Regional disaster risk assessment method and device based on deep neural network, equipment and medium

Abstract

This application discloses a method, apparatus, equipment, and medium for regional disaster risk assessment based on deep neural networks, relating to the fields of artificial intelligence and deep learning model technology. The method includes: dividing multi-source disaster monitoring data of a target area into grids and synchronizing them over time to obtain a disaster isomorphic tensor sequence; inputting this data into a disaster situational awareness deep model to extract initial latent disaster features; calculating the sensitivity gradient of the latent features to determine the main controlling factors of the disaster, and extracting evolution extreme values ​​to obtain disaster evolution extreme value features; generating a vulnerability tensor and critical stable state impact data based on the distribution of disaster-bearing bodies and geological environment data; generating an outbreak probability matrix based on the impact data, main controlling factors, and extreme value features; and inputting the probability matrix, vulnerability tensor, and impact data into a risk assessment deep model for spatial topological fusion and feature mapping to obtain the disaster risk assessment result. This application utilizes neural networks to analyze monitoring data and the dynamic characteristics of disaster-bearing bodies, achieving a quantitative assessment of regional disaster risk.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence and deep learning model technology, and in particular to regional disaster risk assessment methods, devices, equipment and media based on deep neural networks. Background Technology

[0002] With the increasing frequency of regional natural disasters, there is an urgent need in the field of disaster prevention and mitigation engineering to dynamically and quantitatively assess the disaster risks of target areas. To address this need, current practices often involve using geographic information systems (GIS) combined with the analytic hierarchy process (AHP) to perform multi-criteria weighted superposition of static geological factors; or employing long short-term memory networks (LSTM) to predict the evolutionary trends of one-dimensional time-series data such as surface displacement; in addition, basic convolutional neural networks are sometimes introduced to extract spatial features from remote sensing images for static identification of potential disaster sites.

[0003] However, existing methods have significant limitations when processing massive amounts of heterogeneous disaster data. First, shallow networks lack the ability to mine high-dimensional spatiotemporal nonlinear features, making it difficult to achieve deep mapping of spatiotemporal characteristics. Second, existing multi-source data processing models lack feature attribution and sensitivity calculation mechanisms, making it difficult to effectively extract the main controlling factors and extreme values ​​of disaster evolution. Furthermore, existing technologies often treat geological stability and the vulnerability of disaster-bearing bodies as isolated factors, failing to achieve cross-integration at the spatial topological level, and thus making it difficult to characterize complex disaster risks. Therefore, how to achieve quantitative assessment of regional disaster risk by analyzing multi-source disaster monitoring data and the dynamic characteristics of disaster-bearing bodies has become an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this application is to provide a method, device, equipment and medium for regional disaster risk assessment based on deep neural networks, which aims to solve the technical problem of how to achieve quantitative assessment of regional disaster risk by analyzing multi-source disaster monitoring data and dynamic characteristics of disaster-bearing bodies.

[0005] To achieve the above objectives, this application proposes a regional disaster risk assessment method based on deep neural networks, the method comprising: The multi-source disaster monitoring data of the target area are divided into grids and time-synchronized to obtain a disaster isomorphic tensor sequence. The disaster isomorphic tensor sequence is input into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features; The feature sensitivity gradient corresponding to the initial hidden features of the disaster situation is calculated by the disaster situation awareness depth model to determine the main controlling factors of the disaster occurrence, and the evolution extreme value is extracted according to the initial hidden features of the disaster to obtain the disaster evolution extreme value feature; Based on the distribution data of disaster-bearing bodies in the target area, a vulnerability tensor of disaster-bearing bodies is generated, and based on the geological environment data of the target area, critical steady-state influence data is generated. Based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution, a disaster outbreak probability matrix is ​​generated; The disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data are input into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results.

[0006] In one embodiment, the disaster situation awareness deep model includes a spatiotemporal coding network, which includes a three-dimensional convolutional layer, a long short-term memory network layer, and a self-attention mechanism layer. The step of inputting the disaster isomorphic tensor sequence into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features includes: The disaster isomorphic tensor sequence is input into the three-dimensional convolutional layer, and local spatial features are extracted in the grid space dimension to obtain a spatial local feature map; The spatial local feature map is input into the long short-term memory network layer to extract temporal evolution features in the time dimension; The attention weight matrix between the spatial local feature map and the temporal evolution feature is calculated through the self-attention mechanism layer; The spatial local feature map and the temporal evolution feature are weighted and fused according to the attention weight matrix to obtain the initial disaster hidden feature.

[0007] In one embodiment, the disaster situation awareness deep model includes a feature attribution network and a situation evolution network, wherein the situation evolution network includes a multi-scale temporal convolutional network and a multilayer perceptron; The steps of calculating the feature sensitivity gradient corresponding to the initial hidden disaster features through the disaster situation awareness depth model to determine the main controlling factors of disaster occurrence, and extracting evolution extreme values ​​based on the initial hidden disaster features to obtain disaster evolution extreme value features include: The initial disaster latent features are input into the feature attribution network, the partial derivatives of the initial disaster latent features with respect to the input data of each channel in the disaster isomorphic tensor sequence are calculated, and global average pooling is performed to obtain the feature sensitivity gradient corresponding to each channel data. The disaster monitoring features corresponding to channel data whose feature sensitivity gradient is greater than a preset contribution threshold are taken as the main controlling factors of disaster occurrence; The initial disaster latent features are input into the multi-scale temporal convolutional network for temporal dimension expansion to obtain a multi-scale evolution feature map; The multi-scale evolution feature map is nonlinearly mapped using the multi-layer perceptron to obtain a disaster evolution probability distribution sequence. Local extreme points with probability values ​​greater than a preset probability confidence interval are selected from the disaster evolution probability distribution sequence as the initial evolution extreme value set; The global maximum value is selected from the initial set of evolutionary extreme values ​​to obtain the disaster evolution extreme value characteristics.

[0008] In one embodiment, the steps of generating a vulnerability tensor of disaster-bearing bodies based on the distribution data of disaster-bearing bodies in the target area, and generating critical steady-state impact data based on the geological environment data of the target area, include: Obtain building distribution data and infrastructure distribution data of the target area, and map them to a spatial grid with the same spatial grid as the disaster isomorphic tensor sequence to obtain disaster-bearing body distribution data; The disaster-bearing body distribution data is transformed into a disaster-bearing body exposure matrix according to the preset asset exposure mapping rules; Obtain the disaster resistance capacity attenuation coefficient corresponding to each of the spatial grids, and multiply the disaster-bearing body exposure matrix and the disaster resistance capacity attenuation coefficient element by element to obtain the disaster-bearing body vulnerability tensor; By using preset geological stability assessment rules, the geological environment data of the target area are subjected to stability calculation to obtain a geological stability coefficient matrix; The geological stability coefficient matrix is ​​subjected to inverse numerical mapping and normalization to obtain critical steady-state influence data.

[0009] In one embodiment, the step of generating a disaster outbreak probability matrix based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution includes: Extract the channel feature data corresponding to the main controlling factors of disaster occurrence from the disaster isomorphic tensor sequence; After the extreme value features of disaster evolution are expanded by spatial dimension broadcasting, they are concatenated with the channel feature data by channel dimension, and then input into a probability calculation network containing fully connected layers and a sigmoid function for mapping calculation to obtain the original probability matrix of disaster outbreak. Using the critical steady-state impact data as spatial penalty coefficients, the original probability matrix of the disaster outbreak is multiplied element by element to obtain the probability matrix of the disaster outbreak.

[0010] In one embodiment, the risk assessment deep model includes a graph attention network layer, a feature splicing layer, and a fully connected network layer; The step of inputting the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results includes: The vulnerability tensor of the disaster-bearing body and the probability matrix of the disaster outbreak are input into the graph attention network layer for spatial node feature aggregation and topological fusion to obtain a multidimensional environmental vulnerability tensor; The critical stable state influence data and the multidimensional environmental vulnerability tensor are channel-stitched together by the feature stitching layer to obtain a cascaded feature tensor. The risk quantization matrix is ​​obtained by performing dimensionality reduction and linear mapping on the cascaded feature tensor through the fully connected network layer. The value range of the risk quantification matrix is ​​mapped to a preset risk level range to obtain the regional disaster risk assessment results.

