Landslide identification method and system for collapsible loess slope, terminal and storage medium

Abstract

The invention provides a landslide identification method and system for a collapsible loess slope, a terminal and a storage medium, and the method comprises the steps: inputting a loess slope collection image into a landslide identification model, carrying out the haze concentration estimation and defogging processing, obtaining a collection recovery image, carrying out the feature extraction of the collection recovery image, and obtaining a collection feature map; carrying out landslide identification on the loess slope collection image according to the collection feature map to obtain a collection identification result, and carrying out parameter updating on the landslide identification model according to the collection identification result; and inputting the to-be-identified slope image into the converged landslide identification model for landslide identification to obtain a landslide identification result. According to the embodiment of the invention, defogging processing is carried out on the loess slope acquisition image through the haze estimation value, so that the converged landslide identification model can effectively carry out defogging on the to-be-identified slope image, and the landslide identification accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of landslide identification, and in particular to a landslide identification method and system for collapsible loess slopes, a terminal, and a storage medium. BACKGROUND

[0002] Collapsible loess landslides are typical geological disasters in the Loess Plateau region, characterized by steep slopes, dense rhombic cracks, pits and through cracks, and other natural phenomena that slide down the slope under the action of gravity. Landslides have strong destructive power and pose a great threat to the ecological environment, transportation, and building sites, resulting in a large number of casualties and property losses. Therefore, the problem of landslide identification for collapsible loess slopes is increasingly valued by people.

[0003] In the existing landslide identification process for collapsible loess slopes, features are identified for collected slope images based on unmanned aerial vehicle remote sensing images combined with deep learning technology. However, the collected slope images are easily affected by fog and haze, resulting in low landslide identification accuracy. SUMMARY

[0004] The purpose of the present application is to provide a landslide identification method and system for collapsible loess slopes, a terminal, and a storage medium to solve the problem of low landslide identification accuracy for collapsible loess slopes in the prior art.

[0005] The present application is implemented as follows: a landslide identification method for collapsible loess slopes, the method comprising: acquiring a loess slope collection image and inputting the loess slope collection image into a landslide identification model for fog and haze concentration estimation to obtain a fog and haze estimation value; performing fog and haze removal processing on the loess slope collection image according to the fog and haze estimation value to obtain a collection recovery image, and performing feature extraction on the collection recovery image to obtain a collection feature map; performing landslide identification on the loess slope collection image according to the collection feature map to obtain a collection identification result, and determining a model loss according to the collection identification result; performing parameter updating on the landslide identification model according to the model loss until the landslide identification model converges; acquiring a to-be-identified slope image and inputting the to-be-identified slope image into the converged collapsible loess landslide identification model for landslide identification to obtain a landslide identification result.

[0006] Preferably, the collection recovery image is subjected to feature extraction to obtain a collection feature map, comprising: perform image segmentation on the collected recovery image to obtain a sample segmentation image, and perform global average pooling on the sample segmentation image according to a channel attention branch in the landslide identification model to obtain a sample pooling feature map; calculate a channel attention feature map according to the sample pooling feature map; perform group normalization processing on the sample segmentation image according to a spatial attention branch in the landslide identification model to obtain a sample normalization feature map, and perform full connection processing on the sample normalization feature map to obtain a spatial attention feature map; combine the spatial attention feature map and the channel attention feature map to obtain a sample combined feature, and perform channel shuffle on the sample combined feature to obtain the collected feature map.

[0007] Preferably, the formula for calculating the channel attention feature map comprises: wherein, denotes the channel attention feature map, denotes the sample segmentation image in the channel attention branch, denotes a weight parameter in the channel attention branch, denotes a bias parameter in the channel attention branch, denotes the sample pooling feature map, denotes a Sigmoid activation function.

[0008] Preferably, the formula for calculating the spatial attention feature map comprises: wherein, denotes the spatial attention feature map, denotes the sample segmentation image in the spatial attention branch, denotes a weight parameter in the spatial attention branch, denotes a bias parameter in the spatial attention branch, denotes group normalization processing.

