Hidden danger identification method, device and equipment for slope type geological disasters and medium

Abstract

The invention discloses a hidden danger identification method, device and equipment for slope type geological disasters and a medium. The method comprises the following steps: acquiring multi-source remote sensing data of a target slope area, and performing layered preprocessing on the multi-source remote sensing data; carrying out adaptive screening of fusion wavebands on the multi-source remote sensing data after space-time registration processing, and carrying out feature enhancement processing on the fusion wavebands after multi-scale decomposition; obtaining terrain perception information of the target slope area, constructing a heterogeneous feature matrix based on the multi-source remote sensing data after feature enhancement processing and the terrain perception information, inputting the heterogeneous feature matrix into a deep learning model, and extracting a slope hidden danger feature vector; and inputting the slope hidden danger feature vector into a pre-trained hidden danger detection model, positioning a hidden danger area of the geological disaster, and identifying the geological disaster type of the positioned hidden danger area to obtain a hidden danger identification result of the slope type geological disaster. The method improves the recognition precision of the hidden danger of the slope under the complex terrain.

Description

Technical Field

[0001] This invention relates to the field of geological hazard identification technology, specifically to a method, device, equipment, and medium for identifying potential hazards of slope-type geological hazards. Background Technology

[0002] Slope-type geological hazard sites are widely distributed, highly concealed, and pose a high risk of disaster. Accurate and efficient hazard identification is fundamental to geological disaster prevention and control. Traditional slope-type geological hazard identification relies heavily on manual on-site surveys, which are not only inefficient and costly but also limited by factors such as terrain complexity and climatic conditions, making it difficult to achieve large-scale, detailed hazard investigation. Furthermore, surveys of dangerous areas pose safety risks to personnel. With the development of remote sensing technology, single remote sensing data sources such as optical remote sensing data, synthetic aperture radar interferometry data, and lidar data have been gradually applied to geological hazard identification. However, single data sources have inherent limitations: optical remote sensing is easily affected by weather and vegetation obstruction, making it difficult to capture terrain deformation information; synthetic aperture radar interferometry data suffers from speckle noise, resulting in insufficient accuracy in identifying small-scale hazards; while lidar data can acquire high-precision elevation data, it lacks spectral information and cannot comprehensively characterize the environmental features of hazards. Simultaneously, existing multi-source remote sensing data fusion methods are mostly simple pixel-level methods, failing to fully consider the spatiotemporal heterogeneity of different data sources. The fused images suffer from feature blurring and spectral distortion, making it difficult to meet the needs of refined identification. In addition, traditional machine learning algorithms are weak in capturing heterogeneous features in complex terrain when extracting features of slope hazards. They have problems such as missing small-scale hazards, inaccurate boundary positioning, and large type discrimination errors. Furthermore, they lack a complete accuracy verification and model iteration optimization system, resulting in insufficient reliability of the identification results.

[0003] Therefore, improving the accuracy of identifying potential slope hazards in complex terrain is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] This invention provides a method, device, equipment, and medium for identifying potential hazards in slope-type geological disasters, in order to solve the technical problem of how to improve the accuracy of identifying potential hazards in complex terrain.

[0005] In a first aspect, the present invention provides a method for identifying potential hazards of slope-type geological disasters, including: Acquire multi-source remote sensing data of the target slope area, perform layered preprocessing on the multi-source remote sensing data, and perform spatiotemporal registration processing on the multi-source remote sensing data after layered preprocessing. Adaptive selection of fusion bands is performed on the spatiotemporal registration-processed multi-source remote sensing data. The selected fusion bands are then decomposed into multiple scales using an image decomposition algorithm, and feature enhancement processing is applied to the multi-scale decomposed fusion bands. The terrain perception information of the target slope area is obtained, and a heterogeneous feature matrix is ​​constructed based on the multi-source remote sensing data and terrain perception information after feature enhancement processing. The heterogeneous feature matrix is ​​then input into a deep learning model to extract slope hazard feature vectors. The slope hazard feature vector is input into a pre-trained hazard detection model to locate the hazard area of ​​the geological disaster and identify the geological disaster type of the located hazard area, thus obtaining the hazard identification result of slope-type geological disasters.

[0006] Optionally, the step of acquiring multi-source remote sensing data of the target slope area, performing layered preprocessing on the multi-source remote sensing data, and performing spatiotemporal registration processing on the layered preprocessed multi-source remote sensing data includes: Acquire optical remote sensing data, synthetic aperture radar interferometry data, and lidar data for the target slope area; An enhanced directional smoothing filter algorithm is used to reduce the speckle of the synthetic aperture radar interferometry data to obtain speckle-reduced synthetic aperture radar interferometry data. The optical remote sensing data is radiometrically calibrated, and atmospheric correction is performed on the radiometrically calibrated optical remote sensing data based on the radiative transfer model to obtain atmospherically corrected optical remote sensing data. The lidar data is denoised based on the elevation anomaly detection algorithm to obtain denoised lidar data. Based on the digital elevation model generated from the denoised lidar data, a second polynomial transformation is performed on the atmospheric-corrected optical remote sensing data and the speckle-reduced synthetic aperture radar interferometry data to obtain multi-source remote sensing data with preliminary geometric correction. A feature matching model was used to register the multi-source remote sensing data after preliminary geometric correction, resulting in registered multi-source remote sensing data.

[0007] Optionally, the adaptive selection of fusion bands for the spatiotemporally registered multi-source remote sensing data, the multi-scale decomposition of the selected fusion bands using an image decomposition algorithm, and the feature enhancement processing of the multi-scale decomposed fusion bands include: Adaptive selection of fusion bands is performed on the spatiotemporal registration-processed multi-source remote sensing data, and multi-scale joint decomposition is performed on the selected fusion bands using non-subsampling contour wave transform to obtain the low-frequency subband coefficients and high-frequency subband coefficients of each band. The low-frequency subband is weighted and fused based on the regional correlation coefficient, and the high-frequency subband is adaptively fused based on the local directional energy. The fused low-frequency subband and the fused high-frequency subband are subjected to inverse contour wave transformation without subsampling to reconstruct the preliminary fused image of the multi-source remote sensing data. Based on the slope information in the lidar data and the deformation rate information in the synthetic aperture radar interferometry data, a feature enhancement weight matrix is ​​constructed, and the preliminary fused image is adaptively enhanced based on the feature enhancement weight matrix to obtain feature-enhanced multi-source remote sensing data.

[0008] Optionally, the step of acquiring terrain sensing information of the target slope area, constructing a heterogeneous feature matrix based on the multi-source remote sensing data and terrain sensing information after feature enhancement processing, and inputting the heterogeneous feature matrix into a deep learning model to extract slope hazard feature vectors includes: Heterogeneous dimension mapping is performed on the feature-enhanced multi-source remote sensing data, and the mapped fused features are fused with the terrain perception information to construct a heterogeneous feature matrix; The heterogeneous feature matrix is ​​filtered by a terrain-adaptive attention gating mechanism to obtain the filtered feature matrix. Construct a terrain-aware location encoding vector, and fuse the terrain-aware location encoding vector with the filtered feature matrix to obtain a fused feature vector; A Transformer model containing a multi-scale window attention module and a terrain feature enhancement module is constructed, and the fused feature vector is input into the constructed Transformer model to extract multi-level encoded features; The multi-level encoded features are subjected to cross-layer weighted fusion processing, and the dimensionality of the multi-level encoded features after cross-layer weighted fusion processing is reduced by attention pooling to output the slope hazard feature vector.

[0009] Optionally, the step of inputting the slope hazard feature vector into a pre-trained hazard detection model to locate the hazard area of ​​the geological disaster and identify the geological disaster type of the located hazard area to obtain the hazard identification result of slope-type geological disasters includes: Based on the slope hazard feature vector, an adaptive anchor point generation strategy is used to generate candidate hazard areas; The feature vector of each candidate hazard area is extracted by the target detection algorithm, and the feature vector of each candidate hazard area is input into the pre-trained hazard detection model to calculate the category probability and boundary coordinate offset of each candidate hazard area. Based on the category probability and boundary coordinate offset of each candidate hazard area, the initial hazard area and its confidence level of the target slope area are determined; A flexible nonmaximum suppression algorithm is used to screen the initial hidden danger areas, and the confidence of the screened hidden danger areas is corrected to obtain the hidden danger areas and corresponding geological disaster types of the target slope area.