[0011] In one embodiment, the step of performing grid division and time synchronization processing on multi-source disaster monitoring data of the target area to obtain a disaster isomorphic tensor sequence includes: Acquire the spatial boundary information of the target area, and divide the spatial boundary information into grids according to a preset spatial resolution to obtain the target spatial grid matrix; Multi-source disaster monitoring data of the target area are mapped to the grid nodes corresponding to the target spatial grid matrix to obtain spatially discrete monitoring data; The spatial discrete monitoring data is resampled and linearly interpolated and aligned in the time dimension according to a preset time step to obtain spatiotemporally aligned monitoring data. The spatiotemporal aligned monitoring data is divided into feature channels according to different disaster monitoring data types to obtain multi-channel spatial feature slices; The multi-channel spatial feature slices are spliced ​​together according to the temporal order of occurrence to obtain a disaster isomorphic tensor sequence.

[0012] Furthermore, to achieve the above objectives, this application also proposes a regional disaster risk assessment device based on deep neural networks, the device comprising: The data sequence construction module is used to perform grid division and time synchronization processing on multi-source disaster monitoring data of the target area to obtain a disaster isomorphic tensor sequence. The feature extraction module is used to input the disaster isomorphic tensor sequence into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features; The disaster feature analysis module is used to calculate the feature sensitivity gradient corresponding to the initial hidden disaster features through the disaster situation awareness depth model to determine the main controlling factors of disaster occurrence, and to extract the evolution extreme values ​​based on the initial hidden disaster features to obtain the disaster evolution extreme value features; The disaster-bearing body feature generation module is used to generate a disaster-bearing body vulnerability tensor based on the disaster-bearing body distribution data of the target area, and to generate critical steady-state influence data based on the geological environment data of the target area. The probability calculation module is used to generate a disaster outbreak probability matrix based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution; The risk assessment module is used to input the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results.

[0013] In addition, to achieve the above objectives, this application also proposes a regional disaster risk assessment device based on a deep neural network. The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the regional disaster risk assessment method based on a deep neural network as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the regional disaster risk assessment method based on deep neural networks as described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the regional disaster risk assessment method based on deep neural networks as described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: By gridding and time-synchronized processing of multi-source disaster monitoring data in the target area, a disaster isomorphic tensor sequence was obtained, unifying the spatiotemporal scale of heterogeneous monitoring data and providing a standardized data input format for deep network models. Subsequently, the disaster isomorphic tensor sequence was input into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features, uncovering the spatiotemporal nonlinear correlation patterns hidden in complex geological environments. Next, the feature sensitivity gradient corresponding to the initial disaster latent features was calculated using the disaster situation awareness deep model to determine the main controlling factors of disaster occurrence, and evolutionary extreme values ​​were extracted based on the initial disaster latent features to obtain disaster evolution extreme value features, quantifying the induced contribution of each environmental factor and clarifying the critical boundary of physical abrupt changes in disaster state. Simultaneously, a disaster-bearing body vulnerability tensor was generated based on the distribution data of disaster-bearing bodies in the target area, and based on the target... Regional geological environmental data is used to generate critical steady-state impact data, which calculates the vulnerability of regional physical assets and establishes the spatial constraint penalty basis for the stable state of the geological base. Furthermore, a disaster outbreak probability matrix is ​​generated based on the critical steady-state impact data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution. This combines static geological penalties with dynamic evolution characteristics, improving the numerical rationality of disaster occurrence probability prediction. Finally, the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data are input into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results. This application comprehensively considers the spatial cascading amplification effect between environmental vulnerability and asset exposure, and ultimately enables a quantitative assessment of regional disaster risk by analyzing multi-source disaster monitoring data and the dynamic characteristics of the disaster-bearing body. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the regional disaster risk assessment method based on deep neural networks provided in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the regional disaster risk assessment method based on deep neural networks provided in this application; Figure 3 This is a schematic diagram of the module structure of the regional disaster risk assessment device based on deep neural networks according to an embodiment of this application; Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the regional disaster risk assessment method based on deep neural networks in the embodiments of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or disaster risk assessment system capable of realizing the above functions. The following uses a disaster risk assessment system as an example to describe this embodiment and the subsequent embodiments.

[0024] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0025] Based on this, embodiments of this application provide a regional disaster risk assessment method based on deep neural networks, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the regional disaster risk assessment method based on deep neural networks in this application.

[0026] In this embodiment, the regional disaster risk assessment method based on deep neural networks includes steps S10 to S60: Step S10: Perform grid division and time synchronization processing on the multi-source disaster monitoring data of the target area to obtain the disaster isomorphic tensor sequence; It should be noted that the target area refers to a specific geographic spatial range where disaster risk situation awareness and quantitative assessment are required. Multi-source disaster monitoring data can include stratigraphic lithology data, topographic slope data, and soil moisture content data. Disaster isomorphic tensor sequences refer to multi-dimensional feature matrices formed by unifying data of different dimensions and sampling frequencies into the same spatiotemporal coordinate system.

[0027] Understandably, the process involves acquiring the latitude and longitude spatial boundary information of the target area and dividing it into grids according to a preset spatial resolution. The preset spatial resolution is based on the base resolution of high-resolution remote sensing imagery (e.g., 30m x 30m) to obtain the target spatial grid matrix. A cubic spline interpolation algorithm is then used to align monitoring data from different sampling rates to the same preset time step, which can be set to 1 hour. The spatiotemporally aligned monitoring data is then divided into channel dimensions according to different types and stitched together in chronological order to form a multidimensional tensor sequence containing time, spatial dimensions, and feature channels. This step overcomes the technical obstacle of inconsistencies in spatial and temporal scales between multi-source heterogeneous monitoring data, providing a standardized data input format for subsequent deep network model processing.

[0028] Step S20: Input the disaster isomorphic tensor sequence into the pre-trained disaster situation awareness deep model for feature extraction to obtain the initial disaster latent features; It should be noted that the deep disaster situation awareness model can be a deep learning feature extraction architecture that includes three-dimensional convolutional layers, long short-term memory network layers, and self-attention mechanism layers. The initial disaster latent features refer to high-dimensional dense vectors extracted from multi-dimensional spatiotemporal sequences that characterize the potential evolutionary patterns of disasters.

[0029] Understandably, the isomorphic disaster tensor sequence is input into the 3D convolutional layer of the disaster situation awareness deep model. Local geometric features are extracted in the grid space dimension to obtain a spatial local feature map. This spatial local feature map is then input into a long short-term memory network layer to calculate temporal dependencies in the time dimension to obtain temporal evolution features. Next, an attention weight matrix is ​​calculated between the spatial local feature map and the temporal evolution features through a self-attention mechanism layer. Based on this attention weight matrix, the two types of features are weighted and fused. This step uncovers the deep nonlinear correlation patterns of multi-source geological environmental data in spatial topology and time series, obtaining feature expressions that objectively reflect the characteristics of geological evolution.

[0030] Step S30: Calculate the feature sensitivity gradient corresponding to the initial hidden disaster features using the disaster situation awareness depth model to determine the main controlling factors of disaster occurrence, and extract evolution extreme values ​​based on the initial hidden disaster features to obtain disaster evolution extreme value features; It should be noted that the feature sensitivity gradient refers to the mathematical partial derivative matrix of the network output state with respect to the input channel features. The controlling factor refers to the specific monitoring feature channel that plays a major inducing or controlling role in the formation of regional disasters. Disaster evolution extreme value features can be high-dimensional hidden layer feature vectors that characterize physical abrupt changes in the disaster's evolution state.

[0031] Understandably, the integral gradient algorithm is used to calculate the partial derivatives of the initial disaster latent features with respect to each input channel of the multi-source disaster monitoring data. Channels with absolute values ​​of feature sensitivity gradients greater than a preset contribution threshold are selected as the corresponding monitoring data types and used as the main control factors. The preset contribution threshold can be set to a range of 0.75 to 0.85 to filter out the influence of conventional environmental background noise. Simultaneously, the initial disaster latent features are input into a multi-scale temporal convolutional network for feature expansion. A disaster evolution probability distribution sequence is obtained through nonlinear mapping using a multilayer perceptron. Local extreme points with probability values ​​greater than a preset probability confidence interval are selected as the initial evolution extreme value set. The preset probability confidence interval is [0.85, 1.0]. Finally, the global maximum value is selected from this set. This step quantifies the relative contribution of different environmental monitoring factors to the disaster outbreak, clarifies the influence weights of the input multi-source disaster monitoring data on the model output state, and defines the critical numerical conditions for sudden changes in disaster state.

