Complex terrain landslide detection method based on CTL-YOLO model

Abstract

The invention relates to a complex terrain landslide detection method based on a CTL-YOLO model. The method comprises the following steps: carrying out image acquisition on a complex terrain landslide to obtain a landslide image, and making a terrain remote sensing data set; landslide area labeling is carried out on the collected landslide image; randomly dividing the landslide image and the corresponding annotation file into a training set, a verification set and a test set; performing data enhancement on the images in the training set; a CTL-YOLO model is constructed; training and verifying the CTL-YOLO model by adopting a terrain remote sensing data set; according to the invention, the trunk feature extraction network captures the local spatial features of the landslide area, and adaptively adjusts the weight of the important features, and the feature fusion network improves the fusion effect of the multi-scale features; the LSDECD detection head improves the positioning and classification performance of the model, reduces the dependence of batch normalization on the batch size, and enhances the stability in different environments; the DEConv module effectively recognizes detail features of the landslide boundary through multi-direction differential convolution, and the prediction precision of the bounding box is further improved.

Description

Technical Field

[0001] The present invention relates to a landslide detection method on complex terrain based on a CTL-YOLO model, belonging to the technical field of geological identification. Background Art

[0002] Landslides are common natural disasters triggered by factors such as rainfall and earthquakes. They are characterized by suddenness and widespread destruction, posing a serious threat to human life and often causing enormous social and economic losses. After a landslide occurs, quickly and accurately determining its size and distribution using remote sensing imagery is crucial for risk management, rescue efforts, and the development of landslide databases.

[0003] Current remote sensing landslide detection methods mainly include visual interpretation, pixel-based analysis, object-oriented recognition, and artificial intelligence. Among traditional methods, visual interpretation relies on expert experience to analyze shape / texture features, which is inefficient and has large errors; although pixel-based analysis reduces the false detection rate, it is difficult to distinguish spectrally similar objects; object-oriented recognition has limited applicability because it requires manual setting of feature thresholds; among artificial intelligence methods, machine learning relies on manual feature engineering and has poor cross-regional generalization capabilities; deep learning achieves end-to-end feature learning through convolutional neural networks, significantly improving detection accuracy and adaptability; CNN-based target detection frameworks can be divided into two-stage and one-stage models; as a representative of one-stage models, the YOLO series of models has been widely used in the field of landslide detection due to its advantages in speed and accuracy balance; among the many versions of the YOLO series, YOLOv8 has achieved a good balance between accuracy and model complexity; such as Figure 2 As shown in Figure 1, the YOLOv8 model consists of three main components: the backbone network (Backbone), the neck network (Neck), and the detection head (Head). The backbone network (Backbone) is responsible for feature extraction and contains multiple Conv and C2f modules. The neck network (Neck) adopts a feature pyramid structure and achieves multi-scale feature fusion through upsampling and feature concatenation. The detection head (Head) performs predictions for bounding box regression (Bbox_Loss) and class classification (Cls_Loss), achieving end-to-end landslide detection.

[0004] A Chinese invention patent application, publication number CN118505969A, discloses a method for rapid landslide identification in remote sensing imagery based on an improved YOLOv8 model. The method comprises the following steps: replacing the original bounding box loss function from CIOU with EIOU to improve computational speed and optimize positioning accuracy; introducing a scalable residual attention (DWRSeg) module based on the C2F module to optimize the model structure, reduce the number of parameters and computational complexity; and conducting a comprehensive evaluation through extensive experiments on real remote sensing datasets. However, the aforementioned approach suffers from insufficient multi-scale feature fusion and lacks an effective fusion strategy for the large scale variation of landslides in remote sensing imagery. The DWRSeg module primarily focuses on optimizing the number of parameters and computational complexity, failing to design a feature extraction and fusion mechanism specifically for landslides of varying sizes. This results in uneven recognition of landslides of varying sizes. This inadequate multi-scale feature representation may limit the model's practical application, particularly in complex terrain environments where small and large landslides coexist.

[0005] At the same time, the use of the YOLOv8 model for remote sensing landslide detection still has the following problems:

[0006] (1) Feature similarity: The spectral and texture characteristics of landslide areas are highly similar to those of exposed soil, vegetation-covered areas, and other landforms, making it difficult for the model to effectively distinguish between landslide and non-landslide areas;

[0007] (2) Boundary fuzziness: The landslide boundary area often contains a mixed coverage of vegetation, exposed soil and debris accumulation, and is affected by the terrain undulation. The landslide boundary presents a highly irregular shape. Traditional convolution is difficult to capture edge details, resulting in errors in the model's boundary positioning;

[0008] (3) Environmental interference: Imaging conditions such as cloud obstruction and uneven lighting lead to the loss of image feature information, resulting in missed detections in the model.

[0009] In summary, the existing YOLO model has problems such as insufficient detection accuracy, missed detection and false detection in landslide detection under complex backgrounds, which reduces its reliability and application effect in complex scenes. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a landslide detection method based on the CTL-YOLO model, which has a strong ability to identify landslide features, high boundary positioning accuracy, low calculation amount and simple network structure.

[0011] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0012] The present invention provides a complex terrain landslide detection method based on the CTL-YOLO model, which comprises the following steps:

[0013] S1. Collect images of landslides on complex terrain to obtain landslide images and create terrain remote sensing datasets;

[0014] S2. marking the landslide area on the collected landslide image;

[0015] S3, randomly dividing the landslide images and their corresponding annotation files into training set, validation set and test set;

[0016] S4. Perform data augmentation on images in the training set;

[0017] S5. Build the CTL-YOLO model;

[0018] S6. Use terrain remote sensing dataset to train and verify the CTL-YOLO model.

