Highway slope landslide debris flow disaster identification method, device, equipment and medium

Abstract

The present application relates to the field of geological disaster identification, and particularly relates to a highway slope landslide debris flow disaster identification method, device, equipment and medium. The present application realizes deep fusion of spectral features of optical images and DEM elevation features through a cross-modal cycle fusion module by using high spatial resolution optical remote sensing images of high mountain and canyon area highway slope and DEM data of the same area, and can simultaneously capture texture damage and surface deformation information of landslide debris flow. Meanwhile, when the optical image is seriously covered by cloud and snow, the DEM data branch can be used as an independent data source for identification, ensuring the continuity and stability of the identification work. The present application organically combines two different modal and information complementary remote sensing data, improves the accuracy of landslide debris flow geological disaster identification, can meet the demand of rapid survey of disaster points along the highway in the high mountain and canyon area and major engineering area, can reduce the human workload, and improve the work efficiency.

Description

Technical Field

[0001] This invention relates to the field of disaster detection, and in particular to a method, device, equipment and medium for identifying debris flow disasters caused by landslides on highway slopes. Background Technology

[0002] Landslide debris flows are among the most common and highly dangerous geological hazards in my country, posing a serious threat to people's lives and causing significant property damage. After a disaster, timely and accurate location and identification of the affected area using information technology and intelligent technologies has become a crucial aspect of disaster prevention and early warning systems. With the rapid development of remote sensing Earth observation technology, the spatial and temporal resolution of remote sensing images has continuously improved, data acquisition cycles have been significantly shortened, and costs have continued to decrease. Compared to traditional manual survey methods, remote sensing technology, with its advantages of wide coverage, speed, objectivity, and efficiency, has gradually become the main technical means for regional-scale landslide identification and monitoring.

[0003] Currently, most existing landslide identification studies employ artificial intelligence methods supported by single-source data to achieve large-scale identification. These methods typically use high-resolution optical remote sensing imagery or digital elevation models (DEMs) as the main data source to construct a landslide training sample database, and then directly achieve intelligent identification of landslide targets through deep learning network models. However, in high mountain and canyon areas, i.e., mountainous areas with an average altitude of over 3500 meters, steep slopes, and large areas of exposed bedrock, as well as deeply incised river valleys with a "V"-shaped cross-section, steep slopes (usually >25°), narrow valley bottoms, and large longitudinal gradients, single-source data methods still have significant limitations.

[0004] Taking the Tibet Autonomous Region of my country as an example, this region has a complex geographical environment and drastically varied terrain. Major transportation routes such as the Sichuan-Tibet Highway, Qinghai-Tibet Highway, and Xinjiang-Tibet Highway traverse high mountain and canyon areas, with a dense distribution of landslide disaster points along these routes, particularly in steep terrain areas such as Nyingchi and Chamdo. Due to the year-round snow cover and strong topographical shadows in high-altitude areas, snow cover, shadows, and landslide bodies are extremely similar in color and texture in optical remote sensing images. If feature extraction and training are based solely on a single optical image, it is highly likely to result in missed or false positives. On the other hand, if a single synthetic aperture radar (SAR) data is used for identification, the vast topographical differences in the Tibet Autonomous Region will cause severe problems such as overlay, shadows, and perspective distortion in radar side-view imaging, thereby distorting surface information and affecting the accurate identification of landslide features.

[0005] Therefore, whether using a single optical image or a single radar image, it is difficult to overcome the common problems caused by the special terrain effects in high mountain and canyon areas, and it is impossible to achieve large-scale, accurate and stable identification of landslide debris flows in such complex areas.

[0006] Therefore, there is an urgent need for a method, device, equipment, and medium for identifying debris flow hazards on highway slopes that can combine cross-modal remote sensing data with different characteristics. Summary of the Invention

[0007] The purpose of this invention is to overcome the problem that single-source data in the prior art is difficult to overcome the special terrain effects of high mountain and canyon areas, resulting in weak generalization ability and inability to accurately identify the characteristics of landslide debris flow. This invention provides a method, device, equipment and medium for identifying landslide debris flow disasters on highway slopes.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] A method for identifying debris flow hazards on highway slopes, characterized by the following steps:

[0010] S1: Input the remote sensing data of the landslide debris flow to be identified in the high mountain canyon area into the pre-trained landslide debris flow disaster area identification network; wherein, the remote sensing data includes high spatial resolution optical images and DEM data of the same period; the high spatial resolution optical images are optical remote sensing images with a spatial resolution of ≤5 meters and including four bands: blue, green, red and near-red.

