Deep learning-based reservoir area landslide disaster intelligent identification method and system, and storage medium

Abstract

The invention discloses a deep learning method-based reservoir area geological disaster intelligent identification method and system, and a storage medium, and the method comprises the steps: S1, analyzing basic geological data and monitoring data, and preliminarily studying and judging the landslide susceptibility; s2, obtaining a reservoir area slope orthoimage and a three-dimensional digital elevation model by using an unmanned aerial vehicle; s3, performing feature enhancement on the orthoimage data set, and constructing and training a deep learning model for landslide disasters; s4, comprehensively identifying the landslide mass in combination with multi-stage ortho-image comparison and displacement analysis, and rechecking by using a three-dimensional digital elevation model; and S5, analyzing a landslide deformation failure mode according to a prediction result, drawing up a support mode in combination with numerical simulation analysis software, and providing support for disaster control. Based on unmanned aerial vehicle remote sensing data and ground base station monitoring data, in combination with a deep learning algorithm, the efficiency of reservoir area landslide disaster risk prediction and subsequent stability processing is effectively improved, disaster control guidance can be provided, and effective guarantee is provided for major hydropower engineering reservoir bank engineering safety.

Description

Technical Field

[0001] The present invention relates to an intelligent identification system for landslide geological disasters, and in particular to an intelligent identification method, system and storage medium for reservoir landslide disasters based on deep learning, belonging to the intersection of remote sensing data intelligent processing technology and geological disaster early identification technology. Background Art

[0002] my country has a vast territory, complex geological and geographical environment, and significant differences in climatic conditions in time and space. Mountains and hills account for 65% of the country's land area, but they carry 56% of the country's population. The geological conditions in mountainous and hilly areas are very complex overall, with varied topography and frequent tectonic activities, which leads to diverse and widespread geological disasters such as collapse, landslides, mudslides, ground collapse, ground fissures, and ground subsidence. As one of the most influential geological disasters in the world, landslides have long occupied a large proportion of the types of geological disasters in my country. In a narrow sense, landslides refer to the phenomenon that rock and soil bodies slide along a certain sliding surface as a whole due to natural or human factors, which has strong suddenness and uncertainty. The construction of high dams and large reservoirs will lead to a substantial increase in the groundwater level in the reservoir area, which will not only affect the stability of the existing landslide body, but also bring potential landslide hazards. The periodic changes in rainfall in the reservoir area and the reservoir water level will also have a certain impact on the stress state of the landslide body, the properties of the rock and soil body, and the dynamic and static water pressure, thereby affecting the stability of the reservoir bank during the operation of the project.

[0003] The identification of geological disasters such as landslides benefits from the increasingly developed remote sensing technology and interpretation capabilities. In recent years, interferometric synthetic aperture radar deformation monitoring technology and high-resolution optical remote sensing image interpretation have become important methods for early identification of geological disaster hazards. Research by domestic and foreign scholars has shown that drone-based airborne remote sensing technology has great advantages in identifying detailed deformation features of landslides, mudslides and other geological disasters with its high image resolution and high mobility. Traditional landslide disaster investigation methods are limited by factors such as long manual investigation time, high cost, and small identification range, and the investigation data has time and space limitations. In the past 20 years, the development of remote sensing and geographic information technology has provided support for the acquisition and processing of large-scale, high-precision geological disaster data. The application and development of machine learning algorithms provide effective solutions for big data computing, high-precision image recognition and data mining. Deep learning methods are suitable for multi-scale landslide identification due to their in-depth feature extraction of samples. However, they have not yet been widely used in the field of landslide prevention and control. Considering the key role of the quality of landslide datasets and the applicability of deep learning algorithms in prediction results, the following problems still exist in current research: (1) Landslide image datasets are regional, and the characteristic morphology of image data from different regions is quite different; (2) The applicability of landslide image recognition algorithms based on deep learning needs to be improved; (3) The comparison and discrimination methods for multi-period landslide orthophoto data need to be improved; (4) Landslide identification methods are independent of each other and urgently need to be integrated and interconnected.

[0004] In terms of automatic identification methods of landslide hazards based on remote sensing data, the Chinese invention patent with publication number CN 114821376 A discloses "a method for automatic extraction of geological hazards from drone images based on deep learning".

[0005] The Chinese invention patent with publication number CN 115661681 A discloses "a method and system for automatic identification of landslide hazards based on deep learning".

[0006] The above two patents both propose automatic identification methods and systems for landslide geological disasters. The methods described in the above two patents can, to a certain extent, solve the problems of complex identification process, low efficiency and unclear extraction of specific landslide areas in the prior art, and realize automatic identification of remote sensing data. However, there are the following shortcomings:

[0007] 1. None of the above-mentioned automatic identification methods and systems for landslide disasters have built an intelligent identification system for landslide disasters in reservoir areas. Due to the long-term water level fluctuations in the reservoir area, bank collapse and collapse occur frequently, and the bank slope drawdown zone is widely distributed. Deep learning can only be carried out through orthophoto data. It is necessary to build a complete reservoir landslide dataset covering sample sets, verification sets, test sets and datasets suitable for landslide disasters in the reservoir area. Automatic identification research based on the existing landslide sample library has problems such as insufficient matching and low identification success rate.

[0008] 2. There are certain limitations in the intelligent identification of landslide disasters based only on a single batch of orthophoto data. The damage modes of landslide bodies are divided into creep damage and collapse damage. Under different damage modes, the impact consequences of landslide disasters and the corresponding warning levels are quite different. In addition, for landslides in densely vegetated areas of the reservoir area, their representative characteristics are unclear. It is difficult to effectively identify landslide bodies damaged by creep in the reservoir area only through a single batch of orthophotos. It is necessary to compare and identify orthophoto data of multiple periods of landslides, and compare the displacement of landslide hazard bodies at different times to effectively make up for the limitations of identification based on orthophoto data.