[0010] Optionally, after inputting the slope hazard feature vector into a pre-trained hazard detection model to locate the hazard area and identify the type of geological hazard in the located hazard area to obtain the hazard identification result for slope-type geological hazards, the method further includes: A multi-dimensional accuracy verification index is constructed, which includes the classification accuracy, spatial accuracy, and feature consistency of the hazard area. The hazard identification results of the slope-type geological disaster are quantitatively evaluated, and the hazard areas where the accuracy verification index of each dimension does not reach the preset accuracy threshold are identified. For potential areas where the accuracy verification indicators of each dimension do not reach the preset accuracy threshold, an uncertainty sampling strategy is used to screen samples that need to be supplemented with annotations. The samples to be annotated are added to the training set of the hazard detection model, and the class weights of the loss function are adjusted. The hazard detection model after adjusting the category weights of the loss function is iteratively optimized until the accuracy indicators of each dimension meet the preset accuracy threshold, and the optimized slope geological hazard identification results are obtained.

[0011] Optionally, after iteratively optimizing the hazard detection model after adjusting the category weights of the loss function until the accuracy indicators of each dimension meet the preset accuracy threshold, and obtaining the optimized slope-type geological hazard identification result, the method further includes: Based on the optimized hazard identification results, multi-dimensional core parameters related to the terrain, deformation, and environment of the hazard area are extracted. Based on the aforementioned multi-dimensional core parameters, a hazard susceptibility evaluation index is constructed, and a hazard distribution map and risk level zoning map of the target slope area are generated. The digital elevation model generated from the denoised lidar data is fused with the risk level zoning map to construct a three-dimensional visualization model of the target slope area. Based on the aforementioned hazard distribution map, risk level zoning map, and 3D visualization model, standard visualization results for hazard identification in the target slope area are generated.

[0012] Secondly, the present invention provides a hazard identification device for slope-type geological hazards, comprising: The acquisition module is used to acquire multi-source remote sensing data of the target slope area, perform layered preprocessing on the multi-source remote sensing data, and perform spatiotemporal registration processing on the multi-source remote sensing data after layered preprocessing. The decomposition module is used to adaptively select fusion bands for multi-source remote sensing data after spatiotemporal registration. It performs multi-scale decomposition on the selected fusion bands using an image decomposition algorithm and then performs feature enhancement processing on the multi-scale decomposed fusion bands. The construction module is used to acquire terrain perception information of the target slope area, construct a heterogeneous feature matrix based on the multi-source remote sensing data and terrain perception information after feature enhancement processing, and input the heterogeneous feature matrix into a deep learning model to extract slope hazard feature vectors. The localization module is used to input the slope hazard feature vector into the pre-trained hazard detection model to locate the hazard area of ​​the geological disaster and identify the geological disaster type of the located hazard area, so as to obtain the hazard identification result of slope geological disaster.

[0013] Thirdly, the present invention provides a hazard identification device for slope-type geological hazards, comprising a memory and a processor, wherein: The memory is used to store computer programs; The processor is used to read the computer program in the memory and execute the steps of the method for identifying potential hazards of slope-type geological disasters as described in any of the first aspects above.

[0014] Fourthly, a computer-readable storage medium having a readable computer program stored thereon, which, when executed by a processor, implements the steps of a method for identifying potential hazards of slope-type geological hazards as described in any of the first aspects.

[0015] Compared with existing technologies, the present invention provides a method, device, equipment, and medium for identifying potential hazards in slope-type geological disasters, which has the following beneficial effects: This invention adaptively selects fusion bands from multi-source remote sensing data after spatiotemporal registration, performs multi-scale decomposition of the selected fusion bands using an image decomposition algorithm, and then enhances the features of the decomposed fusion bands. This significantly improves the information richness, clarity, and spectral fidelity of the fused image, strengthens the saliency of features in slope hazard areas, and effectively solves the problems of feature blurring and detail loss in traditional fusion methods. Simultaneously, a heterogeneous feature matrix is ​​constructed based on the feature-enhanced multi-source remote sensing data and terrain perception information. This heterogeneous feature matrix is ​​then input into a deep learning model to extract slope hazard feature vectors, greatly improving the identification accuracy of slope hazards in complex terrain and effectively reducing the false negative rate and type discrimination error of small-scale hazards. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and not all embodiments. For those skilled in the art, other drawings obtained from these drawings without creative effort are all within the scope of protection of this application.

[0017] Figure 1 A flowchart illustrating a method for identifying potential hazards in slope-type geological disasters, as provided in an embodiment of the present invention.

[0018] Figure 2 This is a flowchart of a method for preprocessing multi-source remote sensing data provided in an embodiment of the present invention.

[0019] Figure 3 This is a flowchart of a method for feature enhancement of multi-source remote sensing data provided in an embodiment of the present invention.

[0020] Figure 4 This is a flowchart for extracting feature vectors of potential slope hazards, provided as an embodiment of the present invention.

[0021] Figure 5 This is a flowchart of a method for identifying the type of geological disaster in a potential hazard area, provided as an embodiment of the present invention.

[0022] Figure 6 This invention provides a device for identifying potential hazards of slope-type geological disasters.

[0023] Figure 7 This is a schematic diagram of the structure of a hazard identification device for slope-type geological disasters provided in an embodiment of the present invention.

[0024] Figure 8 This is a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] To make the description of this disclosure more detailed and complete, illustrative descriptions of embodiments and specific examples of the present invention are provided below; however, these are not the only forms of implementing or utilizing the specific examples of the present invention. The embodiments encompass features of multiple specific examples and the method steps and their order for constructing and operating these specific examples. However, other specific examples may also be used to achieve the same or equivalent functions and order of steps. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0028] In the description of the embodiments of the present invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The word "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more. Other quantifiers should be understood similarly. The preferred embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0029] Example 1 To address the technical challenge of improving the accuracy of slope hazard identification in complex terrain, this invention provides a method for identifying slope-related geological hazards, such as... Figure 1 , Figure 1 This is a flowchart illustrating a method for identifying potential hazards of slope-type geological disasters according to an embodiment of the present invention. The method specifically includes: S10: Acquire multi-source remote sensing data of the target slope area, perform hierarchical preprocessing on the multi-source remote sensing data, and perform spatiotemporal registration processing on the preprocessed multi-source remote sensing data. Specifically, in this embodiment of the invention, by acquiring multi-source remote sensing data of the target slope area and performing hierarchical preprocessing and spatiotemporal registration processing, the heterogeneity of different data sources in terms of imaging mechanisms, noise characteristics, and geometric deformation can be effectively eliminated, providing a data foundation with high consistency and high alignment accuracy for subsequent multi-source data fusion and feature extraction. Specifically, hierarchical preprocessing adopts appropriate noise reduction and correction algorithms for different types of data such as optical, radar, and lidar, preserving the effective information of each data source; while spatiotemporal registration ensures the accurate correspondence of multi-source data in spatial location, avoiding misidentification problems caused by geometric deviations, thereby improving the reliability and accuracy of the entire hazard identification process from the source.

[0030] S20: Adaptive selection of fusion bands is performed on the spatiotemporally registered multi-source remote sensing data. The selected fusion bands are then decomposed at multiple scales using an image decomposition algorithm, followed by feature enhancement processing. Specifically, in this embodiment of the invention, by adaptively selecting fusion bands, performing multi-scale decomposition, and enhancing features on the registered multi-source remote sensing data, efficient fusion of heterogeneous multi-source information and significant enhancement of hazard features are achieved. Adaptive selection of fusion bands automatically selects the optimal band combination based on data characteristics, avoiding information redundancy. The combination of multi-scale decomposition and image decomposition algorithms can separate and reconstruct the contour and detail information of images at different spatial scales, fully exploring the multi-scale geometric features of slope hazards. Feature enhancement further strengthens the subtle deformation or texture features related to geological disasters. This effectively solves the technical problems of low utilization rate of multi-source information and lack of prominent hazard features in traditional methods, significantly improving the sensitivity of subsequent hazard identification and location.