[0009] Preferably, the formula for inputting the loess slope collected image into the landslide identification model for haze concentration estimation comprises: wherein, denotes the haze estimation value, denotes the loess slope collected image, A denotes the illumination intensity of atmospheric light, denotes image transmittance, b denotes a constant bias; wherein, β denotes a scattering coefficient, d denotes a distance between an imaging object and an imaging device in the loess slope collection image, x denotes a natural constant. e

[0010] Preferably, the formula for dehazing processing of the loess slope collection image according to the haze estimation value comprises: wherein, denotes the collection recovery image.

[0011] Preferably, the model loss is determined according to the collection recognition result, comprising: performing non-maximum suppression on a target bounding box in the collection recognition result to obtain a sample prediction bounding box, and performing loss calculation according to a bounding box similarity between the sample prediction bounding box and a true landslide bounding box of the loess slope collection image to obtain a first loss; performing loss calculation according to a predicted probability distribution in the collection recognition result and a true predicted probability distribution of the loess slope collection image to obtain a second loss; performing weighted value operation on the first loss and the second loss to obtain the model loss.

[0012] Another purpose of the embodiment of the present application is to provide a landslide recognition system for collapsible loess slope, comprising: an image collection module, configured to obtain a loess slope collection image; a haze estimation module, configured to input the loess slope collection image into a landslide recognition model to estimate haze concentration and obtain a haze estimation value; a dehazing processing module, configured to perform dehazing processing on the loess slope collection image according to the haze estimation value to obtain a collection recovery image; a feature extraction module, configured to perform feature extraction on the collection recovery image to obtain a collection feature map; a model training module, configured to perform landslide recognition on the loess slope collection image according to the collection feature map to obtain a collection recognition result, and determine a model loss according to the collection recognition result; performing parameter update on the landslide recognition model according to the model loss until the landslide recognition model converges; a landslide recognition module, configured to obtain a to-be-recognized slope image, and input the to-be-recognized slope image into the converged landslide recognition model to perform landslide recognition and obtain a landslide recognition result. ​

[0013] The loess slope image collected by the loess slope image collection device is input into the landslide identification model to estimate the haze concentration, so that the haze estimation value of the loess slope image collected can be effectively calculated, the loess slope image collected is processed by the haze estimation value, the converged landslide identification model can effectively dehaze the slope image to be identified, the influence of the haze on the landslide identification is prevented, and the landslide identification accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a flowchart of the landslide identification method of the collapsible loess slope provided by the first embodiment of the present application; Figure 2 is a schematic diagram of the K(x) value estimation module provided by the first embodiment of the present application; Figure 3 is a schematic diagram of the extracted collection feature map provided by the first embodiment of the present application; Figure 4 is a structural schematic diagram of the landslide identification system of the collapsible loess slope provided by the second embodiment of the present application; Figure 5 is a structural schematic diagram of the terminal device provided by the third embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0016] In order to illustrate the technical solutions of the present application, the following specific embodiments are used to illustrate the technical solutions of the present application.

[0017] Embodiment One Please refer to Figure 1 is a flowchart of the landslide identification method of the collapsible loess slope provided by the first embodiment of the present application. The landslide identification method of the collapsible loess slope can be applied to any device or system. The landslide identification method of the collapsible loess slope includes the following steps: Step S10, acquiring a loess slope image, and inputting the loess slope image into a landslide identification model to estimate the haze concentration, to obtain a haze estimation value; The data set is made up of unmanned aerial vehicle image data, small-size simulation experiment pictures, and network landslide pictures. Data augmentation is performed on the data set using Matlab, including enlargement, reduction, rotation, adjustment of brightness and contrast, mirroring, addition of noise, and pixel distortion. The data is expanded to 1600 images, and the loess slope collection images are obtained. For the missing parts of the loess slope collection images, white filling is used to restore them to their original size and shape, enriching the data set. In this step, the loess slope collection images that are blurred and whitened due to water mist after rainfall and light are also dehazed.

