A geological disaster risk intelligent prediction method based on deep learning

Abstract

The application discloses a kind of geological disaster risk intelligent pre-judgment methods based on deep learning, it is related to geological disaster monitoring and early warning technical field, the present application can automatically carry out high-precision alignment and feature extraction to optical remote sensing, radar measurement, laser point cloud and so on different space-time resolution, different physical meaning monitoring data by deep feature extraction and attention weighted fusion mechanism, generate unified and information-rich representation, lay a solid data foundation for subsequent accurate prediction, overcome the limitation of data fragmentation application in traditional method;By constructing the space-time diagram with slope unit as node, and using the attention mechanism graph convolution network introduced by hydrological direction, the model can accurately describe the spatial propagation process of the migration, accumulation and blockage of slope materials along the flow path under rainfall conditions;At the same time, the time series module effectively learns the influence of past hydrological state on future evolution trend.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological disaster monitoring and early warning, and in particular to a geological disaster risk intelligent prediction method based on deep learning. BACKGROUND

[0002] In the prior art, in view of the characteristics of large terrain undulation, complex structure and concentrated rainfall intensity in high mountain and canyon areas, multi-source remote sensing data and machine learning are often used for risk identification and prediction. Typical methods include: time series change detection and classification based on satellite optical / radar images to identify landslide bodies, and combining rainfall monitoring to build a rainfall threshold model. In the disaster chain scenario, some existing solutions attempt to couple slope stability models with watershed hydrological models, and integrate optical remote sensing, interferometric synthetic aperture radar (InSAR) and ground sensor data through multi-modal data fusion. Some solutions also introduce graph neural networks to describe the spatial correlation between slope units.

[0003] However, in the prediction of disaster chains in high mountain and canyon areas, different sensors have significant differences in spatial / temporal resolution and observation mechanism, making it difficult to achieve high-precision spatio-temporal alignment and uncertainty transmission, which affects the description of the entire trigger-evolution process. Landslide occurrence, accumulation body stability, river blockage and dammed lake formation are often modeled separately, lacking continuous coupling and feedback description of rock mass rupture, material transport, river blockage and breaching. The processes of rainfall infiltration, surface / underground runoff collection and river erosion and deposition have large parameter uncertainties under canyon micro-topographic conditions, leading to prediction bias of dammed lake formation location and scale, and further affecting the lead time and reliability of early warning.

[0004] The above problems result in the need for a technical solution that can achieve high-precision spatio-temporal fusion under heterogeneous data conditions and dynamically couple the chain evolution process. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a geological disaster risk intelligent prediction method based on deep learning to solve the three major pain points of data fusion difficulty, disaster chain process modeling fragmentation and weak physical mechanism description in existing methods, resulting in inaccurate prediction.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] In a first aspect, the present application provides a geological disaster risk intelligent prediction method based on deep learning, which includes,

[0009] obtaining multi-source monitoring data of a target area;

[0010] Preprocessing and spatio-temporal alignment of multi-source monitoring data, and constructing a spatio-temporal graph containing slope units;

[0011] Feature-level fusion of multi-modal features of each slope unit;

[0012] The fused multi-modal features are input into a spatio-temporal graph neural network containing a spatial graph convolution module and a time series module to learn spatial correlation and temporal evolution features; and a training loss based on seepage and runoff process is introduced during training; the spatio-temporal graph neural network is applied to output prediction results including slope unit landslide occurrence probability, potential barrier point and its formation probability, and barrier body breaching risk level.

[0013] As a preferred scheme of the intelligent geological disaster risk prediction method based on deep learning, the multi-source monitoring data includes:

[0014] Optical remote sensing images, synthetic aperture radar interferometric data, laser radar point clouds and digital elevation models, geological structure and lithology data, and gridded rainfall and surface hydrological monitoring data;

[0015] The preprocessing includes radiation and atmospheric correction, geometric correction, terrain correction, registration and resampling in the same coordinate system.

[0016] As a preferred scheme of the intelligent geological disaster risk prediction method based on deep learning, the slope unit is segmented according to the terrain undulation and the runoff direction, the nodes of the spatio-temporal graph correspond to the slope units, the edge relationship includes at least the contact relationship between adjacent units and the hydrological connectivity relationship along the runoff direction, and the edge weight can be dynamically updated according to the spatio-temporal distribution of rainfall and geological properties.

