslide set is expressed as: MSL = {(P k ,A k ,V k )|1≤k≤K}; where (Pk ,A k ,V k ) represents the information of the kth landslide area of ​​the target mountain; P k A is represented as the coordinate position of the k-th landslide area, and the coordinate position is obtained based on the spatial coordinates established by the target mountain; k Represents the regional information of the kth landslide area; V k It is represented as the landslide volume of the kth landslide area; K is represented as the total number of landslide areas of the target mountain;

[0140] Generate a landslide hazard level weight for each landslide area; use a landslide hazard assessment function to assess the target mountain to obtain a landslide hazard assessment value;

[0141] The landslide hazard assessment value is obtained in the following way: Wherein, LHAV represents the landslide hazard assessment value of the target mountain; f() represents the landslide hazard assessment function; R k It is expressed as the landslide hazard level weight of the k-th landslide area; wherein, the landslide hazard level weight is obtained according to the location analysis results and regional analysis results of the landslide area; the location analysis results include: slope influence and main structure distance, and the calculation formula is: Among them, PR k Represented as the location analysis result of the kth landslide area; NS k Denotes the normalized slope of the kth landslide area; D k It is expressed as the distance between the kth landslide area and the main structure; w p1 Expressed as the slope weight coefficient; w p2 It is expressed as the distance influence coefficient; the regional analysis results include: regional area and geological conditions, and the calculation formula is: A k =w a1 *NA k +w a2 *S k Among them, A k Represents the regional analysis result of the kth landslide area; NA k It is represented as the normalized area of ​​the kth landslide area; S k Expressed as the geological stability coefficient of the kth landslide area; w a1 Expressed as area weight coefficient; w a2 It is expressed as a geological weight coefficient; the landslide hazard level weight is calculated according to the location analysis results and the regional analysis results, and the formula is: R k =w p *PR k +w a *A k ; Among them, wp Expressed as position influence weight; w a It is expressed as regional impact weight;

[0142] Comparing the landslide hazard assessment value with an assessment threshold to obtain a landslide hazard level;

[0143] Furthermore, the causes of group landslides were analyzed.

[0144] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for rapid volume extraction of cluster landslide disasters based on low-altitude UAV remote sensing, characterized in that: include: Use low-altitude drones to obtain high-resolution image data; Preprocessing the image data to obtain standard image data; Generate a mountain digital surface model using the standard image data; Training a cluster landslide recognition model; inputting a remote sensing image of a target mountain into the trained cluster landslide recognition model to obtain a cluster landslide disaster recognition area of ​​the target mountain; The landslide volume of the mass landslide disaster area is calculated using the mountain digital surface model to obtain the landslide volume result of the target mountain; wherein the calculation step of the landslide volume result includes: Step 1: Acquire digital model data of the target mountain before and after the landslide occurs; Step 2: Processing the digital model data before and after the landslide occurs to obtain standard digital model data; Step 3: performing joint registration based on the group landslide hazard identification area and the standard digital model data; Step 4: Optimizing the digital model data according to the joint registration result; Step 5: Calculate the volume before and after the landslide according to the optimized digital model data; Step 6: Draw a graph based on the volume before the landslide and the volume after the landslide, and determine whether an intersection point appears; when an intersection point appears, the calculation is terminated; when no intersection point appears, repeat the step 5; Step 7: Output calculation results; Conducting landslide hazard level assessment based on the landslide volume results; Analyze the causes of mass landslides.

2. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 1 is characterized in that: The preprocessing includes: performing geometric correction on the image data to obtain geometrically corrected image data; performing atmospheric correction on the geometrically corrected image data to obtain atmospherically corrected image data; removing clouds and shadows from the atmospherically corrected image data to obtain de-imaged data; performing image mosaicking and splicing on the de-imaged data to obtain complete image data; and performing resolution enhancement on the complete image data to obtain the standard image data.

3. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 1 is characterized in that: The generation process of the mountain digital surface model includes: extracting features from the standard image data to obtain ground object information and mountain features respectively; generating an initial digital elevation model using the ground object information and the mountain features; rasterizing the initial digital elevation model to convert it into raster data in a specified format; establishing a three-dimensional point cloud using the raster data; wherein the three-dimensional point cloud includes discrete point data on the surface, each point having spatial coordinates and corresponding elevation values; interpolating the three-dimensional point cloud to fill the space between the data to obtain continuous elevation data; and generating the mountain digital surface model based on the continuous elevation data.

4. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 1 is characterized in that: The training process of the cluster landslide identification model includes: Acquire a remote sensing image of the target mountain based on historical landslide records; perform data enhancement on the remote sensing image to increase the number of the remote sensing images; perform landslide annotation on the remote sensing image after data enhancement, and construct a landslide data set; wherein the landslide annotation includes: landslide location and landslide area; The images in the landslide data set are input into the backbone feature extraction module to obtain the first feature map, the second feature map and the third feature map respectively; wherein the backbone feature extraction module is composed of a convolution layer and five residual structures; each of the residual structures contains a different number of residual blocks; wherein the first residual structure contains a basic residual block; the second residual structure contains two residual blocks; the third residual structure and the fourth residual structure contain eight residual blocks, which are respectively used to obtain the first feature map and the second feature map; the fifth residual structure contains four residual blocks, which are used to obtain the third feature map; The first feature map, the second feature map and the third feature map are input into the neck feature fusion module to obtain the first fused feature map, the second fused feature map and the third fused feature map; the neck fusion module includes: a first branch, a second branch and a third branch; wherein the first branch is used for the third feature map to pass through five convolution layers to obtain the third fused feature map; the second branch is used for the third fused feature map to pass through a convolution layer and an upsampling layer, and then to splice with the second feature map to obtain a spliced ​​feature map, and the spliced ​​feature map is passed through five convolution layers to obtain the second fused feature map; the third branch is used for the second fused feature map to be used for convolution and upsampling, and then to be spliced ​​with the first feature map, and the first fused feature map is obtained through five convolution layers; Inputting the first fused feature map, the second fused feature map and the third fused feature map into a landslide identification module to identify a landslide location and a landslide area; The landslide position and the landslide area are output to an output module to obtain a landslide identification result; Calculating the identification error between the landslide identification result and the landslide annotation; The group landslide identification model is optimized according to the identification error; when the preset training conditions are reached, the training is ended to obtain the trained group landslide identification model.

5. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 4 is characterized in that: The landslide identification module includes: a first identification branch, a second identification branch and a third identification branch; wherein the first identification branch is used to identify the third fused feature map to obtain a first identification result; the second identification branch is used to identify the second fused feature map to obtain a second identification result; the third identification branch is used to identify the first fused feature map to obtain a third identification result; the first identification branch, the second identification branch and the third identification branch all include: a convolutional layer and a dynamic landslide feature attention mechanism.

6. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 5 is characterized in that: The dynamic landslide feature attention mechanism includes: inputting the fused feature map into the convolution layer to obtain the convolution feature map; inputting the convolution feature map into the average pooling layer to obtain the average pooling feature map; passing the average pooling feature map through two fully connected layers in sequence to obtain the fully connected feature map; passing the fully connected feature map through the normalization layer to obtain the normalized feature map; inputting the normalized feature map into the hyperfunction to generate four learning parameters; inputting the convolution feature map and the four learning parameters into the activation function to obtain the activated feature map.

7. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 6 is characterized in that: In step five, the density adjuster adjusts the volume calculation; wherein, the specific setting process of the density adjuster includes: dividing the mountain digital surface model into N areas; setting different density standards according to the N areas; wherein, the density standards include: low density and high density; setting the density coefficient according to the actual situation of the area; adjusting the density coefficient according to the data difference of the mountain digital surface model before and after the landslide; recalculating the volume before and after the landslide using the adjusted density coefficient; verifying whether the adjusted volumes before and after the landslide are consistent; when the results are consistent, the density coefficient is the current optimal density coefficient; when the results are inconsistent, continuing to optimize the density coefficient according to the error.

8. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 1 is characterized in that: The specific process of landslide hazard level assessment includes: According to the result of the mass landslide identification model and the landslide volume result, a landslide set is established; wherein the landslide set is expressed as: MSL = {(P k ,A k ,V k )|1≤k≤K}; where (P k ,A k ,V k ) represents the information of the kth landslide area of ​​the target mountain; P k A is represented as the coordinate position of the k-th landslide area, and the coordinate position is obtained based on the spatial coordinates established by the target mountain; k Represents the regional information of the kth landslide area; V k It is represented as the landslide volume of the kth landslide area; K is represented as the total number of landslide areas of the target mountain; Generate a landslide hazard level weight for each landslide area; use a landslide hazard assessment function to assess the target mountain to obtain a landslide hazard assessment value; The landslide hazard assessment value is obtained in the following way: Wherein, LHAV represents the landslide hazard assessment value of the target mountain; f() represents the landslide hazard assessment function; R k It is expressed as the landslide hazard level weight of the kth landslide area; The landslide hazard assessment value is compared with an assessment threshold to obtain a landslide hazard level.

9. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 8 is characterized in that: The landslide hazard level weight is obtained based on the location analysis results and regional analysis results of the landslide area; the location analysis results include: slope influence and main structure distance, and the calculation formula is: Among them, PR k Represented as the location analysis result of the kth landslide area; NS k Denotes the normalized slope of the kth landslide area; D k It is expressed as the distance between the kth landslide area and the main structure; w p1 Expressed as the slope weight coefficient; w p2 It is expressed as the distance influence coefficient; the regional analysis results include: regional area and geological conditions, and the calculation formula is: A k =w a1 *NA k +w a2 *S k Among them, A k Represents the regional analysis result of the kth landslide area; NA k It is represented as the normalized area of ​​the kth landslide area; S k Expressed as the geological stability coefficient of the kth landslide area; w a1 Expressed as area weight coefficient; w a2 It is expressed as a geological weight coefficient; the landslide hazard level weight is calculated according to the location analysis results and the regional analysis results, and the formula is: R k =w p *PR k +w a *A k ; Among them, w p Expressed as position influence weight; w a Expressed as regional influence weight.

10. The method for rapid volume extraction of mass landslide disasters based on low-altitude unmanned aerial vehicle remote sensing according to claim 1 is characterized in that: The specific process of the landslide cause analysis includes: obtaining relevant data of the landslide area of ​​the target mountain; wherein the relevant data includes: landslide information, environmental information, rock information, characteristic information and meteorological data before the landslide; characterizing the relevant data to establish landslide characteristic data; obtaining historical mass landslide information of the target mountain; adjusting parameters in the landslide analysis function according to the mass landslide information; wherein the landslide analysis function is expressed as: Wherein, S represents the probability of mass landslides occurring in the target mountain; LAF() represents the landslide analysis function; It is represented as the weight vector of the landslide analysis function; HALF is represented as the characteristic data of the historical cluster landslide information; The landslide characteristic data is input into the landslide analysis function after parameter adjustment to obtain the landslide contribution.

Images