[0011] S2: The landslide debris flow hazard identification network outputs the debris flow identification result of the landslide to be identified; the landslide debris flow hazard identification result includes the boundary and area of ​​the hazard area corresponding to the landslide debris flow;

[0012] The landslide debris flow hazard identification network includes a dual-branch encoder, a cross-modal cyclic fusion module, and a residual attention decoder. The dual-branch encoder is used to extract features from high spatial resolution optical images and DEM data respectively. The cross-modal cyclic fusion module is used to fuse the features from high spatial resolution optical images and DEM data to generate dual-modal features. The residual attention decoder is used to identify the landslide debris flow to be identified based on the dual-modal features.

[0013] As a preferred embodiment of the present invention, the pre-training of the landslide debris flow hazard identification network includes the following steps:

[0014] Remote sensing data of highway slopes in high mountain and canyon areas were acquired, georeferenced, and manually annotated before being output as labeled sample data.

[0015] The labeled sample data is subjected to data augmentation processing to generate a landslide debris flow labeled sample dataset; wherein, the data augmentation processing includes any one or more of random flipping, random rotation, random cropping, random brightness adjustment, and random color enhancement;

[0016] The landslide debris flow hazard area identification network is trained using the labeled sample dataset of the landslide debris flow. After the model training is completed, the output is the pre-trained landslide debris flow hazard area identification network.

[0017] As a preferred embodiment of the present invention, the method further includes S3:

[0018] Verify the identification results of the landslide debris flow to be identified, annotate the remote sensing data of the landslide debris flow to be identified based on the verification results, and store it in the landslide debris flow annotation sample dataset;

[0019] After the landslide debris flow labeled sample dataset is increased by a set number of labeled sample data, the landslide debris flow disaster area identification network is trained using the current landslide debris flow labeled sample dataset.

[0020] As a preferred embodiment of the present invention, the dual-branch encoder includes two parallel gated cyclic context modules; the gated cyclic context module includes a gated unit and a context-aware unit connected in sequence, which are used to adaptively fuse multi-scale context information of remote sensing data through a gated mechanism.

[0021] As a preferred embodiment of the present invention, the cross-modal cyclic fusion module includes a comparison-sharing fusion module and a consistent feature representation module; and includes the following processing steps:

[0022] The comparison, sharing, and fusion module:

[0023]

[0024] in, For the fused multimodal weight information, C() is the shared convolution function. To train complementary parameters; W i and W j represents the i-th and j-th modal parameters of the input; E() is the fusion metric function.

[0025] The consistent feature representation module:

[0026]

[0027]

[0028] Where W represents the extracted contrast features. For the output consistent feature map, σ and It is an activation function. as well as Here, F() is the training consistency parameter, F() is the feature extraction function, and R() is the information reconstruction function.

[0029] As a preferred embodiment of the present invention, the comparison sharing fusion module further includes a bidirectional cross-attention mechanism;

[0030] The bidirectional cross-attention mechanism comprises two attention mechanism branches; weight extraction is performed based on the high spatial resolution optical image and the DEM data, respectively; and the extracted weight maps are then fused; the expression is as follows:

[0031] Branch 1:

[0032]

[0033]

[0034]

[0035]

[0036] in, and These are the feature maps of the i-th and j-th modes after being processed by the shared convolutional module, respectively. It is the linear projection matrix of mode i onto the query matrix. , These are the linear projection matrices of mode j onto the bond matrix and value matrix, respectively; It is the query matrix generated by modality i. It is the key matrix generated by mode j. It is the value matrix generated by mode j; Representative to Perform matrix transpose; It is the scaling factor; It is a normalization function; It is the attention weight matrix for branch one; It is the attention output feature sequence of branch one. It is a layer normalization function. This represents the modal characteristics of the enhanced i-th mode;

[0037] Branch Two:

[0038]

[0039]

[0040]

[0041]

[0042] in, It is the linear projection matrix of mode j onto the query matrix. , These are the linear projection matrices of mode i onto the bond matrix and value matrix, respectively; It is the query matrix generated by modality j. It is the bond matrix generated by mode i. It is the value matrix generated by mode i; Representative to Perform matrix transpose; It is the attention weight matrix for branch two; It is the attention output feature sequence of branch two. This represents the characteristics of the enhanced j-th mode;

[0043] Feature fusion:

[0044]

[0045] in, This indicates feature concatenation, and E() is the fusion metric function.

