Intelligent identification and grading early warning method for regional geological disasters based on multi-element data deep fusion

Abstract

This invention belongs to the field of geological disaster monitoring and early warning technology, specifically a method for intelligent identification and hierarchical early warning of regional geological disasters through deep fusion of multi-source data. The method includes the following steps: collecting multi-source data and constructing a three-dimensional voxel array; calculating the data density of multi-source data in the three-dimensional voxel array and determining a standardized voxel fusion array; constructing an enhanced multimodal feature set based on the determined standardized voxel fusion array and performing modal consistency verification; determining voxel-level disaster identification results based on the verification results; calculating the displacement ratio and rainfall intensity ratio; performing multi-level early warning judgment based on the calculation results; simultaneously, voting and fusing the voxel early warning results within the region to generate a regional geological disaster hierarchical early warning map; comparing the early warning results with actual disaster data and performing feature coupling parameter correction. This invention achieves structured expression of multi-source data under a unified spatiotemporal benchmark by adopting spatiotemporal consistency calibration technology for multi-source heterogeneous data.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster monitoring and early warning technology, specifically a method for regional geological disaster intelligent identification and hierarchical early warning through deep fusion of multi-source data. Background Technology

[0002] With the increasing frequency of geological disasters such as landslides, debris flows, and ground subsidence, traditional monitoring methods are gradually shifting from single-sensor monitoring to multi-source data collaborative analysis. Existing technologies typically use remote sensing images, surface displacement monitoring data, meteorological data, and geological exploration data for disaster identification and risk assessment, which to some extent improves the coverage and information dimensions of disaster identification.

[0003] However, existing technologies are mainly limited by core technical bottlenecks such as inconsistent spatiotemporal resolution of data and insufficient cross-modal feature coupling ability when achieving collaborative fusion of multi-source heterogeneous data and high-precision disaster identification. This makes it difficult to balance the real-time performance of disaster identification with the accuracy and stability of the identification results in practical applications, thereby affecting the response efficiency and reliability of the overall early warning system. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for intelligent identification and hierarchical early warning of regional geological disasters through deep fusion of multi-source data, thereby resolving issues such as inconsistent spatiotemporal resolution of data and insufficient cross-modal feature coupling capability in existing technologies.

[0005] A method for intelligent identification and hierarchical early warning of regional geological disasters through deep fusion of multi-source data includes the following steps: collecting multi-source data of the study area, setting the basic voxel granularity based on the spatial resolution of each data, constructing a three-dimensional voxel array, calculating the data density of multi-source data in the three-dimensional voxel array, performing adaptive resampling based on the calculated data density, verifying the structural connectivity of the resampled three-dimensional voxel array, and determining the standardized voxel fusion array.

[0006] Based on the defined standardized voxel fusion array, the corresponding modal features are extracted, and the attention weight matrix between each pair of modal feature maps is calculated based on the attention mechanism. Coupled features are generated based on the calculated attention weight matrix, and the coupled features are residually connected with the original modal features to construct an enhanced multimodal feature set. Modal consistency verification is performed on the enhanced multimodal feature set, and the voxel-level disaster identification result is determined based on the verification results.

[0007] A time-sliding window is constructed based on the voxel-level disaster identification results, and the displacement ratio and rainfall intensity ratio are calculated. Multi-level early warning judgments are performed based on the calculation results. At the same time, the voxel early warning results in the region are fused by voting to generate a regional geological disaster graded early warning map.

[0008] The early warning results are compared with the actual disaster data, and the accuracy rate, missed rate and false alarm rate are calculated. Based on the calculation results, the feature coupling parameters are corrected, and the optimization results are fed back to the aforementioned steps to form a closed-loop update mechanism.