[0019] The CTL-YOLO model of the present invention includes a backbone feature extraction network, a feature fusion network and a detection head;

[0020] The backbone feature extraction network includes Conv module, C2f_RVB module and C2f_RVB_EMA module;

[0021] The feature fusion network adopts the reparameterized global dynamic sampling feature pyramid network (RepGDSFPN);

[0022] The detection head adopts the Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD).

[0023] The C2f_RVB module of the present invention is composed of replacing the Bottleneck module of the C2f module in the YOLOv8 backbone network; wherein, the Bottleneck modules of the first, second, and fourth C2f modules are replaced with RepViTBlock modules, and the Bottleneck module of the third C2f module is replaced with RepViTEMABlock module; the replacement is achieved by adjusting the number of input and output channels of RepViTBlock and RepViTEMABlock to be consistent with the original Bottleneck module, while maintaining the overall structure and shortcut connection of the C2f module unchanged;

[0024] The RepViTBlock module consists of a 3×3 depthwise convolutional branch, a 1×1 depthwise convolutional branch, an identity mapping branch, and a feedforward network (FFN). The identity mapping branch is a residual connection structure that directly passes the input feature map to the subsequent layer without any transformation. The feedforward network (FFN) consists of a sequential structure consisting of two Conv2d convolutional layers and two BatchNorm2d normalization layers, which are connected in sequence and finally perform a residual addition operation with the input features. The RepViTEMABlock module is an enhanced structure formed by introducing an EMA module before the feedforward network (FFN) on the basic architecture of the RepViTBlock module. The EMA module is an efficient multi-scale attention module. By introducing a multi-scale feature processing mechanism after fusing the output features of the 3×3 depthwise convolution, 1×1 depthwise convolution, and identity mapping branch and before the feedforward network (FFN) processing, it captures long-range dependencies in the feature map and performs channel-level feature importance weighting, thereby enhancing the model's multi-scale perception of complex terrain landslide characteristics.

[0025] The EMA module includes branch 1, branch 2, and branch 3;

[0026] Branch 1 performs initial dimensionality reduction as follows:

[0027] F1(X)=Conv1(X)

[0028] Where X∈R C×H×W is the input feature map, where Conv1 reduces the channel dimension to C / r, where r is the reduction ratio, and F1 represents branch 1;

[0029] Branch 2 captures global context information through bidirectional average pooling as follows:

[0030] F2(X)=σ(Conv2([Pool h (X),Pool w (X)]))

[0031] Where, Pool h and Pool w Represents the average pooling operation in the horizontal and vertical directions respectively, Conv2 aggregates the features after pooling, σ represents the Sigmoid activation function, and F2 represents branch 2;

[0032] Branch 3 focuses on the extraction of local features, as follows:

[0033] F3(X)=Conv3(X)

[0034] Where Conv3 is a depth-wise separable convolution that maintains spatial information while having high computational efficiency, and F3 represents branch 3.

[0035] For multi-scale feature processing, the three branches of EMA are fused through a weighted fusion mechanism:

[0036] Y=X⊙(α·F r (X)+β·F ctx (X)+γ·F l (X)

[0037] where α, β, and γ are learnable weights used to balance the contribution of each branch.

[0038] The present invention reparameterized global dynamic sampling feature pyramid network (RepGDSFPN) includes a RepGFPN structure and a Dysample module;

[0039] The RepGFPN structure includes several CSPStage modules connected in series;

[0040] The input of the CSPStage module is two or three feature layers, which are divided into two branches after channel splicing;

[0041] One branch adjusts the number of channels through 1×1 convolution.

[0042] The other branch is the ELAN feature aggregation module, which consists of n 3×3Rep convolutions and 3×3 standard convolutions;

[0043] After the two branches are spliced ​​through the channels, a final convolution is performed to obtain the final output.

[0044] The Dysample module of the present invention includes a sampling point generator, a sampling set, and a grid sampling function; the workflow of the Dysample module is as follows:

[0045] S51. Given an input feature map of size C×H×W, it is converted into a sampling set S of size 2×sH×sW through the sampling point generator. The 2 in the first dimension of the sampling set S represents the x and y coordinates.

[0046] S52, using the gridsample function to resample the input features to obtain an upsampled feature map of size C×sH×sW;

[0047] Where H and W represent height and width, and C represents the number of channels;

[0048] S53, perform bilinear interpolation on X through the gridsample function to obtain the upsampled features Figure X ', the formula is as follows:

[0049] X′=grid_sample(X,S)

[0050] The sampling point generator generates an offset O through a linear layer, a dynamic range factor, and a pixel shuffle operation, which is then added to the original grid position G to generate a sampling set S. The generation process of the sampling set S is defined as:

[0051] S=G+O;

[0052] The dynamic range factor has a value range of 0.25 to 0.5.

[0053] The lightweight shared detail enhancement convolutional detection head (LSDECD) of the present invention shares the convolutional layer weights through a decoupled detection head to reduce parameters. The decoupled detection head separates the detection classification task and the bounding box regression task, and by making the classification task and the bounding box regression task branches share the pre-convolutional layer parameters, only using independent parameters in the task-specific layer, introduces a DEConv module in the shared convolutional layer. The DEConv module includes five parallel branches: conventional convolution (VC), center difference convolution (CDC), angle difference convolution (ADC), horizontal difference convolution (HDC) and vertical difference convolution (VDC). The output feature maps of these five branches are fused by weighted summation in the channel dimension. The lightweight shared detail enhancement convolutional detection head (LSDECD) uses a group normalization (GN) layer instead of a batch normalization (BN) layer.