[0018] To meet the high-precision data collection needs of collapsible loess areas, unmanned aerial vehicles are used to collect image data, with a flight height of 50-80 meters (adjusted dynamically according to terrain undulations) to ensure a ground resolution of 2-3 centimeters per pixel, accurately capturing surface micro-cracks and collapsible deformation features. The flight path planning uses a "grid zigzag" automatic cruising mode with a heading overlap rate of ≥85% and a lateral overlap rate of ≥75%. A 15% buffer zone is added to the target area to eliminate edge distortion. For gully and steep slope areas, segmented flight (local height reduced to 30 meters) is added, and key hidden danger points are manually supplemented.

[0019] The training set and test set in the loess slope collection images are allocated in a 2:8 ratio. Based on the characteristics of each stage of landslide hazards, all processed images are labeled. Due to liquefaction effects, collapsible loess landslides are different from ordinary landslides. They directly enter the high-speed motion stage without a noticeable expansion stage. Liquefaction is mainly due to the special physical properties of loess. When loess is saturated with water, soil particles lose their original contact structure under external forces, resulting in a decrease in shear strength. This allows the landslide to quickly transition from an unstable state to a high-speed motion state, skipping the slow expansion stage commonly seen in ordinary landslides. Therefore, the self-made data set has only four labels: steep slope, pit-through crack, and dense diamond-shaped crack. The label details are as follows: (1) Steep slope label: Based on the characteristics of the slope, when the slope is less than 45 degrees, the dominant deformation mode is collapsible deformation; when the slope is greater than 45 degrees, the landslide becomes the dominant failure mode. This label represents the potential landslide area.

[0020] (2) Dense diamond-shaped crack label: In the early stage of loess collapsible landslide, the soil body begins to show signs of internal stress redistribution. The formation of diamond-shaped cracks is a manifestation of this stress change, which is formed by the soil body experiencing uneven settlement. The appearance of such cracks is an early warning signal of the gradual damage to the stability of the slope, and can be used to determine that the loess collapsible landslide is in the premonitory stage.

[0021] (3) Pit and through crack label: Pit is formed by local collapse and erosion of soil under the action of water. Through crack is the evidence of continuous shear and separation of soil. It shows that the soil has lost overall stability and started to move, which is a clear sign of active landslide and can be used to determine the sliding stage of loess collapsible landslide.

[0022] In this embodiment, the landslide identification model adopts an improved YOLO v8 network structure (YOLO v8 pro). A large-scale detection head is added to the YOLOv8 pro model, which can improve the detection effect of small size targets. The principle is mainly based on multi-scale feature fusion and detail enhancement mechanism. Small targets are easily weakened or even lost by the downsampling operation of deep network in the conventional detection head due to less pixel information and low feature abstraction level. By introducing a large-scale detection head, the landslide identification model can retain more spatial detail information in the shallow network, and the large-scale detection head corresponds to a larger anchor density. It can generate small size anchor boxes densely on high resolution feature maps, improve the sensitivity of target positioning, and reduce the possibility of missing.

[0023] In this step, a dehazing enhanced detection framework based on UnfogNet network is introduced to improve the robustness and accuracy of the model in foggy scenes by combining dehazing and detection tasks. UnfogNet network is directly embedded in the landslide identification model, which is an end-to-end deep learning dehazing network that can effectively learn the mapping relationship from foggy images to clear images without relying on intermediate parameter estimation in the atmospheric scattering model. UnfogNet network includes K(x) value estimation module and clear image generation module.

[0024] Please refer to Figure 2 , K ( x ) value estimation module is used to model the fog density distribution, K ( x ) value prediction network adopts a hierarchical architecture containing 5 convolution modules and 3 fully connected layers. Through multi-scale convolution kernel design, different size filters are deployed in parallel in each feature extraction layer to capture diverse spatial information. Especially embedding feature fusion channels between adjacent convolution layers, this structure not only compensates for feature loss in the convolution process through cross-layer connection, but also builds a progressive conversion path from primary texture features to high-level semantic features, effectively enhancing the network's ability to handle complex fog distribution.