[0009] 3. With the rapid development of remote sensing technology in recent years, airborne 3D laser scanning technology has been widely used in the investigation of high-risk areas of geological disasters and major geological disaster risk points due to its high penetration of dense vegetation and the ability to construct 3D digital models. Therefore, intelligent identification of landslide hazards based solely on single orthophoto data is increasingly unable to meet the needs of high-precision investigation of geological disaster risk points. It is necessary to combine multiple remote sensing technologies including airborne 3D laser scanning technology to build a comprehensive landslide hazard identification system.

[0010] Therefore, a method, system and storage medium based on deep learning methods that can effectively realize the intelligent identification of landslide disasters in the reservoir area under the conditions of dense vegetation in the covering layer and high spatiotemporal distribution density of the reservoir bank drawdown zone has become a scientific problem that needs to be urgently solved in the field of disaster prevention and mitigation. Summary of the invention

[0011] The technical problem to be solved by the present invention is to provide a method, system and storage medium for intelligent identification of landslide disasters in reservoir areas based on deep learning in view of the deficiencies of the above-mentioned prior art. The method, system and storage medium for intelligent identification of landslide disasters in reservoir areas are based on UAV remote sensing data and ground base station monitoring data, combined with deep learning algorithms, which can effectively improve the efficiency of risk identification, prediction and subsequent stability processing of landslide disasters in reservoir areas, and can provide guidance on disaster control, which can effectively provide effective protection for the safety of reservoir and bank engineering of major hydropower projects.

[0012] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0013] A deep learning-based intelligent identification method for landslide disasters in reservoir areas, comprising:

[0014] Step S1: Obtain basic geological information and monitoring data of the proposed study area, and combine it with the two-dimensional geological profile to make a preliminary assessment of the potential landslide susceptibility of the area;

[0015] Step S2: obtaining an orthophoto of the reservoir bank slope and a three-dimensional digital elevation model of the reservoir area within the proposed study area through UAV orthophotography and three-dimensional laser scanning technology;

[0016] Step S3: The orthophoto obtained in step S2 is subjected to feature enhancement using a sparse representation method, and on this basis, a deep learning model for reservoir area landslide disaster prediction and analysis is constructed and trained;

[0017] Step S4: inputting the orthophoto data after feature enhancement into the deep learning model trained in step S3 to obtain a preliminary recognition result; then performing comparative analysis on the orthophotos of multiple periods, calculating the displacement of the landslide body, and further identifying the landslide hazard body in the reservoir area; combining the preliminary recognition result with the displacement calculation result, obtaining the identification and prediction conclusion of the landslide body in the reservoir area, and verifying the conclusion using the three-dimensional digital elevation model;

[0018] Step S5: Based on the prediction results of step S4, the deformation and failure mode of the landslide disaster in the reservoir area is determined, and the corresponding support mode is designed in combination with the numerical simulation analysis software to provide theoretical support for subsequent disaster management.

[0019] Further, step S1 includes:

[0020] Step S11: compile and sort out the existing basic geological data and monitoring data of the proposed study area, where the basic geological data include: ① regional topography; ② regional stratum lithology; ③ regional geological structure; ④ hydrogeology; ⑤ rock weathering and unloading; ⑥ adverse geological effects; ⑦ human disturbance, etc. The monitoring data mainly include: ① reservoir water level data; ② groundwater level monitoring data; ③ surface and deep displacement monitoring data; ④ ground tilt monitoring data, etc.

[0021] Furthermore, the regional topography includes the elevation of the top of the reservoir bank slope, the elevation of the top and foot of each slope and the corresponding slope, the horizontal distance from the slope to the center of the river valley, etc.; the regional stratum lithology includes rock type, composition, structure, structure, lithofacies change, genetic type, thickness and stratum age, etc.; the regional geological structure includes earthquakes, faults, rock formation occurrence, etc.; the hydrogeology includes groundwater, surface runoff and reservoir water level, among which the groundwater in the reservoir area is mainly divided into bedrock fissure water and porous groundwater; the weathering and unloading degree of the rock mass is affected by factors such as topography, stratum lithology, geological structure, climatic conditions and groundwater; adverse geological effects mainly include landslides, reservoir bank collapse, collapse, slope deformation and debris flow, etc., mainly recording the landslide range, landslide volume, surge height and other data of historical landslides in chronological order. Human disturbance mainly focuses on material yards, tunnels and high road slopes.

[0022] Furthermore, the monitoring data takes the monthly average value of the proposed study area. For historical landslides with monitoring data, for different landslide types, the monitoring data of creep-destruction landslides takes the monthly average value, and the monitoring data of collapse-slip-destruction landslides takes the daily average value, forming monitoring data containing pressure data, deformation data and displacement data.

[0023] Step S12: After completing the above data processing and aggregation, the local geographical characteristics of the processed data are refined and analyzed in combination with the 1:50,000 geological map and the two-dimensional geological profile map, and the topography, geological structure, stratigraphic lithology, climatic conditions and vegetation coverage factors are comprehensively considered to preliminarily establish a landslide susceptibility zoning table for the proposed study area.

[0024] Further, step S2 includes:

[0025] Step S21: Select a suitable UAV model according to research needs and set camera parameters; the UAV model is preferably a compound wing UAV with vertical take-off and landing and large-range cruising functions, and can be equipped with a visible light aerial survey module, a visible light tilt module, a visible light video module, a thermal infrared camera module, a thermal infrared video module, a multi-spectral module, a hyperspectral module, a lidar module, a synthetic aperture radar module and some combined modules, and has a 1:500 large-scale mapping capability without control points, realizing image control-free applications and meeting remote sensing data acquisition requirements; the camera is preferably a high-pixel camera, which can realize remote sensing data acquisition of high-resolution orthophotos and high-density LiDAR point clouds.