[0017] As a preferred scheme of the intelligent geological disaster risk prediction method based on deep learning, the spatial graph convolution module adopts a multi-head self-attention mechanism, and the mechanism includes calculation of attention weights, which takes the spatio-temporal distribution features of rainfall intensity as query vectors, and takes the terrain and geological features as keys and values to realize the importance reweighting of connectivity;

[0018] The calculation of the attention weights includes:

[0019] For each slope unit node at time Extract rainfall spatio-temporal embedding and terrain-geological features, the rainfall spatio-temporal embedding as the query source, and the terrain-geological features as the key and value sources; set independent projection parameter groups for the multi-head mechanism, and limit the neighborhood to the set of adjacent contact and runoff direction connected units;

[0020] Attention scoring and normalization are performed:

[0021] ,

[0022] in, For a moment In the head Next node To the neighbors Attention weights; The index of the aggregated node; Index the neighboring nodes; It is a multi-head index; For time indexing; For nodes The neighborhood set; For query vector; The key vector; For the head The vector dimension; For the head Learnable bias coefficients; As a hydrological directional indicator, it is taken as 1 with the confluence of currents and 0 with the opposite direction or no direction. Indicates neighborhood index Normalization on;

[0023] In the formula:

[0024] , ,

[0025] in, and Head respectively The query and key projection matrix; For nodes At any moment The rainfall embedding vector; For nodes Topographic-geological feature vector;

[0026] Defined value aggregation and multi-head output:

[0027] ,

[0028] in, For the head Next node At any moment The neighborhood aggregation results; In this expression, it represents the pair of The weights; It is a value vector; For index quantity;

[0029] In the formula: ,

[0030] In the formula: is the value projection matrix of the head ; the multi-head output represents the node time when the channel dimension is spliced;

[0031] Pay attention to the dynamic graph edge weight and inject the graph convolution:

[0032] ,

[0033] In the formula: is the final edge weight at time ; is the convergence result of the multi-head attention; is the baseline structure edge weight, which is pre-normalized by the contact relationship and hydrological connectivity; is the fusion gate coefficient of node at time ;

[0034] In the formula:

[0035] , ,

[0036] In the formula: is the number of attention heads; is the Sigmoid function; is the learnable scaling coefficient; is the normalized rainfall intensity of node at time ; the graph convolution layer takes as the weight of neighborhood message passing to participate in spatial aggregation and propagation.

[0037] As a preferred scheme of the deep learning-based geological disaster risk intelligent prediction method, the time sequence module is a long short-term memory network, which is used to represent the time sequence dependent relationship of rainfall infiltration, water content change and runoff collection, and the influence of hydrological state on hidden state update is explicitly modeled through the forget gate, input gate and output gate.

[0038] As a preferred scheme of the deep learning-based geological disaster risk intelligent prediction method, the training loss includes a data-driven loss and a physical constraint loss, and the physical constraint loss at least includes a mass conservation constraint term and a geometric compatibility constraint term, wherein the mass conservation constraint term includes residual calculation of the mass conservation constraint term, and the residual is constructed based on the mass balance error of the infiltration, runoff, inflow and outflow of the slope unit and the river section;

[0039] The residual calculation step of the mass conservation constraint term comprises:

[0040] 1) At time step , define the mass conservation residual for a slope cell as:

[0041] ,

[0042] where, is the mass conservation residual for cell at time step ; , is the water storage at time step ; is the time step length; is the rainfall conversion volumetric input rate; is the evapotranspiration loss rate; is the infiltration rate; is the total outflow rate pointing to neighbors; is the total inflow rate from neighbor set ; is the river outflow flux for cell at time step ; is the slope cell index; is the neighbor index; is the time step index; 2) Define the discrete mass conservation residual for a reach, define the residual for reach

[0043] as:

[0044] ,

[0045] where, is the mass conservation residual for reach at time step ; , is the reach water storage; is the total lateral inflow from slopes; is the inflow from upstream reach set ; is the outflow to downstream; is the boundary term; is the reach index; is the upstream reach index; is the time step index;

[0046] ​3) Discrete mapping of the slope-river flux coupling, lateral inflow aggregated by connectivity weight:

[0047] ,

[0048] where, is the number of slope elements; is the element connectivity weight of the river reach; is the river reach index;

[0049] 4) Explicit push-forward calculation and ; no flux boundary for divide and cliff; water level-discharge or given outflow sequence boundary for watershed outlet; initial storage from antecedent moisture and river stage inversion; non-negative clipping and upper volume constraint for ;

[0050] 5) Physical constraint loss constructed from normalized residual:

[0051] ,

[0052] where, is the physical constraint loss; is the time step number; and are the number of slopes and river reaches; , are the scale coefficients; is the penalty order;

[0053] 6) Training with the weighted sum of data items and physical items:

[0054] ,

[0055] where, is the total loss; is the data-driven loss; is the physical item weight, which is non-negative: ,where, is the learnable log-variance parameter, which automatically adjusts the physical item proportion during training;

[0056] 7) The training process of this step is as follows:

[0057] The time field and flux field are obtained at each step, the mass conservation residual is calculated according to steps 1) to 3), and the data-driven loss and physical constraint loss are combined by weighted combination according to step 6).

[0058] As a preferred scheme of the intelligent geological disaster risk prediction method based on deep learning, in the scheme, a multi-modal feature level fusion adopts a deep feature extraction network to process optical, SAR / InSAR and point cloud elevation features respectively, and generates a unified feature representation through an attention weighted fusion mechanism.

[0059] For data of different spatial resolutions, a learnable up / down sampling module is used to realize cross-scale alignment.