[0046] As a preferred embodiment of the present invention, the landslide debris flow hazard area identification network includes the following steps:

[0047] The dual-branch encoder extracts features from the remote sensing data respectively;

[0048] The cross-modal cyclic fusion module fuses the extracted features to generate an adaptive attention weight map;

[0049] The residual attention decoder decodes based on the adaptive attention weight map and outputs the landslide debris flow disaster identification result.

[0050] A device for identifying debris flow hazards on highway slopes includes:

[0051] The data input module is used to input remote sensing data of the landslide debris flow to be identified; the remote sensing data includes high spatial resolution optical images and DEM data from the same period.

[0052] The landslide debris flow hazard identification module is used to execute any of the above-mentioned methods for identifying landslide debris flow hazards on highway slopes based on the remote sensing data, and output the identification result of the landslide debris flow hazard area corresponding to the landslide debris flow to be identified.

[0053] A highway slope landslide debris flow hazard identification device includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform any of the above-described highway slope landslide debris flow hazard identification methods.

[0054] A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for identifying debris flow hazards on highway slopes as described above.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] This invention utilizes high spatial resolution optical remote sensing imagery of highway slopes in high mountain and canyon areas, along with DEM data from the same region. Through a cross-modal fusion module, it achieves deep fusion of the spectral features of the optical imagery and the elevation features of the DEM. This allows for the simultaneous extraction of texture variations and surface deformation information of landslide debris flows, thereby enhancing the identification capability of debris flows from landslides with incomplete morphology and weak spectral response (such as shallow landslides and reactivated old landslides), thus improving identification accuracy. Furthermore, when the optical imagery is heavily obscured by clouds or snow, the DEM data branch can serve as an independent identification source, ensuring the continuity and stability of the identification process. This invention organically combines two different modalities of complementary remote sensing data, improving the accuracy of landslide debris flow geological hazard identification. It can meet the needs of rapid surveys of disaster sites along highways and in major engineering areas in high mountain and canyon regions, reducing manual workload and improving operational efficiency. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a method for identifying debris flow hazards on highway slopes as described in Embodiment 1 of the present invention.

[0058] Figure 2 This is a schematic diagram of the landslide debris flow hazard identification network in the method for identifying landslide debris flow hazards on highway slopes as described in Embodiment 2 of the present invention;

[0059] Figure 3 This is a schematic diagram of the remote sensing data input in the method for identifying debris flow hazards on highway slopes as described in Embodiment 4 of the present invention.

[0060] Figure 4 This is a schematic diagram of the labeled sample data in the method for identifying debris flow hazards on highway slopes as described in Embodiment 4 of the present invention;

[0061] Figure 5 This is a schematic diagram of the landslide debris flow hazard identification results of a section of highway located in a high mountain canyon area in the highway slope landslide debris flow hazard identification method described in Embodiment 4 of the present invention;

[0062] Figure 6 This is a schematic diagram of the structure of a highway slope landslide debris flow disaster identification device as described in Embodiment 6 of the present invention. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0064] Example 1

[0065] like Figure 1 As shown, a method for identifying debris flow hazards on highway slopes includes the following steps:

[0066] S1: Input remote sensing data of landslide debris flows to be identified in high mountain and canyon areas into a pre-trained landslide hazard area identification network; wherein, the remote sensing data includes high spatial resolution optical imagery and DEM data from the same period; the high spatial resolution optical imagery has a spatial resolution ≤ 5 meters. The DEM data is a ground digital elevation model with a spatial resolution not greater than 12.5 meters.

[0067] Furthermore, the high spatial resolution optical imagery includes multiple bands; for example, when using Google imagery or UAV visible light imagery, it is a three-band imagery of blue, green, and red; when using Gaofen satellite imagery, such as Planet, Gaofen-2, or Gaofen-7, it is an optical remote sensing imagery of four bands: blue, green, red, and near-red.

[0068] Furthermore, the higher the spatial resolution accuracy of the high spatial resolution optical image and the DEM data, the higher the accuracy of landslide debris flow identification.

[0069] S2: The landslide debris flow hazard identification network outputs the landslide debris flow hazard identification result to be identified; the landslide debris flow hazard identification result includes the landslide debris flow boundary and the landslide debris flow area.

