Landslide disaster intelligent identification method and equipment based on unmanned aerial vehicle image deep learning, and medium

Abstract

The invention provides a landslide disaster intelligent identification method and device based on unmanned aerial vehicle image deep learning, and a medium, and the method comprises the steps: S1, obtaining the image data of a research region through the aerial survey of an unmanned aerial vehicle, and constructing a live-action three-dimensional model based on the image data of the research region; s2, obtaining an optical image and a topographic factor based on the live-action three-dimensional model, and carrying out multiband combination; s3, constructing a strip mine layered landslide classification sample data set based on a multi-scale-spectral difference segmentation method and threshold classification; s4, constructing a landslide recognition model based on the strip mine layered landslide classification sample data set in combination with a U-net model and a ResNet network; and S5, accurately identifying the mine landslide disaster through the landslide identification model. According to the invention, powerful technical support is provided for safety management of the surface mine.

Description

Technical Field

[0001] This invention belongs to the field of open-pit mine slope landslide identification, specifically involving a method, equipment, and medium for intelligent identification of landslide disasters based on deep learning from UAV images. Background Technology

[0002] my country is currently in a crucial stage of industrialization, and the surge in energy demand has accelerated the development of the mining industry. With the exploitation of open-pit mines, the number of steep slopes is gradually increasing, leading to frequent safety accidents in open-pit mines. Slope landslides rank among the top causes of mine safety accidents in terms of both total number of accidents and fatalities. Therefore, the rapid and accurate identification of landslide hazards in their early stages, and the guarantee of the safety and stability of mine slopes, are of significant theoretical and practical importance.

[0003] The use of high-precision remote sensing imagery combined with deep learning algorithms for landslide disaster identification is relatively mature both domestically and internationally. Landslide identification based on remote sensing imagery and deep learning has been widely applied in the identification of various geological landslides. However, the datasets based on remote sensing imagery used by existing deep learning network models are basically manually labeled. Traditional manual labeling relies mainly on the accuracy of human visual recognition, which is not only highly subjective and has low recognition ability, but also reduces data accuracy and information quantity, which is not conducive to the feature learning of subsequent network models. In addition, the network degradation problem caused by the increase of network depth in convolutional neural networks and the existence of redundant information in the feature space that is irrelevant to landslide identification limit the efficiency and accuracy of network models in landslide identification.

[0004] In summary, accurate identification of open-pit mine landslides requires an efficient method. First, it is essential to ensure that the dataset preparation process is complete and efficient. Accurate and effective dataset processing is crucial for the accurate identification of open-pit mine landslide disasters. Subsequent training models should retain as much semantic information as possible while avoiding redundant information, thereby increasing detection accuracy. Summary of the Invention

[0005] This invention proposes an intelligent landslide disaster identification method based on deep learning from UAV imagery. It aims to address the problems of low data accuracy and information quantity in the data annotation process of remote sensing imagery for open-pit mine landslide disaster identification, severe degradation of deep learning network models, and excessive redundant information leading to poor identification efficiency and accuracy.

[0006] The technical solution of the present invention is as follows:

[0007] A method for intelligent identification of landslide hazards based on deep learning from UAV imagery, the method comprising:

[0008] S1 uses drone aerial surveying to acquire image data of the study area, and constructs a real-scene 3D model based on the image data of the study area;

[0009] S2 acquires optical images and terrain factors based on the real-scene 3D model and performs multi-band merging;

[0010] S3 constructs a classification sample dataset of open-pit mine layered landslides based on a multi-scale-spectral difference segmentation method and threshold classification.

[0011] S4 constructs a landslide identification model based on the open-pit mine stratified landslide classification sample dataset, combining the U-net model and ResNet network.

[0012] S5 accurately identifies mine landslide hazards using the landslide identification model.

[0013] Furthermore, in step S1, when using UAV aerial survey to acquire image data of the study area, it is necessary to plan the flight route, aerial survey altitude, heading and lateral overlap rate, and ensure that the elevation error between the aerial survey data and the image control point at the same location is greater than the horizontal error; by setting up plane and elevation connection points and checkpoints, the horizontal and elevation accuracy of the checkpoints are statistically analyzed, and aerial survey statistics and error results are calculated.