[0009] Preferably, the adaptive resampling is performed as follows:

[0010] The data density within each voxel of the initial 3D voxel array for the remote sensing image data layer, surface displacement data layer, meteorological data layer, and geological data layer in the multi-source dataset is calculated separately, and adaptive resampling is performed based on the calculation results, specifically as follows:

[0011] If the data density of a data layer within a voxel is less than 30% of the global average density of the corresponding data layer, the granularity of the corresponding voxel in the corresponding dimension is doubled, and the average data density within the expanded voxel is recalculated until the data density within all voxels is not less than 30% of the global average density.

[0012] Preferably, the structural connectivity verification of the resampled three-dimensional voxel array is performed as follows:

[0013] The resampled 3D voxel array is divided into K connected sub-regions, and one sub-region is randomly selected from the divided sub-regions as the initial region for structural connectivity verification. At the same time, all adjacent sub-regions of the initial region are located.

[0014] Determine the boundary voxel set of the initial region and the boundary voxel set of the adjacent region, calculate the minimum Euclidean distance between the two boundary voxel sets, and verify the structural connectivity based on the calculation results.

[0015] Preferably, the structural connectivity verification based on the calculation results is performed as follows:

[0016] If the minimum Euclidean distance between two boundary voxel sets is zero or the total length of the voxel connectivity path within adjacent regions, it indicates that the data structure between the initial region and the adjacent region is connected; otherwise, it indicates that the data structure between the two regions is not connected.

[0017] Using adjacent regions as the initial regions for the next generation, we continue to calculate the minimum Euclidean distance between the initial regions and their adjacent regions until all sub-regions have been traversed. We count the number of structural disconnections between all pairs of regions and verify the overall structural connectivity based on the number of disconnections. Then:

[0018] Set the threshold for the number of times the structure is disconnected as follows: If the number of disconnections counted during the verification of the 3D voxel array is greater than If the verification fails, the basic voxel granularity is reduced to 80% of the current granularity, and adaptive resampling and structural connectivity verification are re-executed until the verification passes.

[0019] If the number of disconnections is less than or equal to If the structure of the current three-dimensional voxel array is verified, the current three-dimensional voxel array will be output as a normalized voxel fusion array.

[0020] Preferably, the modality consistency verification of the enhanced multimodal feature set is performed as follows:

[0021] Set the difference threshold The KL divergence of all modal feature pairs is greater than 1. If the logarithm is greater than the preset maximum inconsistency logarithm threshold, then... If the modal coupling is insufficient and there is a feature conflict, based on the spatial distribution of the current coupled features, the voxel granularity corresponding to the conflict region is further refined by 20% in the adaptive resampling, and the normalized voxel fusion array is reconstructed.

[0022] Conversely, if the logarithm is less than or equal to If the modal coupling verification is successful, the enhanced multimodal feature set will be input into a pre-trained classifier, which will output the geological hazard identification results for each voxel, including four categories: landslide, debris flow, ground collapse, and no hazard.

[0023] Preferably, the calculated displacement ratio to rainfall intensity ratio is as follows:

[0024] For each voxel of the disaster category in the voxel-level disaster identification results, including landslides, debris flows and ground collapses, the displacement features and meteorological features of each voxel of the disaster category are extracted in time series sequences within N consecutive time windows.

[0025] For each disaster voxel, calculate the ratio of its cumulative displacement in the current time window to the average displacement over the previous M windows. And the ratio of the cumulative rainfall intensity in the current window to the historical average rainfall intensity for the same period. .

[0026] Preferably, the step of performing multi-level early warning determination based on the calculation results is as follows:

[0027] Define four levels of early warning thresholds: Attention level (blue warning), Warning level (yellow warning), Alert level (orange warning), and Risk avoidance level (red warning);

[0028] like If so, it is determined to be a blue alert;

[0029] like If so, it is determined to be a yellow alert;

[0030] like If so, it is determined to be an orange alert;

[0031] like If so, it is determined to be a red alert.

[0032] Preferably, the step of performing feature coupling parameter correction based on the calculation results is as follows:

[0033] Calculate the accuracy rate for each warning level. underreporting rate and false alarm rate And set a lower limit for accuracy tolerance. Upper limit of tolerance for underreporting rate And the upper limit of false alarm rate tolerance ;

[0034] like If the current warning fails, the feature coupling parameter correction is performed according to the failure type.

