Near-ground landslide detection method and system and computer readable storage medium

Abstract

The invention relates to the technical field of computer vision, and particularly discloses a near-earth landslide detection method and system and a computer readable storage medium, a near-earth landslide detection model is constructed based on a YOLOv8 neural network model, and the model adopts a full-dimensional dynamic convolution module to dynamically adjust the convolution kernel and weight, so that the detection accuracy is improved. Therefore, the adaptive capacity of the model under different illumination conditions and landslide forms and states is enhanced; a convolution block attention module is adopted to accurately capture landslide precursor features by utilizing channel and space attention, interference of background noise is effectively inhibited, and robustness and accuracy of landslide detection are enhanced; a generalized feature pyramid module is adopted, and the capability of combining high-layer semantic information and low-layer space details is utilized at the same time, so that key features of the landslide are effectively focused, irrelevant background noise is inhibited, the landslide precursor detection capability is enhanced, and the performance of the model in multi-scale landslide detection is improved. The method has high accuracy and robustness for near-earth landslide in a complex environment.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a near-ground landslide detection method, system and computer-readable storage medium. Background Art

[0002] Landslides are a sudden geological disaster that is not only highly destructive but also difficult to prevent, causing serious economic losses and casualties to human society. Especially in complex geological environments such as mountainous areas, landslides are unpredictable and sudden, and their disasters have a wide range of impacts, covering transportation, construction facilities, agricultural land, water sources and other aspects.

[0003] In recent years, with the rapid development of computer vision and artificial intelligence technologies, landslide detection technology based on images and videos has gradually become a research hotspot. These technologies rely on real-time analysis of surveillance videos and can quickly identify landslides before or in the early stages, thus providing strong support for prevention and disaster reduction. However, landslide detection still faces multiple challenges in complex geological environments. First, the manifestations of landslides under different terrain conditions vary significantly, and the scale, speed, and morphology of landslides vary, which places extremely high demands on the versatility of detection algorithms. Second, changes in vegetation coverage and lighting conditions increase the complexity of image processing, especially in low-light or high-contrast environments, where the performance of traditional detection algorithms is often difficult to guarantee. In addition, the precursory features of landslides are usually not obvious, and it is necessary to accurately capture small changes in surveillance images, which further increases the difficulty of detection.

[0004] Traditional landslide detection methods mainly rely on ground sensors, satellite remote sensing and other technologies. Although these methods can provide certain monitoring capabilities, they are insufficient in real-time, flexibility and accuracy. With the rise of deep learning technology, landslide detection methods based on convolutional neural networks (CNN) have gradually become the mainstream research direction. The YOLO (You Only Look Once) series of models are widely used in the field of landslide detection due to their high efficiency and accuracy in target detection tasks. As the latest version of the YOLO series, YOLOv8 has excellent performance in target detection, but it is still difficult to detect near-ground landslides in real time in complex landslide environments (different terrains, vegetation coverage, and lighting conditions), variable landslide shapes (irregular, different sizes), and multi-state landslide conditions (weak landslide precursors, rapid landslide dynamics, and stable end of landslides). Summary of the invention

[0005] The present invention provides a near-ground landslide detection method, system and computer-readable storage medium, which solve the technical problem that the existing near-ground landslide real-time detection method has poor detection performance in complex environments (variable landslide shapes, landslide states, vegetation coverage and lighting conditions, etc.).

[0006] In order to solve the above technical problems, the present invention provides a near-ground landslide detection method, comprising:

[0007] Construct a near-ground landslide detection model based on the YOLOv8 neural network model;

[0008] Training and verifying the constructed near-ground landslide detection model;

[0009] The trained and verified near-ground landslide detection model is used in actual near-ground landslide monitoring;

[0010] The constructed near-ground landslide detection model includes backbone network, neck network and head network;

[0011] The backbone network includes a first-layer convolution module, a second-layer convolution module, a first-layer C2f module, a third-layer convolution module, a second-layer C2f module, a first-layer convolution block attention module, a fourth-layer convolution module, a third-layer C2f module, a second-layer convolution block attention module, a fifth-layer convolution module, a fourth-layer C2f module, a third-layer convolution block attention module, and an SPPF module, which are sequentially connected; the second-layer convolution module, the third-layer convolution module, the fourth-layer convolution module, and the fifth-layer convolution module adopt a full-dimensional dynamic convolution module; the neck network adopts a generalized feature pyramid module, whose input is connected to the first-layer C2f module, the second-layer C2f module, the third-layer C2f module and the SPPF module, and its output is connected to the four prediction modules of the head network.

[0012] Furthermore, the full-dimensional dynamic convolution module is different from the convolution module in the backbone network of the YOLOv8 neural network model in that:

[0013] Replace the first convolutional layer in the convolutional module in the backbone network of the YOLOv8 neural network model with a full-dimensional dynamic convolutional layer;

[0014] The input of the full-dimensional dynamic convolution layer is x, the output is y, and the operation performed is:

[0015] Perform global average pooling on x;

[0016] Extract the features after global average pooling through the fully connected layer;

[0017] The output of the fully connected layer is nonlinearly transformed through the rectified linear unit to map the features to a low-dimensional space;

[0018] The output of the rectified linear unit is dynamically convolved using a linear combination of n convolution kernels;

[0019] Multiply the result of the dynamic convolution with the input x to get the output y.