[0014] Furthermore, the multi-band merging operation in step S2 is as follows:

[0015] Optical images are divided into three channels based on the three primary colors; topographic factor data are divided into four channels based on four types of topographic factors.

[0016] The optical imagery and topographic factors are cropped, and the 3-channel optical imagery data is merged with the 4-channel topographic data to generate 7-band data. The topographic information of each channel is visualized as grayscale images for subsequent image segmentation.

[0017] Furthermore, the multi-scale-spectral difference segmentation method specifically includes:

[0018] For each pixel, the difference between its reflectance or intensity values ​​is calculated across all bands. Treating RGB values ​​as coordinate points in three-dimensional space, the spectral difference between two pixels is calculated using the following formula:

[0019]

[0020] Where R, G, and B are the pixel RGB values;

[0021] The overall spectral dissimilarity is obtained by calculating the average spectral difference between all corresponding pixel pairs:

[0022]

[0023] Where n is the number of paired pixels in the two regions, and D ij This represents the spectral difference between the i-th and j-th pixels.

[0024] Furthermore, the threshold classification method for constructing a stratified landslide classification sample dataset for open-pit mines is as follows:

[0025] First, the input data is classified into roads and unclassified data A based on three parameters: slope, brightness, and terrain relief.

[0026] Unclassified data A is classified using mean, brightness, shape index, and NDVI index, and is divided into other artificial facilities, green vegetation, and unclassified data B.

[0027] Unclassified data B is classified based on slope, curvature, and mean, and is divided into stepped slopes and landslide areas.

[0028] Furthermore, S4 specifically refers to:

[0029] The landslide identification model is a Res-U-Net model. The input layer and residual module of ResNet are used to replace the input layer and encoding block of the U-Net network. The optimized structure is divided into two parts: encoding and decoding, containing a total of 9 modules, which are used to perform 5 convolutions and 4 deconvolutions. The last 3 layers of ResNet are discarded in the Res-U-Net feature extraction part, and the encoding convolutional layers of the U-Net network are replaced with conv2_x, con3_x, con4_x, and con5_x. In the decoding stage, bilinear interpolation is first used to replace the deconvolutional layers, gradually increasing the size of the output feature map while reducing the number of channels. Then, a skip structure is used to combine the feature maps of the same scale in downsampling and upsampling by channels, and the information is superimposed and fused. The result of each layer is output by the ReLU activation function.

[0030] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the landslide disaster intelligent identification method based on deep learning of UAV imagery described above.

[0031] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent landslide disaster identification method based on deep learning of UAV images.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] 1. The data source of this application comes from optical image maps and topographic factors in UAV imagery. Compared with other remote sensing image acquisition methods, it can collect topographic information and spectral data more quickly and flexibly, effectively improving the accuracy and reliability of landslide disaster identification.

[0034] 2. By adopting a multi-scale-spectral difference segmentation method and threshold classification principle, combined with the Res-U-Net model, we achieved efficient identification of stratified landslides in open-pit mines, significantly improving the model's generalization ability and robustness.

[0035] 3. By constructing a high-precision, object-oriented landslide sample dataset, the quality of the basic data for model training was ensured, further enhancing the accuracy of landslide disaster prediction.

[0036] 4. This method is easy to operate in practical applications, has an optimized data processing flow, and can quickly respond to the needs of landslide disaster identification, providing strong technical support for the safety management of open-pit mines. Attached Figure Description

[0037] The accompanying drawings illustrate various embodiments generally by way of example rather than limitation, and are used, together with the specification and claims, to explain embodiments of the invention. Where appropriate, the same reference numerals are used in all drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the apparatus or method.

[0038] Figure 1 This is a flowchart of a landslide disaster intelligent identification method based on deep learning of UAV images provided by the present invention;

[0039] Figure 2 This is a flowchart of the multi-band data synthesis visualization process provided by the present invention;

[0040] Figure 3 This is a flowchart of the multi-condition threshold hierarchical landslide classification provided by the present invention;

[0041] Figure 4 This is a diagram of the Res-U-net network architecture provided by the present invention. Detailed Implementation

[0042] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0043] This application discloses an intelligent landslide disaster identification method based on deep learning from UAV imagery, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0044] S1. Use UAV aerial surveys to obtain data of the research area, construct a 3D model based on the UAV aerial survey results, and obtain DEM and DOM data based on the 3D model.