[0035] like If the current warning is valid, the voxel granularity strategy, attention weight parameters, and temporal criterion thresholds in the current warning process will be used as the optimized initial state for the next cycle of geological disaster warning.

[0036] Preferably, the step of performing feature coupling parameter correction based on the failure type is as follows:

[0037] like If the error is not found, it is determined to be feature coupling overfitting. The voxel features of the false positive and false positive regions are used as negative samples. The initialization parameters of the cross-modal attention weight matrix are adjusted using a contrastive learning adjustment strategy, and the regularization coefficient of the attention weight is increased.

[0038] like If the data fusion is insufficient, the base voxel granularity will be reduced by 50% within a range of twice the number of voxels around the location of the missing voxel region, and the structural connectivity verification and all subsequent steps will be re-executed.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This invention achieves structured representation of multi-source heterogeneous data under a unified spatiotemporal benchmark by adopting spatiotemporal consistency calibration and adaptive voxelization fusion techniques. This effectively solves the problem of inconsistent resolution of different data in the prior art, thus taking into account both spatial accuracy and data integrity in the data fusion process.

[0041] 2. This invention achieves adaptive compensation and optimization for sparse data regions by adopting data density-driven adaptive resampling and voxel granularity dynamic adjustment techniques, effectively solving the problem of missing or unbalanced data in local areas in the prior art, thereby improving the stability and reliability of the overall data fusion.

[0042] 3. By adopting the three-dimensional voxel array structure connectivity verification and iterative correction technology, this invention achieves global constraints on the continuity of the fused data structure, effectively avoiding the spatial breakage problem generated during the data fusion process, thereby improving the structural consistency of the input data of the subsequent disaster identification model.

[0043] 4. This invention achieves dynamic correlation modeling between different modal data by adopting a multi-feature coupling technique driven by a cross-modal attention mechanism, effectively solving the problem of insufficient cross-modal feature fusion capability in the existing technology, thereby improving the discrimination capability of disaster feature extraction.

[0044] 5. This invention achieves dynamic characterization of disaster evolution trends by adopting a multi-dimensional feature joint analysis technique based on time-series stability criteria. This effectively solves the problem of delayed early warning caused by relying solely on single-moment data in existing technologies, thus balancing the real-time nature and forward-looking nature of early warning.

[0045] 6. This invention achieves effective mapping of local identification results to macro-level early warning decisions by adopting a voting fusion and level enhancement technique that integrates voxel-level results with regional-level early warnings, thereby effectively improving the spatial expressiveness and decision-making guidance value of early warning results. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall method steps of the regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data of the present invention. Detailed Implementation

[0047] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0048] Example 1

[0049] Reference Figure 1 As an embodiment of the present invention, a method for intelligent identification and graded early warning of regional geological disasters through deep fusion of multi-source data is provided, comprising the following steps:

[0050] S1: Spatiotemporal consistency calibration and voxelization fusion of multi-source heterogeneous data.

[0051] Specifically, the spatiotemporal consistency calibration and voxelization fusion of the multi-source heterogeneous data is achieved by acquiring remote sensing images, surface displacement monitoring data, meteorological data, and geological exploration data of the study area, and performing spatiotemporal alignment and adaptive voxelization fusion based on the acquired multi-source data. The specific implementation is as follows:

[0052] By utilizing multi-source sensors and data acquisition systems to obtain remote sensing imagery, surface displacement monitoring data, meteorological data, and geological exploration data for the study area, we can obtain:

[0053] ;

[0054] in, Represents remote sensing image data, This represents surface displacement monitoring data. Represents meteorological data. This represents geological exploration data. This represents the constructed multi-source data set;

[0055] Furthermore, for any data, we have:

[0056] ;

[0057] in, Represents spatial coordinates, Indicates the timestamp of the collection. Represents the observed value. Indicates the first Number of samples in class data Indicates the first The data includes: remote sensing imagery of the study area, surface displacement monitoring data, meteorological data, and geological exploration data.