[0054] The DEConv module combines multiple convolution kernels into a single standard convolution kernel through reparameterization. The reparameterization process is as follows:

[0055]

[0056] In the formula, F out represents the output feature map of DEConv, K i＝1:5 Represent the convolution kernels of VC, CDC, ADC, HDC and VDC respectively. The symbol * indicates the convolution operation. K cvt Represents the equivalent convolution kernel obtained by combining these parallel convolutions;

[0057] The Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD) introduces a scale layer to adaptively scale the output features of the regression branch of each detection layer. Specifically, by setting a learnable scaling parameter for each detection layer, the optimal scaling factor is automatically adjusted during model training, thereby dynamically balancing the contributions of different feature layers and maintaining scale consistency between detection heads.

[0058] The initial value of the scaling parameter is 1.0.

[0059] The images in S1 of the present invention include images of landslide, loess landslide and debris flow.

[0060] The data is collected using remote sensing satellites, drones or ground-based photography equipment;

[0061] In S2, the CVAT annotation tool is used to annotate the landslide area on the collected images and generate a standardized rectangular bounding box label file;

[0062] In S3, the landslide images and their corresponding label files are randomly divided according to a uniform ratio of 7:2:1;

[0063] Data augmentation in S3 uses geometric transformation and color transformation; geometric transformation includes random horizontal and vertical translation, rotation, scaling, cropping, and up and down and left and right flipping; color transformation includes random changes in hue, saturation, and brightness.

[0064] The specific operation of S6 of the present invention is as follows:

[0065] S61. Set up the experimental environment: the image input size is set to 640×640, the optimizer is SGD, the initial learning rate is set to 0.01, the momentum factor is 0.937, the weight decay coefficient is 0.0005, the batch size is 16, and the number of training iterations is 200 rounds;

[0066] S62. Select evaluation indicators: The evaluation indicators include precision (P), recall (R), score (F1), and average precision (AP). The specific calculation formula is as follows:

[0067]

[0068] Where TP, FP, and FN represent the counts of correctly predicted positive samples, incorrectly predicted positive samples, and incorrectly predicted negative samples, respectively; P measures the proportion of correctly predicted landslides among all predictions, and R measures the proportion of correctly predicted landslides among actual landslides; the F1 score is the harmonic mean of P and R, and AP is the area under the PR curve, which is used as a comprehensive performance indicator;

[0069] The lightweight CTL-YOLO model is evaluated using three metrics: parameter count (Params), floating-point operations (FLOPs), and frames per second (FPS); Params evaluates spatial complexity, GFLOPs quantifies computational complexity, and FPS measures processing speed.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] 1) The present invention designs a RepViTBlock module in the backbone feature extraction network and integrates the EMA attention mechanism into part of the RepViTBlock. This design captures the local spatial characteristics of the landslide area through a three-branch structure and adaptively adjusts the weights of important features. This solves the problem of missed detection caused by interference factors such as lighting changes and cloud occlusion in complex environments in traditional models, and improves the feature extraction capability of the backbone network.

[0072] 2) This paper introduces the RepGDSFPN feature fusion network into the feature fusion network, improving the fusion of multi-scale features. By cascading multiple CSPStage modules, this network optimizes the interaction between high-level semantic features and low-level spatial features, thereby enhancing the ability to distinguish between landslide and non-landslide areas. Furthermore, the Dysample module replaces traditional upsampling methods, reducing the loss of fine-grained features and improving the model's detection performance and accuracy in complex terrain.

[0073] 3) The proposed LSDECD detection head improves the model's positioning and classification performance, reduces the dependency of batch normalization on batch size, and enhances stability under different environments. The DEConv module effectively identifies detailed features of landslide boundaries through multi-directional differential convolution, further improving the bounding box prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 Flowchart of the present invention;

[0075] Figure 2 This is a diagram of the YOLOv8 model structure in the prior art;

[0076] Figure 3 This is the structure diagram of the CTL-YOLO model of the present invention;

[0077] Figure 4 This is a structural diagram of the C2f_RVB module of the present invention;

[0078] Figure 5 This is a structural diagram of the C2f_RVB_EMA module of the present invention;

[0079] Figure 6 This is the EMA attention structure diagram of the present invention;

[0080] Figure 7 This is the RepGFPN network structure diagram of the present invention;

[0081] Figure 8 This is the Dysample module structure diagram of the present invention;

[0082] Figure 9 This is a structural diagram of the LSDECD of the present invention;

[0083] Figure 10 Schematic diagram of the comparison structure of landslide detection results on complex terrain. DETAILED DESCRIPTION

[0084] The present invention will be described in further detail below with reference to the accompanying drawings.

[0085] As attached Figure 1 As shown in Figure 1, a complex terrain landslide detection method based on the CTL-YOLO model adopts the following steps:

[0086] S1. Use remote sensing satellites, drones, or ground-based imaging equipment to capture images of complex terrain areas, ensuring that the images include both landslide and non-landslide areas under different terrain, lighting, and meteorological conditions. The images collected should cover a variety of landslide types, such as mountain landslides, loess landslides, and debris flows. Delete images that are unclear or that may affect recognition accuracy due to shooting angles to ensure image clarity and quality.

[0087] S2. Cropping the collected landslide image to a standard size of 640 × 640 pixels, and using the CVAT annotation tool to annotate the landslide area on the image to generate a standardized rectangular bounding box label file;

[0088] S3, randomly divide the landslide images and their corresponding annotation files into training set, validation set and test set according to a uniform ratio of 7:2:1;

[0089] S4. Perform data augmentation on the images in the training set by randomly applying two basic data augmentation methods: geometric transformation and color transformation. Geometric transformation includes random horizontal and vertical translation, rotation, scaling, cropping, and up and down and left and right flipping; color transformation includes random changes in hue, saturation, and brightness.