[0070] The landslide debris flow hazard identification network includes a dual-branch encoder, a cross-modal cyclic fusion module, and a residual attention decoder. The dual-branch encoder is used to extract features from high spatial resolution optical images and DEM data respectively. The cross-modal cyclic fusion module is used to fuse the features from high spatial resolution optical images and DEM data to generate dual-modal features. The residual attention decoder is used to identify the landslide debris flow hazard to be identified based on the dual-modal features.

[0071] Example 2

[0072] This embodiment is a specific implementation of the landslide debris flow hazard identification network in the method for identifying highway slope landslide debris flow hazards described in Embodiment 1. Its pre-training includes the following steps:

[0073] a: Obtain remote sensing data of highway slopes in high mountain and canyon areas, perform georegistration, and output labeled sample data after manual annotation.

[0074] In this embodiment, the manual labeling is as follows: through field investigation and manual interpretation of high-resolution images, the confirmed landslide debris flow locations and boundary polygon information are collected, the categories are labeled, and landslide debris flow label samples are produced.

[0075] Furthermore, the georegistration involves uniformly registering all high spatial resolution optical images and DEM data to the same spatial coordinate system.

[0076] b: Perform data augmentation processing on the labeled sample data to generate a landslide debris flow labeled sample dataset.

[0077] The data augmentation process includes any one or more of random flipping, random rotation, random cropping, random brightness adjustment, and random color enhancement to improve model robustness. After data augmentation, a historical landslide debris flow sample database containing the location and boundary attributes of landslide debris flows is constructed and divided into training, testing, and validation sets. The output is a labeled landslide debris flow sample dataset.

[0078] c: The landslide debris flow hazard area identification network is trained using the landslide debris flow labeled sample dataset. After the model training is completed, the output is the pre-trained landslide debris flow hazard area identification network.

[0079] Furthermore, the landslide debris flow disaster area identification network is a dual-modal deep learning network GR-ACCR-Net (Gate Recurrent-Attention Consistent Contrastive Representation Network) driven by high spatial resolution optical images and DEM data, including a dual-branch encoder, a cross-modal recurrent fusion module, and a residual attention decoder.

[0080] Dual-branch encoder:

[0081] This study includes two parallel gated recurrent context modules (GRCMs). These modules extract multi-level features from optical feature cubes and DEM feature cubes, respectively. Each branch replaces the standard convolutional encoder backbone with two GRCMs. By adaptively fusing multi-scale context through a gating mechanism, the long-range spatial dependencies between landslide debris flows and their surrounding environment are effectively modeled. In traditional convolutional neural networks (CNNs), the receptive field is local. This means that a neuron can only process information from a small area around it. For complex scenarios like landslide debris flow disasters on steep slopes, this presents two main problems: First, insufficient contextual information. A patch that appears to be a landslide debris flow disaster may require seeing the steep backwall (crown) above it and the accumulation area (tongue) below it for accurate identification. These crucial clues may be far away, exceeding the receptive field of ordinary convolutional layers. Second, difficulty in modeling spatial dependencies. Landslide debris flow disaster areas have strong spatial and semantic connections with their surrounding environment. Traditional CNNs struggle to explicitly model these long-range dependencies. Therefore, the purpose of introducing this module is to overcome the limitations of local receptive fields, enabling deep networks to actively and adaptively capture and fuse long-range, multi-scale contextual information in images. The core principle of GRCM lies in using a "memory unit" to iteratively integrate information at different scales, and using multiple dilated convolutional layers with different dilation rates or pooling layers with different kernel sizes in parallel to process a feature map of a certain layer, resulting in a set of feature map combinations {C1, C2, C3, ...} with the same spatial size (C1 has a small dilation rate, C2 has a medium dilation rate, and C3 has a large dilation rate). At each time step t, for the feature C at the current scale... t Based on the memory state of the previous time step, GRCM uses "gates" to generate a new memory state. This new memory state is no longer information at a single scale, but rather a dynamic weighted sum and condensation of contextual information from all processed scales. The final memory state is then transformed through convolutional layers to match the input channels.

[0082] Cross-modal cyclic fusion module:

[0083] This includes a comparison and sharing fusion module and a consistent feature representation module;

[0084] The goal of the comparative fusion module is to share complementary weight matrices among multimodal information. On one hand, a built-in module captures the non-linear structure of the current view, and other views train these internal weights to adjust the shared information. On the other hand, a fusion metric function is used to obtain complete information and form a shared multi-view fusion space.