[0060] As a preferred scheme of the intelligent geological disaster risk prediction method based on deep learning, in the scheme, a hybrid model is used to perform multi-peak prediction on the potential damming point position, and a plurality of candidate damming points and their formation probabilities are output.

[0061] Each candidate corresponds to a position assumption on a river center line, and model parameters are optimized through approximate inference.

[0062] As a preferred scheme of the intelligent geological disaster risk prediction method based on deep learning, in the scheme, the method further includes transfer learning, and a domain discriminator is used to realize field self-adaptation between a source field and a target field through an adversarial training and a feature distribution difference minimization strategy.

[0063] In a second aspect, an embodiment of the present application provides a computer device including a memory and a processor, and the memory stores a computer program, wherein when the computer program is executed by the processor, any step of the intelligent geological disaster risk prediction method based on deep learning according to the first aspect of the present application is implemented.

[0064] The present application has the advantages that: the present application effectively solves the problem of multi-source heterogeneous data fusion. Through deep feature extraction and attention weighted fusion mechanism, optical remote sensing, radar measurement, laser point cloud and other monitoring data of different temporal and spatial resolutions and different physical meanings can be automatically aligned and feature extracted with high precision, and unified and information-rich representation is generated, which lays a solid data foundation for subsequent accurate prediction and overcomes the limitations of data fragmentation application in traditional methods.

[0065] The present application realizes dynamic and continuous coupling simulation of the whole process of geological disaster chain triggering-evolution-disaster. By constructing a time-space graph with slope units as nodes, and using an attention mechanism graph convolution network with hydrological directionality, the model can accurately depict the spatial propagation process of slope material migration, accumulation and river blockage along the flow path under rainfall conditions. At the same time, the time series module effectively learns the influence of past hydrological state on future evolution trend. This spatio-temporal coupling modeling method changes the mode of independent segmentation of landslide, river blocking and breach in traditional methods, and reveals the continuous evolution mechanism of the disaster chain.

[0066] The application introduces a physical constraint loss function into the training process of a deep learning model, deeply fuses data driving and mechanism driving, and ensures that the output result of the model strictly follows basic physical laws such as conservation of mass while learning a large amount of data, so as to ensure that the prediction result is not only statistically accurate but also physically reasonable and interpretable. The prediction deviation problem of a pure data-driven model in an extreme scenario is greatly alleviated, and the reliability and stability of the early warning result in a complex environment such as a high mountain and a valley are significantly improved.

[0067] The application outputs a multi-dimensional result including landslide probability, probability distribution of multiple potential barrier points and breaching risk, provides more abundant and reliable decision information, can provide a longer early warning lead time and more scientific decision support for a disaster prevention and reduction department, and has great practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope of the present application.

[0069] Figure 1 A flowchart of the intelligent geological disaster risk prediction method based on deep learning in the embodiments. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0071] All terms (including technical and scientific terms) used in the present application have meanings generally understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted to have meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0072] For example, the terms "first", "second", etc. used in the present application are only used to distinguish similar objects, to distinguish the first object from another object, and are not used to describe a specific order or sequence, and cannot be understood as indicating or implying relative importance.

[0073] The present application proposes an intelligent geological disaster risk prediction method based on deep learning, which combines Figure 1 As shown in the figure, the method comprises the following steps.

[0074] Step S1, acquiring multi-source monitoring data of a target area;

[0075] In this embodiment, the target area refers to the analysis range bounded by river basins or sub-basins in a unified projection coordinate system. The spatial range is obtained by the intersection of the outer rectangular and the administrative district / basin boundary. The time range is mainly in the high-incidence season of disasters and covers the training and verification intervals. The default update cycle is daily batch processing and is overlaid by hourly meteorological driving. It can be adjusted between every 15 minutes and every 3 hours. The length of the time period is not less than 90 consecutive days by default to ensure that the model learns the dry-wet conversion process. The setting is based on the joint distribution of historical rainfall and landslide records and expert experience. Multi-source monitoring data comes from public archives or licensed commercial providers, using programmatic interfaces or batch offline downloads. When some data is missing in the time period, the valid observations in the recent period are filled forward, with a maximum of 72 hours. When the time exceeds, the minimum executable aperture is triggered, and only the necessary elements of weather and terrain are retained to maintain the coherence of the reasoning process.

[0076] Step S2, pre-processing and spatio-temporal alignment of multi-source monitoring data to construct a spatio-temporal graph containing slope units;

[0077] Step S3, feature-level fusion of multi-modal features for each slope unit;

[0078] Step S4, input the fused multi-modal features into a spatio-temporal graph neural network containing a spatial graph convolution module and a time series module to learn spatial correlation and temporal evolution;

[0079] Step S5, introduce training loss based on seepage and runoff processes during training; output prediction results including slope unit landslide occurrence probability, potential barrier point and its formation probability, and barrier body breaching risk level;

[0080] In one embodiment, the multi-source monitoring data includes:

[0081] optical remote sensing images, synthetic aperture radar interferometric data, laser radar point clouds and digital elevation models, geological structure and lithology data, and gridded rainfall and surface hydrological monitoring data;