[0031] S30: Obtain terrain perception information of the target slope area, construct a heterogeneous feature matrix based on the multi-source remote sensing data after feature enhancement processing and the terrain perception information, and input the heterogeneous feature matrix into a deep learning model to extract slope hazard feature vectors. Specifically, in this embodiment of the invention, by introducing terrain perception information and constructing a heterogeneous feature matrix with the multi-source remote sensing data after feature enhancement, and then combining it with a deep learning model to extract slope hazard feature vectors, a deep and high-dimensional expression of slope hazard characteristics is achieved. Terrain perception information is an important controlling factor in the formation of slope geological disasters. By fusing it with remote sensing data to construct a heterogeneous feature matrix, the deep learning model can simultaneously learn the intrinsic relationship between the surface appearance of the slope and the terrain background. Compared with feature extraction methods that rely solely on remote sensing data, this invention significantly enhances the model's ability to distinguish slope hazards, and the extracted slope hazard feature vectors are more representative, robust, and have better generalization ability, providing high-quality feature input for subsequent accurate identification and positioning.

[0032] S40: The slope hazard feature vector is input into a pre-trained hazard detection model to locate the hazard area and identify the type of geological hazard in the located hazard area, thus obtaining the hazard identification result for slope-type geological hazards. Specifically, in this embodiment of the invention, by inputting the extracted slope hazard feature vector into a pre-trained hazard detection model, simultaneous location and type identification of geological hazard areas are achieved. By using the hazard detection model, spatial location regression and category label prediction can be completed simultaneously in one forward propagation, significantly improving identification efficiency. Simultaneously, the pre-trained hazard detection model has learned rich hazard morphology and distribution patterns on a large-scale sample, possessing strong generalization ability and effectively addressing the identification needs of slope hazards with different geological backgrounds and scales. This step directly outputs the spatial location of the hazard area and its corresponding hazard type, providing basic information for early warning and disaster prevention and mitigation of geological disasters.

[0033] As an optional implementation method, such as Figure 2 , Figure 2 This is a flowchart of a method for preprocessing multi-source remote sensing data provided in an embodiment of the present invention. In step S10, the method includes acquiring multi-source remote sensing data of a target slope area, performing layered preprocessing on the multi-source remote sensing data, and performing spatiotemporal registration processing on the layered preprocessed multi-source remote sensing data. S11: Acquire optical remote sensing data, synthetic aperture radar interferometry data, and lidar data for the target slope area. Specifically, in this embodiment of the invention, optical remote sensing data, synthetic aperture radar interferometry data, and lidar data for the target slope area can be acquired first, and then preprocessed according to the inherent characteristics of different data sources.

[0034] S12: The synthetic aperture radar interferometric measurement data is subjected to speckle reduction processing using an enhanced directional smoothing filter algorithm to obtain speckle-reduced synthetic aperture radar interferometric measurement data. Specifically, in this embodiment of the invention, considering the influence of speckle noise on the subsequent fusion accuracy of the synthetic aperture radar interferometric measurement data, an enhanced directional smoothing filter algorithm is used for speckle reduction processing. This algorithm can achieve efficient noise reduction while preserving edge information. The calculation formula for speckle reduction processing of synthetic aperture radar interferometric measurement data is shown in formula (1): (1) in, For pixels An adaptive neighborhood window that dynamically adjusts its size based on terrain complexity. The total number of pixels within the window. For spatial distance weighting parameters, The grayscale difference weighting parameter is used. The grayscale values ​​are those of the original synthetic aperture radar interferometry image. The image grayscale values ​​of the synthetic aperture radar interferometry data after speckle reduction are shown. For local directional contrast, To achieve the maximum global directional contrast, the original synthetic aperture radar interferometry data is subjected to speckle reduction processing to enhance the noise suppression targeting of the edge region, effectively solving the edge blurring problem of traditional filtering algorithms.

[0035] S13: Perform radiometric calibration on the optical remote sensing data, and perform atmospheric correction on the radiometrically calibrated optical remote sensing data based on the radiative transfer model to obtain atmospherically corrected optical remote sensing data. Specifically, in this embodiment of the invention, for optical remote sensing data, radiometric calibration is first performed to convert the digital quantization value into apparent radiance, and atmospheric correction is performed using the radiative transfer model to eliminate the effects of atmospheric absorption, scattering, etc. The calculation formula for atmospheric correction using the radiative transfer model is shown in formula (2): (2) in, This represents the true radiance of ground features. The radiance received by the satellite sensor. This represents the upward radiance of the atmosphere. This represents the downward atmospheric radiance. Atmospheric transmittance, By using ground surface emissivity and dynamically calculating various atmospheric parameters based on synchronously acquired atmospheric profile data, the accuracy of correction under complex meteorological conditions can be improved.

[0036] S14: The lidar data is denoised using an elevation anomaly detection algorithm to obtain denoised lidar data. Specifically, in this embodiment of the invention, an elevation anomaly detection-based denoising algorithm is used for the lidar data to remove anomalies caused by vegetation obstruction and instrument errors, thus preserving the true terrain elevation information.

[0037] S15: Using the digital elevation model generated from the denoised lidar data as a benchmark, a quadratic polynomial transformation is performed on the atmospherically corrected optical remote sensing data and the speckle-reduced synthetic aperture radar interferometry data to obtain multi-source remote sensing data with preliminary geometric correction. Specifically, in this embodiment of the invention, in the adaptive spatiotemporal registration of multi-source data, an adaptive spatiotemporal registration framework based on rational polynomial coefficient models and deep learning is constructed to address the differences in resolution, coordinate system, and temporal phase among different data sources. First, using the digital elevation model generated from the high-resolution lidar data as a benchmark, a quadratic polynomial transformation is used to correct the geometric deviation between the atmospherically corrected optical remote sensing data and the speckle-reduced synthetic aperture radar interferometry data. The quadratic polynomial transformation is shown in formula (3): (3) in, The original data consists of three-dimensional coordinates. To register the plane coordinates, , The registration coefficient is obtained by jointly iteratively solving the data from ground control points and lidar elevation data, and the elevation term Z is introduced to solve the registration error caused by terrain undulation.

[0038] S16: A feature matching model is used to register the multi-source remote sensing data after preliminary geometric correction, resulting in registered multi-source remote sensing data. Specifically, in this embodiment of the invention, to further improve registration accuracy, a Transformer-based feature matching model is constructed to extract invariant feature points from the multi-source data, such as terrain edges and artificial feature edges, and fine registration is achieved by calculating the spatiotemporal similarity of the feature points. The spatiotemporal similarity of the feature points... The calculation formula is shown in formula (4): (4) in, These are feature points from two different data sources. For the corresponding high-dimensional feature vector, For cosine similarity, The imaging time corresponding to the feature point. The time decay coefficient, These are the weighting factors for spatial and temporal features. The value can be greater than or equal to 0.7 and less than or equal to 0.85. This similarity formula enables joint matching in the spatiotemporal dimensions, eliminating erroneous matches caused by excessive temporal differences, and ultimately controlling the registration error to within 1 pixel, providing a high-precision geometric benchmark for subsequent multi-source data fusion.

[0039] As an optional implementation method, such as Figure 3 , Figure 3 This is a flowchart of a method for feature enhancement of multi-source remote sensing data provided in an embodiment of the present invention. In step S20, the multi-source remote sensing data after spatiotemporal registration is adaptively filtered for fusion bands. The filtered fusion bands are decomposed into multiple scales using an image decomposition algorithm, and feature enhancement processing is performed on the multi-scale decomposed fusion bands. Specifically, this includes steps S21-S24: S21: Adaptive selection of fusion bands is performed on the spatiotemporally registered multi-source remote sensing data, and multi-scale joint decomposition is performed on the selected fusion bands using a non-subsampling contourlet transform to obtain the low-frequency sub-band coefficients and high-frequency sub-band coefficients of each band. Specifically, in this embodiment of the invention, the multi-source remote sensing data after preliminary geometric correction is registered. After obtaining the registered multi-source remote sensing data, adaptive selection of fusion bands can be performed first. An optimal index factor model can be constructed to achieve quantitative evaluation and accurate selection of multi-source data band combinations. This model introduces the disaster sensitivity factor correlation coefficient (…). By conducting correlation analysis between each waveband and key influencing factors of slope hazards such as slope gradient, deformation rate, and vegetation cover, the weight of characteristic information closely related to geological hazards is strengthened, and an optimal index factor model is constructed. The calculation formula is shown in formula (5): (5) in, For the first The standard deviation of a band reflects the degree of dispersion and richness of information within that band. For the first The Pearson correlation coefficient between the band and the slope hazard sensitivity factor ranges from [value range missing]. The larger the absolute value, the stronger the ability of the band to characterize disaster features; For the first Band and the The correlation coefficient of a band is used to measure the degree of information redundancy between bands; This is a local minimum value, which can be 10⁻⁶, used to avoid the extreme case where the denominator is zero; The number of bands participating in the screening. This is determined by calculating different band combinations. The maximum value can be selected as the base band for fusion. Simultaneously, targeted optimizations are made based on the band characteristics of different data sources. For optical remote sensing data, considering the synergistic relationship between the normalized vegetation index and topographic slope, near-infrared and short-wave infrared bands sensitive to identifying potential hazards in vegetation-covered areas are prioritized. For synthetic aperture radar interferometry data, C-band or X-band data bands with high coherence and prominent deformation rate information are selected to ensure accurate capture of slope deformation information. For lidar data, the core band corresponding to elevation information is extracted.