[0020] Furthermore, the input image is convolved by linear combination of n convolution kernels and dynamically weighted by the attention mechanism;

[0021] The dynamically weighted attention includes four types of attention scalars, namely the attention scalar assigned to the current convolution kernel, and the attention scalar calculated along the spatial dimension, input channel dimension, and output channel dimension of the current convolution kernel.

[0022] Furthermore, the first-layer convolutional block attention module, the second-layer convolutional block attention module, and the third-layer convolutional block attention module all adopt convolutional attention modules;

[0023] The operations performed by the convolutional attention module are:

[0024] Take the feature map F as input and get the one-dimensional channel attention map M through the channel attention module c (F) is multiplied by F to get F′, and then the spatial attention module extracts the spatial attention of F′ to get the two-dimensional spatial attention map M s (F′) is multiplied by F′ to obtain the final refined output F″.

[0025] Furthermore, the channel attention module performs the following operations:

[0026] The spatial information of the feature map F is aggregated through average pooling and maximum pooling operations to generate average pooling features And the maximum pooling feature

[0027] The average pooling feature And the maximum pooling feature Generate a one-dimensional channel attention map M through a shared network c (F);

[0028] The shared network consists of a multi-layer perceptron, averaging the pooled features And the maximum pooling feature The weights W0 and W1 of the two layers of the multilayer perceptron are shared and activated by the sigmoid activation function, and the ReLU activation function is followed by W0;

[0029] The operations performed by the spatial attention module are:

[0030] Generate the corresponding average pooling feature of the feature map F′ through average pooling and maximum pooling And the maximum pooling feature

[0031] The average pooling feature is performed through a standard convolutional layer. And the maximum pooling feature After concatenation, convolution is performed to generate the required two-dimensional spatial attention map M s (F′).

[0032] Furthermore, a cross-stage local network is introduced after each Concat operation of the generalized feature pyramid module, and the cross-stage local network consists of a partial dense block and a partial transition layer; the input feature map x0=[x′0,x″0] is divided into two parts, wherein a part x′0 is directly connected to the partial transition layer, and the other part x″0 is processed by the partial dense block and then input to the partial transition layer;

[0033] The partial dense block feeds x″0 through k weighted dense layers to obtain the overall output [x″0, x″1, ..., x″ k ], x″1 to x″ k Respectively represent the outputs of the 1st to the kth dense layers, [] represents a connection;

[0034] The partial transition layer is provided with a first transition layer and a second transition layer, wherein the first transition layer converts the weight matrix W T The output of the partial dense block [x″0,x″1,...,x″ k ] and get the output x T ; The first transition layer converts the weight matrix W U and [x′0,x T ] to get the final output x U .

[0035] Furthermore, in each prediction module of the head network, the WIoU loss function is used to replace the Bbox bounding box loss function in YOLOv8.

[0036] Furthermore, during the training and validation process, the constructed training set and validation set include two types of data sets, type one is the near-ground landslide pictures, type two is the light bar pictures with two reflective light bars inserted at fixed positions;

[0037] For a near-ground landslide image, the near-ground landslide detection model marks a landslide area;

[0038] For the light bar image, the near-ground landslide detection model marks the position of the light bar;

[0039] In the practical application of the near-ground landslide detection model, two reflective light strips of the same shape and size and placed in parallel are fixed at fixed positions on the mountain to be detected within the visual range of the camera; the detection results of the light strips and the landslide area by the near-ground landslide detection model are used to determine whether there is a landslide and calculate the landslide distance;

[0040] Light strip detection and positioning may face three situations during landslides:

[0041] In the first case, the upper light bar is not affected by the landslide and remains in the original position, while the lower light bar moves with the landslide. The light bar is in or out of the camera's visible range. In this case, the landslide state is determined by detecting the position and number of the two light bars. In addition, when there are two light bars, the landslide distance is calculated by the position of the two light bars.

[0042] In the second case, both the upper and lower light bars move with the landslide, and there are light bars in the camera's visible range. At this time, the landslide state is determined by detecting the positions and number of the two light bars. At this time, the landslide distance cannot be measured;

[0043] In the third case, both light bars are out of the visible range and are taken away or disappear with the landslide. In this case, the light bar judgment result is directly defined as invalid. It is necessary to determine the landslide situation by detecting the landslide mass within the visible range. If a landslide mass is detected, the landslide state is determined, otherwise there is no landslide.

[0044] The present invention also provides a near-ground landslide detection system, the key of which is that: it is provided with an intelligent agent for executing the near-ground landslide detection method, the intelligent agent constructs a near-ground landslide detection model based on the YOLOv8 neural network model, and or trains and verifies the constructed near-ground landslide detection model, and or uses the trained and verified near-ground landslide detection model in actual near-ground landslide monitoring.

[0045] The present invention also provides a computer-readable storage medium, the key of which is that a computer program executable by a processor is stored therein, and the computer program executes the near-ground landslide detection method.