[0058] Based on the acquired remote sensing image data, surface displacement monitoring data, meteorological data, and geological exploration data, spatiotemporal consistency calibration and voxel fusion are performed, specifically as follows:

[0059] The acquired remote sensing images, surface displacement monitoring data, meteorological data and geological exploration data were time-stamp aligned and spatial coordinates were normalized to form a raw data layer with a unified geographic reference. The time-stamp alignment used a linear interpolation method to interpolate all data to the same time frequency, and the spatial coordinate normalization used the WGS-84 coordinate system and projected all data onto the same plane grid.

[0060] Based on the original spatial resolution of each data layer, a basic voxel granularity is defined. This basic voxel granularity is half the lowest spatial resolution among all data layers. The study area is then discretized into a three-dimensional voxel array using this basic voxel granularity. The three dimensions of the three-dimensional voxel array represent the longitude direction, the latitude direction, and the time window index, respectively. Specifically:

[0061] Set the spatial resolution of each data layer to , Indicates the first Class data in Spatial resolution in a direction refers to the average spatial distance between two adjacent sampling points in that direction. Indicates the first Class data in Spatial resolution in the direction;

[0062] The spatial resolution of the basic voxel granularity is ;

[0063] A three-dimensional voxel array will be constructed based on the spatial resolution of voxel granularity in the study area and the sampling time node.

[0064] The data density within each voxel of the initial 3D voxel array for the remote sensing image data layer, surface displacement data layer, meteorological data layer, and geological data layer is calculated separately, and adaptive resampling is performed based on the calculation results, as follows:

[0065] If the data density of a data layer within a voxel is less than 30% of the global average density of that data layer (which is the arithmetic mean of the data densities of the corresponding data layer in all voxels), then the granularity of that voxel in the corresponding dimension is doubled, and the data mean within the expanded voxel is recalculated until the data density within all voxels is not less than 30% of the global average density.

[0066] The structural connectivity of the resampled 3D voxel array was verified, specifically as follows:

[0067] The resampled three-dimensional voxel array is divided into K connected sub-regions, and one sub-region is randomly selected from the divided sub-regions as the initial region for structural connectivity verification. At the same time, all adjacent sub-regions of the initial region are located, where K is a positive integer set by the implementers based on the area of ​​the study region and the total number of voxels.

[0068] Determine the boundary voxel set of the initial region and the boundary voxel sets of adjacent regions, and calculate the minimum Euclidean distance between the two boundary voxel sets. Based on the calculation results, verify the structural connectivity. Then:

[0069] If the minimum Euclidean distance between two boundary voxel sets is zero or the total length of the voxel connectivity path within adjacent regions, it indicates that the data structure between the initial region and the adjacent region is connected; otherwise, it indicates that the data structure between the two regions is not connected.

[0070] Using adjacent regions as the initial regions for the next generation, we continue to calculate the minimum Euclidean distance between the initial regions and their adjacent regions until all sub-regions have been traversed. We count the number of structural disconnections between all pairs of regions and verify the overall structural connectivity based on the number of disconnections. Then:

[0071] Set the threshold for the number of times the structure is disconnected as follows: If the number of disconnections counted during the verification of the 3D voxel array is greater than If the verification fails, the basic voxel granularity is reduced to 80% of the current granularity, and adaptive resampling and structural connectivity verification are re-executed until the verification passes.

[0072] If the number of disconnections is less than or equal to If the structure of the current three-dimensional voxel array is verified, the current three-dimensional voxel array will be output as a normalized voxel fusion array.

[0073] S2: Geological hazard feature coupling and identification based on cross-modal attention graph network.