[0026] Step S22: Plan the drone remote sensing data collection path, design the flight route and collection plan, and ensure that the drone completes a comprehensive inspection; the cruise route planning of the reservoir area adopts the following principles: The drone cruise route planning aims to ensure that within a specific area, the drone starts from a preset starting point, traverses all preset patrol points or areas, and finally returns to the starting point, while meeting its structural characteristics, including model, wind resistance and endurance, as well as the constraints of complex operating environment conditions. This process aims to efficiently complete specific tasks, including reservoir area monitoring and terrain mapping, and optimize flight time or cost through detailed planning, so as to ensure that the drone completes a comprehensive inspection with minimal resource consumption.

[0027] Step S23: collecting orthophotos and 3D laser scanning point cloud data of multiple times for different landslide types; after the collection is completed, resampling and image registration processing are performed on the images, and the processed orthophoto data is output.

[0028] Step S24: Based on the processed 3D laser scanning point cloud data, a 3D digital elevation model with the required accuracy is established; specifically, first, the measured horizontal coordinate data and elevation coordinate data are accurately imported into the point cloud processing software. Subsequently, the software will automatically construct a high-precision digital elevation model of the test slope based on these horizontal and elevation coordinate information.

[0029] Further, step S3 includes:

[0030] Step S31: Before constructing and training the deep learning model for reservoir landslide disaster prediction and analysis, construct a reservoir orthophoto dataset including a training set, a validation set, and a test set. Considering that the image acquisition area is mainly located along the river, with hidden landslide risk points, numerous engineering activities, and high vegetation coverage, the original data image is cropped according to the common features in this section to fix its resolution. Subsequently, these cropped images can be subdivided into five sub-datasets of "landslide", "water body", "reservoir bank", "vegetation", and "human activity", with more than 500 images selected for each subset. In the training set, 80% of each sub-dataset is randomly selected as the training set, and the validation set and test set are the remaining 20%.

[0031] Step S32: In deep learning recognition tasks, due to the large number of network parameters, a large amount of data is required to complete the training, but the sample size of the training set is often insufficient. Convolutional neural networks are invariant to scale, shift, perspective, and illumination, so it is planned to use data enhancement methods to expand the number of samples. By means of image angle transformation, mirror transformation, brightness transformation, and contrast transformation, not only the number of training samples can be increased, but also the sample diversity can be improved, and the model's dependence on specific attributes can be reduced, thereby improving the generalization ability and robustness of the model. Specifically, it is planned to rotate the original data set clockwise by 90°, 180°, and 270°, and flip it along the clockwise axis of 0°, 45°, 90°, and 315°, and adjust parameters such as brightness and contrast to expand the data set.

[0032] Step S33: Considering the vegetation coverage and surrounding engineering activities in the orthophoto of the reservoir area, the representative morphology of some geological disaster points is difficult to identify. It is necessary to perform image preprocessing through further denoising, enhancement, fusion and other technologies to improve the image quality. Then, the image features are extracted using relevant technologies to form a description and expression of the image, and then the properties of each target object in the area of ​​interest and the correlation between them are studied to achieve a higher level of image analysis and understanding. In the process of image analysis and processing, the discriminative information that is not easily disturbed by degradation factors is abstracted through analysis, and irrelevant or redundant information is removed. The existing technical means mainly include sparse representation, local feature description, etc. The sparse representation theory believes that the image can be represented by a linear combination of the columns of the selected dictionary matrix. The local feature description method extracts information with high discrimination from the area of ​​interest of the image for image representation. Since sparse representation technology has the ability to capture important information about objects of interest using very little data, it can effectively mine the sparse characteristics of the data itself and the proximity relationship between data, and has good natural discrimination ability, adaptability, and robustness to noise, the sparse representation method is preferably used for orthophoto data enhancement.

[0033] Step S34: The deep learning algorithm adopts an algorithm based on the Caffe network architecture and develops an algorithm architecture suitable for intelligent identification of landslides. The Caffe network architecture, as a model of convolutional structure for fast feature embedding networks, is rooted in the deep learning framework implemented in Python. Its core operating mechanism is that all calculation processes are abstracted into layers. Each neural network module exists as an independent layer, and its responsibility is to receive input data and output the corresponding results after complex internal calculations. These layers are connected and combined to form a complete network. When faced with the need to reconstruct the network, developers will splice each layer one by one in units of layers to build a complete network architecture. This design strategy gives us full control from the bottom to the top, from input data to classification loss, allowing us to flexibly build and optimize deep learning models.

[0034] Furthermore, in step S3: based on the feature-enhanced UAV orthophoto data, the deep learning model is optimized, and an automatic landslide identification algorithm based on the Caffe network architecture is developed, wherein the model threshold is set using an adaptive threshold method, and the threshold is dynamically adjusted according to the model performance and input data quality to improve the adaptability and accuracy of model identification.

[0035] Further, step S4 includes:

[0036] Step S41: inputting fixed-resolution and fixed-size orthophoto data into the landslide hazard deep learning model to generate preliminary identification results of landslide hazards.

