A regional landslide prediction method based on static dynamic heterogeneous feature learning

Abstract

The application discloses a regional landslide prediction method based on static and dynamic heterogeneous feature learning, specifically comprising the following steps: obtaining and preprocessing static geological environment factors and dynamic rainfall time series data of a target region; inputting the data into a static feature encoder and a time series feature encoder respectively for independent representation learning; generating scaling and translation coefficients based on static geological representation to linearly modulate features of dynamic rainfall representation bit by bit; splicing the modulated space-time interaction features and original static features without extraction to construct a hybrid representation vector; inputting the hybrid representation vector into a prediction head to output a landslide occurrence probability of the target region. The application restores the real physical adjustment effect of the geological environment on rainfall response by decoupling feature extraction and adaptive feature modulation, thereby improving the stability and reliability of regional landslide prediction under extreme rainfall conditions.

Description

Technical Field

[0001] This invention pertains to geological disaster prediction technology, and particularly relates to a regional landslide prediction method based on static and dynamic heterogeneous feature learning. Background Technology

[0002] Regional landslide prediction plays a crucial role in disaster prevention and mitigation, and in protecting lives and property. Especially against the backdrop of frequent extreme weather events and increased short-duration heavy rainfall, accurate prediction of the probability of regional landslide disasters is a core support for achieving early warning and emergency deployment.

[0003] Against the backdrop of frequent extreme weather events globally, landslides triggered by short-duration heavy rainfall are becoming increasingly serious geological hazards. Current regional landslide prediction methods are mainly classified into the following categories, but all have significant shortcomings:

[0004] 1. Empirical rainfall threshold method based on meteorological triggering factors: It fails to fully consider the spatial heterogeneity of the geological environment and is prone to generating a high false alarm rate in areas with stable terrain.

[0005] 2. Machine learning methods based on static environmental variables: These are mainly used to generate static landslide susceptibility maps, but they are difficult to capture dynamic rainfall-triggered processes and cannot provide specific instability times.

[0006] 3. Physical mechanism model: The calculation process is rigorous, but it requires intensive field infiltration experiments and indoor direct shear tests to obtain soil mechanical parameters. When applied in vast mountainous areas, it is limited by the cost of geotechnical testing and the difficulty of parameter calibration.

[0007] 4. Traditional data-driven machine learning techniques: These techniques calculate static topographic and geological factors of the target area and directly tabulate them with statistical characteristics of previous rainfall. This direct concatenation strategy treats environmental susceptibility and meteorological triggering factors as independent inputs, failing to preserve the numerical characteristics of heterogeneous data and severely limiting its ability to characterize internal physical interaction mechanisms such as the moderating effect of topography on rainfall and hydrological responses.

[0008] However, most of the aforementioned prediction methods have significant shortcomings when dealing with multi-source heterogeneous data. The current mainstream model processing flow is typically: "static spatial geological factors + dynamic rainfall time series → direct concatenation (Concat) → input into the model for prediction." This simplistic and crude data concatenation method leads to severe mixing of heterogeneous information, not only obliterating the unique numerical distribution characteristics of both spatial background data and time series data, but also resulting in a severe lack of expression of static and dynamic features, making it impossible to mathematically accurately characterize the interaction mechanism between the geological environment and the rainfall triggering process.

[0009] Therefore, there is an urgent need in this field for a new method that learns the features of static spatial factors and dynamic rainfall sequences separately, in order to solve the problem of limited feature expression caused by direct splicing of existing heterogeneous data. Summary of the Invention

[0010] To overcome the problems of heterogeneous information interference and insufficient feature representation caused by directly splicing and fusing static geological environmental factors with dynamic rainfall time series in existing technologies, this invention provides a regional landslide prediction method based on static and dynamic heterogeneous feature learning. By decoupling feature extraction and adaptive feature modulation, the true physical regulation effect of the geological environment on rainfall response is restored, thereby improving the stability and reliability of regional landslide prediction under extreme rainfall conditions.

