Intelligent prediction method for strong earthquakes and landslides driven by event metadata

Abstract

This invention belongs to the field of geological hazard risk assessment technology and discloses an intelligent prediction method for strong earthquake landslides based on event metadata. This invention utilizes a combined event database under a unified observation protocol, counterfactual constraints on widespread unobserved landslide events following strong earthquakes, multi-scale feature-level linear modulation driven by event metadata, soft boundary labels in the 0-1 range, event-balanced pixel localization, landslide triggering event range calibration, and counterfactual event false alarm suppression. Under the condition of unknown event outcome categories, it jointly utilizes ground motion information, topographic environment information, and event-scale source information, which can reduce large-area false alarms in low-risk areas of strong earthquakes and improve the model's cross-event prediction stability.

Description

Technical Field

[0001] This invention belongs to the fields of geological disaster risk assessment, remote sensing geoscientific data processing, construction of a global strong earthquake-induced landslide database, geospatial artificial intelligence, and post-earthquake emergency prediction technology. Specifically, it involves a strong earthquake-induced landslide intelligent prediction method based on event metadata. It constructs a counterfactual sample of widely unobserved landslide events based on a unified observation protocol, uses a multi-scale feature-level linear modulation network driven by event metadata to fuse event-scale source information and pixel-scale geoscientific information, and achieves intelligent prediction of strong earthquake-induced landslides through joint training by event equalization pixel positioning, landslide range calibration, and counterfactual event false alarm suppression. Background Technology

[0002] Earthquake-induced landslides are a significant secondary geological hazard in the earthquake disaster chain. They can disrupt transportation corridors, block rivers, damage engineering facilities, and may evolve into chain risks such as barrier lakes, debris flows, or slope instability long after the earthquake. Rapidly identifying potential high-probability landslide areas after an earthquake is a crucial foundation for emergency rescue deployment, road clearing, secondary disaster inspection, and regional risk management.

[0003] Existing methods for predicting strong earthquake-induced landslides typically input factors such as peak ground acceleration, peak ground velocity, intensity, slope, topographic relief, lithology, and land cover into statistical or machine learning models to estimate the probability of coseismic landslides. These methods can achieve some success in typical events where large-scale landslides have already occurred, but their training samples are often concentrated on events involving strong earthquakes, steep mountainous areas, and those already documented in landslide inventories. They lack the constraint of counterfactual events where strong earthquakes exist but widespread, maptable coseismic landslides are not observed.

[0004] In this training domain, the model easily learns strong ground motion factors as an approximate sufficient condition for landslide occurrence. When the model is transferred to areas with less topographic relief, more stable lithology, and unfavorable surface cover or hydrological conditions for slope failure due to strong earthquakes, it is prone to outputting large-scale, high-probability predictions, misleading emergency response resources. In other words, the model not only needs to learn under what conditions landslides will occur, but also under what conditions strong ground motion is insufficient to cause widespread landslides.

[0005] Globally available inventory of earthquake-induced landslides provides a crucial foundation for model training; however, these publicly available resources have long been biased towards large-scale landslide triggering events, with insufficient records of moderate-magnitude events and widespread unobserved landslide events following strong earthquakes. Recently, based on a screening of earthquake catalogs, and utilizing PlanetScope monthly composite imagery or pre- and post-earthquake optical imagery of equivalent spatial resolution, we have compiled new data to supplement the inventory of 24 new enhanced earthquake-induced landslides and 42 widespread unobserved landslide events following strong earthquakes. The newly added inventory contains a total of 14,291 manually interpreted landslide polygons. These 42 events were not pre-recorded by the satellite platform but were manually verified to have no widely mapable coseismic landslides, provided that the strong ground motion threshold, image coverage, and a unified survey protocol were met.

[0006] Existing methods suffer from insufficient fusion of event-scale and pixel-scale information. Moment magnitude, focal depth, and fault mechanism are global conditions shared by the same event, while slope, lithology, ground motion, rainfall, and land cover are local conditions that vary with spatial location. Simply stacking these two types of data in a grid or statically stitching them together at the network end makes it difficult to dynamically adjust the response of different geological feature channels according to different strong earthquake events. Furthermore, landslide pixels are typically far fewer than background pixels, and the survey range and number of tiles vary significantly across different events. Ordinary background pixels only represent non-landslide locations within a single landslide-triggered event, while strong earthquakes without widespread landslides provide event-level counterfactual supervision indicating "strong ground motion exists but no widespread landslide has formed." Both should not be treated as homogeneous negative samples. Therefore, a joint training method is needed, consisting of counterfactual event data constrained by effective observation masks and soft boundary labels, a conditional modulation network driven by event metadata, and a method that directly constrains the event-scale prediction range. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, the intelligent prediction method for strong earthquake-induced landslides based on event metadata provided by this invention solves the problems of existing strong earthquake-induced landslide prediction models, such as excessive dependence on ground motion factors, lack of event-scale counterfactual constraints, insufficient fusion of event-scale source information and pixel-scale terrain environment information, easy misassignment of deterministic negative labels to unreadable areas of counterfactual events, insufficient cross-event transfer stability, and training bias caused by differences in the number of samples from different events.