[0037] Step S42: Combine the comparative analysis of the orthophoto data of multiple periods and the comprehensive processing with the preliminary recognition result of step S41 to generate the identification prediction result of the landslide body, which specifically includes the following steps:

[0038] Step S421: Based on the orthophoto data of multiple periods, the particle image velocimetry method is used for comparative analysis to extract the coordinates and displacement components in the image, and the displacement threshold is calculated accordingly;

[0039] Step S422: extracting the corresponding background image from the image according to the displacement threshold, marking the specific value and direction of the displacement in each processing window, and specially marking the calculation window whose displacement exceeds the threshold;

[0040] Step S423: using a two-dimensional polynomial surface fitting method to depict the moving area, counting the size of the moving area, and synthesizing the information obtained in steps S421 and S422 into an image for intuitively displaying the displacement analysis results;

[0041] Step S424: marking the area with a large displacement change in the grayscale data in the generated image, and delineating the landslide boundary according to the displacement change range;

[0042] Step S425: Combine the above analysis results with the preliminary identification results obtained through the deep learning model to generate the identification and prediction results of the reservoir area landslide body.

[0043] Step S43: export the landslide body identification prediction results, determine the specific coordinates of the landslide hazard point, and import the coordinate information into the three-dimensional digital elevation model to review and verify the landslide body.

[0044] Furthermore, in step S5, based on the three-dimensional digital elevation model, the main structural features of the landslide body are extracted, including the data of the landslide body boundary, the estimated volume of the landslide body, the top and toe elevation of the landslide body. Combined with the engineering geological report, the deformation and stability of the landslide body are modeled and analyzed in the numerical simulation analysis software to provide technical support for the formulation of subsequent landslide disaster control measures.

[0045] On the other hand, a deep learning-based intelligent identification system for reservoir landslide disasters is provided, which uses a deep learning-based intelligent identification method for reservoir landslide disasters for analysis and includes the following modules:

[0046] Landslide susceptibility preliminary assessment module: Through comprehensive analysis of the basic geological data, monitoring data and two-dimensional geological profiles of the proposed study area, the preliminary assessment of the landslide susceptibility in the proposed study area is completed, and the characteristics of the regional landslide susceptibility zoning are obtained;

[0047] Remote sensing data post-processing module: by post-processing the data collected by UAV remote sensing technology, it generates orthophoto data of the reservoir bank slope and three-dimensional digital elevation model with coordinate information;

[0048] Orthophoto data enhancement module: uses sparse representation method to enhance orthophoto data, highlight the characteristic structure of landslide in orthophoto data, and provide pre-processing data for depth recognition module;

[0049] Deep learning module: It uses an algorithm architecture based on the Caffe network architecture that is suitable for intelligent identification of landslides. By inputting orthophoto data after data enhancement, it generates preliminary identification results of landslide hazards in the reservoir area.

[0050] Multi-phase orthophoto comparison and analysis module: Particle image velocimetry is used to compare and analyze multi-phase orthophoto data, marking areas with large displacement changes in the grayscale data of orthophoto images, delineating landslide boundaries based on the range of displacement changes, and combining the preliminary recognition results obtained by the deep learning model to generate reservoir area landslide body identification and prediction results;

[0051] Landslide hazard identification summary and analysis module: Summarize the results obtained by the above modules to obtain the intelligent identification results of landslide hazards in the reservoir area, and the results are output in the form of text reports and images.

[0052] Finally, a computer storage medium is proposed, which stores a computer program executable on a cloud computing platform. When the program is executed by a processor, it can implement a series of key steps of a deep learning-based intelligent identification method for reservoir landslide disasters, thereby providing an efficient and intelligent solution for disaster prevention and management.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The present invention comprehensively considers the difficulties faced by intelligent identification of landslide disasters in reservoir areas. First, the susceptibility of regional landslide disasters including regional geological surveys is preliminarily determined to obtain the characteristics of regional landslide-prone zoning. Further, based on the deep learning method, the orthophoto and three-dimensional laser scanning point cloud remote sensing data are combined to summarize the prediction results of landslide body identification in the reservoir area. Finally, the support mode is formulated through numerical simulation analysis software to provide theoretical guidance for subsequent disaster management. The present invention has a strong pertinence in the intelligent identification of landslide disasters in reservoir areas, and the identification conclusion is more reliable.

[0055] 2. The deep learning-based intelligent identification method, system and storage medium for reservoir area landslide disasters proposed in the present invention have the characteristics of high applicability and high identification success rate for the intelligent identification of reservoir area landslide disasters.

[0056] 3. The remote sensing data involved in the present invention include orthophoto data and three-dimensional laser scanning data, which can effectively combine the advantages of the two remote sensing technologies in the early identification of geological disasters. It has the characteristics of large coverage area and low cost of orthophotos and high accuracy and high texture detail of three-dimensional laser scanning point clouds, and can realize the early identification of wide-area geological disasters in the reservoir area.

[0057] 4. The present invention has greatly enriched the system of early identification methods for geological disasters in reservoir areas in the field of disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the technical roadmap for the invention of the intelligent identification method, system and storage medium for reservoir area landslide disasters based on deep learning.

[0059] Figure 2 It is a flow chart of the invention of the intelligent identification method, system and storage medium of reservoir area landslide disasters based on deep learning.

[0060] Figure 3 This is an example of a sub-dataset image of the reservoir area orthophoto database of the present invention.

[0061] Figure 4 This is an example of image data input according to the present invention.

[0062] Figure 5 This is an example of image data output according to the present invention.

[0063] Figure 6 This is an example of landslide body verification of the three-dimensional digital elevation model of the present invention. DETAILED DESCRIPTION

[0064] Hereinafter, the technical solution of the present invention will be described in detail and clearly with the aid of the accompanying drawings in the embodiments of the present invention. It should be clear that the embodiments described are only a part of the present invention, not all of it. Based on the embodiments presented in the present invention, all other embodiments that can be derived by professionals and technicians in this field without the need for creative work will be included in the protection scope of the present invention.

[0065] The present invention is further described in conjunction with the following application scenarios.

[0066] Figure 1 It is a schematic diagram of the technical process of the present invention, which is used to illustrate the intelligent identification method of reservoir landslide disasters based on deep learning and its specific implementation process.