[0011] The present invention provides a regional landslide prediction method based on static and dynamic heterogeneous feature learning, comprising the following steps:

[0012] Step 1: Construction of multi-source heterogeneous data: Obtain static geological environmental factors and dynamic rainfall time series data of the target area and preprocess them.

[0013] The target prediction area is divided into grids; static spatial environmental factors of each grid node in the target area are extracted as static inputs; at the same time, continuous dynamic rainfall time series data of the corresponding grid node at T time steps before the target prediction time are obtained as dynamic inputs.

[0014] Step 2: Independent coding in a static environment.

[0015] Construct a dedicated static feature encoder to process the static spatial environment factors obtained in step 1. A linear transformation is performed to map it to scaling factors γ and offset factors β for subsequent feature modulation; the specific calculation formula is as follows:

[0016] ;

[0017] ;

[0018] In the formula, , These are the learnable weight matrices used in the static feature encoder to generate scaling and offset coefficients, respectively. , These are the learnable bias vectors used in the static feature encoder to generate scaling and offset coefficients, respectively.

[0019] Step 3: Independent coding of dynamic rainfall.

[0020] A dedicated dynamic sequence encoder is constructed, and the dynamic rainfall time series of length T obtained in step 1 is input into a stacked two-layer Long Short-Term Memory (LSTM) network. The first LSTM layer extracts short-term high-frequency rainfall intensity fluctuation features, and the second LSTM layer extracts long-term precipitation accumulation trend features. Subsequently, a time attention mechanism is introduced to calculate the hidden state at each time step. Attention score and normalized weights :

[0021] ;

[0022] ;

[0023] The hidden states are weighted and summed to generate a high-dimensional dynamic rainfall feature vector F:

[0024] ;

[0025] In the formula, T represents the total number of time steps in the dynamic rainfall time series; t and i are both time step indices. ; Let be the hidden state vector output by the two-layer LSTM network at time step t; This represents the learnable weight matrix in the temporal attention mechanism; This is a learnable bias term in the time attention mechanism; It is the hyperbolic tangent activation function; Let be the unnormalized attention score at time step t; It is an exponential function with the natural constant as its base; is the normalized attention weight at time step t; F is the final generated high-dimensional dynamic rainfall feature vector.

[0026] Step 4: Feature linear modulation: Based on the static geological characterization, scaling and translation coefficients are generated to perform feature-by-bit linear modulation on the dynamic rainfall characterization.

[0027] Using the static scaling factor γ and offset factor β generated in step 2, feature-level linear modulation is performed on the dynamic rainfall feature vector F generated in step 3; through element-level multiplication... The addition operation enables the static spatial background to adaptively scale and offset dynamic temporal features, generating modulation features. :

[0028] ;

[0029] Step 5: Wide and Deep Feature Fusion: The modulated spatiotemporal interactive features are concatenated with the unextracted original static features to construct a hybrid representation vector.

[0030] Using a Wide and Deep fusion architecture, the modulation features output from step 4 are... The Deep path and the original static spatial environment factors and dynamic rainfall time series from step 1 are used as the Wide path for channel-level splicing; by retaining the original input channels that have not been nonlinearly smoothed, a comprehensive feature space containing the original extreme value signal and deep modulation interaction information is constructed.

[0031] Step 6: Landslide probability classification prediction.

[0032] The high-dimensional integrated feature vector fused in step 5 is input into a machine learning classifier. The classifier is used to fit the nonlinear decision boundary of the multidimensional feature space, and finally outputs the probability prediction value of the rain-induced landslide at the grid node.

[0033] Furthermore, static inputs include, but are not limited to, elevation, slope, aspect, curvature, topographic humidity index, topographic ruggedness index, lithology, soil thickness, and distance from faults / rivers / roads.