[0045] S1-1: Data for the study area is acquired through UAV aerial surveying. To improve the quality and usability of the acquired data, it is important to ensure that the takeoff platform is unobstructed during the UAV aerial surveying process. The flight path, aerial surveying altitude, heading, and lateral overlap rate are determined based on the actual accuracy requirements of the measurement task, terrain conditions, equipment performance, and cost factors.

[0046] S1-2, In order to reduce the number of ground control points while ensuring measurement accuracy, a certain number of ground control points are set as horizontal and vertical connection points, and some ground control points are used as check points. The horizontal and vertical accuracy of the check points are statistically analyzed, and aerial survey statistics and error calculations are performed.

[0047] Δx=Xx

[0048] Δy=Yy

[0049]

[0050] ΔH=H0-H

[0051] Among them, Δx is the difference in the X direction, Δy is the difference in the Y direction, ΔS is the overall difference, and ΔH is the difference in elevation.

[0052] S1-3, based on the statistical analysis results, identify the maximum values ​​of the horizontal coordinate error and the elevation error. Further examine whether the elevation error between control points at the same location is generally greater than the horizontal coordinate error to ensure that the accuracy of the control measurements meets the established standards. The collected data should be suitable for subsequent office data processing.

[0053] S1-4: Based on UAV aerial photogrammetry data, a three-dimensional model of the study area is constructed, and high-precision terrain parameters and spectral image information are extracted from it.

[0054] S2 uses high-precision spectral imagery and topographic factor data generated by UAV aerial surveys as its data source and employs multi-band merging preprocessing of UAV images.

[0055] S2-1 generates DEM and DOM based on the real-world 3D model produced by UAV imagery. To allow the deep learning network model to fully learn various landslide information in the study area, the DEM and DOM are cropped and output.

[0056] S2-2 involves cropping and merging spectral and topographic feature data into multiple bands, overlaying different image information describing various feature types. The spectral data is divided into three channels based on the three primary colors. The topographic feature data is divided into four channels based on four types of topographic factors. The multi-band merging of topographic feature data and spectral image data generates data containing seven channels, or seven band values. The topographic information contained in each single channel is visualized in grayscale image form, clearly presenting topographic changes visually, which is helpful for the subsequent image segmentation task. The multi-band data synthesis and visualization process is as follows: Figure 2 As shown.

[0057] S3, based on the multi-scale-spectral difference segmentation method and threshold classification principle, constructs a classification process for open-pit mine stratified landslides with multiple conditional thresholds, and completes the construction of sample dataset annotation.

[0058] S3-1, After data cropping and fusion, multi-scale spectral difference segmentation is performed. The mean brightness across different phase categories is calculated. Considering the data characteristics of the RGB three channels, for each pixel, the difference in reflectance or intensity values ​​across all bands is calculated. Treating RGB values ​​as coordinate points in three-dimensional space, the spectral difference between two pixels is calculated using the following formula:

[0059]

[0060] Among them, D ij This represents the spectral difference between the i-th color vector and the j-th pixel.

[0061] For the entire regions A and B, the overall spectral difference can be obtained by calculating the average spectral difference between all corresponding pixel pairs:

[0062]

[0063] Where n is the number of paired pixels in the two regions.

[0064] S3-2, Debug and set multi-scale segmentation parameters, compactness, shape index, and band weights. Based on the parameters, and considering the characteristics of the study area, to ensure that the segmentation does not destroy the topographic relief, the weights of the three RGB spectral bands are set to 1, and the weight of the grayscale band for topographic relief is set to 0.5. Multiple iterations and tests are conducted using different parameters for maximum spectral difference. When the segmentation effect achieves the merging of slopes and planes into one, the segmentation parameters should not be increased further. This ensures that most of the UAV imagery's ground feature characteristics are preserved while reducing over-segmentation.