[0025] Optionally, the formula used for fog density estimation by inputting the loess slope image into the landslide identification model includes: wherein, represents the fog estimation value, This indicates that the image of the loess slope was captured in foggy weather, and therefore is a foggy image. A This represents the intensity of atmospheric light, and is generally a constant. t ( x () represents image transmittance, used to describe the ability of light to penetrate fog, and is generally between [0,1]. b Indicates constant bias; in, β This represents the scattering coefficient, which is constant under homogeneous atmospheric conditions. d ( x () represents the distance between the imaged object and the imaging device in the image of the loess slope. e Represents the natural constant.

[0026] Step S20: Defogging the loess slope image based on the haze estimate to obtain a recovered image, and extracting features from the recovered image to obtain a feature map. In this step, the clear image generation module restores image details through multi-scale feature fusion to achieve a dehazing effect on images of loess slopes. It estimates the transfer matrix from the blurred image using a deep learning network, performs K-value estimation, and then outputs a clear image after dehazing. The theoretical basis stems from the theory of atmospheric scattering effects, where natural light undergoes scattering when it penetrates suspended particles in the atmosphere.

[0027] In this step, the formula used to defog the images of the loess slope based on the estimated haze value includes: in, This refers to the acquired and restored image.

[0028] In this embodiment, a Shuffle Attention (SA) mechanism is added to the backbone of the YOLO v8 pro model. The SA mechanism achieves a high-efficiency balance between computational efficiency and detection accuracy through an innovative "grouping-refinement-reorganization" architecture. Its core advantage lies in dividing the input feature map into multiple subgroups. Within each group, channel attention (selecting key feature channels) and spatial attention (enhancing pixel-level responses in target regions) are performed in parallel. Then, a channel shuffle operation is used to exchange feature information across groups, preserving fine-grained details while enhancing semantic complementarity, thus alleviating the redundant computation problem of traditional attention mechanisms.

[0029] Please see Figure 3The sample segmentation images (sub-feature maps) are created along the channel segmentation of the collected recovery image, each sample segmentation image is subdivided into two branches along the channel during training, one is used for channel attention to capture the dependency relationship between channels, and the other is used for spatial dependency to generate a spatial attention map, so as to realize the attention to semantic and position information.

[0030] Optionally, feature extraction is performed on the collected recovery image to obtain a collection feature map, including: The collected recovery image is subjected to image segmentation to obtain a sample segmentation image, and the sample segmentation image is subjected to global average pooling (GAP) according to a channel attention branch in the landslide identification model to obtain a sample pooling feature map; In order to further fuse feature information, the global average pooling is used to extract the overall feature information of the sample segmentation image to obtain the sample pooling feature map; The channel attention feature map is calculated according to the sample pooling feature map; The sample segmentation image is subjected to group normalization processing according to a spatial attention branch in the landslide identification model to obtain a sample normalized feature map, and the sample normalized feature map is subjected to full connection processing to obtain a spatial attention feature map; The spatial attention feature map and the channel attention feature map are combined to obtain a sample combined feature, and the sample combined feature is subjected to channel shuffle to obtain the collection feature map; wherein the channel shuffle operator in Channel Shuffle is used to make the feature information of each branch communicate and merge between channels.

[0031] Further, the formula for calculating the channel attention feature map includes: wherein, represents the channel attention feature map, represents the sample segmentation image in the channel attention branch, represents a weight parameter in the channel attention branch, represents a bias parameter in the channel attention branch, represents the sample pooling feature map, represents a Sigmoid activation function.

[0032] The formula for calculating the spatial attention feature map includes: wherein, represents the spatial attention feature map, representing the sample segmentation image in the spatial attention branch, representing a weight parameter in the spatial attention branch, representing a bias parameter in the spatial attention branch, representing a group normalization process.