[0090] S5. Build the CTL-YOLO model;

[0091] As attached Figure 3 As shown in the figure, the CTL-YOLO model includes a backbone feature extraction network, a feature fusion network, and a detection head; the backbone feature extraction network includes a Conv module, a C2f_RVB module, and a C2f_RVB_EMA module; the feature fusion network uses a re-parameterized global dynamic sampling feature pyramid network (RepGDSFPN); the detection head uses a lightweight shared detail enhanced convolutional detection head (LSDECD);

[0092] Specifically:

[0093] As attached Figure 4As shown in the figure, the C2f_RVB module is constructed by replacing the Bottleneck module of the C2f module in the YOLOv8 backbone network; the Bottleneck modules of the 1st, 2nd, and 4th C2f modules are replaced by RepViTBlock modules, and the Bottleneck module of the 3rd C2f module is replaced by RepViTEMA Block module; the replacement is achieved by adjusting the number of input and output channels of RepViTBlock and RepViTEMA Block to be consistent with the original Bottleneck module, keeping the overall structure and shortcut connection of the C2f module unchanged, and replacing only the internal parallel submodules;

[0094] The RepViTBlock module consists of a 3×3 depth convolution branch, a 1×1 depth convolution branch, an identity mapping branch and a feedforward network (FFN). Each RepViTBlock module extracts features through 3×3 and 1×1 depth convolution layers to capture local information and reduce computational complexity, and then sends them to the feedforward network (FFN) for further processing; the feedforward network (FFN) is composed of two layers of convolution structures connected in series, each layer contains convolution operations and batch normalization, and the final output is residually connected with the original input. The identity mapping branch is a residual connection structure that directly passes the input feature map to the subsequent layer without any transformation. The feedforward network (FFN) consists of a sequence structure consisting of two Conv2d convolution layers and two BatchNorm2d normalization layers that are sequentially connected and finally perform residual addition operations with the input features; as shown in Figure 5 As shown in the figure, the RepViTEMA Block module is an enhanced structure formed by introducing the EMA module before the feed-forward network (FFN) on the basis of the RepViTBlock module. It uses bidirectional pooling to extract global context, combines standard convolution to extract local features, and refines features through dynamic channel reduction. The outputs of all branches are fused through concatenation, and finally the convolution layer compresses the channels and fuses the global information to generate the final output.

[0095] The EMA module is an efficient multi-scale attention module (Efficient Multi-Scale Attention Module). By introducing a multi-scale feature processing mechanism after the output feature fusion of 3×3 depthwise convolution, 1×1 depthwise convolution and identity mapping branch and before feed-forward network (FFN) processing, it captures long-range dependencies in the feature map and performs channel-level feature importance weighting, thereby enhancing the model's multi-scale perception of complex terrain landslide characteristics.

[0096] like Figure 6As shown in the figure, the EMA module adopts a specialized multi-scale attention mechanism; for low-contrast targets, the EMA module adopts a dual-channel design from a functional perspective to process local and global features simultaneously; the local path uses 3×3 standard convolution to capture fine-grained target details, while the global path aggregates contextual information through selective scanning; this dual processing helps to distinguish subtle target features from similar background patterns; in complex environments, the adaptive feature recalibration mechanism of the EMA module plays a vital role; the module dynamically adjusts the attention weights based on feature importance; in terms of structural implementation, the EMA module consists of three parallel branches, which process the input features separately to capture comprehensive spatial and channel relationships;

[0097] The EMA module includes branch 1, branch 2, and branch 3;

[0098] Branch 1 performs initial dimensionality reduction as follows:

[0099] F1(X)=Conv1(X)

[0100] Where X∈R C×H×W is the input feature map, where Conv1 reduces the channel dimension to C / r, where r is the reduction ratio, and F1 represents branch 1;

[0101] In practical applications, the reduction ratio r typically ranges from 4 to 16, adjusted based on the number of input feature channels and computational resources. Smaller r values ​​(e.g., 4) retain more feature information but require more computation, while larger r values ​​(e.g., 16) are computationally efficient but may lose some feature details. In landslide detection tasks, r = 8 is often used as a balance point, effectively reducing computational complexity while maintaining sufficient feature expression capabilities.

[0102] Branch 2 captures global context information through bidirectional average pooling as follows:

[0103] F2(X)=σ(Conv2([Pool h (X),Pool w (X)]))

[0104] Where, Pool h and Pool w Represents the average pooling operation in the horizontal and vertical directions respectively, Conv2 aggregates the features after pooling, σ represents the Sigmoid activation function, and F2 represents branch 2;

[0105] Branch 3 focuses on the extraction of local features, as follows:

[0106] F3(X)=Conv3(X)

[0107] Where Conv3 is a depth-wise separable convolution that maintains spatial information while having high computational efficiency, and F3 represents branch 3.

[0108] For multi-scale feature processing, the three branches of EMA are fused through a weighted fusion mechanism:

[0109] Y=X⊙(α·F r (X)+β·F ctx (X)+γ·F l (X))

[0110] where α, β, and γ are learnable weights used to balance the contribution of each branch.

[0111] Reparameterized Global Dynamic Sampling Feature Pyramid Network (RepGDSFPN) includes the RepGFPN structure and the Dysample module;

[0112] As attached Figure 7 As shown in Figure 1, the RepGFPN structure consists of several CSPStage modules connected in series. RepGFPN decomposes and convolves features of adjacent scale layers and the same scale in different layers through the CSPStage module. The input of the CSPStage module is two or three feature layers, which are divided into two branches after channel splicing.

[0113] One branch adjusts the number of channels through 1×1 convolution.

[0114] The other branch is the ELAN feature aggregation module, which consists of n 3×3Rep convolutions and 3×3 standard convolutions;

[0115] After the two branches are spliced ​​through the channels, a final convolution is performed to obtain the final output.