[0082] Pre-processing includes radiation and atmospheric correction, geometric correction, terrain correction, registration and resampling in the same coordinate system;

[0083] Specifically, the target resolution of pre-processing is 10 meters by default, which can be adjusted in the range of 5-30 meters according to the matching of computing resources and data resolution; the reference step length of time alignment is 1 hour by default, which can be adjusted between 15 minutes and 3 hours, and is aligned at the whole hour. The resampling adopts area weight or distance weight to process continuous variables, and the mode of the classification variable is kept consistent; the coordinate reference adopts the unified projection coordinate system, and the transverse projection matching the latitude and longitude span of the study area is preferentially selected. The parameter setting is based on the statistical stability of the peak signal-to-noise ratio of the grid and the registration residual; when there is local distortion in the image, which leads to registration error exceeding one pixel, local rigid block displacement compensation is used to ensure the continuity of the boundary of the ground object.

[0084] In one embodiment, the slope unit is divided according to the terrain undulation and the flow direction, the nodes of the space-time graph correspond to the slope unit, the edge relationship includes at least the contact relationship of adjacent units and the hydrological connectivity relationship along the flow direction, and the edge weight can be dynamically updated according to the spatio-temporal distribution of rainfall and geological properties;

[0085] For example, the minimum area of the slope unit is 4 pixels by default, which can be adjusted between 1 and 16 pixels; the adjacent determination adopts the contact relationship with the shared boundary length greater than 0, the hydrological connectivity is determined according to the falling unit coded by the flow direction, and the dominant path obtained from the flow accumulation is preferentially adopted. The dynamic update period of the space-time graph is consistent with the meteorological time step by default, and the upper limit frequency of edge weight update is coupled with the rain intensity level to avoid excessive jitter; when the terrain is extremely fragmented, leading to unit fragmentation, regional merging based on the consistency of terrain undulation and slope direction is adopted to keep the graph structure sparse and coherent.

[0086] In one embodiment, the spatial graph convolution module adopts a multi-head self-attention mechanism to dynamically adjust the connection weight between slope units, and the mechanism includes the calculation of attention weight, which takes the spatio-temporal distribution feature of rainfall intensity as the query vector, and takes the terrain and geological features as the key and value to realize the importance reweighting of connectivity;

[0087] The calculation steps of the attention weight include:

[0088] For each slope unit node at time Extract the rainfall spatio-temporal embedding and the terrain-geological feature, the rainfall spatio-temporal embedding as the query source, and the terrain-geological feature as the key and value source; set independent projection parameter groups for the multi-head mechanism, and limit the neighborhood to the set of adjacent contact and connectivity units along the flow direction;

[0089] Attention scoring and normalization are performed:

[0090] ,

[0091] wherein, is the time In the head Lower node Attention weight to neighbor ; Index for aggregated node Index for neighbor node Index for multi-head Index for time Neighborhood set for node ; Query vector Key vector Vector dimension for head ; Learnable bias coefficient for head ; Hydrological directionality indicator, 1 for confluence, 0 for reverse or undirected Indicates normalization on neighborhood index Similarly, the number of attention heads defaults to 4 to 8, and the single-head channel dimension and the fused node representation dimension are determined according to the total calculation amount constraint. The query vector is obtained by embedding the area average or quantile value of the rainfall in the slope unit by hour or by day, and the key / value vector is obtained by splicing the terrain slope, slope direction, curvature, and geological structure and lithology code. The directionality indicator is derived from the adjacency matrix according to the flow direction relationship, and the pairs that do not form effective hydrological paths are assigned a value of 0 and remain masked during normalization. To enhance numerical stability, the scoring item uses temperature scaling during extreme rainfall, and the temperature defaults to adjustment between 0.8 and 1.2. The scaling is determined based on the entropy index of the attention distribution on the validation set.

[0092] In the formula:

[0093] , ,

[0094] Wherein, And Query and key projection matrix of head ; Rainfall embedding vector of node At time ; Topographic-geological feature vector of node ; Optionally, the composition of the topographic-geological feature vector is subject to the availability of the present embodiment, typically including slope, relative elevation, curvature, geomorphological position index, and one-hot encoding of lithology category, and can include fault distance and joint density derived from geological structure data. Each component is standardized or robustly scaled before input to reduce the attention bias caused by inconsistent dimensions.

[0095] Defined value aggregation and multi-head output:

[0096] ,

[0097] in, For the head Next node At any moment The neighborhood aggregation results; In this expression, it represents the pair of The weights; It is a value vector; For index quantity;

[0098] In the formula: ,

[0099] in, For the head The value of the projection matrix; the multi-head output is concatenated in the channel dimension to form the node time representation for use in subsequent spatial graph convolution; further, the node representation after multi-head concatenation is compressed to the target number of channels by a linear layer to match the bandwidth of subsequent spatial aggregation, and the compression ratio is selected by the validation set performance by default between 0.5 and 1.0; when computational resources are limited, it can degenerate into single-head attention while keeping the input / output form unchanged, only sacrificing the diversity of neighborhood relationship modeling.