[0085]

[0086] in, For the fused multimodal information, C() is a shared convolution function module used to adjust for the influence of complete or incomplete graphs. It trains complementary parameters; W i and W j It is the information of the i-th and j-th modes of the input ( E() is the fusion metric function used to integrate information across all modules. The modules described above capture complete contrast information by training complementary weights and build reliable view completion even in incomplete cases.

[0087] The Consistent Feature Representation module ensures that the contrast images retain the learned original structural information. This module consists of two parts: a feature extraction layer F and a reconstruction layer R. The feature extraction layer F is mainly responsible for anomaly handling and extracting contrastive features W from the contrast images, while the reconstruction layer R is used to reconstruct the infographic.

[0088]

[0089] in as well as These are the training consistency parameters, σ and It is the activation function. After training, the output of the consistent feature representation module yields the consistent feature map. .

[0090] These two modules are used for deep fusion of features from optical imagery and DEM data branches. The feature maps output from the optical imagery branch and the DEM data branch at the same level are aligned, and after feature concatenation, a pair of adaptive attention weight maps are generated through a small convolutional network. Using the generated attention weight maps, the optical and DEM features are weighted and summed separately to generate preliminary fused features. However, this mode is dominated by high-resolution imagery, with DEM depth data as a supplement, and cannot fully utilize the active information that may exist in the DEM that complements the optical imagery. Therefore, this invention embeds a bidirectional cross attention mechanism (BCAM) into the existing fusion module of the model, which establishes an "equal dialogue" mechanism.

[0091] Unlike standard self-attention, the bidirectional cross-attention mechanism is designed for multimodal tasks to achieve bidirectional information flow between two sequences. It includes two attention mechanism branches: one based on the high spatial resolution optical image and the other on the DEM data, to extract weights; and the extracted weight maps are then fused.

[0092] Specifically, in this mechanism, optical imagery and DEM data act as the queryer and reference recipient, respectively, achieving deep bidirectional information fusion through two rounds of cross-attention. The execution is divided into two symmetrical, opposite-direction cross-attention flows. In the first flow, the high-resolution imagery takes the lead, queries terrain information, calculates the similarity matrix between the two, and obtains the cross-attention weight matrix through Softmax. This weighting represents using optical features to formulate a question / requirement (Query), finding matching content (Key) from DEM features, and then aggregating the DEM information (Value) onto the optical features according to the matching weights, thereby enhancing the feature representation of the optical image. Similarly, if topographic information is used as the primary factor in the flow, and the optical image is queried, the weight matrix is ​​obtained in the same way. This can effectively support the attention weighted summation process of CCR mentioned above. The specific mathematical expression is as follows:

[0093] To facilitate attention calculations, the space is flattened here:

[0094]

[0095] ,

[0096] ,

[0097] Where reshape is the spatial flattening function. and For the information of the i-th and j-th modes of input, and It is the feature map after the two modalities have been processed by a shared convolutional module.

[0098] 1. Branch 1: Define from (Using optical image data for queries and DEM data for key / value pairs):

[0099]

[0100]

[0101]

[0102] in, , It is the linear projection matrix of mode i; It is the query matrix generated by modality i. It is the key matrix generated by mode j. It is the value matrix generated by mode j; Representative to Perform matrix transpose; It is a scaling factor, representing the dimensions of the Key and Query; It is a normalization function; It is the attention weight matrix for branch one; It is the attention output feature sequence of branch one.

[0103] 2. Branch Two: Definition from (Using DEM data for queries and optical image information for key / value pairs):

[0104]

[0105]

[0106]

[0107] in, It is the linear projection matrix of mode j onto the query matrix. , These are the linear projection matrices of mode i onto the bond matrix and value matrix, respectively; It is the query matrix generated by modality j. It is the bond matrix generated by mode i. It is the value matrix generated by mode i; Representative to Perform matrix transpose; It is the attention weight matrix for branch two; It is the attention output feature sequence of branch two. This represents the feature of the enhanced j-th mode.

[0108] 3. Add residuals and normalization:

[0109] ,

[0110] in, It is a layer normalization function. and This represents the features of the enhanced modes i and j.