[0116] RepGFPN enhances the interaction between high-level semantic features and low-level spatial features by fusing features of adjacent scale layers and different scales in the same layer, thereby improving the model's ability to distinguish between landslide and non-landslide areas; RepConv in the CSPStage module merges multi-branch structures into a single branch through structural reparameterization, improving the model's detection efficiency without increasing the amount of computation.

[0117] As attached Figure 8 As shown in the figure, the Dysample module includes a sampling point generator, a sampling set, and a grid sampling function; the specific steps are as follows:

[0118] S51. Given an input feature map of size C×H×W, it is converted into a sampling set S of size 2×sH×sW through the sampling point generator, where the first dimension 2 in S represents the x and y coordinates;

[0119] S52, using the gridsample function to resample the input features to obtain an upsampled feature map of size C×sH×sW;

[0120] Where H and W represent height and width, and C represents the number of channels;

[0121] S53, perform bilinear interpolation on X through the gridsample function to obtain the upsampled features Figure X ', the formula is as follows:

[0122] X′=grid_sample(X,S)

[0123] The sampling point generator generates an offset O through a linear layer, a dynamic range factor, and a pixel shuffle operation, which is then added to the original grid position G to generate a sampling set S. The generation process of the sampling set S can be defined as:

[0124] S=G+O; the dynamic range factor value range is: 0.25~0.5, which is determined adaptively through network training.

[0125] Dysample is based on position-aware dynamic offset, which can retain more fine-grained features during upsampling. The dynamic range factor can improve the flexibility of the offset, alleviate the overlap problem between offsets, and make the model more accurate in distinguishing landslides from similar backgrounds.

[0126] The Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD) reduces parameters by sharing convolutional layer weights through a decoupled detection head. The decoupled detection head separates the detection classification task and the bounding box regression task. By making the classification task and the bounding box regression task branches share the pre-convolutional layer parameters, only independent parameters are used in the task-specific layers, thereby significantly reducing the total number of model parameters. The DEConv module is introduced in the shared convolutional layer. The DEConv module includes five parallel branches: conventional convolution (VC), center difference convolution (CDC), angular difference convolution (ADC), horizontal difference convolution (HDC), and vertical difference convolution (VDC). The output feature maps of these five branches are fused by weighted summation in the channel dimension, where the weight coefficient of each branch is trainable, realizing the effective combination of intensity features and multi-directional gradient features, thereby improving the accuracy of landslide boundary positioning. The Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD) uses a group normalization (GN) layer instead of a batch normalization (BN) layer to reduce the dependence of detection performance on batch size and improve the training stability of the model under different hardware conditions.

[0127] The DEConv module combines multiple convolution kernels into a single standard convolution kernel through reparameterization. The reparameterization process is as follows:

[0128]

[0129] In the formula, F out represents the output feature map of DEConv, K i＝1:5 Represent the convolution kernels of VC, CDC, ADC, HDC and VDC respectively. The symbol * indicates the convolution operation. K cvt Represents the equivalent convolution kernel obtained by combining these parallel convolutions;

[0130] The Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD) introduces a scale layer to adaptively scale the output features of the regression branch of each detection layer. Specifically, a learnable scaling parameter (initial value is 1.0) is set for each detection layer, and the optimal scaling factor is automatically adjusted during model training. This dynamically balances the contributions of different feature layers and maintains scale consistency between detection heads. This effectively solves the scale mismatch problem in multi-scale feature fusion and improves the detection accuracy of landslide targets at different resolutions.

[0131] As attached Figure 9 As shown in the figure, the lightweight shared detail enhanced convolutional detection head (LSDECD) receives three feature layers (P3, P4, P5) from the feature fusion network and generates detection results. The specific operations are as follows:

[0132] First, each feature layer is passed through a 1×1 Conv_GN module to adjust the number of channels;

[0133] Subsequently, the adjusted feature map is fed into two 3×3 DEConv_GN modules to further process and enhance the spatial information of the feature map. In this process, the convolution operation adopts a weight sharing strategy, which helps to reduce the number of model parameters and improve computational efficiency.

[0134] Finally, the feature map processed by the DEConv module is processed through two branches, namely regression convolution (Conv_Reg) and classification convolution (Conv_Cls), to generate the regression and classification outputs of the target.

[0135] S6. Use terrain remote sensing dataset to train and validate the CTL-YOLO model;

[0136] The specific operations for training the CTL-YOLO model are as follows:

[0137] S61. Set up the experimental environment: the image input size is set to 640×640, the optimizer is SGD, the initial learning rate is set to 0.01, the momentum factor is 0.937, the weight decay coefficient is 0.0005, the batch size is 16, and the number of training iterations is 200 rounds;

[0138] S62. Select evaluation indicators: The evaluation indicators include precision (P), recall (R), score (F1), and average precision (AP). The specific calculation formula is as follows:

[0139]

[0140] Among them, TP, FP and FN represent the counts of correctly predicted positive samples, incorrectly predicted positive samples and incorrectly predicted negative samples, respectively; P measures the proportion of correctly predicted landslides in all predictions, and R measures the proportion of correctly predicted landslides in actual landslides; F1 score is the harmonic mean of P and R, and AP is the area under the PR curve, which is used as a comprehensive performance indicator.

[0141] The lightweight CTL-YOLO model is evaluated using three metrics: parameter count (Params), floating-point operations (FLOPs), and frames per second (FPS); Params evaluates spatial complexity, GFLOPs quantifies computational complexity, and FPS measures processing speed.