[0100] Focus attention on dynamic graph edge weights and inject them into graph convolution:

[0101] ,

[0102] in, For a moment The final edge weight; This is the result of the convergence of multiple heads' attention; As the baseline structural boundary weight, it is pre-normalized based on contact relationship and hydrological connectivity; For nodes At any moment The fusion gating coefficient;

[0103] In the formula:

[0104] , ,

[0105] in, For the number of attention heads; For the Sigmoid function; The learnable scaling factor; For nodes At any moment Normalized rainfall intensity; convolutional layer in figure As a weight for neighborhood message passing, it participates in spatial aggregation and propagation. In this embodiment, the fusion gating coefficient is obtained by normalizing the rainfall intensity through a single-layer mapping, and the default S-shaped mapping is used and restricted between 0 and 1. When there is no rainfall or the rainfall intensity is extremely low, the gating coefficient automatically decays, so that the structural edge weights dominate spatial propagation. Under short-term heavy rainfall, the gating coefficient increases to emphasize the data-driven path, thereby improving the sensitivity to temporary connectivity.

[0106] Masking is applied to non-adjacent pairs to prevent them from participating in normalization; during extreme rainfall, temperature scaling and small constant stabilization terms are added to the scoring terms to suppress sharp distributions and numerical overflow; specifically, the small constant is set to the machine precision level of numerical types by default to avoid overflow from division by zero or logarithmic operations; when the local neighborhood is too sparse, causing attention to focus on a single edge, upper limit pruning and minimum weight backfilling are used to maintain propagation continuity; if rainfall data is missing at a single moment, weighted interpolation of the two most recent moments is used to supplement it and the attention weight at that moment is reset with reduced confidence to avoid misleading dynamic edge weights.

[0107] Specifically, using rainfall as the query source and topography and geological carrying capacity as the key value source, an attention allocation consistent with the hydrological direction is constructed. The scoring item superimposes similarity measurement and directional indicator, so that downstream connections receive higher weights under the same similarity, which is consistent with the action path of runoff generation and confluence. The multi-head mechanism observes the same neighborhood from different subspaces, first weighted aggregation within the head, and then forming a robust convergence result by the mean, reducing the adverse effects of single-head bias. The gating coefficient adjusts the relative proportion of learning edge weights and structural edge weights according to the rainfall intensity. When the rainfall intensity is high, the data-driven channel is highlighted, and when the rainfall intensity is low, the structure-dominated channel is regressed, thus maintaining stable propagation between dry and wet periods. The final dynamic graph edge weights are seamlessly entered into graph convolution, realizing interpretable modeling of spatial correlation and providing a consistent input representation for subsequent temporal modules.

[0108] In one embodiment, the time series module is a long short-term memory network to characterize the temporal dependencies between rainfall infiltration, water content changes and runoff collection, and to explicitly model the impact of hydrological state on hidden state updates through forget gates, input gates and output gates.

[0109] For example, the time step is set to 1 hour by default, consistent with weather-driven timing, and can be adjusted between 15 minutes and 3 hours; the input sequence length is set to 48 to 168 steps by default to cover short durations and intra-week variations; the hidden state dimension is set between 64 and 256 based on training stability and resource constraints. When there are discontinuities in the input sequence, time-scale alignment and zero-padding with a mask are used to ensure that the gating state updates follow the chronological order and are not affected by missing data.

[0110] In one embodiment, the training loss includes data-driven loss and physical constraint loss. The physical constraint loss includes at least a mass conservation constraint term and a geometric compatibility constraint term, wherein the mass conservation constraint term includes the residual calculation of the mass conservation constraint term, which is constructed based on the mass balance error of infiltration, runoff generation, inflow and outflow of the slope unit and the river segment.

[0111] The steps for calculating the residuals of the mass conservation constraint terms include:

[0112] 1) At the time step Above, for the slope unit The mass conservation residual is defined as:

[0113] ,

[0114] in, For unit At time step The quality conservation residual; , For time step and The water storage capacity; For time step; The volumetric input rate of rainfall conversion; Evaporation loss rate; Infiltration rate; This refers to the total outflow rate pointing to the neighbor; For the set of neighbors Total inflow rate; For unit At time step The inflow flux into the river; Index for slope units; Index for neighbors; For time step index;

[0115] Similarly, water storage is obtained from node states through volume transformation, with the transformation coefficient taken from an empirical range of unit area and effective soil layer thickness; the rainfall volume input rate is mapped from the unit average rainfall intensity and takes into account the reduction due to surface resistance; infiltration and evapotranspiration terms are obtained based on the latent state of the time series module and regression of surface / geological attributes; outflow and inflow terms are jointly determined by the dynamic graph edge weights and the allocation ratio of outflow from adjacent nodes. The parameter value range is limited by the accessibility of historical water balance and training stability.