[0033] Step S30, according to the collection feature map, the loess slope collection image is subjected to landslide identification, a collection identification result is obtained, and a model loss is determined according to the collection identification result; Optionally, the model loss is determined according to the collection identification result, comprising: The target bounding box in the collection identification result is subjected to non-maximum suppression to obtain a sample predicted bounding box, and a bounding box similarity between the sample predicted bounding box and a real landslide bounding box of the loess slope collection image is calculated to obtain a first loss; wherein the similarity between the sample predicted bounding box and the real landslide bounding box is calculated to obtain the bounding box similarity, and the loss is calculated based on the bounding box similarity to obtain the first loss, and the first loss is used to represent the boundary box recognition error of the sample data of the landslide identification model; A second loss is obtained by calculating the difference between the predicted probability distribution and the real predicted probability distribution of the loess slope collection image; wherein the second loss is obtained by calculating the difference between the predicted probability distribution and the real predicted probability distribution, and the second loss is used to represent the recognition error of the sample data type recognition of the landslide identification model; The first loss and the second loss are subjected to weighted value operation to obtain the model loss; wherein in the weighted value operation process, the weighted coefficients of the first loss and the second loss can be set according to requirements.

[0034] Step S40, according to the model loss, the landslide identification model is subjected to parameter updating until the landslide identification model converges; Wherein, when the iteration number of the landslide identification model is greater than the number threshold or the model loss is less than the loss threshold, it is determined that the landslide identification model converges. In this step, the training process of the landslide identification model has been carried out for 1500 iterations. In order to ensure the stability of the model, the warm-up learning strategy is adopted in the first 3 iterations of training, and the initial learning rate is set to 0.01. Subsequently, the cosine annealing algorithm is used to control the decay of the learning rate, and the hyperparameter lrf of the cosine annealing algorithm is set to 0.2, and the minimum learning rate is 0.002.

[0035] Step S50, obtaining a to-be-identified slope image, and inputting the to-be-identified slope image into the converged landslide identification model for landslide identification to obtain a landslide identification result; The converged landslide identification model performs haze concentration estimation and haze removal processing on the to-be-identified slope image to obtain a to-be-identified restored image, performs feature extraction on the to-be-identified restored image to obtain a target feature map, performs landslide identification on the target feature map, and obtains a landslide identification result.

[0036] In this embodiment, the loess slope collected image is input into the landslide identification model for haze concentration estimation, which can effectively calculate the haze estimation value of the loess slope collected image. The haze estimation value is used for haze removal processing on the loess slope collected image, so that the converged landslide identification model can effectively remove haze from the to-be-identified slope image, prevent the influence of haze on landslide identification, and improve the landslide identification accuracy. Through haze removal preprocessing, small target detection head and attention mechanism optimization, an efficient detection model suitable for collapsible loess landslide is constructed, and the robustness deficiency and small target missing detection problem of the prior art in bad weather are solved. The improved model has significantly improved detection accuracy, and provides reliable technical support for early identification, stage division and prevention decision of landslide hazards. The UnfogNet network is introduced to remove haze and restore image clarity, and the new detection head and Shuffle Attention attention mechanism are added to solve the small target feature extraction problem.

[0037] Embodiment two Please refer to Figure 4 is a structural schematic diagram of a landslide identification system 100 of a collapsible loess slope provided by the second embodiment of the present application, which comprises: An image acquisition module 10 is configured to acquire a loess slope collected image.

[0038] A haze estimation module 11 is configured to input the loess slope collected image into a landslide identification model for haze concentration estimation to obtain a haze estimation value.

[0039] Optionally, the formula used for inputting the loess slope collected image into the landslide identification model for haze concentration estimation comprises: wherein, K the haze estimation value is represented by haze estimation value, x the loess slope collected image is represented by loess slope collected image, I the light intensity of atmospheric light is represented by atmospheric light intensity, x the image transmittance is represented by image transmittance, A the constant bias is represented by constant bias; t x b wherein, the scattering coefficient is represented by scattering coefficient, wherein, β d x ​​​represents a distance between an imaging object and an imaging device in the loess slope collected image, e represents a natural constant.

[0040] The haze removal processing module 12 is configured to perform haze removal processing on the loess slope collected image according to the haze estimation value, to obtain a collected recovery image.

[0041] Optionally, the formula used for haze removal processing on the loess slope collected image according to the haze estimation value comprises: wherein, J(x) represents the collected recovery image.

[0042] The feature extraction module 13 is configured to perform feature extraction on the collected recovery image, to obtain a collected feature map.