[0046] The near-ground landslide detection method, system and computer-readable storage medium provided by the present invention construct a near-ground landslide detection model based on the YOLOv8 neural network model, and the model integrates a full-dimensional dynamic convolution module, a convolution block attention module, a generalized feature pyramid module and a WIOU loss function. The full-dimensional dynamic convolution module dynamically adjusts the convolution kernel and weights, thereby enhancing the adaptability of the model under different lighting conditions, landslide forms and states; the convolution block attention module uses channel and spatial attention to accurately capture landslide precursor features such as surface cracks and displacements, and effectively suppresses the interference of background noise, thereby enhancing the robustness and accuracy of the model for landslide detection; the generalized feature pyramid module simultaneously utilizes the ability to combine high-level semantic information with low-level spatial details, effectively focuses on key landslide features and suppresses irrelevant background noise, thereby enhancing the detection capability of landslide precursors and improving the performance of the model in multi-scale landslide detection; the WIoU loss function is used to replace the traditional IoU loss function, to enhance the focus on high-quality candidate frames, and to suppress harmful gradients of low-quality samples, thereby improving the positioning accuracy of the model and the model's detection capability for small targets. In general, the present invention provides a near-ground landslide detection method, system and computer-readable storage medium, which have high accuracy and robustness for near-ground landslides in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a structural diagram of a near-ground landslide detection model provided by an embodiment of the present invention;

[0048] Figure 2 is an improved position diagram of the ODConv module provided in an embodiment of the present invention;

[0049] Figure 3 is a structural diagram of the ODConv layer provided in an embodiment of the present invention;

[0050] Figure 4 Schematic diagram of spatial dimension direction multiplication operation provided by an embodiment of the present invention;

[0051] Figure 5 Schematic diagram of input channel dimensional direction multiplication operation provided by an embodiment of the present invention;

[0052] Figure 6 Schematic diagram of clockwise multiplication operation of output channel dimension provided by an embodiment of the present invention;

[0053] Figure 7 Schematic diagram of convolution kernel spatial dimension multiplication operation provided by an embodiment of the present invention;

[0054] Figure 8 is a schematic diagram of a CBAM provided by an embodiment of the present invention;

[0055] Fig. 9is a schematic diagram of a channel attention module provided by an embodiment of the present invention;

[0056] Fig.10 is a schematic diagram of a spatial attention module provided by an embodiment of the present invention;

[0057] Fig.11 is a comparison diagram of the traditional feature pyramid structure and the GFPN structure provided by an embodiment of the present invention;

[0058] Fig.12 is a schematic diagram of a CSP module provided by an embodiment of the present invention;

[0059] Fig.13 It is a YOLOv8 bounding box loss function replacement comparison diagram provided by an embodiment of the present invention;

[0060] Fig.14 This is an example diagram of the landslide mass detection effect provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0062] The near-ground landslide detection method provided by the embodiment of the present invention includes:

[0063] Construct a near-ground landslide detection model based on the YOLOv8 neural network model;

[0064] Training and validation of near-ground landslide detection models;

[0065] The trained and verified near-ground landslide detection model is used in actual near-ground landslide monitoring.

[0066] The architecture of the near-ground landslide detection model constructed in this embodiment is as follows: Figure 1 As shown in the figure, it includes the backbone network (Backbone), the neck network (Neck) and the head network (Head). Compared with the YOLOv8 neural network model, the improvements of the near-ground landslide detection model are:

[0067] Replace all convolution modules (layers 2 to 5) in the backbone network of the YOLOv8 neural network model except the first convolution module (Conv) with full-dimensional dynamic convolution modules (ODConv);

[0068] Add a convolutional block attention module (CBAM) after the second-layer C2f module, the third-layer C2f module, and the fourth-layer C2f module of the backbone network of the YOLOv8 neural network model;

[0069] Replace the Neck network (Neck) of the YOLOv8 neural network model with the Generalized Feature Pyramid Module (GFPN).

[0070] After the above improvements, the architecture of the near-ground landslide detection model is as follows: Figure 1 As shown. The backbone network (Backbone) of the near-ground landslide detection model includes the first layer of convolutional modules (Conv), the second layer of convolutional modules (full-dimensional dynamic convolutional modules ODConv), the first layer of C2f modules, the third layer of convolutional modules (full-dimensional dynamic convolutional modules ODConv), the second layer of C2f modules, the first layer of convolutional block attention modules (CBAM), the fourth layer of convolutional modules (full-dimensional dynamic convolutional modules ODConv), the third layer of C2f modules, the second layer of convolutional block attention modules (CBAM), the fifth layer of convolutional modules (full-dimensional dynamic convolutional modules ODConv), the fourth layer of C2f modules, the third layer of convolutional block attention modules (CBAM), and the SPPF module (fast spatial pyramid pooling layer). The neck network (Neck) of the near-ground landslide detection model adopts the generalized feature pyramid module (GFPN), whose input is connected to the first layer of C2f modules, the second layer of C2f modules, the third layer of C2f modules and the SPPF module, and its output is connected to the four prediction modules (Detect) of the head network (Head) of the near-ground landslide detection model.