[0067] See also Figure 2 The embodiment shows a method, system and storage medium for intelligent identification of landslide disasters in reservoir areas based on deep learning, including:

[0068] Step S1: Obtain basic geological information and monitoring data of the proposed study area, and combine it with the two-dimensional geological profile to make a preliminary assessment of the potential landslide susceptibility of the area;

[0069] In one implementation, step S1 includes:

[0070] Step S11: cleaning, classifying and structuring the existing basic geological data and monitoring data in the proposed study area, and summarizing and archiving the processed data;

[0071] The basic geological data include: ① regional topography; ② regional stratum lithology; ③ regional geological structure; ④ hydrogeology; ⑤ rock weathering and unloading; ⑥ adverse geological effects; ⑦ human disturbance, etc. The monitoring data mainly include: ① reservoir water level data; ② groundwater level monitoring data; ③ surface and deep displacement monitoring data; ④ ground tilt monitoring data, etc.

[0072] The specific contents of each part of the basic geological data are as follows: ① Regional topography includes: the elevation of the top of the reservoir bank slope, the elevation of the top and foot of each slope and the corresponding slope, the horizontal distance from the slope to the center of the valley, etc.;

[0073] ② Regional stratigraphic lithology includes: rock type, composition, structure, texture, lithofacies change, genetic type, thickness and stratigraphic age, etc.;

[0074] ③ Regional geological structure including earthquakes, faults, rock formation occurrence, etc.;

[0075] ④ Hydrogeology includes: groundwater, surface runoff and reservoir water level, among which the groundwater in the reservoir area is mainly divided into bedrock fissure water and porous groundwater;

[0076] ⑤ The degree of rock weathering and unloading is affected by factors such as topography, stratum lithology, geological structure, climatic conditions and groundwater;

[0077] ⑥ Adverse geological effects include: landslides, reservoir bank collapse, collapse, slope deformation and debris flow, etc., mainly recording the landslide range, landslide volume, surge height and other data of historical landslides in chronological order;

[0078] ⑦ Human disturbance mainly focuses on material yards, tunnels and high slopes of highways;

[0079] Among them, the monitoring data should be collected according to the following principles. The main monitoring data include ① reservoir water level data, ② groundwater level monitoring data, ③ surface and deep displacement monitoring data, and ④ ground tilt monitoring data. The monthly average value of the proposed study area is taken. For historical landslides with monitoring data, for different landslide types, the creep-destruction landslide monitoring data is taken as the monthly average value, and the collapse-destruction landslide monitoring data is taken as the daily average value, forming a monitoring data set containing pressure data, deformation data and displacement data.

[0080] Step S12: After completing the above data processing and aggregation, the local geographical characteristics of the processed data are refined and analyzed in combination with the 1:50,000 geological map and the two-dimensional geological profile map, and the topography, geological structure, stratigraphic lithology, climatic conditions and vegetation coverage factors are comprehensively considered to preliminarily establish a landslide susceptibility zoning table for the proposed study area.

[0081] Step S2: Obtain the orthophoto image and 3D digital elevation model of the reservoir bank slope in the proposed study area through UAV orthophotography and 3D laser scanning technology, including:

[0082] Step S21: Select a suitable drone model according to research requirements and set camera parameters;

[0083] The drone model selected is Pegasus V10 composite wing drone, with a wingspan of 4150m and a length of 1750mm. It has vertical take-off and landing and wide-range cruising functions, and can be equipped with visible light aerial survey module, visible light tilt module, visible light video module, thermal infrared camera module, thermal infrared video module, multi-spectral module, hyperspectral module, lidar module, synthetic aperture radar module and some combination modules. It has the ability to generate large-scale maps of 1:500 without control points, realizes image control-free applications, and meets the requirements of remote sensing data collection;

[0084] The camera selected is the Phase One 100-megapixel camera, which can realize remote sensing data acquisition of high-resolution orthophotos and high-density 3D laser scanning point clouds.

[0085] Step S22: planning the drone remote sensing data collection path, designing the flight route and collection plan, and ensuring that the drone completes a comprehensive inspection;

[0086] The purpose of drone cruise route planning is to ensure that within a specific area, the drone starts from a preset starting point, traverses all preset patrol points or areas, and finally returns to the starting point, while meeting its structural characteristics, including aircraft type, wind resistance and endurance, as well as the constraints of complex operating environment conditions. This process is designed to efficiently complete specific tasks, including reservoir area monitoring and terrain mapping, and optimize flight time or cost through detailed planning, thereby ensuring that the drone completes a comprehensive inspection with minimal resource consumption.

[0087] Step S23: collecting orthophotos and 3D laser scanning point cloud data of multiple times for different landslide types;

[0088] Orthophotos of different periods are collected for the proposed study area, and orthophotos of different periods are collected according to different landslide types. In this embodiment, the number of periods is 2. After the collection is completed, 0.08m tif format orthophoto data is output after resampling and image registration.

[0089] Step S24: establishing a three-dimensional digital elevation model with accuracy meeting the requirements based on the processed three-dimensional laser scanning point cloud data;

[0090] The acquisition of the 3D digital elevation model relies on point cloud processing software, which can generate a 3D digital elevation model from the collected 3D laser scanning point cloud data. Specifically, first, the measured horizontal coordinate data and elevation coordinate data are accurately imported into the point cloud processing software. Then, the software will automatically construct a high-precision 3D digital elevation model of the test slope based on these horizontal and elevation coordinate information.