[0032] Step S40: Generate a vulnerability tensor of disaster-bearing bodies based on the distribution data of disaster-bearing bodies in the target area, and generate critical steady-state influence data based on the geological environment data of the target area; It should be noted that the data on the distribution of disaster-bearing bodies can include population density distribution data, building distribution data, and critical infrastructure distribution data within the target area. The vulnerability tensor of disaster-bearing bodies refers to a quantitative matrix characterizing the potential damage level of various disaster-bearing bodies when subjected to geological disasters. Critical steady-state impact data can be standardized environmental indicators reflecting the inherent resilience of the regional geological base.

[0033] Understandably, the distribution data of various disaster-bearing bodies are mapped to the same spatial grid. Based on a pre-defined asset exposure mapping rule, this data is transformed into a disaster-bearing body exposure matrix. The disaster resistance capacity attenuation coefficient corresponding to each spatial grid is obtained, and the disaster-bearing body exposure matrix is ​​multiplied element-wise by the disaster resistance capacity attenuation coefficient. The disaster resistance capacity attenuation coefficient ranges from 0 to 1 and is set according to the seismic fortification standards for buildings. Additionally, a geological stability coefficient matrix is ​​obtained by performing an exponential calculation on the geological environment data using a single-factor evaluation function. This geological stability coefficient matrix is ​​then subjected to inverse numerical mapping and maximum-minimum normalization. This step quantifies the vulnerability of various physical assets within the region and the stability of the basic geological environment, providing necessary environmental base penalty constraints for subsequent comprehensive disaster probability calculations.

[0034] Step S50: Generate a disaster outbreak probability matrix based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution; It should be noted that the disaster outbreak probability matrix refers to the set of predicted probability values ​​of the corresponding disaster occurring at each spatial grid node in the future time period.

[0035] Understandably, the process involves extracting channel feature data from the isomorphic tensor sequence of disasters, where the controlling factors reside. This channel feature data, along with the extreme value features of disaster evolution, is then input into a probability prediction branch containing a fully connected layer and a sigmoid function to calculate the original probability matrix of the disaster outbreak. Finally, the critical steady-state influence data is used as a spatial constraint penalty coefficient to multiply the original probability matrix of the disaster outbreak element-wise. This step effectively combines static geological penalty constraint coefficients with dynamically evolving characteristic states, improving the reliability and numerical stability of the disaster outbreak probability prediction results.

[0036] Step S60: Input the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data into the pre-trained risk assessment deep model, perform spatial topology fusion and feature mapping, and obtain the regional disaster risk assessment results.

[0037] It should be noted that the deep risk assessment model can be a graph neural network model containing graph attention network layers, feature splicing layers, and fully connected network layers. The regional disaster risk assessment result can be a spatial thermal distribution map that divides different quantitative risk levels.

[0038] Understandably, the vulnerability tensor of the disaster-bearing body and the disaster outbreak probability matrix are input into the graph attention network layer of the risk assessment deep model to aggregate spatially adjacent node features to obtain a multidimensional environmental vulnerability tensor. A feature splicing layer then concatenates the critical steady-state impact data and the multidimensional environmental vulnerability tensor to form a cascaded feature tensor. A fully connected network layer then performs dimensionality reduction on the cascaded feature tensor to obtain a risk quantification matrix. Finally, the values ​​of the risk quantification matrix are mapped to a preset risk level range, which can be divided into four gradient levels based on geological disaster reduction standards. This step comprehensively considers the spatial cascading amplification effect between environmental vulnerability, physical asset vulnerability, and disaster outbreak probability, outputting a quantitative assessment result of regional disaster risk.

[0039] This embodiment constructs a homogeneous tensor sequence by unifying the spatiotemporal dimensions of multi-source heterogeneous environmental monitoring data. It utilizes a deep network architecture to mine deep spatiotemporal evolution characteristics and combines an integral gradient attribution algorithm to extract the main controlling factors and critical thresholds for disaster occurrence. Simultaneously, it introduces the geological base stability coefficient and the vulnerability index of the disaster-bearing body for graph attention space topological fusion. This technical solution overcomes the technical shortcomings of conventional mathematical statistical models in mining high-dimensional spatiotemporal nonlinear features and the lack of physical attribution explanations for shallow network results. It quantifies the regional compound disaster risk under multi-factor coupling, improving the objectivity and reliability of geological disaster prevention, mitigation, and graded early warning decisions.

[0040] As an example, the step of performing grid partitioning and time synchronization processing on multi-source disaster monitoring data of the target area to obtain a disaster isomorphic tensor sequence includes: acquiring spatial boundary information of the target area, and dividing the spatial boundary information into grids according to a preset spatial resolution to obtain a target spatial grid matrix; mapping the multi-source disaster monitoring data of the target area to the grid nodes corresponding to the target spatial grid matrix to obtain spatial discrete monitoring data; resampling and linear interpolation alignment of the spatial discrete monitoring data in the time dimension according to a preset time step to obtain spatiotemporally aligned monitoring data; dividing the spatiotemporally aligned monitoring data into feature channels according to different disaster monitoring data types to obtain multi-channel spatial feature slices; and splicing the multi-channel spatial feature slices in tensor dimension according to the time occurrence order to obtain a disaster isomorphic tensor sequence.

[0041] It should be noted that, in this example, spatial boundary information refers to the extreme range of latitude and longitude coordinates of the target area. The target spatial grid matrix can be a two-dimensional planar coordinate node array formed after dividing the geographic space. Spatially discrete monitoring data refers to physical monitoring values ​​that are scattered across various grid nodes and not aligned in time. Spatiotemporally aligned monitoring data can be a continuous feature matrix with a uniform sampling frequency on the time axis after interpolation processing. Multi-channel spatial feature slices refer to a set of two-dimensional spatial feature layers containing multiple monitoring data types at a single time point.

[0042] Understandably, the process begins with acquiring the spatial boundary information of the target area. This spatial boundary information is then divided into grids according to a preset spatial resolution, resulting in a target spatial grid matrix. The preset spatial resolution can be set to 30 meters × 30 meters based on the basic accuracy requirements of the remote sensing imagery, balancing spatial representation granularity with equipment computing resources. Multi-source disaster monitoring data for the target area is then mapped to the corresponding grid nodes in the target spatial grid matrix based on geographic coordinates, yielding spatially discrete monitoring data.

[0043] Next, the spatially discrete monitoring data is resampled along the temporal dimension according to a preset time step. The preset time step is typically set to 1 hour or 2 hours based on the regular observation frequency of geological disasters. A linear interpolation algorithm is used to numerically estimate and align the missing data nodes from the resampling process, resulting in spatiotemporally aligned monitoring data.

[0044] Spatiotemporal aligned monitoring data is divided into feature channels according to different disaster monitoring data types, so that different disaster monitoring data occupy independent channels, resulting in multi-channel spatial feature slices. The multi-channel spatial feature slices are then concatenated according to the temporal order of occurrence, resulting in a four-dimensional matrix structure containing the time dimension, channel dimension, and two-dimensional spatial dimension, thereby outputting a disaster isomorphic tensor sequence.

[0045] This example overcomes the technical obstacles of scattered spatial distribution and inconsistent temporal sampling rates of multi-source monitoring data by dividing the spatial boundary into grids and aligning the time steps using a linear interpolation algorithm. This process unifies the dimensions and specifications of complex geological monitoring data, constructs a four-dimensional tensor sequence with strict spatiotemporal correspondence, lays a standardized data foundation for deep network models to mine spatial topology and temporal evolution features, and improves the numerical stability of the overall model inference process.

[0046] As an example, the steps of generating a disaster-bearing body vulnerability tensor based on the disaster-bearing body distribution data of the target area and generating critical stable state impact data based on the geological environment data of the target area include: acquiring building distribution data and infrastructure distribution data of the target area, and mapping them to a spatial grid with the same spatial grid as the disaster isomorphic tensor sequence to obtain disaster-bearing body distribution data; converting the disaster-bearing body distribution data into a disaster-bearing body exposure matrix according to a preset asset exposure mapping rule; acquiring the disaster resistance capacity attenuation coefficient corresponding to each spatial grid, and multiplying the disaster-bearing body exposure matrix and the disaster resistance capacity attenuation coefficient element-wise to obtain a disaster-bearing body vulnerability tensor; performing stability calculation on the geological environment data of the target area through a preset geological stability assessment rule to obtain a geological stability coefficient matrix; and performing reverse numerical mapping and normalization processing on the geological stability coefficient matrix to obtain critical stable state impact data.