[0040] After adaptively filtering the fused bands of the spatiotemporally registered multi-source remote sensing data, a non-subsampling transform can be used to decompose the filtered multi-source data into multiple scales and directions, fully exploring the spatiotemporal spectral characteristics of the data. Based on the multi-source remote sensing data with preliminary geometric correction in step S15, the accurate spatial matching of optical remote sensing, synthetic aperture radar interferometry, and lidar data is ensured. Subsequently, the three types of data are processed separately. Level-scale decomposition and Level directional decomposition. Among them, Decomposition levels range from 4 to 6, adaptively adjusted based on the terrain complexity of the study area. Higher-level decomposition is used for fragmented terrain areas, and lower-level decomposition is used for gently sloping terrain areas. Decomposition of directions at each level by: The directional decomposition accuracy is gradually increased from fine to coarse scales to ultimately obtain the low-frequency subband coefficients. and high-frequency directional subband coefficient .in, Representing different data sources, it mainly reflects the overall contour and low-frequency information of the image, and the high-frequency directional subband coefficients. It mainly reflects the details of the image, such as edges and textures.

[0041] S22: The low-frequency sub-band is weighted and fused based on the regional correlation coefficient, and the high-frequency sub-band is adaptively fused based on the local directional energy. Specifically, in this embodiment of the invention, in order to solve the problems of spectral distortion in the fusion of low-frequency coefficients and loss of detailed features in the fusion of high-frequency coefficients in the traditional non-subsampled contour wave transform decomposition, the regional correlation coefficient and local directional energy are introduced as fusion quantity measurement operators to construct an adaptive fusion strategy. For the low-frequency sub-band, which contains the main energy and spectral information of the image, a weighted fusion strategy is adopted. The terrain trend surface derived from the lidar elevation data is used as the fusion benchmark. The weighting coefficient is assigned by calculating the regional correlation coefficient between the low-frequency coefficients of each data source and the terrain trend surface to ensure that the fused image can retain the real terrain contour and spectral features. The calculation formula for the weighted fusion of the low-frequency sub-band is shown in formula (6): (6) in, The data is the result of weighted fusion of low-frequency subbands. This is the data before weighted fusion of low-frequency subbands. For the first The regional correlation coefficient between the low-frequency coefficient of the data source and the topographic trend surface is calculated using the following formula: (7) in, As an adaptive neighborhood window, the size is dynamically adjusted according to the complexity of the terrain around the pixel. The window takes a smaller value in areas with drastic terrain changes and a larger value in areas with gentle terrain. The elevation value of the topographic trend surface is obtained by polynomial fitting using lidar data. and These represent the mean low-frequency coefficient and the mean topographic elevation within the window, respectively.

[0042] For high-frequency subbands, which carry detailed information such as image edges and textures, the accuracy of boundary recognition for potential hazard areas is directly affected. Adaptive fusion of detailed information is achieved based on local directional energy. Local directional energy can effectively characterize the intensity of detail in an image in different directions. The calculation formula is shown in formula (8): (8) in, Representing different data sources, by comparing the local directional energy of high-frequency coefficients from different data sources, coefficients with richer detailed information are adaptively selected for fusion, and high-frequency subbands are analyzed based on local directional energy. The calculation formula for adaptive fusion is shown in formula (9): (9) in, The energy threshold can be determined by maximizing the inter-class variance adaptive thresholding method, which is used to distinguish effective detail information from noise. The weighting coefficients are adaptively allocated by the band information entropy, with data sources having higher weights for those with higher information entropy.

[0043] S23: Perform inverse contourlet transform without subsampling on the fused low-frequency and high-frequency subbands to reconstruct a preliminary fused image of the multi-source remote sensing data. Specifically, in this embodiment of the invention, by performing inverse contourlet transform without subsampling on the fused low-frequency and high-frequency subbands, the image structure and detail information retained during the multi-scale, multi-directional decomposition and adaptive fusion process can be completely reconstructed into a preliminary fused image. Since the inverse contourlet transform without subsampling has the advantages of translation invariance and anisotropic expression, the reconstruction process will not introduce spectral aliasing or ringing artifacts. Thus, while maintaining the geometric accuracy of the original remote sensing data, it maximizes the preservation of the key geometric features and texture details of landslide hazards, providing a fused image foundation with rich information, high fidelity, and good spatial consistency for subsequent feature enhancement processing.

[0044] S24: Based on the slope information in the lidar data and the deformation rate information in the synthetic aperture radar interferometry data, a feature enhancement weight matrix is ​​constructed. Adaptive enhancement is then applied to the preliminary fused image based on this feature enhancement weight matrix to obtain feature-enhanced multi-source remote sensing data. Specifically, in this embodiment of the invention, after obtaining the preliminary fused image, secondary feature enhancement processing is performed to address the common problems of feature blurring and insufficient contrast in slope-type geological hazard hazard areas. An adaptive enhancement model based on an attention mechanism is introduced. A feature enhancement weight matrix is ​​constructed using the slope information in the lidar data and the deformation rate information in the synthetic aperture radar interferometry data. This matrix can highlight the characteristics of high-risk hazard areas such as high slope and high deformation, and suppress interference information from areas without hazards. The calculation formula is shown in formula (10): (10) in, The slope value is extracted from the lidar data. The deformation rate is obtained from synthetic aperture radar interferometry. and These represent the maximum slope and maximum deformation rate of the study area, respectively. As a balancing coefficient, it is used to highlight the dominant influence of topographic factors on disaster risks, while also taking into account the supplementary role of deformation information.

[0045] Adaptive enhancement of the fused image is performed based on the feature enhancement weight matrix. The enhancement intensity is dynamically adjusted using a logarithmic function to avoid noise amplification caused by over-enhancement. The calculation formula is shown in formula (11): (11) in, To initially fuse the grayscale values ​​of the images, The local contrast is calculated using the Sobel operator. This represents the global average contrast.

[0046] Finally, a multi-dimensional quality evaluation system was used to comprehensively verify the enhanced fused images, ensuring that they achieved optimal results in terms of information richness, clarity, and spectral fidelity, providing high-quality data support for subsequent landslide hazard feature extraction. Evaluation indicators included information entropy. Average gradient And spectral angle matching degree SAM, where information entropy The calculation formula is shown in formula (12), the average gradient The calculation formula is shown in formula (13), and the calculation formula for spectral angle matching degree SAM is shown in formula (14).

[0047] (12) (13) (14) in, The probability is at the gray level. Image size, for Orientation gradient, To fuse image band values, These are the band values ​​of the original multispectral image. Information entropy is used to measure the information richness of the fused image, the average gradient reflects the image's sharpness and detail representation ability, and the spectral angle matching degree is used to evaluate the spectral fidelity of the fused image.