[0008] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a strong earthquake landslide intelligent prediction method driven by event metadata, comprising the following steps: S100. Select dual-temporal remote sensing images before and after the earthquake event to construct vector labels for landslides induced by strong earthquakes, and screen for widespread landslide events that were not observed during strong earthquakes. S200, combined with landslide triggering events and widely unobserved landslide events during strong earthquakes, construct a combined event database; S300: Extract predictors from the combined event database, perform feature engineering on them, and vectorize the available event metadata. Process the predictors, soft boundary labels, effective observation masks, event vectors, and event category identifiers into a tile dataset. S400: Construct a multi-scale conditional affine modulation network driven by event metadata, and based on the tile dataset, use event equalization pixel localization, event calibration and counterfactual event false alarm mechanism to jointly train the multi-scale conditional affine modulation network to obtain a strong earthquake-induced landslide intelligent prediction model. S500 uses a strong earthquake-induced landslide intelligent prediction model to predict earthquake events, combining a sliding window and edge-ignoring methods for spatial prediction, and then performs raster stitching on the prediction results to output a landslide occurrence probability map in vector form.

[0009] The beneficial effects of this invention are as follows: 1. Under the same ground motion survey boundary, dual-temporal remote sensing image quality control and three-dimensional terrain verification protocol, this invention simultaneously constructs landslide triggering events and widely unobserved landslide events during strong earthquakes; it eliminates cloud, snow, shadow and image missing areas through effective observation masking, and uses soft boundary labels in the range of 0 to 1 to characterize the transition pixels of the landslide boundary, avoiding treating uncertain pixels as deterministic positive or negative samples.

[0010] 2. This invention constructs a list of newly enhanced earthquake-induced landslide events by following a unified process of earthquake event screening, seismic ground motion survey boundary determination, dual-temporal remote sensing image quality control, and three-dimensional terrain verification. It then standardizes the spatial reference, label format, survey scope, and event metadata of the list of historical strong earthquake-induced landslide events. This expands the coverage of landslide triggering events under different source conditions, landform types, lithological conditions, and surface environments, providing a unified and comparable positive fact training sample for event-conditionalized networks to learn cross-event landslide response patterns.

[0011] 3. This invention distinguishes between ordinary background pixels and event-level counterfactual samples, and organizes landslide triggering events and widespread landslide events that were not observed during strong earthquakes into a unified positive-fact-counterfact event domain. This allows the model to not only learn the spatial differences within a single landslide triggering event, but also to learn the event conditions where there is strong ground motion but not enough to form a widespread landslide, thus reducing the model's over-reliance on ground motion factors.

[0012] 4. This invention utilizes moment magnitude, focal depth, and fault mechanisms to generate channel scaling and offset parameters at multiple network scales, and performs event-adaptive affine modulation on ground motion, topography, lithology, hydrology, and land cover features; the event result category is not input into the prediction network, thereby avoiding the direct differentiation of events using the target result category.

[0013] 5. This invention uses event-balanced pixel localization loss, landslide-triggered event range calibration loss, and strong earthquake-unobserved widespread landslide false alarm suppression loss for joint training to constrain the spatial location of landslides, the prediction range of positive events, and the high probability coverage range of counterfactual events, thereby improving the stability of cross-event prediction.

[0014] 6. This invention outputs pixel-level probability maps and vector results with geographic coordinates, which can be used for post-earthquake emergency investigation, key area screening, disaster chain risk assessment, and cross-event model benchmark testing. Attached Figure Description

[0015] Figure 1 The flowchart of the intelligent prediction method for strong earthquake landslides based on event metadata provided by the present invention is shown.

[0016] Figure 2 This is a schematic diagram of the event metadata-driven multi-scale conditional affine modulation network structure provided by the present invention.

[0017] Figure 3 This is a schematic diagram of the prediction results provided by the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0019] The intelligent prediction method for strong earthquake landslides based on event metadata provided by this invention aims to organize historical lists, newly added high-resolution remote sensing interpretation lists, and widely unobserved landslide events from strong earthquakes into a unified database framework to construct an effective observation mask. During the label rasterization stage, soft boundary labels ranging from 0 to 1 are generated in one go. Moment magnitude, focal depth, and fault mechanism are encoded into event vectors. Channel scaling parameters and offset parameters at different network scales are generated from these event vectors to implement event-adaptive modulation of pixel-scale ground motion, topography, lithology, hydrology, and land cover features. Furthermore, event-scale landslide extent calibration and counterfactual event false alarm suppression are used to reduce large-area, high-probability false alarms in low-risk areas of strong earthquakes.

[0020] refer to Figure 1 A smart prediction method for strong earthquake landslides based on event metadata includes the following steps: S100. Select dual-temporal remote sensing images before and after the earthquake event to construct vector labels for landslides induced by strong earthquakes, and screen for widespread landslide events that were not observed during strong earthquakes. S200, combined with landslide triggering events and widely unobserved landslide events during strong earthquakes, construct a combined event database; S300: Extract predictors from the combined event database, perform feature engineering on them, and vectorize the available event metadata. Process the predictors, soft boundary labels, effective observation masks, event vectors, and event category identifiers into a tile dataset. S400: Construct a multi-scale conditional affine modulation network driven by event metadata, and based on the tile dataset, use event equalization pixel localization, event calibration and counterfactual event false alarm mechanism to jointly train the multi-scale conditional affine modulation network to obtain a strong earthquake-induced landslide intelligent prediction model. S500 uses a strong earthquake-induced landslide intelligent prediction model to predict earthquake events, combining a sliding window and edge-ignoring methods for spatial prediction, and then performs raster stitching on the prediction results to output a landslide occurrence probability map in vector form.