[0047] It should be noted that, in this example, the disaster-bearing body distribution data refers to the collection of spatial location and density information of population, buildings, and infrastructure potentially threatened by geological disasters within the target area. The preset asset exposure mapping rule can be a mapping function that transforms physical entities into dimensionless exposure weights, set according to building type, spatial distribution density, and infrastructure fortification level. The disaster-bearing body exposure matrix is ​​a feature matrix that quantifies the distribution density and structural importance of disaster-bearing bodies at each node on a two-dimensional spatial grid. The disaster resistance attenuation coefficient can be a weighted penalty parameter reflecting the decrease in the disaster-bearing body's ability to resist disaster impacts due to structural aging or environmental corrosion. The geological stability coefficient matrix is ​​a two-dimensional numerical tensor that quantifies the stability of the physical foundation of each grid node in the target area.

[0048] Understandably, the process begins by acquiring population density distribution data, building distribution data, and infrastructure distribution data for the target area. A geographic coordinate alignment algorithm is then used to map these three types of data to a spatial grid identical to the disaster isomorphic tensor sequence obtained in the previous steps, yielding disaster-bearing body distribution data. This data is then transformed into a disaster-bearing body exposure matrix according to a pre-defined asset exposure mapping rule. Finally, the service life of buildings within each spatial grid is obtained, and a pre-defined negative exponential decay function is used to calculate the disaster resistance capacity attenuation coefficient corresponding to each spatial grid. The negative exponential decay function is as follows: in, This is the disaster resilience attenuation coefficient. To preset the material aging constant, The service life is used as the basis for calculating the vulnerability tensor of the disaster-bearing body by weighting the exposure matrix of the disaster-bearing body using the disaster resistance capacity attenuation coefficient.

[0049] Next, the stability of the geological environment data of the target area is calculated using preset geological stability assessment rules. The analytic hierarchy process (AHP) combined with geological engineering experience evaluation functions is used to perform exponential mapping and weighted summation of various characteristic parameters in the geological environment data, outputting a comprehensive stability rating for each grid node, thus obtaining the geological stability coefficient matrix.

[0050] The geological stability coefficient matrix is ​​numerically inverted and normalized. Since the more stable the geological base, the smaller its promoting effect on disaster outbreaks should be, the inverse numerical mapping rule is used to calculate the values ​​of each grid node. Then, the max-min normalization algorithm is used to proportionally compress the mapped values ​​to the range of 0 to 1 to obtain the critical steady-state influence data.

[0051] This example objectively quantifies the vulnerability of disaster-bearing bodies in a target area to disaster impacts by mapping the spatial distribution of population and building assets to a unified grid and combining this with the conversion of asset economic value and disaster resilience decay characteristics. Simultaneously, stability calculations and reciprocal mappings are performed on geological environmental data, transforming the macroscopic geological stability state into spatial penalty parameters that conform to probabilistic calculation logic. This step establishes a quantitative mathematical correlation between physical asset distribution and the environmental base state, providing a unified constraint for subsequent analysis of the cascading amplification effect of disaster risks, and improving the objectivity and practical guiding value of regional disaster risk assessment results.

[0052] As an example, the step of performing stability calculations on the geological environment data using preset geological stability assessment rules to obtain a geological stability coefficient matrix includes: extracting topographic slope data, stratigraphic lithology data, and soil moisture content data from the geological environment data; performing index mapping calculations on the topographic slope data, stratigraphic lithology data, and soil moisture content data respectively according to the single-factor evaluation function in the preset geological stability assessment rules to obtain a single-factor stability index matrix corresponding to each data; obtaining the evaluation weight parameters corresponding to each single-factor stability index matrix in the preset geological stability assessment rules; and performing a weighted summation calculation on each single-factor stability index matrix according to the evaluation weight parameters to obtain the geological stability coefficient matrix.

[0053] It should be noted that, in this example, terrain slope data refers to the set of tilt angle characteristics of surface units within the target area relative to the horizontal plane. Stratigraphic lithology data can be geological survey classification values ​​characterizing the physical and mechanical properties and strength of rocks within the area. Soil moisture content data refers to the spatial distribution matrix of the percentage of water volume or mass in the shallow soil of the area. The single-factor evaluation function can be a mathematical mapping relationship set for a single physical characteristic variable, converting its original dimensions into a dimensionless safety interval. The evaluation weight parameter refers to the coefficient reflecting the proportion of the influence of different physical factors on the overall geological structure stability.

[0054] Understandably, the first step is to extract topographic slope data, stratigraphic lithology data, and soil moisture content data from geological environmental data. In practice, this is achieved by separating multi-channel environmental raster files from the geographic information system and extracting data layers containing surface slope, soil and rock type, and moisture saturation, thus forming spatially grid-aligned topographic slope data, stratigraphic lithology data, and soil moisture content data.

[0055] Next, based on the single-factor evaluation function in the preset geological stability assessment rules, index mapping calculations were performed on topographic slope data, stratigraphic lithology data, and soil moisture content data. For the continuously varying slope and moisture content parameters, an inverse proportional decay evaluation function was used to nonlinearly map their original observed values ​​into stability indices ranging from 0 to 1. The closer the value is to 1, the more stable the geological state under that single factor. For the discretely distributed stratigraphic lithology data, corresponding mechanical constants were assigned based on the empirical table of uniaxial compressive strength of rocks and normalized, thereby outputting a single-factor stability index matrix corresponding to each physical characteristic independently.

[0056] Obtain the evaluation weight parameters for each single-factor stability index matrix in the preset geological stability assessment rules. The values ​​of these evaluation weight parameters are based on the statistical induction frequency of various environmental factors in historical geological landslides or collapses in the region. For example, the weights for topographic slope, soil moisture content, and stratigraphic lithology can be set to 0.4, 0.35, and 0.25, respectively. Based on the evaluation weight parameters, perform a weighted summation of the corresponding grid node values ​​in each single-factor stability index matrix, summarize multiple physical influencing factors, and output a geological stability coefficient matrix representing the comprehensive stability rating of each spatial node.

[0057] This example introduces a single-factor evaluation function to uniformly convert slope, moisture, and lithology monitoring data with different physical dimensions and variation patterns into a dimensionless stability index. This process restores the mechanical mechanism of the nonlinear influence of various physical variables on foundation stability in the real geological environment, eliminates dimensional conflicts and calculation biases caused by directly weighting multi-source heterogeneous raw data, and avoids calculation biases caused by directly weighting data with different dimensions.

[0058] As an example, the step of generating a disaster outbreak probability matrix based on the critical stable state influence data, the main controlling factors of disaster occurrence, and the extreme value features of disaster evolution includes: extracting channel feature data corresponding to the main controlling factors of disaster occurrence from the disaster isomorphic tensor sequence; after spatial dimension broadcasting expansion of the extreme value features of disaster evolution, concatenating them with the channel feature data in terms of channel dimension, and inputting them into a probability calculation network containing fully connected layers and a sigmoid function for mapping calculation to obtain the original probability matrix of disaster outbreak; using the critical stable state influence data as a spatial penalty coefficient, multiplying the original probability matrix of disaster outbreak element by element to obtain the disaster outbreak probability matrix.

[0059] It should be noted that, in this example, channel feature data refers to data slices specifically recording the core physical variables that induce disasters within a multidimensional spatiotemporal tensor structure. The probability computation network can be a branch of a deep neural network used to reduce the dimensionality of high-dimensional abstract features and map them to a probability distribution space. The original disaster outbreak probability matrix refers to the initial disaster occurrence probability distribution grid, derived solely from the controlling physical factors and extreme characteristics of disaster evolution, without constraints from the geological environment. The spatial penalty coefficient can be a weight reduction term used to suppress the theoretical disaster outbreak probability at grid nodes with a relatively stable geographical base. The disaster outbreak probability matrix refers to the final quantified probability set of the occurrence of a real disaster at each spatial grid node within a specific future time period.

[0060] Understandably, firstly, based on the channel dimension labels corresponding to the main controlling factors of disaster occurrence determined in the previous steps, tensor slicing operations are used to separate data layers of specific dimensions from the disaster isomorphic tensor sequence, and channel feature data that specifically characterizes the core inducing variable is extracted.

[0061] After spatial dimensionality broadcasting expansion of the extreme value features of disaster evolution, they are concatenated with channel feature data to construct a joint feature vector containing environmental monitoring variables and evolutionary mutation boundary information. This joint feature vector is then input into a probabilistic computation network containing fully connected layers and a sigmoid activation function. The fully connected layers perform nonlinear dimensionality reduction and feature space transformation on the joint feature vector. Subsequently, the sigmoid activation function is used to transform the dimensionality-reduced output values ​​of each grid node into probability prediction values ​​between 0 and 1, yielding the original probability matrix of the disaster outbreak.