[0048] As an optional implementation method, such as Figure 4 , Figure 4 A flowchart for extracting slope hazard feature vectors provided in this embodiment of the invention includes steps S31-S35, which involve acquiring terrain sensing information of the target slope area, constructing a heterogeneous feature matrix based on multi-source remote sensing data after feature enhancement processing and terrain sensing information, and inputting the heterogeneous feature matrix into a deep learning model to extract slope hazard feature vectors. S31: Perform heterogeneous dimension mapping on the feature-enhanced multi-source remote sensing data, and fuse the mapped fused features with the terrain perception information to construct a heterogeneous feature matrix. Specifically, in this embodiment of the invention, a heterogeneous feature matrix including spectrum, deformation, and terrain is first constructed to achieve deep correlation of multi-source information. An adaptive dimension mapping module is introduced, and combined with the elevation information extracted by the lidar in the preprocessing step, the dimensionality reduction and enhancement of high-dimensional features are achieved through dynamic adaptive weight matrix allocation. Heterogeneous dimension mapping is performed on the feature-enhanced multi-source remote sensing data. The calculation formula for heterogeneous dimension mapping of the feature-enhanced multi-source remote sensing data is shown in formula (15), and the calculation formula for the adaptive weight matrix is ​​shown in formula (16). (15) (16) in, To generate feature vectors from multi-source remote sensing data after feature enhancement and heterogeneous dimensionality mapping. The output is a high-dimensional feature vector of the fused image. It is a feature vector composed of topographic parameters such as slope, aspect, and distance from the fault. The deformation rate eigenvector obtained from synthetic aperture radar interferometry. The adaptive weight matrix is ​​dynamically generated through a two-layer neural network. It is the Sigmoid activation function. This is an activation function that enables effective fusion and dimensional unification of heterogeneous features.

[0049] S32: The heterogeneous feature matrix is ​​filtered using a terrain-adaptive attention gating mechanism to obtain a filtered feature matrix. Specifically, in this embodiment of the invention, to enhance the saliency of features in potential hazard areas, a terrain-adaptive attention gating mechanism is introduced, based on the feature enhancement weight matrix constructed in the steps. The features mapped by the normalized vegetation index are filtered to obtain the feature matrix. The calculation formula is shown in formula (17): (17) in, For the gated weight matrix, The vegetation cover index, This is the balance coefficient (taken as 0.65).

[0050] S33: Construct a terrain-aware location encoding vector and fuse it with the filtered feature matrix to obtain a fused feature vector. Specifically, in this embodiment of the invention, to address the problem of Transformer's insensitivity to spatial location information, a terrain-aware location encoding mechanism is further introduced. This mechanism combines the spatial coordinates of pixels with terrain attributes such as slope and elevation to construct a two-dimensional location encoding vector. The calculation formula is shown in formula (18): (18) in, As a scale factor, For pixels The slope value, For elevation, This is the balancing coefficient (ranging from 0.5 to 0.8). The position encoding is then superimposed with the gated features to obtain the input feature matrix of the Transformer encoder. .

[0051] S34: Construct a Transformer model containing a multi-scale window attention module and a terrain feature enhancement module, and input the fused feature vector into the constructed Transformer model to extract multi-level encoded features. Specifically, in this embodiment of the invention, a Transformer encoder containing a multi-scale attention module and a terrain feature enhancement branch is constructed. A hierarchical attention mechanism can be designed for the multi-dimensional features of slope hazards, improving the model's ability to capture hazard features under complex terrain. First, the input feature matrix... Perform block processing to generate feature patch sequences ( (Number of patches), and convert each patch into a dimension through an embedding layer. The feature vectors. To adapt to slope hazards of different scales, a multi-scale window attention module is introduced, which divides the feature patch into three scale window sets: small (3x3), medium (5x5), and large (7x7). Self-attention is calculated within each window, and the fusion of features of different scales is achieved through cross-window attention interaction. The calculation formulas for the fusion of the multi-scale window attention module and the terrain feature enhancement module are shown in formulas (19) and (20): (19) in, This represents data after multi-scale feature fusion. ( The self-attention calculations correspond to small, medium, and large window sizes, respectively. This is a query, key, and value matrix at the corresponding scale. Weights for multi-scale feature fusion.

[0052] (20) in, The terrain weighting factor is dynamically allocated based on the sensitivity of different scale windows to terrain features.

[0053] After constructing a Transformer model that includes a multi-scale window attention module and a terrain feature enhancement module, residual connections and adaptive layer normalization can be inserted into the encoder, and a terrain feature enhancement branch can be introduced. Local terrain features are extracted through a convolutional neural network and fused with the global features of the Transformer to introduce a terrain-enhanced feature vector. The calculation formula is shown in formula (21): (twenty one) in, This is a 3x3 convolution operation used to extract local terrain features. For gating parameters, For element-wise multiplication, For adaptive layer normalization, the normalization parameters are dynamically adjusted based on the input features.

[0054] Feature depth extraction can be achieved by stacking encoder layers. Each encoder layer contains a multi-scale window attention module, a terrain feature enhancement branch, and a feedforward neural network. The calculation formula is shown in formula (22): (twenty two) in, This is the weight matrix. As a bias term, it achieves deep integration and abstract expression of multi-dimensional features such as terrain, texture, and deformation of slope hazards through multi-layer coding.

[0055] S35: Perform cross-layer weighted fusion processing on the multi-level encoded features, and reduce the dimensionality of the multi-level encoded features after cross-layer weighted fusion processing by attention pooling to output the slope hazard feature vector. Specifically, in this embodiment of the invention, in order to further improve the discriminative ability of the features, a cross-layer feature fusion mechanism is introduced to perform weighted fusion of the output features of each encoder layer of Transformer, making full use of the feature information of different levels. Shallow features focus on detailed textures, while deep features focus on semantic globality. The calculation formula for cross-layer weighted fusion processing of multi-level encoded features is shown in formula (23): (twenty three) in, This represents the feature vector obtained after cross-layer weighted fusion processing of multi-level encoded feature vectors. For the first Layer encoder output features For encoder layers, This refers to the layer weight.

[0056] The layer weight The calculation formula is shown in formula (24): (twenty four) By normalizing the feature response and feature importance score Dynamic allocation For normalization operations, and The correlation between features and slope hazard sensitivity factors is calculated to highlight the hierarchical weight of features related to disasters.

[0057] To address the issue of feature blurring in complex terrains such as vegetated areas and mountainous regions, a feature refinement module is introduced, based on local directional contrast. Maximum distance from water system in the study area The cross-layer fusion features are optimized, and the calculation formula for optimizing the cross-layer fusion features is shown in formula (25): (25) in, This represents the feature vector after optimizing the cross-layer fusion features. This is the refining coefficient (taken as 0.3~0.5). This is the weight matrix. To determine the maximum distance from the water system in the study area, a distance-to-water system factor is introduced to enhance the characteristic details of the near-water system area.

[0058] Finally, attention pooling is used for feature dimensionality reduction. Important feature dimensions are filtered by adaptive threshold. The calculation of dimensionality reduction of multi-level encoded features after cross-layer weighted fusion processing by attention pooling is shown in formula (26). The calculation formula of attention pooling is shown in formula (27). (26) (27) in, This represents the feature vector after attention pooling reduction. For the cross-layer fusion feature matrix, the first dimensional features, A two-layer perceptron is used to calculate the dimensional importance score. and These are the dimensionality reduction parameters. The final output dimension is... Key feature vectors of slope hazards (256~512 dimensions).

[0059] As an optional implementation method, such as Figure 5 , Figure 5 This is a flowchart of a method for identifying geological hazard types in a potential hazard area, provided by an embodiment of the present invention. In step S40, the slope hazard feature vector is input into a pre-trained hazard detection model to locate the potential hazard area and identify the geological hazard type in the located hazard area to obtain the hazard identification result for slope-type geological hazards, including: S41: Based on the slope hazard feature vector, an adaptive anchor point generation strategy is used to generate candidate hazard regions. Specifically, in this embodiment of the invention, candidate hazard regions can be generated based on the slope hazard feature vector using an adaptive anchor point generation strategy, dynamically adjusting the scale and aspect ratio of the anchor points. The formula for generating candidate hazard regions is shown in formula (28): (28) in, This represents the generated candidate region. For anchor point coordinates, Basic scaling factor Based on the aspect ratio factor, This is the slope weighting coefficient (taken as 0.4~0.6). This represents the slope value at the location of the anchor point. The maximum slope in the study area is defined as the maximum slope. By introducing a slope factor, the anchor points are made to better match the morphological characteristics of high-slope potential hazard areas, thereby improving the accuracy of suggestions for small-scale, steep terrain hazards.