[0071] In the YOLOv8 network structure, the basic feature extraction module is the Conv convolution module. Through a series of convolution, normalization, and activation operations, the network's ability to capture and represent image features is enhanced, so that feature information at all levels can be better integrated, thereby improving the detection accuracy and efficiency of the model. However, the Conv convolution module of the YOLOv8 basic model has low accuracy and is not flexible enough for target detection in low-light environments, variable landslide areas, and different landslide states. Therefore, in the Backbone part of the YOLOv8 basic network, the original Conv convolution module (layers 2 to 5) is replaced by the ODConv convolution module. The specific replacement position is shown in 1. The detailed comparison of the Conv convolution module and the ODConv convolution module before and after the improvement is shown in Figure 1. Figure 2 As shown in the figure, the first convolution layer (Conv2d) in the original Conv convolution module is replaced by the ODConv convolution layer (full-dimensional dynamic convolution layer). The improved ODConv convolution module effectively enhances the ability to focus on the landslide area and dynamics of the near-ground video in complex environments and improves the detection accuracy.

[0072] The structure of the ODConv convolutional layer is as follows Figure 3 As shown, its input is x, the output is y, and the operation performed by the ODConv convolutional layer is:

[0073] Perform global average pooling (GAP) on x to convert it into a feature vector with a length equal to the number of input channels. This feature vector represents the global information of the entire input feature map.

[0074] Through the fully connected layer (FC), the features of the output of the global average pooling layer are further extracted;

[0075] The output of the fully connected layer is nonlinearly transformed through the Rectified Linear Unit (ReLU) to map it to a low-dimensional space;

[0076] The output of the rectified linear unit is dynamically convolved using a linear combination of n convolution kernels;

[0077] Multiply the result of the dynamic convolution with the input x to get the output y.

[0078] In dynamic convolution, the spatial dimension is used to adjust the sampling position, stride and padding of the convolution kernel, the input channel dimension is used to dynamically calculate and assign the weight of each input channel, the output channel dimension is used to dynamically calculate and apply the weight of each output channel, and the convolution kernel dimension is used to assign attention weights to multiple convolution kernels. Finally, the feature map is convolved with the dynamically adjusted convolution kernel and the processed feature map is output, thereby enhancing the model's feature extraction capability and performance.

[0079] Dynamic convolution uses a linear combination of n convolution kernels and dynamically weights them with an attention mechanism to make the convolution operation dependent on the input features. Mathematically, the dynamic convolution operation using a linear combination of n convolution kernels can be defined as:

[0080] Y=(α w1 W1+…+α wn W n )*X (1)

[0081] in, and They are the input feature map and output feature map of the dynamic convolution respectively (h and w represent the height and width of the input feature map and the output feature map, c in 、c out Respectively represent the number of channels of the input feature map and the output feature map), W i (i=1,2,…,n) represents the convolution kernel of the i-th convolution (given by c outFilters The kernel composed of k×k represents the size of the filter, that is, the spatial dimension of the convolution kernel, m=1,…,c out ), It is the attention scalar of the i-th convolution.

[0082] Based on formula (1), the dynamic convolution in this example is defined as follows:

[0083] Y=(α w1 ⊙α f1 ⊙α c1 ⊙α s1 ⊙W1+…+α wn ⊙α fn ⊙α cn ⊙α sn ⊙W n )*X (2)

[0084] in, and are the three newly introduced attentions, which are respectively along the convolution kernel W i The spatial dimension (k×k) and input channel dimension (c in ) and the output channel dimension (c out ) is calculated; ⊙ is the multiplication operation of different dimensions of the kernel space. Here, α si , α ci , α fi and α wi is through the corresponding multi-head attention module π i (x) is calculated.

[0085] In the ODConv layer, for the convolution kernel W i , α si Assign different attention scalars, α, to the convolution parameters (each filter) at k×k spatial locations. ci For each convolution filter W i m c in Different attention scalars are assigned to each channel; α fi c out Convolutional transformers assign different attention scalars; α wi Assign an attention scalar to the entire convolution kernel.

[0086] like Figure 3 As shown in the figure, before performing each attention dynamic convolution, the features need to be fully connected (FC) and processed with Sigmoid function.

[0087] Figures 4 to 7 The process of multiplying four types of attention by n convolution kernels is shown. Figure 4This is a schematic diagram of multiplication operation in the spatial dimension direction. Figure 5 This is a schematic diagram of the multiplication operation in the channel direction of the input channel dimension. Figure 6 This is a schematic diagram of clockwise multiplication operation in the output channel dimension. Figure 7 Schematic diagram of the multiplication operation of the convolution kernel space dimension. In principle, these four attentions are complementary, and they are gradually multiplied by the convolution kernel W in the order of position, channel, filter, and kernel. i , can make the convolution operation different. All spatial positions of input X, all input channels, all filters and all kernels provide performance guarantee for capturing rich context clues. Therefore, the ODCony layer can significantly enhance the feature extraction ability of the basic convolution operation of the neural network. In addition, when n=1, there is α si , α ci , α wi are all 1, ODConv using a single convolution kernel can be comparable to or even better than the standard CondConv and DyConv, while the additional parameters introduced to the final model are greatly reduced.