[0142] The CTL-YOLO model was validated using a terrain remote sensing dataset:

[0143] The comparative experimental results of the CTL-YOLO model of the present invention and other target detection models are shown in Table 1. The experimental results in Table 1 show that the average precision (AP) of the Faster-RCNN and SSD models are 72.6% and 74.8%, respectively, which are both the lowest; the number of parameters of these models is also relatively high, 28.3M and 23.8M, respectively, and the inference speed is slow, 28.7 and 83.3, respectively, resulting in poor real-time performance in landslide detection. Relatively speaking, the YOLO series model is superior to the above two models in terms of parameter amount, calculation amount and inference speed. In the YOLO series, YOLOXs has the lowest average precision (AP) value of 85.4%, while the average precision (AP) values ​​of YOLOv5s, YOLOv7-tiny and YOLOv8s are similar, but YOLOv8s is better than YOLOv5s and YOLOv7-tiny in precision (P) and recall (R).

[0144] Compared to the previously improved YOLOv8 model, the CTL-YOLO model of this invention achieves the highest average precision (AP) of 97.7%. P, R, and AP improve by 2.3%, 5.2%, and 5.4%, respectively. The number of parameters and computational complexity are reduced by 19.8% and 20.8%, respectively, and the inference speed is increased by 13.3%. Although the CTL-YOLO model is slightly slower than YOLOv5 and YOLOv7-tiny, it still meets the real-time requirements of landslide detection. Experimental data demonstrates that this invention maintains high accuracy while maintaining low computational overhead.

[0145] Table 1 Model comparison experimental results

[0146]

[0147] Detection results of landslides with fuzzy boundaries by different detection models, such as Figure 10 As shown, from left to right are true values, Faster-RCNN, SSD, YOLOXs, YOLOv5s, YOLOv7-tiny, YOLOv8s and CTL-YOLO; Figure 10 a and Figure 10 b Comparison of landslide detection results with fuzzy boundaries. Figure 10 c and Figure 10 d compares the landslide detection results with similar backgrounds. Figure 10 e and Figure 10 f Comparison of landslide detection results affected by environmental interference; the green and red boxes represent the true labels and the detection results of different models, respectively, and the blue and yellow boxes are used to mark the missed and falsely detected landslide areas.

[0148] Analyze the landslide detection results with fuzzy boundaries ( Figure 10 a and Figure 10 b) shows that all models can detect the landslide area; however, Faster-RCNN has a large error in positioning and Figure 10 False detection occurs in b. Although SSD, YOLOXs, and YOLOv7-tiny can detect landslide areas, their confidence levels are low when dealing with landslides with blurred boundaries, and the predicted boxes deviate from the true values. SSD's confidence levels are 0.55 and 0.54, YOLOXs's are 0.62 and 0.60, and YOLOv7-tiny's are 0.78 and 0.77. In comparison, YOLOv5s, YOLOv8s, and CTL-YOLO perform better in terms of positioning accuracy and confidence. YOLOv5s' confidence level is 0.82, YOLOv8s's are 0.82 and 0.84, and CTL-YOLO's are 0.91 and 0.90. CTL-YOLO outperforms other models in positioning accuracy, confidence, and handling of blurred boundaries, and achieves the best results on the landslide dataset.

[0149] Analyze the landslide detection results with similar backgrounds ( Figure 10 c and Figure 10 d) shows that some models miss detection and make false detection when the landslide and background features are similar. Faster-RCNN, SSD and YOLOXs miss detection when distinguishing landslide from background, and Faster-RCNN Figure 10 In the image d, exposed soil is mistakenly detected as a landslide. The confidence levels for YOLOv5s were 0.85 and 0.86, for YOLOv7-tiny 0.75 and 0.82, and for YOLOv8s 0.87. The CTL-YOLO model achieved the highest confidence levels of all models, at 0.95 and 0.92. These results demonstrate that the CTL-YOLO model is capable of handling situations where landslides and non-landslide areas have similar characteristics, outperforming other models.

[0150] exist Figure 10 In the foggy scene of e, Faster-RCNN, SSD, YOLOXs, and YOLOv7-tiny failed to detect the landslide. Although YOLOv5s and YOLOv8s detected the landslide, their confidence levels were 0.70 and 0.78, respectively, which were low. The CTL-YOLO model not only successfully detected the landslide, but also achieved the highest confidence level of 0.90. In the shadowy scene ( Figure 10 f), Faster-RCNN mistakenly identifies the shadow as a landslide, and the confidence of SSD is 0.51. YOLOXs has a large positioning error, while the confidences of YOLOv5s, YOLOv7-tiny, and YOLOv8s are 0.81, 0.77, and 0.85, respectively; the CTL-YOLO model performs best in positioning accuracy and confidence, with a confidence of 0.90; In summary, the CTL-YOLO model shows good robustness under environmental interference and outperforms other models.

[0151] This paper optimizes and enhances the YOLOv8s model by replacing the C2f module in the backbone feature extraction network with the C2f_RVB and C2f_RVB_EMA modules; employing the RepGDSFPN as the feature fusion network; and using the LSDECD as the detection head. Ablation experiments verified the effect of each module on improving model performance. The experimental results are shown in Table 2.

[0152] As shown in Table 2, the YOLOv8s baseline model has 11.1M parameters and 28.4GFLOPs of computational overhead. The introduction of the C2f_RVB_EMA module improves the model's recall (R) and average precision (AP) by 3% and 2.7%, respectively. Furthermore, the number of parameters and computational overhead decreases by 13.5% and 14.4% compared to the baseline model, demonstrating that this module not only optimizes feature extraction capabilities but also effectively reduces the model's computational overhead. Despite the increase in computational overhead and parameters, the adoption of the Reparameterized Global Dynamic Sampling Feature Pyramid Network (RepGDSFPN) improves detection accuracy, achieving an average precision (AP) of 95.7%. The adoption of the Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD) significantly reduces the number of parameters and computational overhead by 1.7M and 2.6GFLOPs, respectively, achieving performance similar to that of the baseline model. This demonstrates that this module optimizes computational efficiency while maintaining good performance.