[0060] S42: Extract the feature vector of each candidate hidden danger region using the target detection algorithm, and input the feature vector of each candidate hidden danger region into the pre-trained hidden danger detection model to calculate the category probability and boundary coordinate offset of each candidate hidden danger region. Specifically, in this embodiment of the invention, for the generated candidate regions, the feature vector of each candidate hidden danger region can be extracted using the target detection algorithm. Specifically, the feature vector can be extracted using the target detection algorithm, and the feature mapping can be performed by combining the multi-level encoded features after cross-layer weighted fusion processing. The feature mapping formula is shown in formula (29): (29) in, This represents the data after feature mapping of multi-level encoded features. For the mapping weight matrix, As a bias term, a two-dimensional convolution operation is used to unify the feature dimensions, providing high-quality regional features for subsequent classification and regression.

[0061] S43: Based on the category probability and boundary coordinate offset of each candidate hazard area, determine the initial hazard area and its confidence level of the target slope area. Specifically, in this embodiment of the invention, the pre-trained hazard detection model can be a dual-branch detection network, which can construct a dual-output head structure. The classification head is used to identify the hazard type, such as landslide, collapse, and potentially unstable slope; the regression head is used to optimize the boundary coordinates of the hazard area. The classification head adopts a structure combining a fully connected layer and an attention mechanism, and introduces a cross-entropy loss function weighted by hazard sensitivity factors to optimize the model parameters. The calculation formula for the cross-entropy loss function to optimize the model parameters is shown in formula (30): (30) in, This represents the cross-entropy loss function after incorporating a hazard sensitivity factor for weighting. This represents the number of candidate region samples. For real labels, To predict probabilities, As a sensitivity factor for hazard types, landslides are assigned a value of 1.2, collapses 1.1, and potentially unstable slopes 1.0. Differentiated regularization is used to improve the accuracy of identifying high-risk hazard types. The regularization coefficient is . This is the class head weight matrix.

[0062] The regression head uses a smooth L1 loss function to optimize the boundary coordinates, while also introducing a terrain constraint term to ensure that the regression results conform to actual terrain conditions. The calculation formula is shown in formula (31): (31) in, For the actual boundary coordinates, To predict boundary coordinates, This is the terrain constraint coefficient, which can be taken as 0.3 to 0.5. The slope aspect value of the boundary point is used to avoid the boundary prediction from deviating from the actual terrain trend through slope aspect constraints.

[0063] S44: The initial hidden danger area is screened using a flexible non-maximum suppression algorithm, and the confidence of the screened hidden danger area is corrected to obtain the hidden danger area of ​​the target slope area and the corresponding geological disaster type. Specifically, in this embodiment of the invention, in order to further improve the identification accuracy, the initial hidden danger area is screened using a flexible non-maximum suppression algorithm to solve the problem of repeated detection of overlapping hidden danger areas. At the same time, the confidence of the detection result is corrected by combining the local directional energy in step S22. The formula for correcting the confidence of the screened hidden danger area is shown in formula (32): (32) in, This indicates the confidence level of the corrected potential hazard area. The original prediction confidence level of the model. This is the balance coefficient. It can be taken as 0.7. This refers to the localized energy in the area of ​​potential hazards. The system outputs the location, type, and confidence level information of the potential hazard area to determine the maximum local directional energy, thereby enabling intelligent identification of slope-type geological hazard risks.

[0064] As an optional implementation, after inputting the slope hazard feature vector into a pre-trained hazard detection model to locate the hazard area and identify the type of geological hazard in the located hazard area to obtain the hazard identification result for slope-type geological hazards, the method further includes: S51: Construct a multi-dimensional accuracy verification index that includes classification accuracy, spatial accuracy, and feature consistency of the hidden danger area, and quantitatively evaluate the hidden danger identification results of the slope-type geological disasters to determine the hidden danger areas where the accuracy verification index of each dimension has not reached the preset accuracy threshold. Specifically, in this embodiment of the invention, after obtaining the hidden danger area and the corresponding geological disaster type of the target slope area, a multi-dimensional accuracy verification system can be constructed. Combined with field survey data, historical disaster records, and GIS spatial analysis data, a comprehensive quantitative evaluation of the identification results can be achieved. First, the confusion matrix is ​​used as the basic evaluation tool to calculate the core indicators such as overall accuracy (OA), Kappa coefficient, recall, precision, and F1 score. At the same time, the Kappa coefficient is improved by incorporating terrain complexity weight to improve the pertinence of the accuracy evaluation of complex terrain areas. The calculation formula for improving the Kappa coefficient is shown in formula (33): (33) in, This represents the improved Kappa coefficient. The total number of samples, For the first The number of true positive samples for this type of potential hazard. The first line, number Summation of columns, The terrain complexity weight can be calculated based on slope and vegetation coverage. The weight for high slope and high vegetation coverage areas is 1.2 to 1.5, and for flat areas it is 1.0. This formula enhances the reliability of the assessment of the accuracy of hazard identification under complex terrain.

[0065] Secondly, to improve the spatial positioning accuracy of potential hazard areas, spatial consistency (SC) and boundary of set error (BOE) indicators are introduced. Spatial consistency is calculated by the ratio of the intersection area to the union area of ​​the identified area and the field survey area. The calculation formula is shown in formula (34): (34) in, To identify the area of ​​intersection between the region and the actual area, For the area of ​​the union, To identify the root mean square error of the coordinates of the boundary and the true boundary, The error adjustment coefficient can be taken as 1 / 2 of the accuracy of the on-site survey.

[0066] Boundary offset error is obtained by calculating the average Euclidean distance between the identified boundary and the true boundary. The calculation formula is shown in formula (35): (35) in, To identify the boundary Coordinates of each sampling point The coordinates of the sampling points corresponding to the true boundary. The number of sampling points can be selected evenly according to the boundary length, with one sampling point every 10 meters.

[0067] In this embodiment of the invention, the concept of regional correlation coefficient is combined with the supplementary spectral and topographic consistency index. By comparing the correlation between the spectral features and topographic parameters of the identified area and the original remote sensing data, the physical rationality of the identification results can be ensured, forming a three-in-one verification system of classification accuracy, spatial accuracy, and feature consistency.

[0068] S52: For potential hazard areas where the accuracy verification indicators for each dimension do not reach the preset accuracy threshold, an uncertainty sampling strategy is used to screen samples that need to be supplemented with annotations. Specifically, in this embodiment of the invention, firstly, an uncertainty sampling strategy can be used to screen high-value samples, and the sample uncertainty is calculated by combining the confidence level and entropy value of hazard identification. The sample uncertainty... The calculation formula is shown in formula (36): (36) in, For the first The sample belongs to the first Predicted probability of such hidden dangers To maximize the prediction probability, this formula prioritizes selecting samples with ambiguous predictions and low confidence as supplementary labeled samples, which are then added to the training set after on-site verification.

[0069] S53: Add the samples to be annotated to the training set of the hazard detection model and adjust the class weights of the loss function. Specifically, in this embodiment of the invention, during the iterative optimization of the model based on active learning, according to the multi-dimensional evaluation results, for low-precision areas, such as OA < 90% or <0.85, high boundary offset error regions and regions with low feature consistency, for example, BOE>5 meters, initiate the active learning-based model iterative optimization process.

[0070] S54: Iteratively optimize the hazard detection model after adjusting the category weights of the loss function until the accuracy indicators of each dimension meet the preset accuracy threshold, thus obtaining the optimized slope-type geological hazard identification results. Specifically, in this embodiment of the invention, a dynamic weight adjustment mechanism is introduced to optimize the calculation formula of the category weights of the model's loss function based on the various hazard identification errors in the accuracy evaluation results. As shown in formula (37): (37) in, For the first The dynamic weight of each type of potential hazard is calculated as the reciprocal of the recall rate. , The regularization coefficient is used to enhance the learning of low-recall potential problem types through this loss function. Simultaneously, combining the weight allocation idea of ​​the Analytic Hierarchy Process (AHP), the feature extraction and recognition modules of the model are optimized hierarchically. The contribution of different modules is calculated through feature importance evaluation, and the learning rate is adjusted according to the contribution of different modules. Modules with high feature contribution use a smaller learning rate (0.0001~0.0005), and modules with low contribution use a larger learning rate (0.001~0.005), achieving precise iteration of model parameters. After each optimization, the multi-dimensional accuracy verification of S51 is repeated until the recognition accuracy meets the requirements. In this embodiment of the invention, it can be set to simultaneously meet the requirements. , , At a given time, the system outputs the hazard detection model and the optimized identification results for slope-type geological hazards.