[0065] S3-3, Analyzing the manifestations of landslides and extracting feature values. Landslides are mainly shallow landslides, mostly occurring on slope surfaces, exposing shallow soil layers and forming accumulations of loose rocks. To describe the various characteristics of landslides, based on spectral image data and topographic thematic maps, four types of features were extracted: spectral, geometric shape, texture, and topography. The feature values ​​extracted based on the influence values ​​of seven bands are shown in Table 1.

[0066] Table 1. Feature Extraction Table for Landslide Area

[0067]

[0068]

[0069] S3-4, combining landslide feature information and utilizing the threshold classification principle, constructs a multi-condition threshold-based classification process for mine landslides. A hierarchical classification rule is adopted to systematically eliminate features other than landslide features, reducing the likelihood of confusion with adjacent features during landslide classification. The steps for multi-condition value-based landslide classification are as follows: Figure 3 As shown.

[0070] First, the input data is classified into roads and unclassified data A based on three parameters: slope, brightness, and terrain undulation.

[0071] The formula for calculating the slope parameter is as follows:

[0072]

[0073] in, and These represent the rates of change of height along the x-axis (east-west direction) and the y-axis (north-south direction), respectively.

[0074] The formula for calculating the brightness parameter is as follows:

[0075] Brightness = w R ×R+w G ×G+w B ×B

[0076] Among them, w R w G w B These are the weights for the red, green, and blue bands, respectively.

[0077] The formula for calculating terrain relief is as follows:

[0078] RF=z mxx -z min

[0079] Among them, z max It is z minMaximum and minimum elevation values ​​in DEM cells within each window.

[0080] Unclassified data A is classified using mean, brightness, shape index, and NDVI index, and is divided into other artificial facilities, green vegetation, and unclassified data B.

[0081] The formula for calculating the mean parameter is as follows:

[0082]

[0083] Where N represents the total number of pixels in the region, z i It is the value of the i-th pixel.

[0084] The formula for calculating the brightness parameter has been described and explained above.

[0085] The formula for calculating the shape index parameter is as follows:

[0086]

[0087] Where P is the perimeter of the landslide area and A is the area of ​​the landslide area.

[0088] The formula for calculating the NDVI index is as follows:

[0089]

[0090] Wherein, NIR represents near-infrared reflectance; and Red represents the reflectance in the red light band.

[0091] Unclassified data B is classified based on slope, curvature, and mean, and is divided into stepped slopes and landslide areas.

[0092] Curvature parameters are divided into profile curvature and planar curvature.

[0093] The formula for calculating the curvature of a cross section is as follows:

[0094]

[0095] Where z is the elevation value of the terrain surface, and s is the distance along the profile line.

[0096] The formula for calculating the curvature of a plane is as follows:

[0097]

[0098] Where z is the elevation value of the terrain surface, and x and y are the horizontal coordinates.

[0099] After setting the multi-scale-spectral difference parameters, the image data and the topographic factor thematic map were combined as input data and divided into roads, other man-made facilities, green vegetation, terraces and slopes, and landslide areas. The landslide areas in the dataset were labeled with high accuracy, and the landslide label samples of the study area were completed.

[0100] S4, Construction of an intelligent landslide disaster identification model. Based on a high-precision landslide sample dataset with object-oriented annotation, the U-net model and ResNet network are organically combined to obtain the Res-U-Net model. This deepens the network and enables it to extract more representative landslide image information features, thus constructing a network model suitable for landslide disaster identification in open-pit mines.

[0101] S4-1, the Res-U-Net model, is an improvement on U-Net, inspired by the image classification network ResNet. It replaces the input layer and encoding blocks of the U-Net network with the input layer and residual modules of ResNet. The optimized structure consists of two parts: encoding and decoding, containing a total of 9 modules used for 5 convolutions and 4 deconvolutions. The Res-U-Net feature extraction part discards the last 3 layers of ResNet and replaces the encoding convolutional layers of the U-Net network with conv2_x, con3_x, con4_x, and con5_x. In the decoding stage, bilinear interpolation is first used to replace the deconvolutional layers, gradually increasing the output feature map size while reducing the number of channels. Then, a skip structure is used to combine feature maps of the same scale from downsampling and upsampling, and the information is superimposed and fused before the ReLU activation function outputs the result of each layer. The Res-U-Net network architecture is as follows: Figure 4 As shown.