[0088] By applying the attention mechanism on multiple dimensions of the convolution kernel, the feature extraction and representation capabilities of the convolutional neural network are significantly improved. Traditional static convolution kernels perform the same operation on all input samples, while dynamic convolution combines multiple convolution kernels and dynamically adjusts the size, shape and weight of the convolution kernel according to the input features, so that the convolution operation can adapt to different landslide characteristics. The ODConv layer dynamically adjusts the four dimensions of the convolution kernel space, including the spatial dimension, input channel dimension, output channel dimension and convolution kernel dimension, thereby improving the flexibility and accuracy of the model.

[0089] In this example, a convolutional block attention module (CBAM) is added to the YOLOv8 backbone network. Its specific location information is as follows: Figure 1 shown. Figure 8 The schematic diagram of CBAM is shown as follows. (c represents the number of channels of the intermediate feature map) as input, and a one-dimensional channel attention map is obtained through the channel attention module. After multiplying by F, we get F′, and then use the spatial attention module to extract the spatial attention of F′ to get a two-dimensional spatial attention map Multiply it with F′ to get the final refined output F″. The entire attention process of CBAM can be summarized as:

[0090]

[0091] In the formula, Denotes element-wise multiplication. During the multiplication, the attention values ​​are broadcasted (copied) accordingly: channel attention values ​​are broadcasted along the spatial dimension and vice versa.

[0092] Fig. 9 The following is a schematic diagram of the channel attention module, which uses the maximum and average pooling outputs to share the network. First, the spatial information of the feature map is aggregated through average and maximum pooling operations (AvgPool, MaxPool) to generate two different spatial context descriptors: and Represent the average and maximum pooling features respectively. Then, these two descriptors are forwarded to the shared network to generate the channel attention map The shared network consists of a multilayer perceptron (MLP) with one hidden layer. To reduce parameter overhead, the hidden activation size is set to Where r is the reduction rate. After applying the shared network to each descriptor, the output feature vectors are merged using element-wise summation. The channel attention calculation formula is as follows:

[0093]

[0094] Where σ(·) represents the sigmoid activation function, which acts on and Note that both inputs share the MLP weights W0 and W1, and the ReLU activation function is followed by W0.

[0095] Fig.10 The figure shows the schematic diagram of the spatial attention module. The spatial attention module uses similar two outputs to pool along the channel axis and forward them to the convolutional layer. The channel information of the feature map is summarized through two pooling operations (MaxPool and AvgPool) to generate two two-dimensional feature maps: and represents the average, maximum pooled features of the entire channel. These features are then concatenated and convolved through a standard convolutional layer to generate the required two-dimensional spatial attention map. The spatial attention is calculated as follows:

[0096]

[0097] In the formula, f 7×7 (·) denotes a convolution operation with a filter size of 7×7.

[0098] CBAM realizes the interaction between feature map dimensions through its two branch structures of channel attention and spatial attention, so as to accurately calculate the attention weights and better capture the local and global contextual information in the image. In the channel attention module, the global average and maximum pooling operations are used to activate and strengthen the interdependence between channels, highlight the precursor features of landslides such as surface cracks and displacement features, and enable the model to focus on these key features more accurately. In the spatial attention module, the spatial attention map is generated through convolution operations, which can effectively locate the specific area where the landslide occurs and suppress irrelevant background noise. Through this sequential processing, CBAM adaptively optimizes the feature representation, significantly improving the robustness and accuracy of the model for landslide detection.

[0099] This example introduces the GFPN (Generalized Feature Pyramid Network) structure in YOLOv8, and its improvements are as follows Figure 1 GFPN enhances the effect of feature fusion through dense-link and Queen-fusion. Fig.11 As shown in the comparison chart of the traditional feature pyramid and GFPN, compared with the traditional feature pyramid network (FPN) ( Fig.11 (a) shows that GFPN ( Fig.11 (b) shows that a cross-level feature fusion method is used to achieve a more refined feature fusion of features extracted at different scales. This feature fusion shows excellent adaptability in variable landslide shapes, different terrain conditions and multi-state landslide situations.

[0100] Efficient multi-level feature fusion: GFPN ensures smoother and more complete information transfer between features at different levels by introducing dense connections and Queen fusion structures. This design can effectively integrate detailed information from shallow layers and semantic information from deep layers, thereby enhancing the model's ability to detect small targets (such as landslide precursors).

[0101] Enhanced feature reuse and information flow: The dense connections in GFPN ensure that each layer of features can be integrated with the features of all previous layers, which not only improves the efficiency of information flow, but also strengthens the ability of feature reuse, thereby improving the detection accuracy of the model, especially when dealing with complex geological disasters such as landslides, it can more accurately capture landslide precursors and dynamic changes.

[0102] Optimized cross-scale feature fusion: GFPN uses the Queen fusion module to enhance the model's adaptability to features of different scales by using cross-scale connections, which can reduce information loss during the detection process and effectively integrate landslide features of different scales. This capability is particularly important in complex and changing landslide environments because it can help the model better cope with detection tasks under different terrains, lighting, and landslide conditions.

[0103] Higher robustness and accuracy: Through the improvement of GFPN, the YOLOv8 model shows higher robustness and accuracy in detecting small objects and handling variable landslide shapes. GFPN not only performs well in reducing feature information loss, but also improves the overall performance of the model through a more effective feature fusion mechanism, especially in better coping with the challenges of landslide detection in practical applications.