[0034] Furthermore, step 7 is included: introducing the SHAP algorithm to calculate the marginal contribution of variables, generating a response surface and performing interpretability analysis.

[0035] The beneficial technical effects of this invention compared to the prior art are as follows:

[0036] (1) Complete preservation of static feature distribution: By establishing an independent static environment coding path, the original numerical distribution characteristics of spatial background data such as topography and geology as the landslide susceptibility baseline are effectively maintained, avoiding dilution by dynamic features in the early feature splicing.

[0037] (2) Precisely preserve dynamic sequence features: The two-layer LSTM and time attention mechanism are used to process the rainfall time series independently, which ensures that the model can keenly capture the short-term extreme precipitation peak signal and prevents excessive smoothing of key trigger information.

[0038] (3) Effectively reduce interference from heterogeneous information: By using feature-level linear modulation (FiLM) to replace the traditional direct data splicing, the "regulatory effect of static geological environment on dynamic hydrological process" is simulated, so that static data can adaptively control the expression of dynamic features, fundamentally eliminating the direct numerical conflict and interference of spatiotemporal heterogeneous data.

[0039] (4) Significantly improve prediction stability: Thanks to the modular recurrent neural network memory mechanism and Wide and Deep architecture, the model can still rely on the historical hidden state to smoothly decay in the "disaster decay lag period" after the end of rainfall, providing a stable risk estimate output and preventing the early warning signal from failing prematurely when the rainfall stops.

[0040] (5) Decoupled architecture for easy and flexible expansion: The feature extraction module and the prediction head of this invention are completely decoupled, which has a high degree of modularity. In practical business applications, it can be easily extended and connected to more dimensions of continuous environmental observation data such as radar precipitation inversion, soil moisture monitoring, and surface runoff index, and has excellent engineering scalability. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the execution of the regional landslide prediction method based on static and dynamic heterogeneous feature learning according to the present invention.

[0042] Figure 2 This is a block diagram illustrating the principle of the prediction model network architecture of this invention.

[0043] Figure 3 This is a block diagram illustrating the principle of the feature modulation module of the present invention.

[0044] Figure 4 This is a schematic diagram of the dynamic temporal feature learning module of the present invention. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0046] The flowchart of a regional landslide prediction method based on static and dynamic heterogeneous feature learning according to the present invention is as follows: Figure 1 As shown, Figure 1 The document demonstrates the complete execution steps from multi-source heterogeneous data acquisition and construction, independent representation learning of static and dynamic features, feature-by-bit linear modulation based on geological features, to wide & deep hybrid feature construction and final output of landslide risk prediction probability. Specifically, it includes the following steps:

[0047] Step 1: Obtain and preprocess static geological environmental factors and dynamic rainfall time series data of the target area.

[0048] The target prediction area is divided into grids; static spatial environmental factors of each grid node in the target area are extracted as static inputs; at the same time, continuous dynamic rainfall time series data of the corresponding grid node at T time steps before the target prediction time are obtained as dynamic inputs.

[0049] Static inputs include, but are not limited to, elevation, slope, aspect, curvature, topographic humidity index, topographic ruggedness index, lithology, soil thickness, and distance from faults / rivers / roads.

[0050] The predictive model network architecture of this invention is as follows: Figure 2 As shown, Figure 2The paper demonstrates the overall network hierarchy of the prediction model and its data flow process, which mainly includes the input layer, encoding layer, modulation layer, fusion layer and prediction layer. It reflects the complete computational logic of static spatial factors and dynamic rainfall sequences from independent extraction, feature modulation to hybrid splicing.

[0051] Step 2: Input the data into the static feature encoder and the temporal feature encoder respectively for independent representation learning.

[0052] Step 2.1: Independent coding in static environment.