[0043] Optionally, the feature extraction module 13 is further configured to perform image segmentation on the collected recovery image, to obtain a sample segmented image, and perform global average pooling on the sample segmented image according to a channel attention branch in the landslide recognition model, to obtain a sample pooled feature map. calculate a channel attention feature map according to the sample pooled feature map; perform group normalization processing on the sample segmented image according to a spatial attention branch in the landslide recognition model, to obtain a sample normalized feature map, and perform full connection processing on the sample normalized feature map, to obtain a spatial attention feature map; combine the spatial attention feature map and the channel attention feature map, to obtain a sample combined feature, and perform channel shuffle on the sample combined feature, to obtain the collected feature map.

[0044] Further, the formula for calculating the channel attention feature map comprises: wherein, represents the channel attention feature map, represents the sample segmented image in the channel attention branch, represents a weight parameter in the channel attention branch, represents a bias parameter in the channel attention branch, represents the sample pooled feature map, represents a Sigmoid activation function.

[0045] The formula for calculating the spatial attention feature map comprises: wherein, This represents the spatial attention feature map. X k2 This represents the sample segmentation image in the spatial attention branch. W 2 represents the weight parameter in the spatial attention branch. b 2 represents the bias parameter in the spatial attention branch. GN This indicates grouping and normalization processing.

[0046] The model training module 14 is used to perform landslide identification on the loess slope images based on the acquired feature maps, obtain the acquisition and identification results, and determine the model loss based on the acquisition and identification results; The parameters of the landslide identification model are updated based on the model loss until the landslide identification model converges.

[0047] Optionally, the model training module 14 is further configured to: perform non-maximum suppression on the target bounding box in the acquisition and recognition results to obtain the sample predicted bounding box, and calculate the loss based on the bounding box similarity between the sample predicted bounding box and the real landslide bounding box in the loess slope acquisition image to obtain the first loss; The second loss is obtained by calculating the loss based on the predicted probability distribution in the acquisition and recognition results and the actual predicted probability distribution of the loess slope acquisition images; The model loss is obtained by weighting the first loss and the second loss.

[0048] The landslide identification module 15 is used to acquire an image of the slope to be identified, and input the image of the slope to be identified into the converged landslide identification model to perform landslide identification and obtain the landslide identification result.

[0049] In this embodiment, by inputting the loess slope images into the landslide recognition model to estimate the haze concentration, the haze estimate of the loess slope images can be effectively calculated. The haze estimate is then used to defog the loess slope images, so that the converged landslide recognition model can effectively defog the images of the slopes to be identified, preventing the influence of haze on landslide recognition and improving the accuracy of landslide recognition.

[0050] Example 3 Figure 5 This is a structural block diagram of a terminal device 2 provided in the third embodiment of this application. For example... Figure 5 As shown, the terminal device 2 in this embodiment includes a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program for a landslide identification method for collapsible loess slopes. When the processor 20 executes the computer program 22, it implements the steps in each embodiment of the landslide identification method for collapsible loess slopes described above.

[0051] For example, the computer program 22 can be divided into one or more modules stored in the memory 21 and executed by the processor 20 to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program 22 in the terminal device 2. The terminal device can include, but is not limited to, the processor 20 and the memory 21.

[0052] The processor 20 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0053] The memory 21 can be an internal storage unit of the terminal device 2, such as a hard disk or a memory of the terminal device 2. The memory 21 can also be an external storage device of the terminal device 2, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 21 can include both the internal storage unit and the external storage device of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 can also be used to temporarily store data that has been output or will be output.

[0054] In addition, each functional module in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0055] The integrated module, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Among them, the computer readable storage medium can be non-volatile or volatile. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can be executed by a processor to implement the steps of each method embodiment described above. Among them, the computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form, etc. The computer readable storage medium can include any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable storage medium does not include electric carrier signal and telecommunication signal.