[0116] 2) Define the discrete mass conservation residuals of the river segment, for the river segment Define the residual as follows:

[0117] ,

[0118] in, For river section At time step The quality conservation residual; , This refers to the water storage capacity of the river section. This represents the total lateral inflow from the slope; Collection of upstream river sections The influx of people; For downstream outflow; For boundary terms; For river segment indexing; Index for the upstream river section; For time step index; optionally, the water storage capacity of the river section is obtained by approximation through equivalent cross section and river storage coefficient, and the outflow term is mapped by empirical or semi-empirical relationship based on the current water level or flow status of the river section; the boundary term is directly substituted when the scheduling or water intake records are known, and set to zero mean disturbance when unknown to avoid systematic bias.

[0119] 3) Perform discrete mapping of slope-river segment flux coupling, with lateral inflow aggregated by connectivity weights:

[0120] ,

[0121] in, Number of slope units; For unit River confluence section The connectivity weights; (non-negative, given by spatial topology and confluence path, often taking the pairwise values) (Return to One) and These are unit and river segment indexes, respectively; The time step index is used; furthermore, the connectivity weights are derived from the normalization of the hydrological distance or the shortest confluence path length between the slope unit and the nearest river segment, with column normalization performed by default for each k; in areas with abrupt topographic changes or extremely narrow canyons, to avoid all weights being concentrated on a single river segment, a minimum apportionment coefficient is set to promote limited diffusion of lateral inflows between adjacent river segments, thereby improving numerical stability.

[0122] 4) Employing explicit propagation calculations and ; No flux boundaries are set for watersheds and fault scarps; Level-discharge or given outflow sequences are set for watershed outlets; Initial storage is obtained by inversion from previous water cut and channel level; Non-negative truncation and upper volume constraint are applied to avoid numerical anomalies; small threshold is used to enable wetting for dry river segments; in this embodiment, initial conditions are set to the meteorological and hydrological records of the previous 7-30 days to perform inversion; the outlet boundary is approximated as an open boundary when observations are lacking, and a maximum outflow level constraint is applied to avoid non-physical water storage accumulation in the early stage of the model; when the wetting threshold is triggered frequently, the stability term of the time step is automatically increased to reduce the oscillation caused by repeated wetting and drying.

[0123] 5) The physical constraint loss is constructed with normalized residuals:

[0124] ,

[0125] wherein, is the physical constraint loss; is the number of time steps; and are the slope and the number of river segments; , are the scale coefficients; is the penalty order; specifically, the scale coefficients can be set according to the median absolute deviation or quantile range of the corresponding residuals on the validation set to ensure that different positions and different types of residuals contribute to the loss at the same order of magnitude; the penalty order is selected between 1 and 2, and the default is 1 to enhance the robustness to outliers, and 2 can be taken when the observation noise is significantly reduced to improve smoothness.

[0126] 6) Training is performed with the weighted sum of data items and physical items as the target:

[0127] ,

[0128] wherein, is the total loss; is the data-driven loss; is the physical item weight, which is a non-negative weight: , is a learnable logarithmic variance parameter that automatically adjusts the proportion of physical items during training; optionally, the total loss can be limited by limiting the change rate of to avoid imbalance between the gradient amplitudes of physical items and data items in the early stages of training; the upper limit of the change rate is determined according to the stable interval of the gradient norm ratio of the validation set; when the supervised data is sparse, the upper limit of the proportion of physical items is temporarily increased to maintain an identifiable convergence path.

[0129] 7) The training process of this step is as follows:

[0130] The time field and flux field are obtained by each forward step, and , and are calculated according to steps 1-3, and Combination according to step 6 gives ; the network parameters are updated by backpropagation and ; early stopping and validation set are used to monitor convergence; similarly, the early stopping criterion is based on the improvement of the comprehensive index on the validation set by default, and the window length is 5 to 10 evaluation cycles; the evaluation cycle is triggered according to a number of training rounds, and the interval can be set to half to two rounds according to the data size; in the case of limited resources, the evaluation frequency is reduced instead of shortening the input sequence to avoid damaging the learning of time dependence.

[0131] Specifically, the residual construction takes volume conservation as the core, discretizes the slope and river section respectively, and connects the lateral inflow and unit confluence through the connection weight; in this way, it can not only follow the topology and confluence direction in space, but also correspond to the dynamic state field of the network output in time; after normalization and aggregation of the residual, the magnitude difference between different scale units and different river section can be weakened, and the abnormality of a single position can be avoided. The training; the weighting of physical and data items adopts learnable weights, so that the model can increase the dependence on conservation constraints when the observation is sparse, and return to data fitting when the observation is sufficient, thereby forming a transferable learning strategy; the boundary and initial condition are integrated into the calculation process with simple and operable rules, which is convenient for reuse in different basins and different data sources;

[0132] In this embodiment, in order to ensure the stability during cross-regional migration, it is recommended to freeze part of the spatial attention projection parameters and scale coefficients in the initial adaptation stage of the new target area, and only fine-tune the time series module and the output layer, and then unfreeze all parameters for further joint optimization when the conservation residual reaches the stable range.