[0091] Step S3: The orthophoto obtained in step S2 is subjected to feature enhancement using a sparse representation method, and on this basis, a deep learning model for reservoir area landslide disaster prediction and analysis is constructed and trained, including:

[0092] Step S31: constructing orthophoto dataset;

[0093] Before building and training a deep learning model for reservoir landslide disaster prediction and analysis, it is necessary to build a complete reservoir remote sensing dataset including training set, validation set and test set. Considering that the image acquisition area is mainly located along the river bank, with hidden landslide risk points, numerous engineering activities and high vegetation coverage, the original data image is cropped according to the common features in this section to fix its resolution. Subsequently, these cropped images can be subdivided into five sub-datasets: "landslide", "water body", "reservoir bank", "vegetation" and "human activity", such as Figure 3 As shown. Each subset has more than 500 images. In the training set, 80% of each sub-dataset is randomly selected as the training set, and the validation set and test set are the remaining 20%.

[0094] Step S32: expanding the orthophoto dataset;

[0095] In deep learning recognition tasks, due to the large number of network parameters, a large amount of data is required to complete the training, but the sample size of the training set is often insufficient. Convolutional neural networks are invariant to scale, shift, perspective, and illumination, so it is planned to use data enhancement methods to expand the number of samples. By means of image angle transformation, mirror transformation, brightness transformation, and contrast transformation, not only can the number of training samples be increased, but also the sample diversity can be improved, and the model's dependence on specific attributes can be reduced, thereby improving the generalization ability and robustness of the model. Specifically, it is planned to rotate the original data set clockwise by 90°, 180°, and 270°, and flip it along the clockwise axis of 0°, 45°, 90°, and 315°, and adjust parameters such as brightness and contrast to expand the data set.

[0096] Step S33: enhancing the orthophoto dataset;

[0097] Considering that the orthophotos of the reservoir area are covered by vegetation and affected by surrounding engineering activities, the representative morphology of some geological disaster points is difficult to identify. It is necessary to perform image preprocessing through further denoising, enhancement, fusion and other technologies to improve the image quality. Then, relevant technologies are used to extract image features to form a description and expression of the image, and then study the properties of each target object in the area of ​​interest and the correlation between them, so as to achieve a higher level of image analysis and understanding. In the process of image analysis and processing, the discriminative information that is not easily disturbed by degradation factors is abstracted through analysis, and irrelevant or redundant information is removed. The existing technical means mainly include sparse representation, local feature description, etc. Sparse representation theory believes that images can be represented by linear combinations of columns of selected dictionary matrices. The local feature description method extracts information with high discrimination from the area of ​​interest of the image for image representation. Since sparse representation technology has the ability to capture important information of the object of interest with very little data, it can effectively mine the sparse characteristics of the data itself and the neighboring relationship between data, and has good natural discrimination ability, adaptability and robustness to noise. Sparse representation method is used to enhance orthophoto data.

[0098] Sparse representation is based on the basic fact that a signal can be represented by some non-zero sparse coefficients in a suitable dictionary matrix. For a given signal x and an underdetermined matrix D, the sparse representation of x is defined as searching for the sparsest possible representation that satisfies α:

[0099] x=D·α (1)

[0100] The local linear model between the input image I and the corresponding dictionary (single image) M is defined as follows:

[0101] I(i)=α k M(i)+ε k ,i∈ω k (2)

[0102] Where i = (x, y) represents the image block ω centered at pixel k k Therefore, I(i) and M(i) represent the pixel points from the input image and the corresponding image patch in the dictionary, respectively. k and ε k is the linear representation coefficient, assuming that the representation error is in ω k is a constant. Further optimization of the objective equation can be performed to find each ω k inside a k and ε k The optimal solution is:

[0103]

[0104] in Through this algorithm, an efficient method for enhancing the features of landslide orthophotos is realized.

[0105] Step S34: constructing a deep learning model for landslide disasters;

[0106] The deep learning algorithm adopts an algorithm based on the Caffe network architecture and develops an algorithm architecture suitable for intelligent identification of landslides. Based on the Caffe network architecture, as a model of convolutional structure for fast feature embedding network, it is rooted in the deep learning framework implemented in Python. Its core operating mechanism is that all calculation processes are abstracted into layers. Each neural network module exists as an independent layer, and its responsibility is to receive input data, and after complex internal calculations, output the corresponding results. These are constructed into a complete network through carefully designed connections and combinations. When faced with the need to reconstruct the network, developers will splice each layer one by one to build a complete network architecture. This design strategy gives us full control from the bottom to the top, from input data to classification loss, allowing us to flexibly build and optimize deep learning models.

[0107] Furthermore, the deep learning model is optimized based on the data source, and the data source is selected as the feature-enhanced UAV orthophoto, and the landslide automatic identification algorithm based on the Caffe network architecture is developed. The adaptive threshold method is selected for model threshold setting. The threshold can be dynamically adjusted to adapt to the current model performance and input data quality.

[0108] Step S4: Input the orthophoto data after feature enhancement into the deep learning model trained in step S3 to obtain preliminary recognition results; then compare and analyze the orthophotos of multiple periods, calculate the displacement of the landslide body, and further identify the landslide hazard body in the reservoir area; combine the preliminary recognition results with the displacement calculation results to obtain the identification and prediction conclusion of the landslide body in the reservoir area, and use the three-dimensional digital elevation model to review the conclusion, including:

[0109] Step S41: Input Figure 4 The fixed resolution and fixed size orthophoto data shown are fed into the landslide hazard deep learning model to obtain preliminary recognition results;

[0110] Step S42: by comparing the orthophoto impact data of multiple periods and integrating the preliminary identification results of step S41, a landslide identification prediction result is obtained, which specifically includes the following steps:

[0111] Step S421: Based on the orthophoto data of multiple periods, the particle image velocimetry method is used for comparative analysis to extract the coordinates and displacement components in the image, and the displacement threshold is calculated accordingly;

[0112] Step S422: extracting the corresponding background image from the image according to the displacement threshold, marking the specific value and direction of the displacement in each processing window, and specially marking the calculation window whose displacement exceeds the threshold;

[0113] Step S423: using a two-dimensional polynomial surface fitting method to depict the moving area, counting the size of the moving area, and synthesizing the information obtained in steps S421 and S422 into an image for intuitively displaying the displacement analysis results;

[0114] Step S424: marking the area with a large displacement change in the grayscale data in the generated image, and delineating the landslide boundary according to the displacement change range;

[0115] Step S425: Combine the above analysis results with the preliminary recognition results obtained through the deep learning model to obtain the following Figure 5 The landslide identification and prediction results in the reservoir area are shown.