[0153] Combining any two of the three improved modules can significantly improve the detection performance of the model compared to using any one of them alone, with the average precision (AP) value increased by about 4%. Ultimately, by integrating the combination of C2f_RVB_EMA, the Reparameterized Global Dynamic Sampling Feature Pyramid Network (RepGDSFPN), and the Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD), the CTL-YOLO model achieved optimal performance, improving accuracy while maintaining low computational overhead. Compared with the YOLOv8s baseline model, the precision (P) increased by 2.3%, the recall (R) increased by 5.2%, and the AP increased by 5.4%, while the number of parameters and computational complexity were reduced by 19.8% and 20.8%, respectively. Experimental results verify the effectiveness of the present invention in module design and combination strategy.

[0154] Table 2 Ablation experiment results

[0155]

[0156] The EMA module's attention mechanism dynamically adjusts attention weights based on feature saliency to effectively mitigate the impact of environmental noise and accurately capture landslide characteristics. To verify the EMA module's improved model performance, ablation experiments and comparative experiments with different attention mechanisms were conducted. Five attention mechanisms, CA, SE, CBAM, ECA, and EMA, were introduced into the C2f_RVB module to analyze their impact on model performance. The experimental results are shown in Table 3.

[0157] Table 3 Comparison of different attention mechanisms in RepViT Block

[0158]

[0159] Experimental results show that introducing different attention mechanisms into the C2f_RVB module can improve model performance to varying degrees; the model using the EMA attention mechanism performs best in all indicators; compared with the model without the attention mechanism, the model using the EMA attention mechanism improves the P, R, F1 and AP indicators by 2.6%, 7.5%, 5% and 4.1% respectively; experimental results prove that the EMA attention mechanism can enhance the feature extraction ability of the model and improve its adaptability in complex terrain.

[0160] The present invention designs a RepViTBlock module in the backbone feature extraction network and integrates the EMA attention mechanism into part of the RepViTBlock. This design captures the local spatial characteristics of the landslide area through a three-branch structure, while adaptively adjusting the weights of important features. This solves the problem of missed detection caused by interference factors such as lighting changes and cloud occlusion in complex environments in traditional models, and improves the feature extraction capability of the backbone network.

[0161] The present invention introduces the RepGDSFPN feature fusion network into the feature fusion network, which improves the fusion effect of multi-scale features. By cascading multiple CSPStage modules, the network optimizes the interaction between high-level semantic features and low-level spatial features, thereby enhancing the ability to distinguish between landslide and non-landslide areas. In addition, the Dysample module replaces the traditional upsampling method, reduces the loss of fine-grained features, and improves the detection effect and accuracy of the model in complex terrain.

[0162] The LSDECD detection head of the present invention improves the positioning and classification performance of the model, reduces the dependence of batch normalization on batch size, and enhances stability in different environments; the DEConv module effectively identifies the detailed features of the landslide boundary through multi-directional differential convolution, further improving the prediction accuracy of the bounding box.

Claims

1. A complex terrain landslide detection method based on the CTL-YOLO model, characterized in that: Use the following steps: S1. Collect images of landslides on complex terrain to obtain landslide images and create terrain remote sensing datasets; S2. marking the landslide area on the collected landslide image; S3, randomly dividing the landslide images and their corresponding annotation files into training set, validation set and test set; S4. Perform data augmentation on images in the training set; S5. Build the CTL-YOLO model; S6. Use terrain remote sensing dataset to train and verify the CTL-YOLO model.

2. A complex terrain landslide detection method based on the CTL-YOLO model according to claim 1, characterized in that: The CTL-YOLO model includes a backbone feature extraction network, a feature fusion network, and a detection head; The backbone feature extraction network includes Conv module, C2f_RVB module and C2f_RVB_EMA module; The feature fusion network adopts the Reparameterized Global Dynamic Sampling Feature Pyramid Network (RepGDSFPN); The detection head adopts the Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD).

3. A complex terrain landslide detection method based on the CTL-YOLO model according to claim 2, characterized in that: The C2f_RVB module is constructed by replacing the Bottleneck module of the C2f module in the YOLOv8 backbone network; the Bottleneck modules of the first, second, and fourth C2f modules are replaced by RepViTBlock modules, and the Bottleneck module of the third C2f module is replaced by RepViTEMABlock modules; the replacement is achieved by adjusting the number of input and output channels of RepViTBlock and RepViTEMABlock to be consistent with the original Bottleneck module, while keeping the overall structure and shortcut connection of the C2f module unchanged; The RepViTBlock module consists of a 3×3 depthwise convolutional branch, a 1×1 depthwise convolutional branch, an identity mapping branch, and a feedforward network (FFN). The identity mapping branch is a residual connection structure that directly passes the input feature map to the subsequent layer without any transformation. The feedforward network (FFN) consists of a sequential structure consisting of two Conv2d convolutional layers and two BatchNorm2d normalization layers, which are connected in sequence and then perform a residual addition operation with the input features. The RepViTEMA Block module is an enhanced structure formed by introducing an EMA module before the feedforward network (FFN) on the basic architecture of the RepViTBlock module. The EMA module is an efficient multi-scale attention module. By introducing a multi-scale feature processing mechanism after fusing the output features of the 3×3 depthwise convolution, 1×1 depthwise convolution, and identity mapping branch and before the feedforward network (FFN) processing, it captures long-range dependencies in the feature map and performs channel-level feature importance weighting, thereby enhancing the model's multi-scale perception of complex terrain landslide characteristics. The EMA module includes branch 1, branch 2, and branch 3; Branch 1 performs initial dimensionality reduction as follows: F1(X)=Conv1(X) Where X∈R C×H×W is the input feature map, where Conv1 reduces the channel dimension to C / r, where r is the reduction ratio, and F1 represents branch 1; Branch 2 captures global context information through bidirectional average pooling as follows: F2(X)=σ(Conv2([Pool h (X),Pool w (X)])) Where, Pool h and Pool w Represents the average pooling operation in the horizontal and vertical directions respectively, Conv2 aggregates the features after pooling, σ represents the Sigmoid activation function, and F2 represents branch 2; Branch 3 focuses on the extraction of local features, as follows: F3(X)=Conv3(X) Where Conv3 is a depth-wise separable convolution that maintains spatial information while having high computational efficiency, and F3 represents branch 3. For multi-scale feature processing, the three branches of EMA are fused through a weighted fusion mechanism: Y=X⊙(α·F r (X)+β·F ctx (X)+γ·F l (X)) where α, β, and γ are learnable weights used to balance the contribution of each branch.