[0111] This invention, based on the original comparison-shared fusion module, first uses function C() for correction, then uses BCAM to obtain the complementary enhanced bimodal representation, and finally inputs the fusion metric function E(). The mathematical expression is:

[0112]

[0113] in, This indicates feature splicing.

[0114] In the consistent feature representation module, no changes are needed; you can... This indicates a shared, integrated space enhanced by BCAM.

[0115] Residual attention decoder:

[0116] This module is used to upsample and restore the deep, abstract features extracted by the encoder and cross-modal fusion features to the original input size, achieving semantic segmentation of landslide debris flow disaster areas. Each upsampling module contains residual connections to ensure the stability of the decoding process and the effective reconstruction of boundary information.

[0117] Furthermore, the landslide debris flow hazard area identification network includes the following steps:

[0118] (1) Low-level feature extraction: A dual-branch encoder is used to extract features from the high-resolution optical image and the DEM respectively;

[0119] (2) Semantic feature extraction: Each branch introduces a gated recurrent context module, which adaptively fuses multi-scale context information through a gating mechanism to effectively model the long-distance spatial dependence between landslide debris flow disaster and the surrounding environment;

[0120] (3) Feature fusion: The cross-modal cyclic fusion module deeply fuses features from optical image and DEM data branches, aligns the feature maps output by the optical image branch and DEM data branch at the same level, and performs feature fusion and refinement.

[0121] (4) Landslide debris flow identification: The residual attention decoder upsamples the deep and abstract features extracted by the encoder and the cross-modal fusion features to restore them to the original input size, so as to realize the semantic recognition of landslide debris flow disaster areas.

[0122] Example 3

[0123] The difference between this embodiment and the previous embodiment is that the method for identifying debris flow hazards on highway slopes further includes S3:

[0124] Verify the identification results of the landslide debris flow to be identified, annotate the remote sensing data of the landslide debris flow to be identified based on the verification results, and store it in the landslide debris flow annotation sample dataset;

[0125] After the landslide debris flow labeled sample dataset is increased by a set number of labeled sample data, the landslide debris flow disaster area identification network is trained using the current landslide debris flow labeled sample dataset.

[0126] Example 4

[0127] This embodiment takes a section of highway located in a high mountain canyon area as an example to verify the method for identifying debris flow from a highway slope in a high mountain canyon area as described in the above embodiment of the present invention. Specific examples are as follows:

[0128] Equipment preparation: Prepare a Windows computer with a TensorFlow virtual environment and Python version 3.8 or higher installed. The implementation steps are as follows:

[0129] Data collection: such as Figure 3 As shown, a 3-meter resolution Planet multispectral satellite image of a section of highway slope located in a high mountain canyon area was acquired (e.g., Figure 3 (as shown in a) and the DEM data of 12.5 meters from the same period (as shown in a) Figure 3 (As shown in b).

[0130] Data preparation: such as Figure 4 As shown, Planet multispectral satellite imagery was georeferenced with 12.5-meter DEM data. Annotated sample data was generated using a combination of expert visual interpretation and field verification to construct a slope landslide debris flow sample library, which was divided into training, testing, and validation sets. The annotated Planet multispectral satellite imagery is shown below. Figure 4 As shown in Figure a, the labeled DEM data is as follows: Figure 4 As shown in b.

[0131] Identification model design: Construct the landslide debris flow hazard identification network described in the above embodiments.

[0132] Recognition model training: Using high spatial resolution optical images and contemporaneous DEM data as inputs and corresponding landslide debris flow sample labels as outputs, the GR-ACCR-Net deep network is trained to obtain a landslide debris flow disaster area recognition network.

[0133] Landslide debris flow identification: 3-meter resolution Planet multispectral satellite imagery of the area to be identified, along with concurrent 12.5-meter DEM data, is input into a trained landslide debris flow hazard area identification network. Figure 5As shown, the geographical location and regional boundaries of landslide debris flow hazards were identified, and the area of ​​the landslide debris flow was estimated. Results Validation: Precision, recall, and F1 score were used to validate and evaluate the accuracy of the identified landslide debris flow boundary data using validation set landslide debris flow sample data or UAV field survey data. Precision refers to the proportion of samples predicted as positive by the model that are actually positive, measuring the accuracy of the model when predicting positive cases. Recall refers to the proportion of samples that are actually positive that are correctly predicted as positive by the model. The F1 score is a comprehensive evaluation metric in classification problems; it is a weighted average of precision and recall, considering both predicted and actual positive examples. Its value ranges from 0 to 1; a value closer to 1 indicates a better balance between positive and negative examples, resulting in better prediction accuracy. Using the aforementioned indicators, along with high-resolution optical remote sensing imagery and DEM data of a section of highway slope located in a high mountain canyon area, as well as landslide debris flow training sample data, ablation comparison experiments were conducted on each module. The impact of different modules on the network model performance is shown in Table 1.