[0053] Construct a dedicated static feature encoder to process the static spatial environment factors obtained in step 1. A linear transformation is performed, mapping it to scaling coefficient γ and shifting coefficient β for subsequent feature modulation; the specific calculation formula is as follows:

[0054] ;

[0055] ;

[0056] In the formula, , These are the learnable weight matrices used in the static feature encoder to generate scaling and offset coefficients, respectively. , These are the learnable bias vectors used in the static feature encoder to generate scaling and offset coefficients, respectively.

[0057] Step 2.2: Independent coding of dynamic rainfall.

[0058] Dynamic temporal feature learning module, such as Figure 4 As shown, Figure 4 The paper demonstrates a cascaded structure of a stacked two-layer long short-term memory network (LSTM) and a temporal attention mechanism, which is used to independently extract long-term cumulative trends and short-term high-frequency extreme value fluctuations in precipitation time series.

[0059] Specifically, a dedicated dynamic sequence encoder is constructed, and the dynamic rainfall time series of length T obtained in step 1 is input into a stacked two-layer Long Short-Term Memory (LSTM) network. The first LSTM layer extracts short-term high-frequency rainfall intensity fluctuation features, and the second LSTM layer extracts long-term precipitation accumulation trend features. Subsequently, a time attention mechanism is introduced to calculate the hidden state at each time step. Attention score and normalized weights :

[0060] ;

[0061] ;

[0062] The hidden states are weighted and summed to generate a high-dimensional dynamic rainfall feature vector F:

[0063] ;

[0064] In the formula, T represents the total number of time steps in the dynamic rainfall time series; t and i are both time step indices. ; Let be the hidden state vector output by the two-layer LSTM network at time step t; This represents the learnable weight matrix in the temporal attention mechanism; This is a learnable bias term in the time attention mechanism; It is the hyperbolic tangent activation function; Let be the unnormalized attention score at time step t; It is an exponential function with the natural constant as its base; is the normalized attention weight at time step t; F is the final generated high-dimensional dynamic rainfall feature vector.

[0065] Step 3: Based on the static geological characterization, scale and translation coefficients are generated to perform feature-by-feature linear modulation on the dynamic rainfall characterization.

[0066] The principle of the feature modulation module is as follows: Figure 3 As shown, Figure 3 The paper demonstrates the specific mathematical operation process of how static environmental factors generate dedicated scaling and offset coefficients through a linear transformation mapping layer, and how they perform element-level adaptive modulation on dynamic rainfall feature vectors.

[0067] Specifically, using the static scaling factor γ and offset factor β generated in step 2.1, feature-level linear modulation is performed on the dynamic rainfall feature vector F generated in step 2.2; element-level multiplication is then used. The addition operation enables the static spatial background to adaptively scale and offset dynamic temporal features, generating modulation features. :

[0068] ;

[0069] Step 4: Concatenate the modulated spatiotemporal interaction features with the unextracted original static features to construct a hybrid representation vector.

[0070] Using a Wide and Deep fusion architecture, the modulation features output from step 3 are... The Deep path, along with the original static spatial environment factors and dynamic rainfall time series from step 1, are used as the Wide path for channel-level concatenation. By retaining the original input channels that have not undergone nonlinear smoothing, a comprehensive feature space containing the original extreme value signal and deep modulation interaction information is constructed.

[0071] Step 5: Input the prediction header and output the probability of landslide occurrence in the target area.

[0072] The high-dimensional integrated feature vector fused in step 4 is input into a machine learning classifier (e.g., an XGBoost prediction head). The classifier is used to fit the nonlinear decision boundary of the multidimensional feature space, and finally outputs the probability prediction value of the rain-induced landslide at the grid node.

[0073] Step 6: Introduce the SHAP algorithm to calculate the marginal contribution of variables, generate the response surface, and perform interpretability analysis.

[0074] Example 1: Landslide prediction in Zixing City based on extreme rainfall from Typhoon "Gemi".

[0075] Taking the actual rainfall and landslide disaster data of Zixing City, Hunan Province during Typhoon Gemi in 2024 as an example, the specific operation steps, process parameters, and test results of the prediction method of this invention are explained:

[0076] 1. Detailed steps and process parameters.