[0074] Specifically, the geological hazard feature coupling and identification based on cross-modal attention map networks is based on a deterministic standardized voxel fusion array. Through dynamic attention mechanism and modality consistency discrimination logic, high-discriminative hazard features are extracted and voxel-level hazard identification is performed. The specific implementation is as follows:

[0075] The determined normalized voxel fusion array is split into remote sensing feature tensors, displacement feature tensors, meteorological feature tensors, and geological feature tensors according to the data source. These are then input into four parallel graph convolutional sub-networks. Each graph convolutional sub-network contains three graph convolutional layers, with the number of output channels for each graph convolutional layer being 64, 128, and 256, respectively. High-order feature maps of each modality are then extracted.

[0076] A cross-modal attention coupling module is constructed, the attention weight matrix between each pair of modal feature maps is calculated, and coupled features are generated based on the calculated attention weight matrix, as follows:

[0077] For any two modes A and B, the attention score of the feature map of mode A relative to the feature map of mode B is determined by calculating the vector dot product of the features of mode A and the features of mode B, and then normalizing the result using Softmax.

[0078] ;

[0079] in, Let A be the feature query matrix of mode A. Let B be the eigenkey matrix. For feature dimensions;

[0080] By performing residual connections between the coupled features and the original modal features to form an enhanced multimodal feature set, we have:

[0081] ;

[0082] in, For original modal features, As a coupling feature, This is the residual weighting coefficient, which is set by the implementers according to the actual application scenario.

[0083] Modality consistency determination is performed by calculating the KL divergence between any two modality features in the enhanced multimodal feature set, and then determining modality consistency based on the calculation results. Specifically:

[0084] ;

[0085] in, and These are the feature probability distributions of two different modalities. The calculated KL divergence is used for modal consistency determination, specifically as follows:

[0086] Set the difference threshold The KL divergence of all modal feature pairs is greater than 1. If the logarithm is greater than the preset maximum inconsistency logarithm threshold, then... If the modal coupling is insufficient and there is a feature conflict, then return to step S1. Based on the spatial distribution of the current coupled features, increase the voxel granularity corresponding to the conflict region by 20% in the adaptive resampling of S1 and reconstruct the normalized voxel fusion array.

[0087] Conversely, if the logarithm is less than or equal to If the modal coupling verification is successful, the enhanced multimodal feature set will be input into a pre-trained classifier, which will output the geological hazard identification results for each voxel, including four categories: landslide, debris flow, ground collapse, and no hazard.

[0088] It should be noted that the pre-trained classifier adopts a three-layer fully connected network structure, with the input dimension being the dimension of the enhanced features and the output dimension being four. The activation function is the ReLU function, and the output layer uses the Softmax function for probability normalization.

[0089] It should be noted that the calculated KL divergence is used to evaluate the degree of distributional difference between different modal features, and thus verify the effectiveness of multimodal feature coupling. Specifically:

[0090] For each modality's feature map, probability distribution normalization is performed, transforming the feature values ​​to the interval between zero and one to form a probability distribution.

[0091] Calculate the KL divergence between the probability distributions of two modes. The smaller the KL divergence value, the closer the feature distributions of the two modes are, and the better the coupling effect.

[0092] S3: Regional geological disaster classification and early warning based on time-series stability criteria.

[0093] Specifically, the regional geological disaster classification and early warning system based on temporal stability criteria is based on voxel-level disaster identification results. Through temporal sliding window and cumulative deformation threshold logic, the identification results are transformed into early warning signals with clear levels. The specific implementation is as follows:

[0094] For each voxel in the voxel-level disaster identification results, the disaster categories include landslides, debris flows, and ground subsidence.

[0095] Extract the displacement and meteorological features time series within N consecutive time windows. Each time window corresponds to the time alignment unit in S1. N is a positive integer set by the implementers based on the geological disaster response cycle, usually ranging from six to twelve.

[0096] Define a temporal stability criterion, and for each disaster voxel, calculate the ratio of its cumulative displacement in the current time window to the average displacement of the previous M windows. And the ratio of the cumulative rainfall intensity in the current window to the historical average rainfall intensity for the same period. Where M is a positive integer set by the implementers based on the length of historical data, typically ranging from three to five, as detailed below:

[0097] The ratio of the cumulative displacement in the current time window to the average displacement of the previous M windows is:

[0098] ;

[0099] in, This represents the cumulative displacement within the current time window. Let be the cumulative displacement over the first i-th time window. This represents the ratio of the cumulative displacement in the current time window to the average displacement of the previous M windows.