[0102] The structural parameters of the S4-2 and Res-U-net models are shown in Table 2. The input image data is 1024x1024 pixels in size, with 3 channels for optical images and 4 channels for terrain data, for a total of 7 channels. After 5 coding layers, the image is reduced to 1 / 16 of its original feature map.

[0103] Table 2 Structural parameters of the Res-U-net model

[0104]

[0105]

[0106] The specific steps are as follows:

[0107] In the input layer, there are two input operations. Input_1 is responsible for inputting image data, with a size of 3×3 / 1×1 and a stride of 1, and an output size of 1024×1024×3. Input_2 is responsible for inputting terrain data, also with a size of 3×3 / 1×1 and a stride of 1, and an output size of 1024×1024×4.

[0108] Operations are performed during encoding, and the encoding layer consists of multiple levels.

[0109] In Level_1, a convolution operation is performed first, with a size of 3×3 / 1×1, a stride of 1, and edge padding of 2 / 0, resulting in an output size of 1024×1024×64. Then, a pooling operation is performed, with a size of 2×2, a stride of 2, and edge padding of 0, resulting in an output size of 512×512×64.

[0110] Level_2 also first performs convolution with a size of 3×3 / 1×1, stride of 1, and edge padding of 2 / 0, resulting in an output size of 512×512×128. Then it performs pooling with a size of 2×2, stride of 2, and edge padding of 0, resulting in an output size of 256×256×128.

[0111] Level_3 also first performs convolution with a size of 3×3 / 1×1, stride of 1, and edge padding of 2 / 0, resulting in an output size of 256×256×256. Then it performs pooling with a size of 2×2, stride of 2, and edge padding of 0, resulting in an output size of 128×128×256.

[0112] Level_4 also first performs convolution with a size of 3×3 / 1×1, stride of 1, and edge padding of 2 / 0, resulting in an output size of 128×128×512. Then it performs pooling with a size of 2×2, stride of 2, and edge padding of 0, resulting in an output size of 64×64×512.

[0113] Level 5: Perform convolution operation with dimensions 3×3 / 1×1, stride 1, edge padding 2 / 0, and output size 64×64×1024.

[0114] The decoding layer also consists of multiple layers. Level 6 first performs an upsampling operation with a size of 2×2, a stride of 2, and edge padding of 0, resulting in an output size of 128×128×512. Next, it performs a skip connection operation, resulting in an output size of 128×128×1024. Then, it performs a convolution operation with a size of 3×3 / 1×1, a stride of 1, and edge padding of 2 / 0, resulting in an output size of 128×128×512.

[0115] Level_7 first performs an upsampling operation with a size of 2×2, a stride of 2, and edge padding of 0, resulting in an output size of 256×256×256. Next, it performs a skip connection operation, resulting in an output size of 256×256×512. Then, it performs a convolution operation with a size of 3×3 / 1×1, a stride of 1, and edge padding of 2 / 0, resulting in an output size of 256×256×256.

[0116] Level_8 first performs an upsampling operation with a size of 2×2, a stride of 2, and edge padding of 0, resulting in an output size of 512×512×128. Next, it performs a skip connection operation, resulting in an output size of 512×512×256. Then, it performs a convolution operation with a size of 3×3 / 1×1, a stride of 1, and edge padding of 2 / 0, resulting in an output size of 512×512×128.

[0117] Level_9 first performs an upsampling operation with a size of 2×2, a stride of 2, and edge padding of 0, resulting in an output size of 1024×1024×64. Next, it performs a skip connection operation, resulting in an output size of 1024×1024×128. Then, it performs a convolution operation with a size of 3×3 / 1×1, a stride of 1, and edge padding of 2 / 0, resulting in an output size of 1024×1024×64.

[0118] Finally, the output layer operation is performed. Output_10 performs output data operation with a size of 1×1, a step of 1, and edge padding of 0. The final output size is 1024×1024×1.