[0077] Step 1: Spatial data sampling and heterogeneous data construction.

[0078] The multi-source heterogeneous spatial database required to construct the prediction model of this invention, and the specific data sources and processing methods are as follows:

[0079] (1) Landslide label data (positive sample construction): Obtain 3-meter resolution optical remote sensing raster images (e.g., Planet Dove satellite images) of the target area before and after the disaster, extract the real landslide contour surface through manual interpretation, and generate SHP vector files of the landslide boundary. When constructing positive samples, the geometric center of the landslide source area in the SHP file is used as the spatial sampling point of the static environmental factor.

[0080] (2) Static environmental factor data (basic susceptibility quantification):

[0081] Topographic features: Using a 12.5-meter resolution digital elevation model (DEM, such as ALOS / PALSAR L-band data), elevation, slope, aspect, curvature, topographic humidity index (TWI), and topographic ruggedness index (TRI) were extracted.

[0082] Geological characteristics: Using a 1:200,000 scale vector geological map (such as data from the China Geological Survey), the lithology and distance to faults of the target grid were extracted.

[0083] Soil characteristics: Soil thickness was extracted using 90-meter resolution soil thickness raster data (such as that provided by ISSCAS).

[0084] Human activity characteristics: Using a 1:250,000 scale vector road network map (such as OpenStreetMap), extract the distance to the road.

[0085] (3) Dynamic rainfall time series data: Access hourly precipitation raster or table data of the target area through meteorological data interface (such as Caiyun Weather API) and extract the continuous dynamic rainfall sequence of each sampling grid before the predicted target time.

[0086] Step 2: Static factor linear mapping.

[0087] The 11 extracted static environmental features are input into the static feature encoder, and the scaling factor γ and offset factor β that match the dimensions of the dynamic features are calculated through linear transformation.

[0088] Step 3: Extraction of spatiotemporal dependencies of dynamic sequences.

[0089] The rainfall time series is input into a two-layer LSTM network. Hyperparameter sensitivity testing shows that in this embodiment, the feature extraction performance reaches its peak when the sliding time window parameter is set to 56 hours (i.e., the hourly rainfall amount within the 56 hours prior to input time t). After LSTM processing, the weights for each time step are calculated and assigned by the time attention mechanism to generate a dynamic rainfall feature vector F.

[0090] Step 4: Characteristic Linear Modulation (FiLM).

[0091] Element-wise multiplication and addition operations are performed on the dynamic rainfall feature vector F using γ and β. This completes the nonlinear modulation of the dynamic hydrological response characteristics by the static geological background.

[0092] Steps 5 and 6: Wide and Deep Fusion and Probabilistic Prediction.

[0093] The modulated features are concatenated with the original 11 static features and the original rainfall sequence. The concatenated comprehensive feature matrix is ​​then input into an XGBoost classifier, which fits and outputs the specific probability value (from 0 to 1) of a landslide occurring in that grid.

[0094] 2. Simulation experiments and test data.

[0095] In the dataset test in Zixing City, the model performance was verified by inputting observational data at the critical moment of the early warning trigger (08:00 on July 27th):

[0096] Classification accuracy data: The model accurately classified 14,801 actual landslide events and 13,409 non-landslide instances. The area under the receiver operating characteristic curve (AUC) reached 0.940, the overall accuracy was 0.8742, the precision was 0.8419, the recall was 0.9204, and the F1 score was 0.8794.

[0097] Threshold response test: Dynamic feature trajectories were extracted using the SHAP framework for cluster analysis. Test data demonstrate that when the rainfall input value at adjacent time steps in the time series exceeds approximately 30 mm / h, the contribution of the features to the prediction probability exhibits a non-linear surge, indicating that the network possesses reliable mathematical sensitivity to short-term extreme precipitation peaks.