[0100] The ratio of the cumulative rainfall intensity in the current window to the historical average rainfall intensity for the same period is:

[0101] ;

[0102] in, This represents the cumulative rainfall intensity for the current time window. The average rainfall intensity for the same historical time window length. This represents the ratio of the cumulative rainfall intensity in the current window to the historical average rainfall intensity for the same period.

[0103] The tiered early warning system is based on the calculated ratio, specifically as follows:

[0104] Define four levels of early warning thresholds: Attention level (blue warning), Warning level (yellow warning), Alert level (orange warning), and Risk avoidance level (red warning);

[0105] like If so, it is determined to be a blue alert;

[0106] like If so, it is determined to be a yellow alert;

[0107] like If so, it is determined to be an orange alert;

[0108] like If so, it is determined to be a red alert;

[0109] In addition, a voting system is used to count the warning levels of all disaster voxels in the same area. If the proportion of red warning voxels in the area exceeds 30% or the proportion of orange warning voxels exceeds 50%, the overall warning level of the area is raised to the next level below the highest voxel warning level. Red warnings remain red and a regional geological disaster classification warning map with geographical boundaries is output. The regional geological disaster classification warning map adopts GIS layer format, and the corresponding warning color is filled in the boundary of each area and the warning level text is marked.

[0110] It should be noted that a voting and statistical analysis of the early warning levels for all disaster voxels within the same region is used to address the decision-making fusion problem when mapping voxel-level early warning results to regional-level early warning results, as detailed below:

[0111] The study area was divided into multiple independent early warning zones according to administrative divisions or geological hazard-prone zones;

[0112] Within each warning area, the distribution of warning levels for all disaster voxels is statistically analyzed, and the percentage of red warning voxels and orange warning voxels in the total number of disaster voxels in the area is calculated.

[0113] If the proportion of red alert voxels exceeds 30%, or the proportion of orange alert voxels exceeds 50%, the level upgrade logic will be triggered, and the final alert level of the area will be set to the next level below the highest level among all voxel alert levels, but the red alert level will remain unchanged as the highest level.

[0114] S4: Adaptive closed-loop feedback correction driven by early warning results.

[0115] Specifically, the adaptive closed-loop feedback correction driven by early warning results compares the regional geological disaster classification early warning map with the actual disaster data, calculates the accuracy, false alarm rate, and false alarm rate for each early warning level based on the comparison results, and dynamically corrects the feature coupling parameters based on the calculation results. The specific implementation is as follows:

[0116] Data on the location and scale of actual geological disasters are collected within a predetermined time period after each warning is issued, serving as a true feedback label. The predetermined time period is set by the implementers themselves according to the timeliness requirements of geological disaster response, and is usually seventy-two hours.

[0117] The regional geological disaster classification early warning map was spatially overlaid with the actual feedback labels to calculate the accuracy, false negative rate, and false alarm rate for each early warning level, as detailed below:

[0118] Accuracy rate, then:

[0119] ;

[0120] The false negative rate is then:

[0121] ;

[0122] The false alarm rate is:

[0123] ;

[0124] in, Indicates the accuracy of the calculation. This represents the calculated false negative rate. This represents the calculated false alarm rate. This indicates the number of disaster events for which warnings were correctly issued. This indicates the number of disaster events that triggered false warnings. This indicates the number of disaster events that actually occurred without prior warning.