[0119] S4-3, accurately identifies mine landslide hazards based on the output of the Res-U-Net model.

[0120] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent identification of landslide hazards based on deep learning from UAV imagery, characterized in that, The method includes: S1 uses drone aerial surveying to acquire image data of the study area, and constructs a real-scene 3D model based on the image data of the study area; S2 acquires optical images and terrain factors based on the real-scene 3D model and performs multi-band merging; S3 constructs a classification sample dataset of open-pit mine layered landslides based on a multi-scale-spectral difference segmentation method and threshold classification. S4 constructs a landslide identification model based on the open-pit mine stratified landslide classification sample dataset, combining the U-net model and ResNet network. S5 accurately identifies mine landslide hazards using the landslide identification model.

2. The intelligent landslide disaster identification method based on deep learning of UAV imagery according to claim 1, characterized in that, In step S1, when using UAV aerial survey to acquire image data of the study area, it is necessary to plan the flight route, aerial survey altitude, heading and lateral overlap rate, and ensure that the elevation error between the aerial survey data and the control point at the same location is greater than the horizontal error. By setting up plane and elevation connection points and checkpoints, the horizontal and elevation accuracy of the checkpoints are statistically analyzed, and aerial survey statistics and error results are calculated.

3. The intelligent landslide disaster identification method based on deep learning of UAV imagery according to claim 1, characterized in that, The multi-band merging operation in step S2 is as follows: Optical images are divided into three channels based on the three primary colors; topographic factor data are divided into four channels based on four types of topographic factors. The optical imagery and topographic factors are cropped, and the 3-channel optical imagery data is merged with the 4-channel topographic data to generate 7-band data. The topographic information of each channel is visualized as grayscale images for subsequent image segmentation.

4. The intelligent landslide disaster identification method based on deep learning of UAV imagery according to claim 1, characterized in that, The multi-scale-spectral difference segmentation method is specifically as follows: For each pixel, the difference between its reflectance or intensity values ​​is calculated across all bands. Treating RGB values ​​as coordinate points in three-dimensional space, the spectral difference between two pixels is calculated using the following formula: Where R, G, and B are the pixel RGB values; The overall spectral dissimilarity is obtained by calculating the average spectral difference between all corresponding pixel pairs: Where n is the number of paired pixels in the two regions, and D ij This represents the spectral difference between the i-th and j-th pixels.

5. The intelligent landslide disaster identification method based on deep learning of UAV imagery according to claim 1, characterized in that, The threshold classification method is used to construct a sample dataset for classifying stratified landslides in open-pit mines. First, the input data is classified into roads and unclassified data A based on three parameters: slope, brightness, and terrain relief. Unclassified data A is classified using mean, brightness, shape index, and NDVI index, and is divided into other artificial facilities, green vegetation, and unclassified data B. Unclassified data B is classified based on slope, curvature, and mean, and is divided into stepped slopes and landslide areas.

6. The intelligent landslide disaster identification method based on deep learning of UAV imagery according to claim 1, characterized in that, S4 specifically refers to: The landslide identification model is a Res-U-Net model. The input layer and residual module of ResNet are used to replace the input layer and encoding block of the U-Net network. The optimized structure is divided into two parts: encoding and decoding, containing a total of 9 modules, which are used to perform 5 convolutions and 4 deconvolutions. The last 3 layers of ResNet are discarded in the Res-U-Net feature extraction part, and the encoding convolutional layers of the U-Net network are replaced with conv2_x, con3_x, con4_x, and con5_x. In the decoding stage, bilinear interpolation is first used to replace the deconvolutional layers, gradually increasing the size of the output feature map while reducing the number of channels. Then, a skip structure is used to combine the feature maps of the same scale in downsampling and upsampling by channels, and the information is superimposed and fused. The result of each layer is output by the ReLU activation function.

7. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program and the processor runs the computer program to enable the electronic device to perform the landslide disaster intelligent identification method based on deep learning of UAV imagery according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the landslide disaster intelligent identification method based on deep learning of UAV imagery as described in any one of claims 1 to 6.

Images