[0134] Table 1. Model ablation experiments and performance evaluation before and after improving the CCR-Net network.

[0135]

[0136] As shown in Table 1, the GR-ACCR-Net model, which integrates the GRCM and BCAM modules, achieved an accuracy of 81.47%, a recall of 79.44%, and an F1 score of 80.44% for identifying debris flows from the landslide on this section of highway. This indicates that the network of the present invention improves the accuracy of debris flow identification from the landslide on the highway slope.

[0137] Results compilation: The location, area and other attribute information of the correctly identified landslide debris flow hazard areas were entered, and the landslide debris flow dataset of the highway slope in the area was updated.

[0138] Example 5

[0139] A device for identifying debris flow hazards on highway slopes includes:

[0140] The data input module is used to input remote sensing data of the landslide debris flow to be identified; the remote sensing data includes high spatial resolution optical images and DEM data from the same period.

[0141] The landslide debris flow hazard identification module is used to execute a method for identifying landslide debris flow hazards on highway slopes according to any one of the above embodiments based on the remote sensing data, and output the landslide debris flow hazard identification result corresponding to the landslide debris flow to be identified.

[0142] Example 6

[0143] like Figure 6 As shown, a highway slope landslide debris flow hazard identification device includes at least one processor, a memory communicatively connected to the at least one processor, and at least one input / output interface communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enables the at least one processor to perform a highway slope landslide debris flow hazard identification method as described in the foregoing embodiments. The input / output interface may include a display, keyboard, mouse, and USB interface for inputting and outputting data.

[0144] Furthermore, the highway slope landslide debris flow hazard identification device can be a desktop computer, mobile phone, tablet computer, wearable highway slope landslide debris flow hazard identification device, or any other highway slope landslide debris flow hazard identification device capable of performing in-depth information identification.

[0145] Furthermore, the processor may include one or more processing cores. The processor connects various parts within the highway slope landslide debris flow hazard identification device using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory, and by calling data stored in memory. Optionally, the processor may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also be implemented separately as a communication chip, without being integrated into the processor.

[0146] The memory may include random access memory (RAM) or read-only memory (ROM). The memory can be used to store instructions, programs, code, code sets, or instruction sets, such as instructions or code sets used to implement a method for identifying debris flow hazards on highway slopes provided in this application. The memory may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function, instructions for implementing the various method embodiments described above, etc. The data storage area may also store data created during the use of the highway slope debris flow hazard identification device (such as a mapping table of modulation sequences and depths, image data, spectrogram data, etc.).

[0147] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0148] When the integrated units of the present invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. The computer-readable storage medium stores program code, which can be called by a processor to execute the methods described in the above method embodiments. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes electronic memories such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, the computer-readable storage medium includes a non-transitory computer-readable storage medium. The computer-readable storage medium has storage space for program code that executes any of the method steps described above. This program code can be read from or written to one or more computer program products. The program code can be compressed, for example, in an appropriate form.