[0116] Step S43: Export the landslide body identification prediction results, determine the specific coordinates of the landslide hazard point, and import the coordinate information into the three-dimensional digital elevation model to review and verify the landslide body. Figure 6 shown.

[0117] Step S5: Based on the prediction results of step S4, the deformation and failure mode of the landslide disaster in the reservoir area is determined, and the corresponding support mode is designed in combination with the numerical simulation analysis software to provide theoretical support for subsequent disaster management.

[0118] Based on the three-dimensional digital elevation model, the main structural characteristics of the landslide body are extracted, including the landslide body boundary, estimated volume, landslide body top and slope foot elevation data. Combined with the engineering geological report, the deformation and stability of the landslide body are modeled and analyzed in the numerical simulation analysis software, providing technical support for the formulation of subsequent landslide disaster control measures.

[0119] On the other hand, a deep learning-based intelligent identification system for reservoir landslide disasters is provided, which uses a deep learning-based intelligent identification method for reservoir landslide disasters for analysis and includes the following modules:

[0120] Landslide susceptibility preliminary assessment module: Through comprehensive analysis of the basic geological data, monitoring data and two-dimensional geological profiles of the proposed study area, the preliminary assessment of the landslide susceptibility in the proposed study area is completed, and the characteristics of the regional landslide susceptibility zoning are obtained;

[0121] Remote sensing data post-processing module: by post-processing the data collected by UAV remote sensing technology, it generates orthophoto data of the reservoir bank slope and three-dimensional digital elevation model with coordinate information;

[0122] Orthophoto data enhancement module: uses sparse representation method to enhance orthophoto data, highlight the characteristic structure of landslide in orthophoto data, and provide pre-processing data for depth recognition module;

[0123] Deep learning module: It uses an algorithm architecture based on the Caffe network architecture that is suitable for intelligent identification of landslides. By inputting orthophoto data after data enhancement, it generates preliminary identification results of landslide hazards in the reservoir area.

[0124] Multi-phase orthophoto comparison and analysis module: Particle image velocimetry is used to compare and analyze multi-phase orthophoto data, marking areas with large displacement changes in the grayscale data of orthophoto images, delineating landslide boundaries based on the range of displacement changes, and combining the preliminary recognition results obtained by the deep learning model to generate reservoir area landslide body identification and prediction results;

[0125] Landslide hazard identification summary and analysis module: Summarize the results obtained by the above modules to obtain the intelligent identification results of landslide hazards in the reservoir area, and the results are output in the form of text reports and images.

[0126] Finally, a computer storage medium is proposed, which stores a computer program executable on a cloud computing platform. When the program is executed by a processor, it can implement a series of key steps of a deep learning-based intelligent identification method for reservoir landslide disasters, thereby providing an efficient and intelligent solution for disaster prevention and management.

[0127] It should be emphasized that the above embodiments are only intended to illustrate the technical solutions of the present invention, and are not intended to limit the protection scope of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent identification of landslide disasters in reservoir areas based on deep learning, characterized in that: include: Step S1: Obtain basic geological information and monitoring data of the proposed study area, and combine it with the two-dimensional geological profile to make a preliminary assessment of the potential landslide susceptibility of the area; Step S2: obtaining an orthophoto of the reservoir bank slope and a three-dimensional digital elevation model of the reservoir area within the proposed study area through UAV orthophotography and three-dimensional laser scanning technology; Step S3: The orthophoto obtained in step S2 is subjected to feature enhancement using a sparse representation method, and on this basis, a deep learning model for reservoir area landslide disaster prediction and analysis is constructed and trained; Step S4: inputting the orthophoto data after feature enhancement into the deep learning model trained in step S3 to obtain a preliminary recognition result; then performing comparative analysis on the orthophotos of multiple periods, calculating the displacement of the landslide body, and further identifying the landslide hazard body in the reservoir area; combining the preliminary recognition result with the displacement calculation result, obtaining the identification and prediction conclusion of the landslide body in the reservoir area, and verifying the conclusion using the three-dimensional digital elevation model; Step S5: Based on the prediction results of step S4, the deformation and failure mode of the landslide disaster in the reservoir area is determined, and the corresponding support mode is designed in combination with the numerical simulation analysis software to provide theoretical support for subsequent disaster management.

2. According to the deep learning-based intelligent identification method for reservoir landslide disasters in claim 1, it is characterized in that: Step S1 includes: Step S11: clean, classify and structure the existing basic geological data and monitoring data in the proposed study area, and summarize and archive the processed data, wherein the basic geological data include: ① regional topography; ② regional stratum lithology; ③ regional geological structure; ④ hydrogeology; ⑤ rock weathering and unloading; ⑥ adverse geological effects; ⑦ human disturbance; monitoring data include: ① reservoir water level data; ② groundwater level monitoring data; ③ surface and deep displacement monitoring data; ④ ground tilt monitoring data; Step S12: After completing the above data processing and aggregation, the local geographical characteristics of the processed data are refined and analyzed in combination with the 1:50,000 geological map and the two-dimensional geological profile map, and the topography, geological structure, stratigraphic lithology, climatic conditions and vegetation coverage factors are comprehensively considered to preliminarily establish a landslide susceptibility zoning table for the proposed study area.