[0062] Furthermore, the critical steady-state influence data generated in the previous calculation is obtained and used as the spatial penalty coefficient. While maintaining the one-to-one correspondence between the spatial coordinates of the two-dimensional grid nodes, the original disaster outbreak probability matrix is ​​multiplied element-wise by the spatial penalty coefficient. The initial grid probability values ​​are attenuated and reduced proportionally using the spatial penalty coefficient, so that the output probability value is lower for regions with more stable geological foundations, ultimately yielding the disaster outbreak probability matrix.

[0063] This example establishes a probabilistic calculation mechanism with dual constraints of dynamic evolutionary characteristics and static environmental base by concatenating the dominant inducing factors of disaster occurrence with the characteristics of evolutionary mutation boundaries, and introducing a spatial penalty coefficient reflecting the objective physical base state for element-by-element multiplication and reduction. This process, by introducing spatial base constraints, suppresses false positive predictions that the model may produce in stable geological regions and reduces the computational bias caused by the lack of physical boundary constraints.

[0064] As an example, the risk assessment deep model includes a graph attention network layer, a feature splicing layer, and a fully connected network layer. The step of inputting the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data into the pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain the regional disaster risk assessment result includes: inputting the vulnerability tensor of the disaster-bearing body and the disaster outbreak probability matrix into the graph attention network layer for spatial node feature aggregation and topology fusion to obtain a multidimensional environmental vulnerability tensor; splicing the critical steady-state impact data and the multidimensional environmental vulnerability tensor through the feature splicing layer to obtain a cascaded feature tensor; performing dimensionality reduction and linear mapping on the cascaded feature tensor through the fully connected network layer to obtain a risk quantization matrix; and mapping the value range of the risk quantization matrix to a preset risk level range to obtain the regional disaster risk assessment result.

[0065] It should be noted that, in this example, the graph attention network layer refers to a graph neural network structure that calculates the feature dependencies between adjacent spatial nodes based on a node attention mechanism. The feature concatenation layer can be a network layer used to concatenate and align feature tensors from different sources or dimensions along the channel dimension. The multidimensional environmental vulnerability tensor refers to a spatial topological feature representation matrix that integrates the tendency of physical assets to be damaged and the probability of potential disasters. The cascaded feature tensor can be a composite data matrix that aggregates environmental vulnerability features and geological foundation stability indicators. The risk quantification matrix refers to a two-dimensional continuous variable matrix that outputs the specific disaster risk assessment scores for each spatial grid node. The preset risk level range can be a risk score grading standard range divided according to the disaster prevention and mitigation emergency response plan.

[0066] Understandably, the vulnerability tensor of the disaster-bearing body and the disaster outbreak probability matrix are first input into the graph attention network layer as node feature sequences. Inside the graph attention network layer, the attention mechanism is used to calculate the association weight between the central grid node and its neighboring grid nodes. Based on this association weight, the vulnerability features and probability features of neighboring nodes are weighted, summed, and spatially aggregated to complete the topological fusion of physical asset status and disaster occurrence probability, outputting a multidimensional environmental vulnerability tensor.

[0067] Next, the critical steady-state impact data and the multidimensional environmental vulnerability tensor generated in the previous steps are directly concatenated and combined along the feature channel dimension through a feature splicing layer. The data layer representing the static penalty index of the geological basement is aligned and merged with the feature layer representing the dynamic risk evolution to obtain a cascaded feature tensor containing multiple environmental constraint attributes.

[0068] Dimensionality reduction and linear mapping calculations are performed on the cascaded feature tensor using a fully connected network layer. The fully connected network layer compresses the high-dimensional cascaded features into single-channel two-dimensional grid data through network weight matrix multiplication, outputting continuous risk scores for each node in the space to obtain the risk quantification matrix. Preset risk level intervals are obtained, which are divided into low-risk intervals [0, 0.25), medium-risk intervals [0.25, 0.6], high-risk intervals [0.6, 0.85], and severe-risk intervals [0.85, 1.0] according to conventional disaster management standards. The values ​​in the risk quantification matrix are compared with the preset risk level intervals and discretized, converting continuous scores into corresponding risk status labels, outputting a regional disaster risk assessment result covering the entire target geographic space.

[0069] This example utilizes the spatial node feature aggregation mechanism of graph attention network layers to capture the topological transmission and cascading amplification effects of disaster risk in adjacent geographic spaces. Through channel feature splicing and linear dimensionality reduction mapping, the static geological environment base constraints and the dynamic vulnerability of disaster-bearing bodies are deeply integrated, ultimately outputting standardized and graded risk quantification results, improving the global spatial coherence of regional disaster risk assessment results and the usability of practical disaster prevention decisions.

[0070] As an example, the training steps of the risk assessment deep model include: acquiring the disaster outbreak probability matrix, the vulnerability tensor samples of disaster-bearing bodies, and the corresponding real risk quantification labels under historical regional disaster scenarios to form a training sample set; inputting the training sample set into the initial risk assessment deep model for forward propagation calculation to obtain the predicted risk assessment result; calculating the network difference loss value between the predicted risk assessment result and the real risk quantification label according to the mean squared error loss function; and performing backpropagation gradient update on the initial risk assessment deep model according to the network difference loss value to obtain the trained risk assessment deep model.

[0071] It should be noted that, in this example, "historical regional disaster scenario" refers to a geographical area and time segment where geological disasters have actually occurred in the past or where there are long-term environmental monitoring records. The true risk quantification label can be an objective hazard score obtained after standardized mapping based on the actual asset losses and physical damage caused by historical disaster events. The initial risk assessment deep model refers to an untrained network architecture where the network topology has been built, but the weight parameters of each layer of nodes have not yet been fitted with historical data. The network difference loss value is a quantitative value that measures the statistical deviation between the model's forward output risk score and the objective true score.

[0072] Understandably, the process begins by acquiring historical disaster probability matrix samples and vulnerability tensor samples of disaster-bearing entities extracted from regional disaster scenarios, and collecting real disaster economic losses and infrastructure damage data at corresponding spatial grid historical nodes. A linear conversion rule is then used to transform the real disaster economic losses and infrastructure damage data into real risk quantification labels with values ​​ranging from 0 to 1. Finally, the aforementioned input feature tensor samples and real risk quantification labels are rigorously aligned and paired in spatial coordinates to form a training sample set for supervised learning of the network.

[0073] The training sample set is divided into batches according to a preset batch size and then input into the initial risk assessment deep model in batches. The data undergoes forward propagation calculations such as internal graph attention aggregation and fully connected dimensionality reduction, outputting the predicted risk assessment results corresponding to each historical scenario grid node. The mean squared error between the predicted risk assessment results and the actual risk quantification labels is calculated using the mean squared error loss function to obtain the network difference loss value.

[0074] Next, the partial derivative gradient of the network difference loss value with respect to the weight matrices of each layer within the model is calculated using the backpropagation algorithm. The Adam optimizer is then used to iteratively update the node parameters of the initial risk assessment deep model based on this gradient. Throughout the model training phase, the optimizer's preset initial learning rate is set to 0.001 to 0.005. This range ensures convergence speed while preventing the loss function from oscillating and deviating from its optimal point near the global optimum. The forward propagation feature calculation and backpropagation parameter update loop is continuously executed until the network difference loss value is below the preset convergence threshold for ten consecutive training epochs, or until the maximum preset number of training epochs (e.g., 500 epochs) is reached. At this point, the parameter iteration update ends, and the network weights are solidified, resulting in the trained risk assessment deep model.

[0075] This example constructs a supervised training sample set by collecting historical real-world disaster loss data. It then utilizes the backpropagation algorithm combined with the Adam optimizer to drive deep network parameter updates, establishing a supervised feedback optimization loop between the risk feature mapping output and objective physical damage records. This training process effectively reduces the numerical bias of high-dimensional environmental feature data during topology fusion and dimensionality reduction, improving the predictive generalization ability and reliability of the quantitative assessment results when facing complex disaster scenarios in unknown areas.