[0021] In S100 of this embodiment, acquiring dual-temporal remote sensing images before and after the earthquake event, constructing vector labels for landslides induced by strong earthquakes, and screening for widespread landslide events not observed during strong earthquakes are performed in the following order: candidate event screening, investigation scope determination, image interpretation, landslide label confirmation, and counterfactual event confirmation. Specifically, it includes the following sub-steps: S101. Select candidate strong earthquake events from the earthquake catalog that are located in land areas, have a focal depth not greater than a preset depth threshold, and whose occurrence time meets the conditions of high-resolution optical image coverage, and record the event number, occurrence time, epicenter location, moment magnitude, and focal depth. S102. Download the ground motion products corresponding to the candidate strong earthquake events, and take the area with peak ground acceleration (PGA) greater than 0.02g as the effective investigation range to obtain the spatial investigation boundary corresponding to the candidate strong earthquake events. S103. Load PlanetScope monthly composite images or pre-earthquake and post-earthquake optical images with equivalent spatial resolution within the boundary of the spatial survey range, and compare and interpret the changes in spectrum, texture, vegetation destruction and exposed surface; at the same time, construct an effective observation mask based on clouds, snow, topographic shadows, image missing, registration error and survey time phase to form a set of candidate landslide patches and reliable interpretation areas. S104. Combined with the three-dimensional terrain scene constructed by the digital elevation model, the morphological and kinematic rationality of the source area, transport area and deposition area of ​​the candidate landslide patches are checked within the effective observation mask. Existing landslides before the earthquake, bare riverbeds, agricultural disturbances and non-gravitational slope damage are eliminated to obtain the vector label of strong earthquake-induced landslides. S105. Based on the vector label of landslides induced by strong earthquakes, strong earthquake events that meet the strong earthquake threshold, image availability, and minimum effective observation ratio, and that do not meet the criteria for widely mapable landslides under the same survey protocol, are recorded as strong earthquake unobserved widespread landslide events.

[0022] Among them, the determination of widely unobserved landslide events in strong earthquakes is based on landslide data, total landslide area, landslide area density, spatial continuity and effective observation ratio in the vector label of landslides induced by strong earthquakes. The determined event does not mean that there are absolutely no landslides, but rather that no widely mappable coseismic landslides were observed under the preset image resolution, investigation time window, effective observation range and interpretation criteria.

[0023] In S200 of this embodiment, landslide triggering events (including the optimized historical list of landslides after strong earthquakes and the list of landslides after new strong earthquakes) and widely unobserved landslide events after strong earthquakes are integrated to construct a combined event database containing fields such as event number, event category, moment magnitude, focal depth, focal mechanism, survey range, effective observation mask, and ground motion product, and then the database is standardized.

[0024] In one specific embodiment, when establishing a combined event database consisting of landslide-triggered events and widely unobserved landslide events from strong earthquakes, the landslide-triggered events include a list of 39 historical landslides induced by strong earthquakes and a list of 24 new landslides induced by strong earthquakes. The widely unobserved landslide events from strong earthquakes include 42 strong earthquake events for which widely mappable coseismic landslides were not observed under a unified survey protocol.

[0025] The 42 events mentioned above are not pre-stored event records in PlanetScope, but are obtained based on the process in step S100, using the earthquake catalog as a candidate source, and after meeting the conditions of strong ground motion threshold and availability of high-resolution optical images, by performing consistency checks using PlanetScope monthly composite images or pre-earthquake and post-earthquake optical images with equivalent spatial resolution.

[0026] For the newly added list of earthquake-induced landslides, candidate events can be obtained from the earthquake catalog. Screening criteria include land area, moment magnitude range, focal depth, event time, image availability, and seismic intensity. For each candidate event, corresponding ground motion information is obtained, with areas having a ground acceleration greater than 0.02g designated as the scope for manual interpretation and investigation.

[0027] Within the survey area, acquire pre- and post-earthquake PlanetScope monthly composite images or other high-resolution optical images with equivalent spatial resolution. During interpretation, identify post-earthquake changes in tone, texture, and vegetation damage, and construct an effective observation mask based on clouds, snow, topographic shadows, image gaps, registration errors, and survey time phase; the effective observation mask is a binary variable used to exclude pixels that cannot be reliably interpreted.

[0028] To reduce misjudgments caused by shadows, bare land, riverbanks, agricultural disturbances, and existing landslides in 2D imagery, a 3D terrain scene was constructed by overlaying a digital elevation model in ArcGIS Pro. The rationality of the source area, transport zone, and deposition zone of candidate landslides was then checked. Confirmed coseismic landslides were stored as polygons, with the coordinate system being the WGS 1984 geographic coordinate system.

[0029] For strong earthquakes without observed widespread landslides, the same PGA threshold, image resolution, pre- and post-earthquake comparison, effective observation quality control, and 3D topographic inspection workflow as for landslide triggering events are employed. These events are not necessarily landslide-free, but rather events where no widespread mappable coseismic landslides were found under the applied image resolution, acquisition time, effective observation range, and interpretation criteria. Ordinary background pixels represent locations within landslide triggering events where no landslides occurred, while these events provide counterfactual monitoring at the event scale for events with strong ground motion but without widespread landslide formation.

[0030] After reviewing the 42 unobserved widespread landslide events from strong earthquakes using a common optical imagery survey protocol, and combining them with a list of 63 landslide events induced by strong earthquakes, the event numbers, event categories, moment magnitudes, focal depths, fault mechanisms, survey ranges, effective observation masks, and seismic products were standardized and unified to construct a standardized combined event database. The event category field was stored as a database label for training grouping, event set partitioning, and loss calculation, but was not used as network input for events to be predicted.