[0104] In the landslide detection task, a cross-stage local network (CSP) is introduced after each Concat operation in GFPN (e.g. Figure 1 As shown in Figure 2, YOLOv8 achieves richer gradient combinations while reducing the amount of computation. Landslide detection requires the model to efficiently extract and process multi-scale features in a complex and changing geological environment while maintaining computational efficiency.

[0105] CSP effectively optimizes the computational bottleneck and significantly improves the network's learning ability and reasoning speed by dividing the input feature map into two parts and fusing them using a cross-stage hierarchical structure. The design of some dense blocks and some transition layers reduces redundant calculations and improves the efficiency of gradient flow. This enables the model to more accurately capture key information such as surface cracks and displacements when detecting landslide precursor features, and effectively locate landslide areas in complex terrain. At the same time, CSP can integrate feature information at different levels and performs well in detecting small targets in landslides, enhancing the adaptability and robustness of the model, thereby improving the overall performance of landslide detection.

[0106] The single-stage architecture of CSP consists of some dense blocks and some transition layers, such as Fig.12 As shown. In the partial dense block, the input feature map x0 = [x′0, x″0] is divided into two parts, one part x′0 is directly connected to the partial transition layer, and the other part x″0 is processed by the partial dense block. In the partial transition layer, the output of the partial dense block x0 = [x″0, x″1, ..., x″ k ] First, it passes through a transition layer to generate an intermediate output x T , then, x T It is connected again with the initially segmented x′0 and passes through another transition layer to finally generate the output x UThis architecture optimizes the feedforward transfer process and weight update through effective feature segmentation and fusion strategies, so that the network achieves a balance between computational efficiency and learning ability. The feedforward transfer calculation formula is as follows:

[0107]

[0108] In the above formula, W k The weight matrix of the dense layer k (here k represents the number of dense layers) of the partial dense block is used to obtain the weight matrix of the dense layer k to the dense layer k-1 ([x″0,x″1,...,x″ k-1 ]) The output x″ of the convolution k .W T Represents the weight matrix of the first transition layer in the partial transition layer, which is used to transform the output feature map [x″0,x″1,...,x″ of the partial dense block k ] is converted to output x T .W U Represents the weight matrix of the second transition layer, which is used to transform x T The concatenated feature map of x′0 is transformed into the final output x U . * indicates matrix multiplication.

[0109] The bounding boxes involved in landslide detection usually have irregular shapes and large scale differences. The traditional IoU loss function is prone to gradient vanishing problems in this case, resulting in the model being unable to effectively capture small targets or areas with low boundary overlap. This example replaces the IoU loss function in the head network (Head) of the YOLOv8 neural network model with the WIoU loss function. Specifically, the WIoU loss function is used to replace the Bbox bounding box loss function in YOLOv8. The specific location is as follows Fig.13 shown.

[0110] The WIoU (Wise Intersection over Union) loss function is a dynamic non-monotonic focusing mechanism based on IoU, which is used to solve the bottleneck problem of bounding box regression in target detection tasks. The WIoU loss function dynamically adjusts the gradient gain, allocates higher attention to anchor boxes of normal quality, and reduces attention to high-quality and low-quality anchor boxes, thereby avoiding the harmful gradient effects caused by low-quality samples and improving the generalization ability of the model. Compared with the traditional IoU loss function, WIoU effectively improves the accuracy of small target detection and significantly improves the performance of the overall model by introducing distance attention and dynamic adjustment mechanisms.

[0111] Based on the above method, an embodiment of the present invention also provides a near-ground landslide detection system, which is provided with an intelligent agent, which constructs a near-ground landslide detection model based on the YOLOv8 neural network model, and or trains and verifies the constructed near-ground landslide detection model, and or uses the trained and verified near-ground landslide detection model in actual near-ground landslide monitoring.

[0112] Based on the above method and system, an embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program executable by a processor, and the computer program executes the above near-ground landslide detection method.

[0113] In summary, the near-ground landslide detection method, system and computer-readable storage medium provided by the embodiments of the present invention construct a near-ground landslide detection model based on the YOLOv8 neural network model. The model integrates a full-dimensional dynamic convolution module, a convolution block attention module, a generalized feature pyramid module and a WIOU loss function. The full-dimensional dynamic convolution module dynamically adjusts the convolution kernel and weights, thereby enhancing the adaptability of the model under different lighting conditions, landslide forms and states; the convolution block attention module uses channel and spatial attention to accurately capture landslide precursor features such as surface cracks and displacements, and effectively suppresses the interference of background noise, thereby enhancing the robustness and accuracy of the model for landslide detection; the generalized feature pyramid module simultaneously utilizes the ability to combine high-level semantic information with low-level spatial details to effectively focus on key landslide features and suppress irrelevant background noise, thereby enhancing the detection capability of landslide precursors and improving the performance of the model in multi-scale landslide detection; the WIoU loss function is used instead of the traditional IoU loss function to enhance the focus on high-quality candidate frames and suppress harmful gradients of low-quality samples, thereby improving the positioning accuracy of the model and the model's ability to detect small targets. In general, the present invention provides a near-ground landslide detection method, system and computer-readable storage medium, which have high accuracy and robustness for near-ground landslides in complex environments.

[0114] The following is an experimental verification.