[0133] In one embodiment, the multi-modal feature level fusion uses a deep feature extraction network to process optical, SAR / InSAR and point cloud elevation features respectively, and generates a unified feature representation through an attention weighted fusion mechanism; further, the optical channel can input a tensor composed of original multispectral and its robust index transformation, the SAR / InSAR channel inputs time series deformation and coherence statistics, and the point cloud and digital elevation channel inputs terrain derived quantities; the query of attention fusion can be taken from the main driving variable embedding of the current task, the key and value correspond to the compressed representation of each modality, and the fusion output dimension is consistent with the expected input dimension of the subsequent network.

[0134] For data of different spatial resolutions, a learnable up / down sampling module is used to realize cross-scale alignment;

[0135] In this embodiment, after the target resolution is aligned with the time step, the mapping of the discrete grid to the slope unit adopts area weight aggregation, and the write-back of the slope unit to the grid adopts distribution weight average; when the resolution difference of the data source exceeds an order of magnitude, alignment is preferentially performed in the feature space rather than the pixel space to reduce artifacts; the parameters of the alignment module are selected according to the reconstruction error of the validation set.

[0136] In one embodiment, a hybrid model is used for multi-peak prediction of potential barrier point positions, and multiple candidate barrier points and their formation probabilities are output; optionally, the number of components of the hybrid model is selected by default between 3 and 8, and is determined according to the trade-off between position deviation and recall on the validation set; initialization can be set heuristically by candidate points of high curvature of the river center line or flow mutation positions to accelerate convergence and reduce invalid components.

[0137] Each candidate corresponds to a position hypothesis on the river center line, and model parameters are optimized by approximate inference;

[0138] Specifically, the convergence tolerance of approximate inference is set to between 1e-4 and 1e-3 by default, and the maximum number of iterations is set to between 50 and 200; when component collapse or high overlap occurs, a minimum variance lower bound and component merging strategy are used to maintain the interpretability of the solution; in the output stage, the final barrier point candidate set is filtered according to the component weight and minimum distance rule.

[0139] In one embodiment, the method further comprises transfer learning, which realizes domain adaptation between the source domain and the target domain through adversarial training of the domain discriminator and a feature distribution difference minimization strategy, and uses pseudo labels and consistency constraints to improve the generalization ability in the target domain.

[0140] The embodiment also provides a computer device, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the deep learning-based geological disaster risk intelligent prediction method according to the above embodiment.

[0141] The computer device can be a terminal, which comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0142] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