[0031] S300 of this embodiment includes the following steps: S301. Extract predictive factors from the effective survey area of ​​each event in the combined event database, including peak ground acceleration, slope, aspect, topographic relief, topographic location index, topographic humidity index, pre-earthquake cumulative rainfall, lithology and surface cover. S302. Perform unified rasterization, resampling and spatial alignment on the predictors, and perform feature redundancy diagnosis and screening based on the Pearson correlation coefficient matrix, variance inflation factor and tolerance. Combine the retained predictors as a multi-channel predictor stack. S303. Rasterize the landslide labels in the landslide triggering event, assign the landslide pixel value to 1, and assign the non-landslide background pixel value to 0 in the effective observation area; assign background labels to the widely observed landslide events of strong earthquakes in the area where the effective observation mask is 1, set ignore labels for unreliable interpretation areas, and convert them into soft boundary labels in the range of 0~1 according to the observation conditions. S304. The multi-channel predictor stack, labels, and effective observation masks are cropped into tiles of a set size with geographic coordinate information according to the sliding window, and the event number, event vector, and event category identifier of the tile are retained to construct the tile dataset.

[0032] In step S301, based on the aforementioned specific embodiments, the events in the combined event database include 63 strong earthquake-induced landslide events and 42 strong earthquake-unobserved widespread landslide events; among the extracted prediction factors, slope, aspect, topographic relief, topographic location index, and topographic humidity index are extracted by digital elevation model, peak ground acceleration is obtained from ground motion products, pre-earthquake cumulative rainfall is obtained from rainfall products, and lithology and surface cover are obtained from corresponding thematic data.

[0033] Specifically, the metadata descriptions of the above predictors and events are shown in Table 1.

[0034] Table 1: Metadata of each predictor and event

[0035] In step S302, the above-mentioned predictive factors are uniformly rasterized, resampled and spatially aligned according to the geographic and spatial information of the digital elevation model using the GDAL library. Before model training, feature redundancy diagnosis and screening are performed based on the Pearson correlation coefficient matrix, variance inflation factor and tolerance to reduce the impact of linear correlation. Then, the retained predictive factors are combined into a multi-channel predictive factor stack.

[0036] Specifically, in this embodiment, Pearson correlation coefficient, variance inflation factor, and tolerance are used to diagnose and screen predictor channels before model training. The calculation object is the predictor channel value corresponding to the sampled pixels in the effective observation area, which is used to identify and screen out highly correlated or strongly collinear input factors. This calculation is not used to measure the correlation between different seismic events or event vectors. The screening results are used to determine the subsequent multi-channel predictor stack.

[0037] In step S304, the multi-channel predictor stack is pruned according to a preset physical range and scaled to a uniform numerical range before input; categorical variables such as lithology and land cover are encoded with integers, and the nearest neighbor method is used during resampling to maintain category boundaries. All predictors, labels, and effective observation masks are aligned according to the same spatial reference, pixel size, and range. Subsequently, a 512×512 sliding window is used to prune the predictor stack, labels, and effective observation masks, while retaining the event number, event vector, and event category identifier of the tile.

[0038] S400 of this embodiment includes the following sub-steps: S401. For the tile dataset, map the moment magnitude, focal depth, and fault mechanism of each earthquake event into an event embedding vector to construct an earthquake event cue encoder. S402. Based on the earthquake event prompt encoder, a multi-scale conditional affine modulation network is constructed using a U-shaped encoder-decoder network as the pixel-level spatial feature extraction skeleton. S403. Calculate the event equilibrium location loss based on the effective observation mask and the soft boundary label in the range of 0~1; calculate the landslide triggering event range calibration loss based on the proportion of observed landslide area of ​​landslide triggering events; calculate the counterfactual event false alarm suppression loss based on the predicted high probability coverage proportion of widely unobserved landslide events during strong earthquakes; and construct a joint loss function by weighting the event equilibrium location loss, range calibration loss, and counterfactual event false alarm suppression loss. S404. Based on the joint loss function, the multi-scale conditional affine modulation network is trained using the data partitioning method of mutual exclusion of events in the tile dataset to obtain the intelligent prediction model of strong earthquake-induced landslides. Among them, event mutual exclusion refers to dividing the tile dataset according to the earthquake event number, so that all remote sensing tiles of the same earthquake event can only enter one set in the training set, validation set, or test set.

[0039] like Figure 2 As shown, the multi-scale conditional affine modulation network includes a feature-level linear modulation module, specifically: Feature-level linear modulation modules are set in at least two network scales after the downsampling layer, bottleneck feature layer and / or upsampling layer of the encoder and the feature fusion layer of the decoder, skipping over each other. In the feature-level linear modulation module, channel scaling parameters and channel offset parameters corresponding to the feature channels of the corresponding scale are generated based on the event embedding vector, and channel affine modulation is performed on the spatial features after normalization of the corresponding convolutional feature blocks.