[0115] The background of this experiment is near-ground video landslide detection. The first detection is to directly detect the landslide area. The second detection is to insert two reflective light strips at fixed positions on the inclined slope. By detecting the position of the light strips, it is used to assist in judging and calculating whether the mountain has a landslide and the displacement of the landslide.

[0116] For detection one, the image size remains consistent, with 3200 images in the training set and 800 images in the validation set.

[0117] For detection 2, the image size remains consistent, with 2000 images in the training set and 600 images in the validation set.

[0118] This example uses precision P, recall R, single-class detection average precision AP, and mixed detection average precision mAP as the main evaluation indicators of the model. Precision P indicates the proportion of samples predicted as positive examples that are correctly predicted. Recall R indicates the proportion of samples actually predicted as positive examples that are correctly predicted. Average precision AP is the area of ​​the curve with precision P of a single-class dataset as the Y-axis and recall R as the X-axis. mAP is the average precision AP value of the detection of two types of datasets.

[0119] In this experiment, a light bar positioning detection and landslide determination process was designed. First, two reflective light bars of the same shape and size and placed in parallel were fixed on the slope of the mountain. The near-ground landslide detection model was used for detection and positioning. The model outputs the bounding box coordinates of each light bar, including the coordinates of the upper left corner and the lower right corner (x min ,y min ) and (x max ,y max ). The displacement of the light bar can be used to determine whether a landslide has occurred and the distance of the landslide.

[0120] Light bar detection and positioning may face three situations during the landslide process, so they need to be defined separately. In the first case, the upper light bar is not affected by the landslide and remains in the original position, while the lower light bar moves with the landslide. The light bar is in or out of the camera's visible range. At this time, the landslide state is determined by detecting the position and number of the two light bars. In addition, the approximate landslide distance can be calculated by the presence of two light bars. In the second case, both the upper and lower light bars move with the landslide, and there is a light bar in the camera's visible range. At this time, the landslide state is determined by detecting the position and number of the two light bars. At this time, the landslide distance cannot be measured. In the third case, both light bars are out of the visible range, have been taken away or disappeared with the landslide. At this time, the light bar judgment result is directly defined as invalid. It is necessary to determine the landslide situation by detecting the landslide mass within the visible range. If a landslide mass is detected, the landslide state is determined, otherwise there is no landslide. Since the experimental background is a landslide, the main task is to detect dynamic landslides. The light bar detection and positioning in front is to assist and measure the details of the landslide, while the direct detection of landslides is the main way to determine the state of the landslide and determine the landslide area. The purpose of improving the YOLOv8 model is to adapt to landslide detection in different regions and with different characteristics, thereby improving accuracy and performance. Examples of the detection effect of landslides are as follows: Fig.14 shown.

[0121] Table 1 shows the comparison results of different model detection.

[0122] Table 1

[0123]

[0124] From the comparison of detection data of different models in Table 1, we can see that Model 4 (our near-ground landslide detection model) performs well in light strip and landslide detection tasks by integrating ODConv convolution, CBAM attention mechanism, GFPN feature fusion structure and WIOU loss function, showing its superior performance in complex environments. Overall, Model 4 performs better. It outperforms other models in most key indicators and shows strong detection ability and robustness in key performance. At the same time, the false alarm rate is significantly reduced.

[0125] Experimental results show that the improved model performs excellently in light streak detection and landslide detection tasks, with an AP of 94.1% for light streak detection, 92.3% for landslide detection, and 93.2% on mAP (0.5), which verifies the superiority of the present invention over other detection algorithms in near-ground landslide detection.

[0126] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A near-ground landslide detection method, characterized in that: include: Construct a near-ground landslide detection model based on the YOLOv8 neural network model; Training and verifying the constructed near-ground landslide detection model; The trained and verified near-ground landslide detection model is used in actual near-ground landslide monitoring; The constructed near-ground landslide detection model includes backbone network, neck network and head network; The backbone network includes a first-layer convolution module, a second-layer convolution module, a first-layer C2f module, a third-layer convolution module, a second-layer C2f module, a first-layer convolution block attention module, a fourth-layer convolution module, a third-layer C2f module, a second-layer convolution block attention module, a fifth-layer convolution module, a fourth-layer C2f module, a third-layer convolution block attention module, and an SPPF module, which are sequentially connected; the second-layer convolution module, the third-layer convolution module, the fourth-layer convolution module, and the fifth-layer convolution module adopt a full-dimensional dynamic convolution module; the neck network adopts a generalized feature pyramid module, whose input is connected to the first-layer C2f module, the second-layer C2f module, the third-layer C2f module and the SPPF module, and its output is connected to the four prediction modules of the head network.

2. The near-ground landslide detection method according to claim 1, characterized in that: Compared with the convolution module in the backbone network of the YOLOv8 neural network model, the full-dimensional dynamic convolution module has the following changes: Replace the first convolutional layer in the convolutional module in the backbone network of the YOLOv8 neural network model with a full-dimensional dynamic convolutional layer; The input of the full-dimensional dynamic convolution layer is x, the output is y, and the operation performed is: Perform global average pooling on x; Extract the features after global average pooling through the fully connected layer; The output of the fully connected layer is nonlinearly transformed through the rectified linear unit to map the features to a low-dimensional space; The output of the rectified linear unit is dynamically convolved using a linear combination of n convolution kernels; Multiply the result of the dynamic convolution with the input x to get the output y.