[0071] As an optional implementation, after iteratively optimizing the hazard detection model after adjusting the category weights of the loss function until the accuracy indicators of each dimension meet the preset accuracy threshold, and obtaining the optimized slope-type geological hazard identification result, the method further includes: S61: Based on the optimized hazard identification results, extract multi-dimensional core parameters related to the terrain, deformation, and environment of the hazard area. Specifically, in this embodiment of the invention, after obtaining the optimized hazard identification results, multi-dimensional core parameters related to the terrain, deformation, and environment of the hazard area can be extracted. Specifically, based on the elevation values ​​generated from the lidar data in step S14, an improved slope and aspect algorithm can be used to extract terrain parameters. The slope calculation introduces a terrain complexity correction factor, and the calculation formula for the terrain complexity correction factor is as follows. As shown in formula (38): (38) in, The elevation values ​​of the DEM. They are respectively The partial derivative of the slope in the direction, The terrain complexity is calculated using the standard deviation of elevations within the window. This correction factor, representing the maximum topographic complexity of the study area, can improve the accuracy of slope extraction in complex terrain regions.

[0072] Furthermore, the slope aspect calculation for the target slope area employs a spherical coordinate system transformation method, and the slope aspect of the target slope area... The calculation formula is shown in formula (39): (39) Simultaneously, based on the image fusion extracted in step S23, environmental parameters such as vegetation cover, distance from water system, and distance from fault are extracted. Soil adjustment factors are introduced to eliminate soil background interference, and vegetation cover... The calculation formula is shown in formula (40): (40) in, For near-infrared reflectivity, Reflectivity in the red light band Soil regulating factors (values ​​ranging from 0.5 to 1.0, adaptively determined based on soil type in the study area); distances from water systems and faults were determined using the Euclidean distance formula in GIS spatial analysis combined with buffer analysis. The Euclidean formula is shown in formula (41): (41) in, For the coordinates of the water system or fracture line, the minimum distance value is obtained by traversing all coordinate points to ensure the accuracy of the distance parameter.

[0073] S62: Construct a hazard susceptibility evaluation index based on the multi-dimensional core parameters, and generate a hazard distribution map and risk level zoning map for the target slope area. Specifically, in this embodiment of the invention, a multi-scale visualization framework can be constructed based on the multi-dimensional core parameters extracted in S61 and the hazard identification results optimized in step S54, and GIS technology can be used to achieve hierarchical display and output of hazard information. An improved susceptibility evaluation model is introduced, integrating terrain, deformation, and environmental parameters to construct an evaluation index. The calculation formula is shown in formula (42): (42) in, The weights of each parameter are determined using the analytic hierarchy process (AHP) combined with the entropy weight method to avoid subjective bias. To normalize the parameters, an extreme value standardization method was used to divide the hazard areas into high, medium, and low risk levels based on the index. A tiered color scheme was used to generate a hazard distribution map, with high-risk areas using a red gradient, medium-risk areas using a yellow gradient, and low-risk areas using a blue gradient. Hazard boundaries, core parameter labels, slope, deformation rate, and normalized values ​​were overlaid. For the risk level zoning map, a combination of vector area symbols and raster rendering was used to highlight the spatial clustering characteristics of high-risk areas. Semi-transparent overlay technology was used to preserve topographic background information and enhance map readability.

[0074] S63: The digital elevation model generated from the denoised lidar data is fused with the risk level zoning map to construct a three-dimensional visualization model of the target slope area. Specifically, in this embodiment of the invention, a three-dimensional visualization model is constructed by combining the digital elevation model generated from the denoised lidar data with the risk level zoning map. The model uses lighting rendering to simulate terrain undulations, which can intuitively display the spatial distribution of the potential hazard area and its relationship with the terrain.

[0075] S64: Based on the aforementioned hazard distribution map, risk level zoning map, and 3D visualization model, generate a standard visualization result for hazard identification in the target slope area. Specifically, in this embodiment of the invention, a standardized result package for hazard identification in the target slope area can be generated based on the aforementioned hazard distribution map, risk level zoning map, and 3D visualization model. This package includes a vector-format hazard boundary file, a raster-format risk level map, an Excel-format hazard parameter statistical table, and a PDF-format analysis report. The report contains key information such as hazard location, risk level, core parameters, and prevention and control recommendations.

[0076] This invention, through adaptive selection of fusion bands from multi-source remote sensing data after spatiotemporal registration, employs an image decomposition algorithm to perform multi-scale decomposition on the selected fusion bands. Feature enhancement processing is then applied to these multi-scale decomposed fusion bands, significantly improving the information richness, clarity, and spectral fidelity of the fused image. This enhances the saliency of features in slope hazard areas and effectively solves the problems of feature blurring and detail loss in traditional fusion methods. Simultaneously, a heterogeneous feature matrix is ​​constructed based on the feature-enhanced multi-source remote sensing data and terrain perception information. This heterogeneous feature matrix is ​​then input into a deep learning model to extract slope hazard feature vectors, greatly improving the identification accuracy of slope hazards in complex terrain and effectively reducing the false negative rate and type discrimination error of small-scale hazards.

[0077] Example 2 Based on the above-mentioned method for identifying potential hazards of slope-type geological disasters, this invention provides a device for identifying potential hazards of slope-type geological disasters. Please refer to [link / reference needed]. Figure 6 , Figure 6A slope-type geological hazard hazard identification device 60 is provided as an embodiment of the present invention. The slope-type geological hazard hazard identification device 60 includes: The acquisition module 61 is used to acquire multi-source remote sensing data of the target slope area, perform layered preprocessing on the multi-source remote sensing data, and perform spatiotemporal registration processing on the multi-source remote sensing data after layered preprocessing. The decomposition module 62 is used to adaptively select fusion bands for multi-source remote sensing data after spatiotemporal registration processing. It performs multi-scale decomposition on the selected fusion bands through an image decomposition algorithm and performs feature enhancement processing on the multi-scale decomposed fusion bands. Module 63 is used to acquire terrain perception information of the target slope area, construct a heterogeneous feature matrix based on the multi-source remote sensing data and terrain perception information after feature enhancement processing, and input the heterogeneous feature matrix into a deep learning model to extract slope hazard feature vectors. The positioning module 64 is used to input the slope hazard feature vector into the pre-trained hazard detection model to locate the hazard area of ​​the geological disaster, identify the geological disaster type of the located hazard area, and obtain the hazard identification result of slope geological disaster.

[0078] For further details regarding the implementation of the above-mentioned technical solutions by each module in the above-mentioned slope geological hazard hazard identification device, please refer to the description in the slope geological hazard hazard identification method provided in the above-mentioned invention embodiments, which will not be repeated here.

[0079] Example 3 Based on the above-mentioned methods for identifying potential hazards in slope-type geological disasters, such as Figure 7 As shown in the diagram, this embodiment of the invention provides a structural schematic of a hazard identification device for slope-type geological hazards. The device includes a processor 71 and a memory 72 coupled to the processor 71. The memory 72 stores a computer program, which, when executed by the processor 71, causes the processor 71 to perform the steps of the hazard identification method for slope-type geological hazards described in the above embodiment.

[0080] For other details regarding the implementation of the above-mentioned technical solution by the processor 71 in the above-mentioned slope geological hazard hazard identification device, please refer to the description in the slope geological hazard hazard identification method provided in the above-mentioned invention embodiment, which will not be repeated here.

[0081] The processor 71 can also be called a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or the processor 71 can be any conventional processor.

[0082] Example 4 like Figure 8 The diagram illustrates the structure of a computer-readable storage medium provided in this embodiment of the invention. The storage medium stores a readable computer program 81. This computer program 81 can be stored in the storage medium as a software product and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in various embodiments of the invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks or optical disks, ROM (Read-Only Memory), RAM (Random Access Memory), or terminal devices such as computers, servers, mobile phones, and tablets.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or modules, and may be electrical, mechanical, or other forms.

[0084] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0086] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0087] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0088] The technical solutions provided in this application have been described in detail above. Specific examples have been used in this application to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

[0089] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for identifying potential hazards of slope-type geological disasters, characterized in that, include: Acquire multi-source remote sensing data of the target slope area, perform layered preprocessing on the multi-source remote sensing data, and perform spatiotemporal registration processing on the multi-source remote sensing data after layered preprocessing. Adaptive selection of fusion bands is performed on the spatiotemporal registration-processed multi-source remote sensing data. The selected fusion bands are then decomposed into multiple scales using an image decomposition algorithm, and feature enhancement processing is applied to the multi-scale decomposed fusion bands. The terrain perception information of the target slope area is obtained, and a heterogeneous feature matrix is ​​constructed based on the multi-source remote sensing data and terrain perception information after feature enhancement processing. The heterogeneous feature matrix is ​​then input into a deep learning model to extract slope hazard feature vectors. The slope hazard feature vector is input into a pre-trained hazard detection model to locate the hazard area of ​​the geological disaster and identify the geological disaster type of the located hazard area, thus obtaining the hazard identification result of slope-type geological disasters.