[0040] Based on the above network structure, during the training process of the multi-scale conditional affine modulation network, training samples are used as input data, and event category identifiers are used to determine the event set to which the training samples belong, thereby outputting a landslide occurrence probability map, including: Read the multi-channel predictor stack and event vector from the training samples; The event vector is processed by the earthquake event cue encoder to output the event embedding vector. , The nonlinear mapping function corresponding to the earthquake event indicator encoder. The encoder prompts the input of earthquake events. The event vector of a strong earthquake event; In the earthquake event prompt encoder, the moment magnitude, focal depth and focal mechanism are spliced ​​together and then passed through two layers of multilayer perceptrons to output the event embedding vector. In the feature-level linear modulation module at each network scale, the corresponding channel scaling parameters and channel offset parameters are generated based on the event embedding vector. The multi-channel predictor stack is convolved and upsampled scale by scale, and the upsampled features are fused with the skip connection features of the same scale at the decoder end. The spatial features are output by the corresponding convolutional feature block and normalized. The spatial features are affine modulated channel by channel based on the channel scaling and channel offset parameters. The modulated features at each scale are fed into the decoder through skip connections to form fused features, and the fused features are directly input into the prediction head, which is activated by Sigmoid to obtain a landslide occurrence probability map. The training samples belong to an event set that includes a landslide-triggered event set and a set of widely observed landslide events that were not observed during strong earthquakes. Each training sample includes a multi-channel predictor stack, labels corresponding one-to-one with the spatial locations of the multi-channel predictor stack, an effective observation mask, a label matrix, an event vector, and an event category identifier, represented as follows:

[0041] In the formula, For the first The first of the strong earthquake events The training samples corresponding to each tile For the first The first of the strong earthquake events The multi-channel predictor stack corresponding to each tile To and Landslide labels with one-to-one spatial location For the first The first of the strong earthquake events The effective observation mask corresponding to each tile For the first The first of the strong earthquake events The label matrix corresponding to each tile The encoder prompts the input of earthquake events. The event vector of a strong earthquake event. For the first Event category identifier for a strong earthquake event.

[0042] Event Vector It does not include event outcome categories, but only includes the standardized moment magnitude and focal depth, as well as the focal mechanism type after one-heat coding or embedding coding, which is represented as:

[0043] In the formula, The encoder prompts the input of earthquake events. The event vector of a strong earthquake event. For the standardized version of the first Moment magnitude of a strong earthquake event. For the standardized version of the first The focal depth of a strong earthquake event. For the first The type of focal mechanism of a strong earthquake event.

[0044] In the aforementioned multi-scale conditional affine modulation network, for the first... For each network scale, in the corresponding feature-level linear modulation module, the channel scaling parameters and channel offset parameters generated from the event embedding vector are expressed as follows:

[0045]

[0046] In the formula, According to the first Channel scaling parameters generated from the event embedding vectors of a strong earthquake event. For the corresponding channel offset parameter; and The first Independent trainable linear projection weight matrices for scaling and offset branches at each network scale. and These are independent trainable bias vectors for the scaling branch and the offset branch, respectively; The encoder for earthquake event alerts is for the first The event embedding vector output by a strong earthquake event Index of strong earthquake events For network-scale indexing.

[0047] The spatial features are represented by channel affine modulation as follows:

[0048] In the formula, For the first After event modulation corresponding to the first strong earthquake event, the th The tile in the first Output feature tensor at the network scale For the first The first strong earthquake event The tile in the first Spatial features at the network scale before event condition modulation. According to the first Channel scaling parameters generated from the event embedding vectors of a strong earthquake event. According to the first Channel offset parameters generated from the event embedding vector of a strong earthquake event. For the first Tile index inside a strong earthquake event Element-wise multiplication by channel This is a normalization operation. Event modulation refers to the process based on event vectors, event embedding vectors, channel scaling parameters, channel offset parameters, and characteristic linear modulation.

[0049] During the training of a multi-scale conditional affine modulation network, for each output landslide probability map, class weights are calculated to mitigate the impact of the imbalance between the number of landslide pixels and background pixels on the training.

[0050] In the landslide occurrence probability diagram, the first... Class weights of each pixel for:

[0051]

[0052] In the formula, For the first A label of 1 pixel, For positive pixel weights, For negative pixel weights, The total number of valid pixels included in the statistics. and These represent the number of positive and negative pixels in the valid pixels, respectively. To prevent stable terms with a denominator of zero.

[0053] Among them, positive pixels are landslide pixels with a label value of 1 in the effective observation area, and negative pixels are background pixels with a label value of 0 in the effective observation area; pixels with an ignore label or an effective observation mask of 0 are not included in the category count.

[0054] In this embodiment, during the training of the multi-scale conditional affine modulation network, the event equalization localization loss is used. for:

[0055]

[0056] In the formula, For the first The first strong earthquake event Label-weighted pixel localization loss for each tile For the first The number of tiles in a strong earthquake event, This represents the number of strong earthquake events in the current training batch. For the first The first strong earthquake event The first of the tiles An effective observation mask for each pixel. For the first The first strong earthquake event The first of the tiles A label of 1 pixel, For the first The first strong earthquake event The first of the tiles Class weights for each pixel, and The first The first strong earthquake event The first of the tiles The predicted probability and label value of each pixel. For binary cross-entropy, To prevent stable terms with a denominator of zero, for The first strong earthquake event A set of pixels for each tile.