[0125] Based on the calculation results, feature coupling parameters are corrected, specifically as follows:

[0126] Set the lower limit of accuracy tolerance Upper limit of tolerance for underreporting rate And the upper limit of false alarm rate tolerance , It is usually set at 70%. It is usually set at 20%. It is usually set at 25%;

[0127] like If the warning is invalid, it is determined that the current warning is invalid and needs to be corrected. Differential corrections are performed based on the type of invalidity, specifically:

[0128] like If the result is not found, it is determined to be feature coupling overfitting. Return to step S2, use the voxel features of the false positive region and the misreported region as negative samples, use the contrastive learning adjustment strategy to adjust the initialization parameters of the cross-modal attention weight matrix, and increase the regularization coefficient of the attention weight. The increase of the regularization coefficient is 50% of the current coefficient.

[0129] like If the data fusion is insufficient, return to step S1. Using the voxel location of the missed region as the center, within a range of twice the voxel size, forcibly reduce the basic voxel granularity in S1 by 50%, and re-execute the structural connectivity verification and all subsequent steps.

[0130] like If the current warning is valid, the voxel granularity strategy of S1, the attention weight parameter of S2, and the temporal criterion threshold of S3 in the current warning process are used as the optimized initial state for the next cycle from S1 to S3, forming a closed-loop positive feedback.

[0131] It should be noted that the contrastive learning adjustment strategy is used to correct the feature coupling overfitting problem, as follows:

[0132] Voxel features are extracted from false alarm and misreport regions to form a negative sample set, and voxel features are extracted from correct warning regions to form a positive sample set.

[0133] Construct a contrastive loss function that reduces the feature distance between positive samples and increases the feature distance between negative samples and positive samples;

[0134] Based on the gradient backpropagation of the contrastive loss function, the initialization parameters of the cross-modal attention weight matrix are updated, thereby optimizing feature coupling.

[0135] Furthermore, if the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0137] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0138] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of protection claimed by the present invention.

Claims

1. A regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data, characterized by: Includes the following steps: Multi-source data of the study area were collected, and the basic voxel granularity was set based on the spatial resolution of each data. A three-dimensional voxel array was constructed, and the data density of the multi-source data in the three-dimensional voxel array was calculated. Adaptive resampling was performed based on the calculated data density. The structural connectivity of the resampled three-dimensional voxel array was verified, and the standardized voxel fusion array was determined. Based on the defined standardized voxel fusion array, the corresponding modal features are extracted, and the attention weight matrix between each pair of modal feature maps is calculated based on the attention mechanism. Coupled features are generated based on the calculated attention weight matrix, and the coupled features are residually connected with the original modal features to construct an enhanced multimodal feature set. Modal consistency verification is performed on the enhanced multimodal feature set, and the voxel-level disaster identification result is determined based on the verification results. A time-sliding window is constructed based on the voxel-level disaster identification results, and the displacement ratio and rainfall intensity ratio are calculated. Multi-level early warning judgments are performed based on the calculation results. At the same time, the voxel early warning results in the region are fused by voting to generate a regional geological disaster graded early warning map. The early warning results are compared with the actual disaster data, and the accuracy rate, missed rate and false alarm rate are calculated. Based on the calculation results, the feature coupling parameters are corrected, and the optimization results are fed back to the aforementioned steps to form a closed-loop update mechanism.

2. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 1, characterized in that: The adaptive resampling process is as follows: The data density within each voxel of the initial 3D voxel array for the remote sensing image data layer, surface displacement data layer, meteorological data layer, and geological data layer in the multi-source dataset is calculated separately, and adaptive resampling is performed based on the calculation results, specifically as follows: If the data density of a data layer within a voxel is less than 30% of the global average density of the corresponding data layer, the granularity of the corresponding voxel in the corresponding dimension is doubled, and the average data density within the expanded voxel is recalculated until the data density within all voxels is not less than 30% of the global average density.

3. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 2, characterized in that: The structural connectivity verification of the resampled three-dimensional voxel array is performed as follows: The resampled 3D voxel array is divided into K connected sub-regions, and one sub-region is randomly selected from the divided sub-regions as the initial region for structural connectivity verification. At the same time, all adjacent sub-regions of the initial region are located. Determine the boundary voxel set of the initial region and the boundary voxel set of the adjacent region, calculate the minimum Euclidean distance between the two boundary voxel sets, and verify the structural connectivity based on the calculation results.

4. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 3, characterized in that: The structural connectivity verification based on the calculation results is as follows: If the minimum Euclidean distance between two boundary voxel sets is zero or the total length of the voxel connectivity path within adjacent regions, it indicates that the data structure between the initial region and the adjacent region is connected; otherwise, it indicates that the data structure between the two regions is not connected. Using adjacent regions as the initial regions for the next generation, we continue to calculate the minimum Euclidean distance between the initial regions and their adjacent regions until all sub-regions have been traversed. We count the number of structural disconnections between all pairs of regions and verify the overall structural connectivity based on the number of disconnections. Then: Set the threshold for the number of times the structure is disconnected as follows: If the number of disconnections counted during the verification of the 3D voxel array is greater than If the verification fails, the basic voxel granularity is reduced to 80% of the current granularity, and adaptive resampling and structural connectivity verification are re-executed until the verification passes. If the number of disconnections is less than or equal to If the structure of the current three-dimensional voxel array is verified, the current three-dimensional voxel array will be output as a normalized voxel fusion array.

5. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 4, characterized in that: The modal consistency verification of the enhanced multimodal feature set is performed as follows: Set the difference threshold The KL divergence of all modal feature pairs is greater than 1. If the logarithm is greater than the preset maximum inconsistency logarithm threshold, then... If the modal coupling is insufficient and there is a feature conflict, based on the spatial distribution of the current coupled features, the voxel granularity corresponding to the conflict region is further refined by 20% in the adaptive resampling, and the normalized voxel fusion array is reconstructed. Conversely, if the logarithm is less than or equal to If the modal coupling verification is successful, the enhanced multimodal feature set will be input into a pre-trained classifier, which will output the geological hazard identification results for each voxel, including four categories: landslide, debris flow, ground collapse, and no hazard.

6. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 5, characterized in that: The calculated displacement ratio and rainfall intensity ratio are as follows: For each voxel of the disaster category in the voxel-level disaster identification results, including landslides, debris flows and ground collapses, the displacement features and meteorological features of each voxel of the disaster category are extracted in time series sequences within N consecutive time windows. For each disaster voxel, calculate the ratio of its cumulative displacement in the current time window to the average displacement over the previous M windows. And the ratio of the cumulative rainfall intensity in the current window to the historical average rainfall intensity for the same period. .

7. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 6, characterized in that: The multi-level early warning determination based on the calculation results is as follows: Define four levels of early warning thresholds: Attention level (blue warning), Warning level (yellow warning), Alert level (orange warning), and Risk avoidance level (red warning); like If so, it is determined to be a blue alert; like If so, it is determined to be a yellow alert; like If so, it is determined to be an orange alert; like If so, it is determined to be a red alert.

8. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 7, characterized in that: The feature coupling parameter correction based on the calculation results is performed as follows: Calculate the accuracy rate for each warning level. underreporting rate and false alarm rate And set a lower limit for accuracy tolerance. Upper limit of tolerance for underreporting rate And the upper limit of false alarm rate tolerance ; like If the current warning fails, the feature coupling parameter correction is performed according to the failure type. like If the current warning is valid, the voxel granularity strategy, attention weight parameters, and temporal criterion thresholds in the current warning process will be used as the optimized initial state for the next cycle of geological disaster warning.

9. The regional geological disaster intelligent identification and hierarchical early warning method based on deep fusion of multi-source data as described in claim 8, characterized in that: The feature coupling parameter correction based on the failure type is performed as follows: like If the error is not found, it is determined to be feature coupling overfitting. The voxel features of the false positive and false positive regions are used as negative samples. The initialization parameters of the cross-modal attention weight matrix are adjusted using a contrastive learning adjustment strategy, and the regularization coefficient of the attention weight is increased. like If the data fusion is insufficient, the base voxel granularity will be reduced by 50% within a range of twice the number of voxels around the location of the missing voxel region, and the structural connectivity verification and all subsequent steps will be re-executed.

Images