[0143] In addition, those skilled in the art can understand that although some embodiments herein include certain features rather than others included in other embodiments, the combination of features of different embodiments means that it is within the scope of the present application and forms different embodiments. For example, all the above embodiments can be used in any combination. The information disclosed in the background section is only intended to deepen the understanding of the overall background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A deep learning-based geological disaster risk intelligent prediction method, characterized in that, The application relates to a landslide prediction method based on a physical-informed spatio-temporal graph neural network. The method comprises the following steps: acquiring multi-source monitoring data of a target area; preprocessing and spatio-temporal alignment of the multi-source monitoring data, and constructing a spatio-temporal graph containing slope units; feature-level fusion of multi-modal features of each slope unit; inputting the fused multi-modal features into a spatio-temporal graph neural network containing a spatial graph convolution module and a time series module to learn spatial correlation and time evolution features; and introducing a training loss based on seepage and runoff and sediment process in the training process; applying the spatio-temporal graph neural network, and outputting prediction results including landslide occurrence probability of the slope unit, potential damming point and its formation probability, and damming body breaching risk level; the training loss comprises a data-driven loss and a physical constraint loss, and the physical constraint loss at least comprises a mass conservation constraint term and a geometric compatibility constraint term, wherein the mass conservation constraint term comprises residual calculation of the mass conservation constraint term, and the residual is constructed based on mass balance errors of infiltration, runoff, inflow and outflow of the slope unit and the river section; 1) At time step The slope cell The mass-conservation residual is defined as: ， where, is the unit at time step ; , is the time step ; is the water storage at time step ; is the volumetric input rate of rainfall conversion; is the evapotranspiration loss rate; is the infiltration rate; is the total outflow rate to neighbors; is the total inflow rate from the neighbor set ; is the unit inflow to the river at time step ; is the slope cell index; is the neighbor index; is the time step index; 2) define a discrete mass-conservation residual for the reach, and solve for the reach Define the residual as follows: ， wherein, is the river reach at time step mass conservation residual; , is the river reach storage; is the total lateral inflow from the slope; is the inflow from the upstream river reach set ; is the outflow to the downstream; is the boundary term; is the river reach index; is the upstream river reach index; the residual calculation step of the mass conservation constraint term comprises: ， wherein, is the number of slope units; is the unit confluence reach connectivity weight; is the reach index; 4) Explicit propagation calculations With ; no flux boundary conditions are applied to watersheds and cliffs; water level-discharge or given outflow hydrograph boundary conditions are applied to outlet of the watershed; initial storage is obtained from antecedent moisture and channel water level inversion; non-negative cut-off and upper volume constraints are applied ​ 3) discrete mapping of slope-river flux coupling, and lateral inflow is aggregated by a connection weight: ， wherein, is a physical constraint loss; is a number of time steps; with is a number of slopes and reaches; , is a scale factor; is a penalty order; 5) constructing a physical constraint loss with a normalized residual: ， wherein, is the total loss; is the data driven loss; is the physical term weight, is a non-negative weight: wherein, is a learnable log-variance parameter, automatically adjusts the physical term proportion with training; 6) training with a weighted sum of data items and physical items as the target: 7) the training process of this step is as follows: 2.The deep learning-based geological disaster risk intelligent prediction method according to claim 1, wherein each step obtains a time field and a flux field, the mass conservation residual is calculated according to steps 1) to 3), and the data-driven loss and the physical constraint loss are combined by weighting according to step 6). The multi-source monitoring data comprises: optical remote sensing images, synthetic aperture radar interferometric measurement data, laser radar point clouds and digital elevation models, geological structure and lithology data, and grid-based rainfall and surface hydrological monitoring data; 3.The deep learning-based geological disaster risk intelligent prediction method according to claim 1, wherein the preprocessing comprises radiation and atmosphere correction, geometric correction, terrain correction, registration in the same coordinate system and resampling. 4.The deep learning-based geological disaster risk intelligent prediction method according to claim 1, wherein, The slope unit is segmented according to the terrain undulation and the flow direction, the nodes of the spatio-temporal graph correspond to the slope units, the edge relationship at least comprises a contact relationship between adjacent units and a hydrological connectivity relationship along the flow direction, and the edge weight can be dynamically updated according to the spatio-temporal distribution of rainfall and the geological properties. The spatial graph convolution module adopts a multi-head self-attention mechanism, the mechanism comprises calculation of attention weight, the calculation takes the spatio-temporal distribution feature of rainfall intensity as a query vector, and takes the terrain and geological features as keys and values to realize reweighting of the importance of connectivity; For each slope unit node at time Extract rainfall spatiotemporal embedding and topography-geology features, rainfall spatiotemporal embedding as query source, topography-geology features as key and value source at the same time; set independent projection parameter group for multi-head mechanism, and limit neighborhood to adjacent contact and unit set connected along confluence direction; the calculation step of the attention weight comprises: ， wherein, is the time instant is the head is the lower node is the attention weight for the neighbor ; is the aggregated node index is the neighbor node index is the multi-head index is the time index is the neighborhood set of nodes ; is the query vector is the key vector is the vector dimension of the head ; is the learnable bias coefficient of the head ; is the hydrological direction indicator, 1 for forward confluence, 0 for reverse or undirected denotes the normalization over the neighborhood index ; attention scoring and normalization: ， ， wherein, with are the head query and key projection matrices, respectively; is the node rainfall embedding vector at time is the node terrain-geology feature vector;​ wherein: ， wherein, is the head is the lower node at time neighborhood aggregation result; is the weight on within the formula; is the value vector; is the index quantity; In the formulae: , in, For the head The value of the projection matrix; the representation of the multi-head output at the moment when the channel dimension is spliced ​​to form a node; value aggregation and multi-head output are defined: ， wherein, is the final edge weight at time is the convergence result of multi-head attention; is the baseline structure edge weight, pre-normalized by contact relationship and hydrological connectivity; is the node is the fusion gate coefficient at time is the fusion gate coefficient at time​ attention is paid to the dynamic graph edge weight, and the graph convolution is injected: ， ， where, is the number of attentions; is the Sigmoid function; is the learnable scaling factor; is the node at time is the normalized rainfall intensity; the graph convolution layer participates in spatial aggregation and propagation with as the weight of neighborhood message passing. 5.The deep learning-based geological disaster risk intelligent prediction method according to claim 1, wherein wherein: 6.The deep learning-based geological disaster risk intelligent prediction method according to claim 1, wherein the time series module is a long short-term memory network, which is used to represent the time sequence dependence relationship of rainfall infiltration, water content change and runoff collection, and the influence of the hydrological state on the hidden state update is explicitly modeled through a forget gate, an input gate and an output gate. The multi-modal feature-level fusion adopts a deep feature extraction network to process optical, SAR / InSAR and point cloud elevation features respectively, and generates a unified feature representation through an attention weighted fusion mechanism; for data with different spatial resolutions, a learnable up / down sampling module is used to realize cross-scale alignment.

7. The geological disaster risk intelligent pre-judgment method based on deep learning according to claim 1, characterized in that, The potential damming point position is predicted by using a mixed model for multi-peak prediction, and multiple candidate damming points and their formation probabilities are output. Each candidate corresponds to a position assumption on the river center line, and the model parameters are optimized by approximate inference. 8.The deep learning-based geological disaster risk intelligent prediction method according to claim 1, wherein, The method also includes transfer learning, which realizes the domain adaptation between the source domain and the target domain through the adversarial training of the domain discriminator and the feature distribution difference minimization strategy. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor implements the steps of the deep learning-based geological disaster risk intelligent prediction method of any one of claims 1-8 when executing the computer program.

Images