[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying debris flow hazards on highway slopes, characterized in that, Includes the following steps: S1: Input the remote sensing data of the landslide debris flow to be identified in the high mountain and canyon area into the pre-trained landslide debris flow disaster area identification network; wherein, the remote sensing data includes high spatial resolution optical imagery and DEM data of the same period; the high spatial resolution optical imagery has a spatial resolution ≤ 5 meters; S2: The landslide debris flow hazard area identification network outputs the identification result of the landslide debris flow to be identified; the landslide debris flow identification result includes the boundary and area of ​​the hazard area corresponding to the landslide debris flow; The landslide debris flow hazard area identification network includes a dual-branch encoder, a cross-modal cyclic fusion module, and a residual attention decoder. The dual-branch encoder is used to extract features from high spatial resolution optical images and DEM data respectively; the dual-branch encoder includes two parallel gated cyclic context modules; the gated cyclic context module includes a gated unit and a context-aware unit connected in sequence, which are used to adaptively fuse multi-scale context information of remote sensing data through a gated mechanism; The cross-modal cyclic fusion module is used to fuse features from high spatial resolution optical images and DEM data to generate dual-modal features; The residual attention decoder is used to identify the landslide debris flow to be identified based on bimodal features; The cross-modal cyclic fusion module includes a comparison-shared fusion module and a consistent feature representation module; it includes the following processing steps: The comparison, sharing, and fusion module: in, For the fused multimodal weight information, C() is the shared convolution function. To train complementary parameters; W i and W j The input parameters are the i-th and j-th modal parameters; E() is the fusion metric function; The consistent feature representation module: Where W represents the extracted contrast features. For the output consistent feature map, σ and It is an activation function. as well as is the training consistency parameter, F() is the feature extraction processing function, and R() is the information reconstruction function; The comparison-sharing fusion module also includes a bidirectional cross-attention mechanism; The bidirectional cross-attention mechanism comprises two attention mechanism branches; weight extraction is performed based on the high spatial resolution optical image and the DEM data, respectively; and the extracted weight maps are then fused; the expression is as follows: Branch 1: in, and These are the feature maps of the i-th and j-th modes after being processed by the shared convolutional module, respectively. It is the linear projection matrix of mode i onto the query matrix. , These are the linear projection matrices of mode j onto the bond matrix and value matrix, respectively; It is the query matrix generated by modality i. It is the key matrix generated by mode j. It is the value matrix generated by mode j; Representative to Perform matrix transpose; It is the scaling factor; It is a normalization function; It is the attention weight matrix for branch one; It is the attention output feature sequence of branch one. It is a layer normalization function. This represents the feature of the enhanced i-th mode; Branch Two: in, It is the linear projection matrix of mode j onto the query matrix. , These are the linear projection matrices of mode i onto the bond matrix and value matrix, respectively; It is the query matrix generated by modality j. It is the bond matrix generated by mode i. It is the value matrix generated by mode i; Representative to Perform matrix transpose; It is the attention weight matrix for branch two; It is the attention output feature sequence of branch two. This represents the characteristics of the enhanced j-th mode; Feature fusion: in, This indicates feature concatenation, and E() is the fusion metric function.

2. The method for identifying debris flow hazards on highway slopes according to claim 1, characterized in that, The pre-training of the landslide debris flow hazard identification network includes the following steps: Remote sensing data of highway slopes in high mountain and canyon areas were acquired, georeferenced, and manually annotated before being output as labeled sample data. The labeled sample data is subjected to data augmentation processing to generate a landslide debris flow labeled sample dataset; wherein, the data augmentation processing includes any one or more of random flipping, random rotation, random cropping, random brightness adjustment, and random color enhancement; The landslide debris flow hazard area identification network is trained using the labeled sample dataset of the landslide debris flow. After the model training is completed, the output is the pre-trained landslide debris flow hazard area identification network.

3. The method for identifying debris flow hazards on highway slopes according to claim 2, characterized in that, The method further includes S3: Verify the identification results of the landslide debris flow to be identified, annotate the remote sensing data of the landslide debris flow to be identified based on the verification results, and store it in the landslide debris flow annotation sample dataset; After the landslide debris flow labeled sample dataset is increased by a set number of labeled sample data, the landslide debris flow disaster area identification network is trained using the current landslide debris flow labeled sample dataset.

4. The method for identifying debris flow hazards on highway slopes according to claim 1, characterized in that, The landslide debris flow hazard area identification network includes the following steps: The dual-branch encoder extracts features from the remote sensing data respectively; The cross-modal cyclic fusion module fuses the extracted features to generate an adaptive attention weight map of the landslide debris flow disaster area; The residual attention decoder decodes based on the adaptive attention weight map and outputs the landslide debris flow identification result.

5. A device for identifying debris flow hazards on highway slopes, characterized in that, include: The data input module is used to input remote sensing data of the landslide debris flow to be identified; the remote sensing data includes high spatial resolution optical images and DEM data from the same period. The landslide debris flow hazard identification module is used to execute a method for identifying landslide debris flow hazards on highway slopes according to any one of claims 1 to 4 based on the remote sensing data, and output the landslide debris flow identification result corresponding to the landslide debris flow to be identified.

6. A device for identifying debris flow hazards on highway slopes, characterized in that, The method includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform a method for identifying debris flow hazards on highway slopes according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a method for identifying debris flow hazards on highway slopes as described in any one of claims 1 to 4.

Images