2. The method for identifying potential geological hazards on slopes according to claim 1, characterized in that, The process of acquiring multi-source remote sensing data of the target slope area, performing layered preprocessing on the multi-source remote sensing data, and performing spatiotemporal registration processing on the preprocessed multi-source remote sensing data includes: Acquire optical remote sensing data, synthetic aperture radar interferometry data, and lidar data for the target slope area; An enhanced directional smoothing filter algorithm is used to reduce the speckle of the synthetic aperture radar interferometry data to obtain speckle-reduced synthetic aperture radar interferometry data. The optical remote sensing data is radiometrically calibrated, and atmospheric correction is performed on the radiometrically calibrated optical remote sensing data based on the radiative transfer model to obtain atmospherically corrected optical remote sensing data. The lidar data is denoised based on the elevation anomaly detection algorithm to obtain denoised lidar data. Based on the digital elevation model generated from the denoised lidar data, a second polynomial transformation is performed on the atmospheric-corrected optical remote sensing data and the speckle-reduced synthetic aperture radar interferometry data to obtain multi-source remote sensing data with preliminary geometric correction. A feature matching model was used to register the multi-source remote sensing data after preliminary geometric correction, resulting in registered multi-source remote sensing data.

3. The method for identifying potential geological hazards on slopes according to claim 2, characterized in that, The adaptive selection of fusion bands for multi-source remote sensing data after spatiotemporal registration processing, the multi-scale decomposition of the selected fusion bands using an image decomposition algorithm, and the feature enhancement processing of the multi-scale decomposed fusion bands include: Adaptive selection of fusion bands is performed on the spatiotemporal registration-processed multi-source remote sensing data, and multi-scale joint decomposition is performed on the selected fusion bands using non-subsampling contour wave transform to obtain the low-frequency subband coefficients and high-frequency subband coefficients of each band. The low-frequency subband is weighted and fused based on the regional correlation coefficient, and the high-frequency subband is adaptively fused based on the local directional energy. The fused low-frequency subband and the fused high-frequency subband are subjected to inverse contour wave transformation without subsampling to reconstruct the preliminary fused image of the multi-source remote sensing data. Based on the slope information in the lidar data and the deformation rate information in the synthetic aperture radar interferometry data, a feature enhancement weight matrix is ​​constructed, and the preliminary fused image is adaptively enhanced based on the feature enhancement weight matrix to obtain feature-enhanced multi-source remote sensing data.

4. The method for identifying potential geological hazards on slopes according to claim 3, characterized in that, The process involves acquiring terrain sensing information of the target slope area, constructing a heterogeneous feature matrix based on the multi-source remote sensing data and terrain sensing information after feature enhancement processing, and inputting the heterogeneous feature matrix into a deep learning model to extract slope hazard feature vectors, including: Heterogeneous dimension mapping is performed on the feature-enhanced multi-source remote sensing data, and the mapped fused features are fused with the terrain perception information to construct a heterogeneous feature matrix; The heterogeneous feature matrix is ​​filtered by a terrain-adaptive attention gating mechanism to obtain the filtered feature matrix. Construct a terrain-aware location encoding vector, and fuse the terrain-aware location encoding vector with the filtered feature matrix to obtain a fused feature vector; A Transformer model containing a multi-scale window attention module and a terrain feature enhancement module is constructed, and the fused feature vector is input into the constructed Transformer model to extract multi-level encoded features; The multi-level encoded features are subjected to cross-layer weighted fusion processing, and the dimensionality of the multi-level encoded features after cross-layer weighted fusion processing is reduced by attention pooling to output the slope hazard feature vector.

5. The method for identifying potential geological hazards on slopes according to claim 1, characterized in that, The process involves inputting the slope hazard feature vector into a pre-trained hazard detection model to locate the hazard area and identify the type of geological hazard in the located hazard area, thereby obtaining the hazard identification result for slope-type geological hazards. This includes: Based on the slope hazard feature vector, an adaptive anchor point generation strategy is used to generate candidate hazard areas; The feature vector of each candidate hazard area is extracted by the target detection algorithm, and the feature vector of each candidate hazard area is input into the pre-trained hazard detection model to calculate the category probability and boundary coordinate offset of each candidate hazard area. Based on the category probability and boundary coordinate offset of each candidate hazard area, the initial hazard area and its confidence level of the target slope area are determined; A flexible nonmaximum suppression algorithm is used to screen the initial hidden danger areas, and the confidence of the screened hidden danger areas is corrected to obtain the hidden danger areas and corresponding geological disaster types of the target slope area.

6. The method for identifying potential hazards of slope-type geological disasters according to claim 5, characterized in that, After inputting the slope hazard feature vector into a pre-trained hazard detection model to locate the hazard area and identify the type of geological hazard in the located hazard area to obtain the hazard identification result for slope-type geological hazards, the method further includes: A multi-dimensional accuracy verification index is constructed, which includes the classification accuracy, spatial accuracy, and feature consistency of the hazard area. The hazard identification results of the slope-type geological disaster are quantitatively evaluated, and the hazard areas where the accuracy verification index of each dimension does not reach the preset accuracy threshold are identified. For potential areas where the accuracy verification indicators of each dimension do not reach the preset accuracy threshold, an uncertainty sampling strategy is used to screen samples that need to be supplemented with annotations. The samples to be annotated are added to the training set of the hazard detection model, and the class weights of the loss function are adjusted. The hazard detection model after adjusting the category weights of the loss function is iteratively optimized until the accuracy indicators of each dimension meet the preset accuracy threshold, and the optimized slope geological hazard identification results are obtained.

7. The method for identifying potential geological hazards on slopes according to claim 6, characterized in that, The method further includes iteratively optimizing the hazard detection model after adjusting the category weights of the loss function until the accuracy indicators of each dimension meet the preset accuracy threshold, and obtaining the optimized slope-type geological hazard identification result. Based on the optimized hazard identification results, multi-dimensional core parameters related to the terrain, deformation, and environment of the hazard area are extracted. Based on the aforementioned multi-dimensional core parameters, a hazard susceptibility evaluation index is constructed, and a hazard distribution map and risk level zoning map of the target slope area are generated. The digital elevation model generated from the denoised lidar data is fused with the risk level zoning map to construct a three-dimensional visualization model of the target slope area. Based on the aforementioned hazard distribution map, risk level zoning map, and 3D visualization model, standard visualization results for hazard identification in the target slope area are generated.

8. A device for identifying potential hazards of slope-type geological disasters, characterized in that, include: The acquisition module is used to acquire multi-source remote sensing data of the target slope area, perform layered preprocessing on the multi-source remote sensing data, and perform spatiotemporal registration processing on the multi-source remote sensing data after layered preprocessing. The decomposition module is used to adaptively select fusion bands for multi-source remote sensing data after spatiotemporal registration. It performs multi-scale decomposition on the selected fusion bands using an image decomposition algorithm and then performs feature enhancement processing on the multi-scale decomposed fusion bands. The construction module is used to acquire terrain perception information of the target slope area, construct a heterogeneous feature matrix based on the multi-source remote sensing data and terrain perception information after feature enhancement processing, and input the heterogeneous feature matrix into a deep learning model to extract slope hazard feature vectors. The localization module is used to input the slope hazard feature vector into the pre-trained hazard detection model to locate the hazard area of ​​the geological disaster and identify the geological disaster type of the located hazard area, so as to obtain the hazard identification result of slope geological disaster.

9. A device for identifying potential hazards in slope-type geological disasters, characterized in that, Includes memory and processor, wherein: The memory is used to store computer programs; The processor is used to read the computer program in the memory and execute the steps of the method for identifying potential hazards of slope-type geological disasters as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a readable computer program that, when executed by a processor, implements the steps of a method for identifying potential hazards of slope-type geological disasters as described in any one of claims 1 to 7.

Images