4. A complex terrain landslide detection method based on the CTL-YOLO model according to claim 3, characterized in that: Reparameterized Global Dynamic Sampling Feature Pyramid Network (RepGDSFPN) includes the RepGFPN structure and the Dysample module; The RepGFPN structure includes several CSPStage modules connected in series; The input of the CSPStage module is two or three feature layers, which are divided into two branches after channel splicing; One branch adjusts the number of channels through 1×1 convolution. The other branch is the ELAN feature aggregation module, which consists of n 3×3Rep convolutions and 3×3 standard convolutions; After the two branches are spliced ​​through the channels, a final convolution is performed to obtain the final output.

5. A complex terrain landslide detection method based on the CTL-YOLO model according to claim 4, characterized in that: The Dysample module includes a sampling point generator, a sampling set, and a grid sampling function. The workflow of the Dysample module is as follows: S51. Given an input feature map of size C×H×W, it is converted into a sampling set S of size 2×sH×sW through the sampling point generator. The 2 in the first dimension of the sampling set S represents the x and y coordinates. S52, using the gridsample function to resample the input features to obtain an upsampled feature map of size C×sH×sW; Where H and W represent height and width, and C represents the number of channels; S53. Perform bilinear interpolation on X through the gridsample function to obtain the upsampled feature map X'. The formula is as follows: X′=grid_sample(X,S) The sampling point generator generates an offset O through a linear layer, a dynamic range factor, and a pixel shuffle operation, which is then added to the original grid position G to generate a sampling set S. The generation process of the sampling set S is defined as: S=G+O; The dynamic range factor has a value range of 0.25 to 0.

5.

6. A complex terrain landslide detection method based on the CTL-YOLO model according to claim 5, characterized in that: The lightweight shared detail enhancement convolutional detection head (LSDECD) shares the convolutional layer weights through the decoupled detection head to reduce parameters. The decoupled detection head separates the detection classification task and the bounding box regression task. By making the classification task and the bounding box regression task branches share the pre-convolutional layer parameters, only independent parameters are used in the task-specific layers. The DEConv module is introduced in the shared convolutional layer. The DEConv module includes five parallel branches: conventional convolution (VC), center difference convolution (CDC), angle difference convolution (ADC), horizontal difference convolution (HDC) and vertical difference convolution (VDC). The output feature maps of these five branches are fused by weighted summation in the channel dimension. The lightweight shared detail enhancement convolutional detection head (LSDECD) uses the group normalization (GN) layer instead of the batch normalization (BN) layer. The DEConv module combines multiple convolution kernels into a single standard convolution kernel through reparameterization. The reparameterization process is as follows: In the formula, F out represents the output feature map of DEConv, K i＝1:5 Represent the convolution kernels of VC, CDC, ADC, HDC and VDC respectively. The symbol * indicates the convolution operation. K cvt Represents the equivalent convolution kernel obtained by combining these parallel convolutions; The Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD) introduces a scale layer to adaptively scale the output features of the regression branch of each detection layer. Specifically, by setting a learnable scaling parameter for each detection layer, the optimal scaling factor is automatically adjusted during model training, thereby dynamically balancing the contributions of different feature layers and maintaining scale consistency between detection heads. The initial value of the scaling parameter is 1.

0.

7. A complex terrain landslide detection method based on the CTL-YOLO model according to claim 5, characterized in that: The images in S1 include images of landslides, loess landslides, and debris flows. The data is collected using remote sensing satellites, drones or ground-based photography equipment; In S2, the CVAT annotation tool is used to annotate the landslide area on the collected images and generate a standardized rectangular bounding box label file; In S3, the landslide images and their corresponding label files are randomly divided according to a uniform ratio of 7:2:1; Data augmentation in S3 uses geometric transformation and color transformation; geometric transformation includes random horizontal and vertical translation, rotation, scaling, cropping, and up and down and left and right flipping; color transformation includes random changes in hue, saturation, and brightness.

8. A complex terrain landslide detection method based on the CTL-YOLO model according to claim 7, characterized in that: The specific operation of S6 is as follows: S61. Set up the experimental environment: the image input size is set to 640×640, the optimizer is SGD, the initial learning rate is set to 0.01, the momentum factor is 0.937, the weight decay coefficient is 0.0005, the batch size is 16, and the number of training iterations is 200 rounds; S62. Select evaluation indicators: The evaluation indicators include precision (P), recall (R), score (F1), and average precision (AP). The specific calculation formula is as follows: TP, FP, and FN represent the counts of correctly predicted positive samples, incorrectly predicted positive samples, and incorrectly predicted negative samples, respectively; P measures the proportion of correctly predicted landslides among all predictions, and R measures the proportion of correctly predicted landslides among actual landslides; the F1 score is the harmonic mean of P and R, and AP is the area under the PR curve, which is used as a comprehensive performance indicator; The lightweight CTL-YOLO model is evaluated using three metrics: parameter count (Params), floating-point operations (FLOPs), and frames per second (FPS); Params evaluates spatial complexity, GFLOPs quantifies computational complexity, and FPS measures processing speed.

Images