[0076] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the regional disaster risk assessment method based on deep neural networks in this application. The disaster situation awareness deep model includes a spatiotemporal coding network, which includes a three-dimensional convolutional layer, a long short-term memory network layer, and a self-attention mechanism layer. Step S20 of the regional disaster risk assessment method based on deep neural networks includes steps S21 to S24: Step S21: Input the disaster isomorphic tensor sequence into the three-dimensional convolutional layer, extract local spatial features in the grid space dimension, and obtain a spatial local feature map; Step S22: Input the spatial local feature map into the long short-term memory network layer to extract temporal evolution features in the time dimension; Step S23: Calculate the attention weight matrix between the spatial local feature map and the temporal evolution feature through the self-attention mechanism layer; Step S24: The spatial local feature map and the temporal evolution feature are weighted and fused according to the attention weight matrix to obtain the initial disaster hidden features.

[0077] It should be noted that in this example, the 3D convolutional layer refers to a deep network component that uses 3D convolutional kernels to simultaneously perform sliding multiply-accumulate operations in the spatial and temporal dimensions of the geographic grid to extract spatiotemporal local correlation features. The spatial local feature map can be a high-dimensional feature matrix reflecting the differences in the physical distribution of disaster monitoring values ​​between adjacent geographic grid nodes. The Long Short-Term Memory (LSTM) network layer refers to a set of recurrent neural network units containing forget gates, input gates, and output gates to handle long-term sequential dependencies. The temporal evolution features can be feature vectors reflecting the dynamic evolution of geological monitoring values ​​at grid nodes over time. The self-attention mechanism layer refers to a network layer that assigns different weights to focus on key features by calculating the dot product correlation between elements within the input sequence. The attention weight matrix can be a matrix of coefficient values ​​that quantifies the importance of the mutual influence between spatial distribution features and temporal variation features.

[0078] Understandably, the disaster isomorphic tensor sequence generated in the preceding steps is first input into the 3D convolutional layer of the disaster situation awareness deep model. A 3×3×3 3D convolutional kernel is set to cover adjacent geographic grid nodes and continuous time steps, performing sliding window convolution operations in both the grid spatial and temporal dimensions. This operation extracts the spatial topological dependencies of different environmental monitoring data within local areas, outputting a spatial local feature map processed by an activation function.

[0079] Next, based on the extracted spatial distribution information, the local spatial feature map is flattened along the time axis and input into the Long Short-Term Memory (LSTM) network layer. Utilizing the gating unit mechanism within the LTM network layer, the hidden states of historical grids are selectively forgotten and updated in the time dimension, capturing the trend of environmental numerical changes across multiple monitoring time steps, thereby extracting temporal evolution features in the time dimension.

[0080] The spatial local feature map and temporal evolution features are synchronously passed as input vectors to the self-attention mechanism layer. Inside the self-attention mechanism layer, a query matrix, a key matrix, and a value matrix are generated through a linear mapping operation, respectively. The scaled dot product of the query matrix and the key matrix is ​​calculated, and the range is mapped through a normalized exponential function to calculate the attention weight matrix between the spatial local feature map and the temporal evolution features.

[0081] Furthermore, based on the calculated attention weight matrix, the spatial local feature map and temporal evolution features are weighted and fused. This operation amplifies the weights of local feature channels that are sensitive to evolutionary abrupt changes, while suppressing the values ​​of redundant environmental background noise channels, ultimately outputting initial disaster latent features that fuse key spatiotemporal information.

[0082] This embodiment overcomes the technical bottleneck of conventional shallow algorithms, which struggle to simultaneously consider both the local spatial topology and long-term temporal evolution patterns of geological environmental data, through a cascaded architecture of three-dimensional convolutional layers and long short-term memory network layers. Combined with the weight allocation mechanism of the self-attention layer, it achieves dynamic focusing of multi-source monitoring features in the spatiotemporal dimension, effectively highlighting deep variables that play a core driving role in disaster formation, improving the representational ability and information density of initial disaster latent features, and laying a high-dimensional feature foundation for the objective extraction of subsequent controlling factors.

[0083] As an example, the disaster situation awareness deep model includes a feature attribution network and a situation evolution network. The situation evolution network includes a multi-scale temporal convolutional network and a multilayer perceptron. The step of calculating the feature sensitivity gradient corresponding to the initial disaster latent features through the disaster situation awareness deep model to determine the main controlling factors of disaster occurrence, and extracting evolution extreme values ​​based on the initial disaster latent features to obtain disaster evolution extreme value features includes: inputting the initial disaster latent features into the feature attribution network, calculating the partial derivatives of the initial disaster latent features with respect to the input data of each channel in the disaster isomorphic tensor sequence, and performing global average pooling to obtain the values ​​of each channel. The feature sensitivity gradient corresponding to the data; the disaster monitoring features corresponding to the channel data whose feature sensitivity gradient is greater than a preset contribution threshold are taken as the main controlling factors of disaster occurrence; the initial disaster latent features are input into the multi-scale temporal convolutional network for time dimension expansion to obtain a multi-scale evolution feature map; the multilayer perceptron performs nonlinear mapping processing on the multi-scale evolution feature map to obtain a disaster evolution probability distribution sequence; local extreme points with probability values ​​greater than a preset probability confidence interval are selected from the disaster evolution probability distribution sequence as the initial evolution extreme value set; the global maximum value is selected from the initial evolution extreme value set to obtain the disaster evolution extreme value feature.

[0084] It should be noted that, in this example, the feature attribution network refers to an algorithm module that uses backpropagation differentiation to quantify the weights of the input features on the model's output. The situational evolution network can be a branch of a composite neural network used for temporal receptive field expansion and abrupt state prediction of temporal hidden layer features. A multi-scale temporal convolutional network refers to a network layer containing multiple one-dimensional convolutional kernels of different sizes to extract local and global trends in a sequence in parallel. The preset contribution threshold refers to the lower bound parameter of the absolute value of the gradient used to distinguish between the dominant inducing factor and environmental background noise. The preset probability confidence interval refers to the numerical range of probability values ​​defining a high-risk state of disaster occurrence.

[0085] Understandably, the initial latent disaster features are first input into the feature attribution network. An integral gradient algorithm is used to calculate the partial derivatives of the initial latent disaster features with respect to the input data of each channel in the disaster isomorphic tensor sequence. This differentiation process directly outputs the high-dimensional gradient tensor corresponding to each monitoring data channel. Subsequently, global average pooling is performed on the high-dimensional gradient tensor in both spatial and temporal dimensions to obtain the feature sensitivity gradient corresponding to each data channel. The absolute value of this gradient reflects the degree of disturbance and influence of relevant environmental physical quantities on the current hidden layer state. The feature sensitivity gradient is then compared one by one with a preset contribution threshold, which can be set to 0.75 to effectively filter low-gradient noise interference from daily minor climate fluctuations. Channel data with gradient values ​​greater than this threshold are selected, and the disaster monitoring features of the corresponding channels are directly used as the main controlling factors for disaster occurrence.

[0086] Next, the initial disaster latent features are input into the multi-scale temporal convolutional network in the situation evolution network. One-dimensional dilated convolutional kernels with sizes of 3, 5, and 7 are used to perform parallel sliding operations on the feature sequence in the temporal dimension to complete the temporal dimension expansion. The temporal features extracted from different receptive fields are then concatenated into feature layers to obtain the multi-scale evolutionary feature map.

[0087] Multi-layer perceptron is used to perform nonlinear mapping on the multi-scale evolutionary feature map. The Sigmoid activation function is then used to map the data from the high-dimensional feature space to a value space between 0 and 1, resulting in a disaster evolution probability distribution sequence with time-series attributes. A preset probability confidence interval is obtained, which can be set to [0.85, 1.0], representing the critical high-risk interval for physical environment instability and damage. A peak-finding algorithm is used to select local maxima points from the disaster evolution probability distribution sequence whose probability values ​​fall within this interval, forming an initial evolution extreme value set. The global maximum value is then selected from the initial evolution extreme value set and output as the disaster evolution extreme value feature.

[0088] This example utilizes a feature attribution network to calculate the partial derivatives of the input channels, enabling the physical source extraction of the main influencing factors of disasters. By combining multi-scale temporal expansion and extreme value determination mechanisms, it can effectively extract the abrupt critical points of disaster evolution across different time spans, filter out redundant temporal fluctuation interference, and output reliable extreme value features of disaster evolution. This provides boundary conditions with clear physical meaning for the subsequent prediction of outbreak probability, improving the scientific rigor and interpretability of disaster situation awareness results.