[0098] 3. Expand the scope of application

[0099] Although the datasets and process parameters in the embodiments are based on specific typhoon rainfall events, the "static-dynamic feature learning and modulation" architecture described in this invention has underlying universality. In practical engineering, by simply replacing the static spatial base data and real-time rainfall sequences of the target application area, this method can be directly extended to other mountainous areas with complex meteorological and topographical conditions (such as the northern mountainous areas of Beijing-Tianjin-Hebei), providing a data-driven paradigm for identifying ecological risks of floods and chain disasters, detecting key nodes, and providing early warning operations in various terrain environments.

[0100] Example 2: Cross-regional migration test based on cluster landslides in the border area of ​​Fujian, Guangdong and Jiangxi provinces.

[0101] To verify the spatial generalization capability of the "static-dynamic decoupling modulation" architecture of the present invention in unseen areas, the model trained on data from Zixing City in Example 1 was directly applied to the cluster landslide event in the Fujian-Guangdong-Jiangxi border area (Longyan, Fujian; Meizhou, Guangdong; and Ganzhou, Jiangxi) on June 16, 2024, for zero-sample independent testing.

[0102] 1. Testing steps.

[0103] Step 1: Constructing heterogeneous data across regions.

[0104] The border area between Fujian, Guangdong, and Jiangxi provinces (the test area) was divided into grids, and static spatial environmental factors (such as topography and geology) of each grid node in the target test area were extracted as a pre-set static input base. At the same time, the continuous dynamic rainfall time series of the corresponding grid nodes during the extreme rainfall event on June 16, 2024, was obtained as a dynamic trigger input.

[0105] Step 2: Zero-shot model migration and deployment.

[0106] Without using any actual landslide samples from the Fujian-Guangdong-Jiangxi border area for model retraining or fine-tuning of internal network parameters, the multi-source heterogeneous data of the test area obtained in step 1 is directly input into the prediction model that has been trained and converged in the source region (Zixing City) in Example 1.

[0107] Step 3: Independent feature extraction and linear modulation execution.

[0108] Without altering existing rules, the prediction model independently encodes the static environmental factors and dynamic rainfall sequences of the test area, and uses the extracted static scaling factor (γ) and offset factor (β) of the test area to perform feature-level linear modulation on the corresponding dynamic hydrological features.

[0109] Step 4: Probability output and dynamic region prediction.

[0110] The high-dimensional integrated feature vector after modulation and fusion is input into the prediction head, and the predicted landslide probability of each grid node in the Fujian-Guangdong-Jiangxi border area during the rainfall evolution process is directly output, generating a macro-disaster risk distribution at the whole-domain scale.

[0111] 2. Test data and evaluation results.

[0112] Classification accuracy data: In the 24-hour peak risk assessment, the overall area under the ROC curve (AUC) of the migration model reached 0.8763, and the Kolmogorov-Smirnov (KS) statistic was 0.651.

[0113] Spatial early warning efficiency test: The results show that by designating the top 10% of spatial areas with the highest prediction probability as early warning zones, 48.8% of actual landslide events were successfully captured, and the risk enrichment (Lift) reached 4.72 times that of the random search baseline.

[0114] This embodiment demonstrates that the decoupled modulation architecture of the present invention enables the model to learn the relative interaction rules of terrain response to rainfall, overcomes the overfitting of traditional direct splicing strategies to local training data, and has excellent cross-regional direct application capabilities.

[0115] 3. Expand the scope of application.