[0056] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for identifying landslides on collapsible loess slopes, characterized in that, The method includes: Images of loess slopes are acquired and input into a landslide identification model to estimate haze concentration, thereby obtaining haze estimates. The haze estimate is used to dehaze the images of the loess slope to obtain the recovered images, and features are extracted from the recovered images to obtain the feature maps. Landslide identification is performed on the loess slope images based on the acquired feature maps to obtain the acquisition and identification results, and the model loss is determined based on the acquisition and identification results. The parameters of the landslide identification model are updated based on the model loss until the landslide identification model converges. The image of the slope to be identified is obtained, and the image is input into the converged landslide identification model to perform landslide identification, thereby obtaining the landslide identification result.

2. The landslide identification method for collapsible loess slopes as described in claim 1, characterized in that, Feature extraction is performed on the acquired and restored image to obtain an acquired feature map, including: The acquired and restored image is segmented to obtain a sample segmentation image, and the sample segmentation image is then subjected to global average pooling based on the channel attention branch in the landslide recognition model to obtain a sample pooling feature map. Calculate the channel attention feature map based on the sample pooling feature map; The sample segmentation images are grouped and normalized according to the spatial attention branch in the landslide identification model to obtain sample normalized feature maps, and the sample normalized feature maps are fully connected to obtain spatial attention feature maps. The spatial attention feature map and the channel attention feature map are combined to obtain the sample combination feature map, and the sample combination feature map is then subjected to channel shuffling to obtain the acquisition feature map.

3. The landslide identification method for collapsible loess slopes as described in claim 2, characterized in that, The formula for calculating the channel attention feature map includes: in, This represents the channel attention feature map. This represents the sample segmentation image in the channel attention branch. This represents the weight parameters in the channel attention branch. This represents the bias parameter in the channel attention branch. This represents the sample pooling feature map. This represents the Sigmoid activation function.

4. The landslide identification method for collapsible loess slopes as described in claim 2, characterized in that, The formula for calculating the spatial attention feature map includes: in, This represents the spatial attention feature map. This represents the sample segmentation image in the spatial attention branch. This represents the weight parameters in the spatial attention branch. This represents the bias parameter in the spatial attention branch. This indicates grouping and normalization processing.

5. The landslide identification method for collapsible loess slopes as described in claim 1, characterized in that, The formula used to estimate haze concentration by inputting the loess slope images into the landslide recognition model includes: in, This represents the estimated value of the haze. This refers to the images captured on the loess slope. A Indicates the intensity of atmospheric light. Indicates image transmittance. b Indicates constant bias; in, β Represents the scattering coefficient. d ( x () represents the distance between the imaged object and the imaging device in the image of the loess slope. e Represents the natural constant.

6. The landslide identification method for collapsible loess slopes as described in claim 5, characterized in that, The formulas used to defog the images of the loess slope based on the estimated haze values ​​include: in, This refers to the acquired and restored image.

7. The landslide identification method for collapsible loess slopes as described in claim 1, characterized in that, The model loss is determined based on the collected and identified results, including: Non-maximum suppression is applied to the target bounding boxes in the acquisition and recognition results to obtain sample predicted bounding boxes. The loss is calculated based on the similarity between the sample predicted bounding boxes and the actual landslide bounding boxes in the loess slope acquisition images to obtain the first loss. The second loss is obtained by calculating the loss based on the predicted probability distribution in the acquisition and recognition results and the actual predicted probability distribution of the loess slope acquisition images; The model loss is obtained by weighting the first loss and the second loss.

8. A landslide identification system for collapsible loess slopes, characterized in that, The system includes: The image acquisition module is used to acquire images of loess slopes. The haze estimation module is used to input the collected images of the loess slope into the landslide recognition model to estimate the haze concentration and obtain the haze estimate value. The defogging module is used to defog the images of the loess slope based on the estimated haze value to obtain the recovered images. The feature extraction module is used to extract features from the acquired and restored image to obtain an acquired feature map; The model training module is used to perform landslide identification on the loess slope images based on the acquired feature maps, obtain the acquisition and identification results, and determine the model loss based on the acquisition and identification results; The parameters of the landslide identification model are updated based on the model loss until the landslide identification model converges. The landslide identification module is used to acquire an image of the slope to be identified, and input the image of the slope to be identified into the converged landslide identification model to perform landslide identification and obtain the landslide identification result.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images