[0057] By using the Huber function to constrain the deviation between the predicted high-probability coverage ratio and the observed landslide area ratio, the calibration loss for the landslide triggering event range is obtained. for:

[0058]

[0059]

[0060] In the formula, For Huber functions, For the first The high probability coverage ratio of predictions for major earthquake events. For the first The proportion of observed landslide area in each strong earthquake event. To predict the discrepancy between the high-probability coverage ratio and the observed landslide area ratio, The deviation threshold for converting from quadratic penalty to linear penalty. This is a set of landslide-triggered events. Among them, the High probability coverage of predictions for major earthquake events The calculation formula is:

[0061]

[0062] In the formula, It is a differentiable soft threshold function. The steepness of the threshold function, To determine the probability threshold for high-probability coverage areas, To prevent stable terms with a denominator of zero; For the set of landslide triggering events, the first... The proportion of observed landslide area in each major earthquake event The calculation formula is:

[0063] For a ensemble of widely unobserved landslide events following strong earthquakes, limiting the predicted high-probability coverage exceeding an allowable threshold yields a counterfactual event false alarm mitigation loss. for:

[0064] In the formula, This is a collection of widely unobserved landslide events following a strong earthquake. This is the allowable threshold.

[0065] In summary, the joint loss function of the multi-scale conditional affine modulation network is obtained. for:

[0066] In the formula, , , These are the event equilibrium location loss, landslide-triggered event range calibration loss, and counterfactual event false alarm suppression loss, respectively.

[0067] In the aforementioned joint loss function, the event equalization localization loss constitutes the pixel localization term of this invention through intra-event tile averaging and inter-event equal-weight averaging. The counterfactual event false alarm suppression loss constitutes the false alarm suppression term of this invention for the high-probability coverage exceeding the threshold of widespread landslide events not observed during strong earthquakes. The landslide triggering event range calibration loss, based on the Huber function, modifies the penalty object to the event-level deviation between the predicted high-probability coverage ratio and the observed landslide area ratio. These three losses respectively address the problems of pixel location recognition, overall range calibration of landslide triggering events, and large-area false alarm suppression of widespread landslide events not observed during strong earthquakes. Geometric augmentation such as rotation, flipping, and transposition can be used during training.

[0068] In this embodiment, a mutually exclusive training set, validation set, and test set are used to prevent spatial tiles of the same seismic event from simultaneously entering the model development set and the test set. Furthermore, the optimizer can employ AdamW, combined with learning rate scheduling, gradient pruning, mixed-precision training, and early stopping strategies.

[0069] S500 of this embodiment includes the following steps: For predicting earthquake events, a multi-channel predictor stack is constructed, and an event vector is generated based on the moment magnitude, focal depth, and fault mechanism of the earthquake event to be predicted. A fixed-size sliding window is used to prune the multi-channel predictor stack to obtain multi-window tiles; among them, there are overlapping areas between consecutive windows; The tiles and their corresponding event vectors are combined to form input samples, which are then input into the intelligent prediction model for landslides induced by strong earthquakes, and the local probability map of each window is output. By setting the edge ignore width and stitching together the local probability maps, a landslide occurrence probability map can be obtained.

[0070] Specifically, in this embodiment, for the earthquake event to be predicted, a multi-channel predictive factor stack is constructed, and an event vector is generated based on the available moment magnitude, focal depth, and fault mechanism of the event. It is not necessary to know in advance whether the event will result in widespread landslides, nor is the event outcome category input into the subsequent model. Subsequently, a fixed-size sliding window is used to prune the large-scale factor stack, and overlapping areas are set between windows to ensure that pixels near the boundary can be covered by the effective central area of ​​at least one window.

[0071] The intelligent prediction model for landslides induced by strong earthquakes outputs a local probability map for each window. Since the intelligent prediction model for landslides induced by strong earthquakes has insufficient available context at the edge of the tile, direct stitching will cause seam artifacts. In this embodiment, the edge width b is ignored, and the predicted value within the range of b pixels at the edge of each tile is assigned 0 weight or a low weight, and only the central region is used for stitching.

[0072] The output is a raster of probabilities of landslides induced by strong earthquakes. Figure 3 Vector files inherit the spatial reference of the input factor stack and can store attributes such as event number, threshold, predicted probability statistics, area, and post-processing parameters for use in post-earthquake emergency investigation and GIS platform display.

[0073] In one specific embodiment, 63 strong earthquake-induced landslide events and 42 strong earthquake-unobserved widespread landslide events were divided into 21 mutually exclusive event groups. Each event group contained 3 strong earthquake-induced landslide events and 2 strong earthquake-unobserved widespread landslide events. In each round of evaluation, one event group was reserved as the test group, and the remaining event groups were used for model development, thus obtaining 21 rounds of outer-layer event grouping cross-validation. This evaluation method can avoid event leakage caused by the simultaneous appearance of spatial tiles of the same earthquake event in both training and testing.

[0074] The above evaluation data is used to illustrate the technical effects of the present invention and should not be construed as limiting specific datasets, thresholds, hardware environments, or model parameters. Those skilled in the art can adjust the training event combinations, tile sizes, network channel numbers, thresholds, and post-processing parameters according to the area to be predicted, available imagery, seismic products, and computing resources.