[0116] Although the datasets and process parameters in the embodiments are based on specific typhoon rainfall events, the "static-dynamic feature learning and modulation" architecture described in this invention has underlying universality. In practical engineering, by simply replacing the static spatial base data and real-time rainfall sequences of the target application area, this method can be directly extended to other mountainous areas with complex meteorological and topographical conditions (such as the northern mountainous areas of Beijing-Tianjin-Hebei), providing a data-driven paradigm for identifying ecological risks of floods and chain disasters, detecting key nodes, and providing early warning operations in various terrain environments.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A regional landslide prediction method based on static and dynamic heterogeneous feature learning, characterized in that, Includes the following steps: Step 1: Construction of multi-source heterogeneous data: Acquire and preprocess static geological environmental factors and dynamic rainfall time series data of the target area; The target prediction area is divided into grids; the static spatial environment factors of each grid node in the target area are extracted as static inputs; at the same time, the continuous dynamic rainfall time series data of the corresponding grid node at T time steps before the target prediction time are obtained as dynamic inputs. Step 2: Independent coding in a static environment; Construct a dedicated static feature encoder to process the static spatial environment factors obtained in step 1. A linear transformation is performed to map it to scaling factors γ and offset factors β for subsequent feature modulation; the specific calculation formula is as follows: ； ； In the formula, , These are the learnable weight matrices used in the static feature encoder to generate scaling and offset coefficients, respectively. , These are the learnable bias vectors used in the static feature encoder to generate scaling and offset coefficients, respectively. Step 3: Independent coding of dynamic rainfall; A dedicated dynamic sequence encoder is constructed to input the dynamic rainfall time series of length T obtained in step 1 into a stacked two-layer long short-term memory network LSTM. The first LSTM layer extracts short-term, high-frequency rainfall intensity fluctuations, while the second LSTM layer extracts long-term precipitation accumulation trends. Subsequently, a temporal attention mechanism is introduced to calculate the hidden state at each time step. Attention score and normalized weights : ； ； The hidden states are weighted and summed to generate a high-dimensional dynamic rainfall feature vector F: ； In the formula, T represents the total number of time steps in the dynamic rainfall time series; t and i are both time step indices. ; Let be the hidden state vector output by the two-layer LSTM network at time step t; This represents the learnable weight matrix in the temporal attention mechanism; This is a learnable bias term in the time attention mechanism; It is the hyperbolic tangent activation function; Let be the unnormalized attention score at time step t; It is an exponential function with the natural constant as its base; Let F be the normalized attention weight at time step t; F is the final generated high-dimensional dynamic rainfall feature vector. Step 4: Feature linear modulation: Based on the static geological characterization, scaling and translation coefficients are generated to perform feature-by-bit linear modulation on the dynamic rainfall characterization; Using the static scaling factor γ and offset factor β generated in step 2, feature-level linear modulation is performed on the dynamic rainfall feature vector F generated in step 3; through element-level multiplication... The addition operation enables the static spatial background to adaptively scale and offset dynamic temporal features, generating modulation features. : ； Step 5: Wide and Deep Feature Fusion: The modulated spatiotemporal interaction features are concatenated with the unextracted original static features to construct a hybrid representation vector; Using a Wide and Deep fusion architecture, the modulation features output from step 4 are... The Deep path and the original static spatial environment factors and dynamic rainfall time series from step 1 are used as the Wide path for channel-level splicing; by retaining the original input channels without nonlinear smoothing, a comprehensive feature space containing the original extreme value signal and deep modulation interaction information is constructed. Step 6: Landslide probability classification and prediction; The high-dimensional integrated feature vector fused in step 5 is input into a machine learning classifier. The classifier is used to fit the nonlinear decision boundary of the multidimensional feature space, and finally outputs the probability prediction value of the rain-induced landslide at the grid node.

2. The regional landslide prediction method based on static and dynamic heterogeneous feature learning according to claim 1, characterized in that, The static inputs in step 1 include, but are not limited to, elevation, slope, aspect, curvature, topographic humidity index, topographic ruggedness index, lithology, soil thickness, and distance from faults / rivers / roads.

3. The regional landslide prediction method based on static and dynamic heterogeneous feature learning according to claim 1, characterized in that, It also includes step 7: introducing the SHAP algorithm to calculate the marginal contribution of variables, generating response surfaces and performing interpretability analysis.

Images