[0089] As an example, the training steps of the disaster situation awareness deep model include: acquiring historical disaster spatiotemporal data and corresponding real disaster evolution labels; inputting the historical disaster spatiotemporal data into the initial disaster situation awareness deep model for forward propagation calculation to obtain the predicted evolution state; calculating the data-driven loss value between the predicted evolution state and the real disaster evolution label according to the cross-entropy loss function; calculating the state difference value of the predicted evolution state between adjacent time steps, and using the state difference value as the evolution continuity loss value; performing a weighted summation of the data-driven loss value and the evolution continuity loss value to obtain the joint optimization total loss value; and performing backpropagation gradient update on the initial disaster situation awareness deep model based on the joint optimization total loss value to obtain the trained disaster situation awareness deep model.

[0090] It should be noted that, in this example, historical disaster spatiotemporal data refers to multi-source environmental monitoring numerical sequences recorded from past regional geological disaster events. Real disaster evolution labels can be objective classification matrices calibrated by geological survey experts, reflecting the actual damage state of the geological base at various geographical time steps. Predicted evolution state refers to the probability distribution of disaster occurrence tendencies output by the forward inference of the network architecture before training is complete. Data-driven loss can be a penalty value that quantifies the statistical deviation between the network output probability distribution and the objective actual labels. State difference value refers to the differential gradient of the predicted probabilities output by the network at adjacent time steps along the time axis. Evolutionary continuity loss value refers to a regularization penalty parameter used to constrain the smoothness of the model's evolution in the time dimension. Joint optimization total loss value can be an overall network error evaluation index that comprehensively considers classification confidence and temporal smoothness characteristics.

[0091] Understandably, the first step is to acquire historical disaster spatiotemporal data and corresponding real disaster evolution labels. In practice, multi-channel environmental monitoring sequences corresponding to past landslides and debris flows are extracted from the geological history disaster database, and the actual damage state classification at the corresponding time node is matched as the real disaster evolution label. The two are then used to construct a supervised learning training sample pair.

[0092] Historical disaster spatiotemporal data are input into the initial disaster situation awareness deep model for forward propagation calculations. The data then sequentially passes through the spatiotemporal coding network and the situation evolution network within the network for feature mapping and nonlinear activation, outputting the continuous predicted evolution state of each geographic grid at different time steps.

[0093] Next, the data-driven loss value between the predicted evolution state and the actual disaster evolution label is calculated based on the cross-entropy loss function, thereby evaluating the classification bias of the network's forward propagation output. Simultaneously, the state difference value between adjacent time steps is calculated using a first-order difference algorithm, and this state difference value is used as the evolution continuity loss value.

[0094] Extract preset dynamic weight coefficients and sum the data-driven loss value and the evolutionary continuity loss value using a weighted average to obtain the joint optimization total loss value. The preset dynamic weight coefficients are dynamically adjusted using a cosine annealing strategy according to the model training epochs. In the early stages of training, a higher weight parameter is assigned to the data-driven loss value (e.g., set to 0.8) to improve the model's data fit. As the training epochs increase, the weight parameter of the evolutionary continuity loss value is gradually increased (e.g., adjusted up to a maximum of 0.8) to prioritize ensuring the overall classification confidence of the model.

[0095] Based on the joint optimization total loss value, the backpropagation algorithm is used to calculate the partial derivatives of the loss function with respect to the weight parameters of each layer within the initial disaster situation awareness deep model. A gradient descent optimizer is then used to iteratively update the network node parameters based on these partial derivatives until the joint optimization total loss value converges to a stable level and meets the stopping condition. At this point, the parameter iteration ends, and the trained disaster situation awareness deep model is output.

[0096] This example establishes a dual-perspective supervised optimization constraint by introducing real disaster evolution labels for cross-entropy calculation and constructing an evolutionary continuity penalty term based on state difference values ​​in the time dimension. This training mechanism not only reduces the classification numerical deviation between the model's prediction results and the actual historical occurrence patterns but also constrains the smoothness of the predicted probabilities over time series, preventing drastic jumps in disaster situation predictions within adjacent time steps that do not conform to objective evolutionary patterns. This improves the prediction stability and generalization ability of the disaster situation awareness deep model in complex spatiotemporal environments.

[0097] This application also provides a regional disaster risk assessment device based on deep neural networks. Please refer to... Figure 3 The regional disaster risk assessment device based on deep neural networks includes: The data sequence construction module 10 is used to perform grid division and time synchronization processing on the multi-source disaster monitoring data of the target area to obtain the disaster isomorphic tensor sequence. Feature extraction module 20 is used to input the disaster isomorphic tensor sequence into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features; The disaster feature analysis module 30 is used to calculate the feature sensitivity gradient corresponding to the initial hidden disaster features through the disaster situation awareness depth model to determine the main controlling factors of disaster occurrence, and to extract the evolution extreme value based on the initial hidden disaster features to obtain the disaster evolution extreme value feature; The disaster-bearing body feature generation module 40 is used to generate a disaster-bearing body vulnerability tensor based on the disaster-bearing body distribution data of the target area, and to generate critical steady-state influence data based on the geological environment data of the target area. The probability calculation module 50 is used to generate a disaster outbreak probability matrix based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution; The risk assessment module 60 is used to input the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state influence data into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results.

[0098] The regional disaster risk assessment device based on deep neural networks provided in this application, employing the regional disaster risk assessment method based on deep neural networks in the above embodiments, can solve the technical problem of how to achieve quantitative assessment of regional disaster risk by analyzing multi-source disaster monitoring data and dynamic characteristics of disaster-bearing bodies. Compared with the prior art, the beneficial effects of the regional disaster risk assessment device based on deep neural networks provided in this application are the same as those of the regional disaster risk assessment method based on deep neural networks provided in the above embodiments, and other technical features in the regional disaster risk assessment device based on deep neural networks are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0099] This application provides a regional disaster risk assessment device based on a deep neural network. The regional disaster risk assessment device based on a deep neural network includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the regional disaster risk assessment method based on a deep neural network in the above embodiment 1.

[0100] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a regional disaster risk assessment device based on a deep neural network, suitable for implementing embodiments of this application. The regional disaster risk assessment device based on a deep neural network in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Android Devices), PMPs (Portable Media Players), and vehicle terminals (e.g., vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The regional disaster risk assessment device based on deep neural networks shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0101] like Figure 4As shown, a regional disaster risk assessment device based on deep neural networks may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM 1002 (Read Only Memory) or a program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the regional disaster risk assessment device based on deep neural networks. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touch screens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the deep neural network-based regional disaster risk assessment device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows a deep neural network-based regional disaster risk assessment device with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0102] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0103] The regional disaster risk assessment device based on deep neural networks provided in this application, employing the regional disaster risk assessment method based on deep neural networks described in the above embodiments, can solve the technical problem of how to achieve quantitative assessment of regional disaster risk by analyzing multi-source disaster monitoring data and dynamic characteristics of disaster-bearing bodies. Compared with the prior art, the beneficial effects of the regional disaster risk assessment device based on deep neural networks provided in this application are the same as those of the regional disaster risk assessment method based on deep neural networks provided in the above embodiments, and other technical features of this regional disaster risk assessment device based on deep neural networks are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0104] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0105] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0106] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the regional disaster risk assessment method based on deep neural networks in the above embodiments.

[0107] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0108] The aforementioned computer-readable storage medium may be included in a regional disaster risk assessment device based on a deep neural network; or it may exist independently and not be assembled into a regional disaster risk assessment device based on a deep neural network.

[0109] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a regional disaster risk assessment device based on a deep neural network, the device performs the following: gridding and time synchronization processing on multi-source disaster monitoring data of the target area to obtain a disaster isomorphic tensor sequence; inputs the disaster isomorphic tensor sequence into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features; calculates the feature sensitivity gradient corresponding to the initial disaster latent features through the disaster situation awareness deep model to determine the main controlling factors of disaster occurrence, and extracts evolution extreme values ​​based on the initial disaster latent features to obtain disaster evolution extreme value features; generates a disaster-bearing body vulnerability tensor based on the disaster-bearing body distribution data of the target area, and generates critical stable state influence data based on the geological environment data of the target area; generates a disaster outbreak probability matrix based on the critical stable state influence data, the main controlling factors of disaster occurrence, and the disaster evolution extreme value features; inputs the disaster outbreak probability matrix, the disaster-bearing body vulnerability tensor, and the critical stable state influence data into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results.

[0110] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0111] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0112] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0113] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned regional disaster risk assessment method based on deep neural networks. This solves the technical problem of how to achieve quantitative assessment of regional disaster risk by analyzing multi-source disaster monitoring data and the dynamic characteristics of disaster-bearing bodies. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the regional disaster risk assessment method based on deep neural networks provided in the above embodiments, and will not be repeated here.