[0075] In summary, this invention, through a combined event database under a unified observation protocol, counterfactual constraints on widespread unobserved landslide events during strong earthquakes, multi-scale feature-level linear modulation driven by event metadata, effective label weighting, event-balanced pixel localization, landslide-triggered event range calibration, and counterfactual event false alarm suppression, jointly utilizes ground motion information, topographic environment information, and event-scale source information under the condition of unknown event outcome categories. This can reduce large-area false alarms in low-risk areas of strong earthquakes and improve the model's cross-event prediction stability. All equivalent substitutions, combinations, or improvements made to the database source, factor resolution, number of modulation scales, network backbone, loss weights, window size, and post-processing algorithm within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0076] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0077] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for intelligent prediction of strong earthquake landslides driven by event metadata, characterized in that, Includes the following steps S100. Select dual-temporal remote sensing images before and after the earthquake event to construct vector labels for landslides induced by strong earthquakes, and screen for widespread landslide events that were not observed during strong earthquakes. S200, combined with landslide triggering events and widely unobserved landslide events during strong earthquakes, construct a combined event database; S300: Extract predictors from the combined event database, perform feature engineering on them, and vectorize the available event metadata. Process the predictors, soft boundary labels, effective observation masks, event vectors, and event category identifiers into a tile dataset. S400: Construct a multi-scale conditional affine modulation network driven by event metadata, and based on the tile dataset, use event equalization pixel localization, event calibration and counterfactual event false alarm mechanism to jointly train the multi-scale conditional affine modulation network to obtain a strong earthquake-induced landslide intelligent prediction model. S500 uses a strong earthquake-induced landslide intelligent prediction model to predict earthquake events, combining a sliding window and edge-ignoring methods for spatial prediction, and then performs raster stitching on the prediction results to output a landslide occurrence probability map in vector form.

2. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 1, characterized in that, S100 includes the following steps: S101. Select candidate strong earthquake events from the earthquake catalog that are located in land areas, have a focal depth not greater than a preset depth threshold, and whose occurrence time meets the conditions of high-resolution optical image coverage, and record the event number, occurrence time, epicenter location, moment magnitude, and focal depth. S102. Download the ground motion products corresponding to the candidate strong earthquake events, and take the area with peak ground acceleration (PGA) greater than 0.02g as the effective investigation range to obtain the spatial investigation boundary corresponding to the candidate strong earthquake events. S103. Load PlanetScope monthly composite images or pre-earthquake and post-earthquake optical images with equivalent spatial resolution within the boundary of the spatial survey range, and compare and interpret the changes in spectrum, texture, vegetation destruction and exposed surface; at the same time, construct an effective observation mask based on clouds, snow, topographic shadows, image missing, registration error and survey time phase to form a set of candidate landslide patches and reliable interpretation areas. S104. Combined with the three-dimensional terrain scene constructed by the digital elevation model, the morphological and kinematic rationality of the source area, transport area and deposition area of ​​the candidate landslide patches are checked within the effective observation mask. Existing landslides before the earthquake, bare riverbeds, agricultural disturbances and non-gravitational slope damage are eliminated to obtain the vector label of strong earthquake-induced landslides. S105. Based on the vector label of landslides induced by strong earthquakes, strong earthquake events that meet the strong earthquake threshold, image availability, and minimum effective observation ratio, and that do not meet the criteria for widely mapable landslides under the same survey protocol, are recorded as strong earthquake unobserved widespread landslide events.

3. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 2, characterized in that, S300 includes the following steps: S301. Extract predictive factors from the effective survey area of ​​each event in the combined event database, including peak ground acceleration, slope, aspect, topographic relief, topographic location index, topographic humidity index, pre-earthquake cumulative rainfall, lithology and surface cover. S302. Perform unified rasterization, resampling and spatial alignment on the predictors, and perform feature redundancy diagnosis and screening based on the Pearson correlation coefficient matrix, variance inflation factor and tolerance. Combine the retained predictors as a multi-channel predictor stack. S303. Rasterize the landslide labels in the landslide triggering event, assign the landslide pixel value to 1, and assign the non-landslide background pixel value to 0 in the effective observation area; assign background labels to the widely observed landslide events of strong earthquakes in the area where the effective observation mask is 1, set ignore labels for unreliable interpretation areas, and convert them into soft boundary labels in the range of 0~1 according to the observation conditions. S304. The multi-channel predictor stack, labels, and effective observation masks are cropped into tiles of a set size with geographic coordinate information according to the sliding window, and the event number, event vector, and event category identifier of the tile are retained to construct the tile dataset.

4. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 1, characterized in that, S400 includes the following steps: S401. For the tile dataset, map the moment magnitude, focal depth, and fault mechanism of each earthquake event into an event embedding vector to construct an earthquake event cue encoder. S402. Based on the earthquake event prompt encoder, a multi-scale conditional affine modulation network is constructed using a U-shaped encoder-decoder network as the pixel-level spatial feature extraction skeleton. S403. Calculate the event equilibrium location loss based on the effective observation mask and the soft boundary label in the range of 0~1; calculate the landslide triggering event range calibration loss based on the proportion of observed landslide area of ​​landslide triggering events; calculate the counterfactual event false alarm suppression loss based on the predicted high probability coverage proportion of widely unobserved landslide events during strong earthquakes; and construct a joint loss function by weighting the event equilibrium location loss, range calibration loss, and counterfactual event false alarm suppression loss. S404. Based on the joint loss function, the multi-scale conditional affine modulation network is trained using the data partitioning method of mutual exclusion of events in the tile dataset to obtain the intelligent prediction model of strong earthquake-induced landslides. The event mutual exclusion refers to dividing the tile dataset according to the earthquake event number, so that all remote sensing tiles of the same earthquake event can only enter one set in the training set, validation set, or test set.

5. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 4, characterized in that, In step S402, the multi-scale conditional affine modulation network includes a feature-level linear modulation module, specifically: Feature-level linear modulation modules are set in at least two network scales after the downsampling layer, bottleneck feature layer and / or upsampling layer of the encoder and the feature fusion layer of the decoder, skipping over each other. In the feature-level linear modulation module, channel scaling parameters and channel offset parameters corresponding to the feature channels of the corresponding scale are generated based on the event embedding vector, and channel affine modulation is performed on the spatial features after normalization of the corresponding convolutional feature blocks.

6. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 4, characterized in that, In S403, the event equalization positioning loss for: In the formula, For the first The first strong earthquake event Label-weighted pixel localization loss for each tile For the first The number of tiles in a strong earthquake event, This represents the number of strong earthquake events in the current training batch. For the first The first strong earthquake event The first of the tiles An effective observation mask for each pixel. For the first The first strong earthquake event The first of the tiles A label of 1 pixel, For the first The first strong earthquake event The first of the tiles Class weights for each pixel, and The first The first strong earthquake event The first of the tiles The predicted probability and label value of each pixel. For binary cross-entropy, To prevent stable terms with a denominator of zero, for The first strong earthquake event A collection of pixels for each tile; The calibration loss of the landslide triggering event range for: In the formula, For Huber functions, For the first The high probability coverage ratio of predictions for major earthquake events. For the first The proportion of observed landslide area in each strong earthquake event. To predict the discrepancy between the high-probability coverage ratio and the observed landslide area ratio, The deviation threshold for converting from quadratic penalty to linear penalty. This is a set of landslide-triggered events. The counterfactual event false alarm suppresses loss for: In the formula, This is a collection of widely unobserved landslide events following a strong earthquake. This is the allowable threshold.

7. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 4, characterized in that, In step S404, during the training process of the multi-scale conditional affine modulation network, training samples are used as input data, and the event set to which the training samples belong is determined using event category identifiers, thereby outputting a landslide occurrence probability map, including: Read the multi-channel predictor stack and event vector from the training samples; The event vector is processed by the earthquake event cue encoder to output the event embedding vector; In the feature-level linear modulation module at each network scale, the corresponding channel scaling parameters and channel offset parameters are generated based on the event embedding vector. The multi-channel predictor stack is convolved and upsampled scale by scale, and the upsampled features are fused with the skip connection features of the same scale at the decoder end. The spatial features are output by the corresponding convolutional feature block and normalized. The spatial features are affine modulated channel by channel based on the channel scaling and channel offset parameters. The modulated features at each scale are fed into the decoder through skip connections to form fused features, and the fused features are directly input into the prediction head, which is activated by Sigmoid to obtain a landslide occurrence probability map. The training samples belong to an event set that includes a landslide-triggered event set and a set of widely observed landslide events that were not observed during strong earthquakes. Each training sample includes a multi-channel predictor stack, labels corresponding one-to-one with the spatial location of the multi-channel predictor stack, an effective observation mask, a label matrix, an event vector, and an event category identifier; The event vector includes the normalized moment magnitude and source depth, as well as the source mechanism type coded either by hot coding or embedding coding.

8. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 7, characterized in that, In the multi-scale conditional affine modulation network, for the first... For each network scale, in the corresponding feature-level linear modulation module, the channel scaling parameters and channel offset parameters generated from the event embedding vector are expressed as follows: In the formula, According to the first Channel scaling parameters generated from the event embedding vectors of a strong earthquake event. For the corresponding channel offset parameter; and The first Independent trainable linear projection weight matrices for scaling and offset branches at each network scale. and These are independent trainable bias vectors for the scaling branch and the offset branch, respectively; The encoder for earthquake event alerts is for the first The event embedding vector output by a strong earthquake event Index of strong earthquake events For network-scale indexing; The spatial features are represented by channel affine modulation as follows: In the formula, For the first After event modulation corresponding to the first strong earthquake event, the th The tile in the first Output feature tensor at the network scale For the first The first strong earthquake event The tile in the first Spatial features at the network scale before event condition modulation. According to the first Channel scaling parameters generated from the event embedding vectors of a strong earthquake event. According to the first Channel offset parameters generated from the event embedding vector of a strong earthquake event. For the first Tile index inside a strong earthquake event For channel-by-channel element-wise multiplication, This is a normalization operation.

9. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 7, characterized in that, During the training of a multi-scale conditional affine modulation network, for each output landslide probability map, the impact of the imbalance between the number of landslide pixels and background pixels on the training is mitigated by calculating class weights. In the landslide occurrence probability diagram, the first... Class weights of each pixel for: In the formula, For the first A label of 1 pixel, For positive pixel weights, For negative pixel weights, The total number of valid pixels included in the statistics. and These represent the number of positive and negative pixels in the valid pixels, respectively. To prevent stable terms with a denominator of zero.

10. The intelligent prediction method for strong earthquake landslides based on event metadata as described in claim 1, characterized in that, The S500 includes the following steps: For predicting earthquake events, a multi-channel predictor stack is constructed, and an event vector is generated based on the moment magnitude, focal depth, and fault mechanism of the earthquake event to be predicted. A fixed-size sliding window is used to prune the multi-channel predictor stack to obtain multi-window tiles; among them, there are overlapping areas between consecutive windows; The tiles and their corresponding event vectors are combined to form input samples, which are then input into the intelligent prediction model for landslides induced by strong earthquakes, and the local probability map of each window is output. By setting the edge ignore width and stitching together the local probability maps, a landslide occurrence probability map can be obtained.

Images