3. The near-ground landslide detection method according to claim 2, characterized in that: The operation of the dynamic convolution is: convolve the input image by linear combination of n convolution kernels and dynamically weight them by attention mechanism; The dynamically weighted attention includes four types of attention scalars, namely the attention scalar assigned to the current convolution kernel, and the attention scalar calculated along the spatial dimension, input channel dimension, and output channel dimension of the current convolution kernel.

4. The near-ground landslide detection method according to claim 3 is characterized in that: The first-layer convolution block attention module, the second-layer convolution block attention module, and the third-layer convolution block attention module all adopt convolution attention modules; The operations performed by the convolutional attention module are: Take the feature map F as input and get the one-dimensional channel attention map M through the channel attention module c (F) is multiplied by F to get F′, and then the spatial attention module extracts the spatial attention of F′ to get the two-dimensional spatial attention map M s (F′) is multiplied by F′ to obtain the final refined output F″.

5. The near-ground landslide detection method according to claim 4 is characterized in that: The operations performed by the channel attention module are: The spatial information of the feature map F is aggregated through average pooling and maximum pooling operations to generate average pooling features And the maximum pooling feature The average pooling feature And the maximum pooling feature Generate a one-dimensional channel attention map M through a shared network c (F); The shared network consists of a multi-layer perceptron, averaging the pooled features And the maximum pooling feature The weights W0 and W1 of the two layers of the multilayer perceptron are shared and activated by the sigmoid activation function, and the ReLU activation function is followed by W0; The operations performed by the spatial attention module are: Generate the corresponding average pooling feature of the feature map F′ through average pooling and maximum pooling And the maximum pooling feature The average pooling feature is performed through a standard convolutional layer. And the maximum pooling feature After concatenation, convolution is performed to generate the required two-dimensional spatial attention map M s (F′).

6. The near-ground landslide detection method according to claim 1, characterized in that: A cross-stage local network is introduced after each Concat operation of the generalized feature pyramid module, and the cross-stage local network consists of a partial dense block and a partial transition layer; the input feature map x0=[x′0, x′0′] is divided into two parts, wherein a part x′0 is directly connected to the partial transition layer, and the other part x′0′ is processed by the partial dense block and then input to the partial transition layer; The partial dense block feeds x′0′ through k weighted dense layers to obtain the overall output [x0″, x1″, ..., x k ″], x1″ to x′ k ′ represents the output of the 1st to the kth dense layers, respectively, and [] represents a connection; The partial transition layer is provided with a first transition layer and a second transition layer, wherein the first transition layer converts the weight matrix W T The output of the partial dense block [x0″, x1″, ..., x k ″] and get the output x T ; The first transition layer converts the weight matrix W U and [x0′,x T ] to get the final output x U .

7. The near-ground landslide detection method according to claim 1 is characterized in that: In each prediction module of the head network, the WIoU loss function is used to replace the Bbox bounding box loss function in YOLOv8.

8. The near-ground landslide detection method according to claim 1 is characterized in that: During the training and validation process, the constructed training and validation sets include two types of data sets. Type one is the image of the near-ground landslide, and type two is the image of light strips with two reflective light strips inserted at fixed positions. For a near-ground landslide image, the near-ground landslide detection model marks a landslide area; For the light bar image, the near-ground landslide detection model marks the position of the light bar; In the practical application of the near-ground landslide detection model, two reflective light strips of the same shape and size and placed in parallel are fixed at fixed positions on the mountain to be detected within the visual range of the camera; the detection results of the light strips and the landslide area by the near-ground landslide detection model are used to determine whether there is a landslide and calculate the landslide distance; Light strip detection and positioning may face three situations during landslides: In the first case, the upper light bar is not affected by the landslide and remains in the original position, while the lower light bar moves with the landslide. The light bar is in or out of the camera's visible range. In this case, the landslide state is determined by detecting the position and number of the two light bars. In addition, when there are two light bars, the landslide distance is calculated by the position of the two light bars. In the second case, both the upper and lower light bars move with the landslide, and there are light bars in the camera's visible range. At this time, the landslide state is determined by detecting the positions and number of the two light bars. At this time, the landslide distance cannot be measured; In the third case, both light bars are out of the visible range and are taken away or disappear with the landslide. In this case, the light bar judgment result is directly defined as invalid. It is necessary to determine the landslide situation by detecting the landslide mass within the visible range. If a landslide mass is detected, the landslide state is determined, otherwise there is no landslide.

9. The near-ground landslide detection system is characterized by: It is provided with an intelligent agent for executing the near-ground landslide detection method described in any one of claims 1 to 8. The intelligent agent constructs a near-ground landslide detection model based on the YOLOv8 neural network model, and or trains and verifies the constructed near-ground landslide detection model, and or uses the trained and verified near-ground landslide detection model in actual near-ground landslide monitoring.

10. A computer-readable storage medium, characterized in that: A computer program executable by a processor is stored therein, and the computer program executes the near-ground landslide detection method according to any one of claims 1 to 8.

Images