[0114] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the regional disaster risk assessment method based on deep neural networks as described above.

[0115] The computer program product provided in this application can solve the technical problem of how to quantitatively assess regional disaster risk by analyzing multi-source disaster monitoring data and the dynamic characteristics of disaster-bearing bodies. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the regional disaster risk assessment method based on deep neural networks provided in the above embodiments, and will not be repeated here.

[0116] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A regional disaster risk assessment method based on deep neural networks, characterized in that, The method includes: The multi-source disaster monitoring data of the target area are divided into grids and time-synchronized to obtain a disaster isomorphic tensor sequence. The disaster isomorphic tensor sequence is input into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features; The feature sensitivity gradient corresponding to the initial hidden features of the disaster situation is calculated by the disaster situation awareness depth model to determine the main controlling factors of disaster occurrence, and the evolution extreme value is extracted based on the initial hidden features of the disaster to obtain the disaster evolution extreme value feature; Based on the distribution data of disaster-bearing bodies in the target area, a vulnerability tensor of disaster-bearing bodies is generated, and based on the geological environment data of the target area, critical steady-state influence data is generated. Based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution, a disaster outbreak probability matrix is ​​generated; The disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data are input into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results.

2. The method as described in claim 1, characterized in that, The disaster situation awareness deep model includes a spatiotemporal coding network, which includes a three-dimensional convolutional layer, a long short-term memory network layer, and a self-attention mechanism layer. The step of inputting the disaster isomorphic tensor sequence into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features includes: The disaster isomorphic tensor sequence is input into the three-dimensional convolutional layer, and local spatial features are extracted in the grid space dimension to obtain a spatial local feature map; The spatial local feature map is input into the long short-term memory network layer to extract temporal evolution features in the time dimension; The attention weight matrix between the spatial local feature map and the temporal evolution feature is calculated through the self-attention mechanism layer; The spatial local feature map and the temporal evolution feature are weighted and fused according to the attention weight matrix to obtain the initial disaster hidden feature.

3. The method as described in claim 1, characterized in that, The disaster situation awareness deep model includes a feature attribution network and a situation evolution network, wherein the situation evolution network includes a multi-scale temporal convolutional network and a multilayer perceptron. The steps of calculating the feature sensitivity gradient corresponding to the initial hidden disaster features through the disaster situation awareness depth model to determine the main controlling factors of disaster occurrence, and extracting evolution extreme values ​​based on the initial hidden disaster features to obtain disaster evolution extreme value features include: The initial disaster latent features are input into the feature attribution network, the partial derivatives of the initial disaster latent features with respect to the input data of each channel in the disaster isomorphic tensor sequence are calculated, and global average pooling is performed to obtain the feature sensitivity gradient corresponding to each channel data. The disaster monitoring features corresponding to channel data whose feature sensitivity gradient is greater than a preset contribution threshold are taken as the main controlling factors of disaster occurrence; The initial disaster latent features are input into the multi-scale temporal convolutional network for temporal dimension expansion to obtain a multi-scale evolution feature map; The multi-scale evolution feature map is nonlinearly mapped using the multi-layer perceptron to obtain a disaster evolution probability distribution sequence. Local extreme points with probability values ​​greater than a preset probability confidence interval are selected from the disaster evolution probability distribution sequence as the initial evolution extreme value set; The global maximum value is selected from the initial set of evolutionary extreme values ​​to obtain the disaster evolution extreme value characteristics.

4. The method as described in claim 1, characterized in that, The steps of generating a vulnerability tensor of disaster-bearing bodies based on the distribution data of disaster-bearing bodies in the target area, and generating critical steady-state impact data based on the geological environment data of the target area, include: Obtain building distribution data and infrastructure distribution data of the target area, and map them to a spatial grid with the same spatial grid as the disaster isomorphic tensor sequence to obtain disaster-bearing body distribution data; The disaster-bearing body distribution data is transformed into a disaster-bearing body exposure matrix according to the preset asset exposure mapping rules; Obtain the disaster resistance capacity attenuation coefficient corresponding to each of the spatial grids, and multiply the disaster-bearing body exposure matrix and the disaster resistance capacity attenuation coefficient element by element to obtain the disaster-bearing body vulnerability tensor; By using preset geological stability assessment rules, the geological environment data of the target area are subjected to stability calculation to obtain a geological stability coefficient matrix; The geological stability coefficient matrix is ​​subjected to inverse numerical mapping and normalization to obtain critical steady-state influence data.

5. The method as described in claim 1, characterized in that, The step of generating a disaster outbreak probability matrix based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution includes: Extract the channel feature data corresponding to the main controlling factors of disaster occurrence from the disaster isomorphic tensor sequence; After the extreme value features of disaster evolution are expanded by spatial dimension broadcasting, they are concatenated with the channel feature data by channel dimension, and then input into a probability calculation network containing fully connected layers and a sigmoid function for mapping calculation to obtain the original probability matrix of disaster outbreak. Using the critical steady-state impact data as spatial penalty coefficients, the original probability matrix of the disaster outbreak is multiplied element by element to obtain the probability matrix of the disaster outbreak.

6. The method as described in claim 1, characterized in that, The risk assessment deep model includes a graph attention network layer, a feature splicing layer, and a fully connected network layer. The step of inputting the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results includes: The vulnerability tensor of the disaster-bearing body and the probability matrix of the disaster outbreak are input into the graph attention network layer for spatial node feature aggregation and topological fusion to obtain a multidimensional environmental vulnerability tensor; The critical stable state influence data and the multidimensional environmental vulnerability tensor are channel-stitched together by the feature stitching layer to obtain a cascaded feature tensor. The risk quantization matrix is ​​obtained by performing dimensionality reduction and linear mapping on the cascaded feature tensor through the fully connected network layer. The value range of the risk quantification matrix is ​​mapped to a preset risk level range to obtain the regional disaster risk assessment results.

7. The method according to any one of claims 1 to 6, characterized in that, The steps of performing grid division and time synchronization processing on multi-source disaster monitoring data of the target area to obtain a disaster isomorphic tensor sequence include: Acquire the spatial boundary information of the target area, and divide the spatial boundary information into grids according to a preset spatial resolution to obtain the target spatial grid matrix; Multi-source disaster monitoring data of the target area are mapped to the grid nodes corresponding to the target spatial grid matrix to obtain spatially discrete monitoring data; The spatial discrete monitoring data is resampled and linearly interpolated and aligned in the time dimension according to a preset time step to obtain spatiotemporally aligned monitoring data. The spatiotemporal aligned monitoring data is divided into feature channels according to different disaster monitoring data types to obtain multi-channel spatial feature slices; The multi-channel spatial feature slices are spliced ​​together according to the temporal order of occurrence to obtain a disaster isomorphic tensor sequence.

8. A regional disaster risk assessment device based on deep neural networks, characterized in that, The apparatus employs the regional disaster risk assessment method based on deep neural networks as described in any one of claims 1 to 7, and the apparatus comprises: The data sequence construction module is used to perform grid division and time synchronization processing on multi-source disaster monitoring data of the target area to obtain a disaster isomorphic tensor sequence. The feature extraction module is used to input the disaster isomorphic tensor sequence into a pre-trained disaster situation awareness deep model for feature extraction to obtain initial disaster latent features; The disaster feature analysis module is used to calculate the feature sensitivity gradient corresponding to the initial hidden disaster features through the disaster situation awareness depth model to determine the main controlling factors of disaster occurrence, and to extract the evolution extreme values ​​based on the initial hidden disaster features to obtain the disaster evolution extreme value features; The disaster-bearing body feature generation module is used to generate a disaster-bearing body vulnerability tensor based on the disaster-bearing body distribution data of the target area, and to generate critical steady-state influence data based on the geological environment data of the target area. The probability calculation module is used to generate a disaster outbreak probability matrix based on the critical steady-state influence data, the main controlling factors of disaster occurrence, and the extreme value characteristics of disaster evolution; The risk assessment module is used to input the disaster outbreak probability matrix, the vulnerability tensor of the disaster-bearing body, and the critical steady-state impact data into a pre-trained risk assessment deep model for spatial topology fusion and feature mapping to obtain regional disaster risk assessment results.

9. A regional disaster risk assessment device based on deep neural networks, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the regional disaster risk assessment method based on a deep neural network as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the regional disaster risk assessment method based on deep neural networks as described in any one of claims 1 to 7.

Images