3. The method for intelligent identification of landslide disasters in reservoir areas based on deep learning according to claim 1 is characterized in that: Step S2 includes: Step S21: Select a suitable drone model according to research requirements and set camera parameters; Step S22: planning the drone remote sensing data collection path, designing the flight route and collection plan, and ensuring that the drone completes a comprehensive inspection; Step S23: collecting orthophotos and 3D laser scanning point cloud data of multiple times for different landslide types; after the collection is completed, resampling and image registration processing are performed on the images, and the processed orthophoto data is output; Step S24: Based on the processed three-dimensional laser scanning point cloud data, a three-dimensional digital elevation model with accuracy that meets the requirements is established.

4. The method for intelligent identification of landslide disasters in reservoir areas based on deep learning according to claim 1 is characterized in that: Step S3 includes: Step S31: constructing a reservoir area orthophoto dataset including a training set, a validation set and a test set; Step S32: expanding the orthophoto data set through data enhancement techniques, including: image angle transformation, mirror transformation, brightness transformation and contrast transformation; Step S33: using a sparse representation method to enhance the orthophoto data, and mining the sparse characteristics of the data and its neighboring relationships; Step S34: construct a deep learning algorithm architecture suitable for intelligent identification of landslides based on the algorithm of the Caffe network architecture, and perform model training and optimization.

5. The method for intelligent identification of landslide disasters in reservoir areas based on deep learning according to claim 1 is characterized in that: In step S3: based on the UAV orthophoto data processed with feature enhancement, the deep learning model is optimized, and an automatic landslide identification algorithm based on the algorithm of the Caffe network architecture is developed, wherein the model threshold is set using an adaptive threshold method, and the threshold is dynamically adjusted according to the model performance and the quality of the input data.

6. The method for intelligent identification of landslide disasters in reservoir areas based on deep learning according to claim 1 is characterized in that: Step S4 includes: Step S41: inputting fixed-resolution and fixed-size orthophoto data into a landslide hazard deep learning model to generate a preliminary recognition result of landslide hazard; Step S42: combining the comparative analysis of the orthophoto data of multiple periods with the preliminary recognition result of step S41, and generating the identification prediction result of the landslide body; Step S43: export the landslide body identification prediction results, determine the specific coordinates of the landslide hazard point, and import the coordinate information into the three-dimensional digital elevation model to review and verify the landslide body.

7. The method for intelligent identification of landslide disasters in reservoir areas based on deep learning according to claim 5 is characterized in that: Step S42 includes: Step S421: Based on the orthophoto data of multiple periods, the particle image velocimetry method is used for comparative analysis to extract the coordinates and displacement components in the image, and the displacement threshold is calculated accordingly; Step S422: extracting the corresponding background image from the image according to the displacement threshold, marking the specific value and direction of the displacement in each processing window, and specially marking the calculation window whose displacement exceeds the threshold; Step S423: using a two-dimensional polynomial surface fitting method to depict the moving area, counting the size of the moving area, and combining the information obtained in steps S421 and S422 into an image; Step S424: marking the area with a large displacement change in the grayscale data in the generated image, and delineating the landslide boundary according to the displacement change range; Step S425: Combine the above analysis results with the preliminary identification results obtained through the deep learning model to generate the identification and prediction results of the reservoir area landslide body.

8. The method for intelligent identification of landslide disasters in reservoir areas based on deep learning according to claim 1 is characterized in that: In step S5, based on the three-dimensional digital elevation model, the main structural features of the landslide body are extracted, including the data of the landslide body boundary, the estimated volume of the landslide body, the top and the toe elevation of the landslide body. Combined with the engineering geological report, the deformation and stability of the landslide body are modeled and analyzed in the numerical simulation analysis software to provide technical support for the formulation of subsequent landslide disaster control measures.

9. A deep learning-based intelligent identification system for landslide disasters in reservoir areas, characterized in that: The method for intelligent identification of landslide disasters in reservoir areas based on deep learning as described in any one of claims 1 to 7 is used for analysis, comprising the following modules: Landslide susceptibility preliminary assessment module: Through comprehensive analysis of the basic geological data, monitoring data and two-dimensional geological profiles of the proposed study area, the preliminary assessment of the landslide susceptibility in the proposed study area is completed, and the characteristics of the regional landslide susceptibility zoning are obtained; Remote sensing data post-processing module: by post-processing the data collected by UAV remote sensing technology, it generates orthophoto data of the reservoir bank slope and three-dimensional digital elevation model with coordinate information; Orthophoto data enhancement module: uses sparse representation method to enhance orthophoto data, highlight the characteristic structure of landslide in orthophoto data, and provide pre-processing data for depth recognition module; Deep learning module: It uses an algorithm architecture based on the Caffe network architecture that is suitable for intelligent identification of landslides. By inputting orthophoto data after data enhancement, it generates preliminary identification results of landslide hazards in the reservoir area. Multi-phase orthophoto comparison and analysis module: Particle image velocimetry is used to compare and analyze multi-phase orthophoto data, marking areas with large displacement changes in the grayscale data of orthophoto images, delineating landslide boundaries based on the range of displacement changes, and combining the preliminary recognition results obtained by the deep learning model to generate reservoir area landslide body identification and prediction results; Landslide hazard identification summary and analysis module: Summarize the results obtained by the above modules to obtain the intelligent identification results of landslide hazards in the reservoir area, and the results are output in the form of text reports and images.

10. A computer storage medium, characterized in that: The computer storage medium stores a computer program executable on a cloud computing platform. When the program is executed by a processor, the key steps of the deep learning-based intelligent identification method for reservoir landslide disasters described in claims 1 to 7 can be implemented.

Images