Spatiotemporal enhancement and heterogeneous integration synergistic landslide susceptibility evaluation method

Abstract

This invention relates to a landslide susceptibility assessment method based on spatiotemporal enhancement and heterogeneous integration, belonging to the field of geological disaster prediction technology. First, spatiotemporal benchmark unification and sample balancing are constructed using multi-source heterogeneous geospatial data of the study area. Then, a three-level progressive framework of "heterogeneous base learner – meta-feature secondary encoding – dual-channel deep cross-fusion" is used as the core. Three heterogeneous base learners—random forest, XGBoost, and LightGBM—are used to output leak-free meta-probability features through temporal cross-folding. Next, the original high-dimensional factors and meta-features are input into a deep sub-network via a "spatial-semantic" dual-channel approach, and cross-fusion is completed using adaptive gating weights. Finally, the landslide occurrence probability is output. Experiments show that this method outperforms traditional single models in multiple metrics such as ROC and PR, and can significantly improve the accuracy and robustness of susceptibility mapping without increasing field survey costs, providing a highly reliable decision-making basis for mountainous land spatial planning and early disaster warning.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prediction technology, and in particular to a susceptibility assessment method for landslide disasters in mountainous areas based on spatiotemporal enhancement and heterogeneous integration synergy. Background Technology

[0002] Landslide susceptibility assessment is a prerequisite for disaster risk management and land spatial planning. Traditional empirical models or single machine learning methods struggle to simultaneously address issues such as factor nonlinearity, sample imbalance, and spatial heterogeneity. While deep learning possesses powerful feature extraction capabilities, its generalization performance is often insufficient due to limited sample size and its "black box" nature. In recent years, stacking ensemble strategies have been introduced into the landslide field, but most studies only employ homogeneous base learners, resulting in single meta-features and failing to explicitly distinguish and fuse "original spatial features with meta-semantic features," thus limiting further accuracy improvements. Furthermore, existing technologies lack a systematic description of unified spatiotemporal benchmarks, balanced samples, and leak-free cross-validation, making it difficult to meet the urgent needs of rapid mapping and engineering deployment in high-altitude areas. Therefore, a new framework for landslide susceptibility assessment that integrates "heterogeneous ensemble + deep cross-validation + spatiotemporal enhancement" is urgently needed to address these shortcomings.

[0003] Beneficial effects

[0004] Accuracy Improvement: Heterogeneous base learners (RF, XGBoost, LightGBM) generate complementary meta-probabilities under temporal cross-folding, and dual-channel deep networks further explore spatial-semantic coupling relationships. Experimental AUC reaches 0.890 and AP reaches 0.862, which is 4–12% higher than the traditional single model on average, and significantly reduces the false negative rate in high-risk areas.

[0005] Enhanced spatiotemporal robustness: Through a three-step preprocessing process of "unified spatiotemporal benchmark + resampling + buffer balanced sampling", spatial biases caused by imbalances in resolution, coordinate system and samples of multi-source data are eliminated, resulting in smaller AUC fluctuations in the model during validation in the same region in different years, thus meeting the needs of rapid mapping updates in mountainous areas.

[0006] Balancing interpretability and engineering usability: The contribution of meta-features and gating weights can be traced back to the original factors, forming a three-level interpretation chain of "factor-meta-feature-susceptibility"; the output of a five-level natural breakpoint zoning map can be directly integrated with land spatial planning, post-disaster reconstruction, and rainfall threshold early warning services, achieving monthly updates without the need for additional field surveys. Summary of the Invention

[0007] S1. Collection of multi-source heterogeneous geospatial data and unification of spatiotemporal benchmarks

[0008] S101. Data Scope Definition: Taking the administrative boundary of Zhangjiajie City, Hunan Province as the research unit, DEM data with a spatial resolution of 30m for the past 10 years of Zhangjiajie City were collected to obtain elevation, slope, aspect, and 1:200,000 scale stratigraphy (vector), land use type (30 m), annual cumulative rainfall (1 km, interpolated to 30m), distance from fault / road / river (1:200,000 scale), NDVI (30 m), and historical landslide records (point / area), totaling 10 factors. S102. Spatiotemporal reference unification: The CGCS2000 coordinate system is adopted, and all grids are resampled to 30 m × 30 m using the nearest neighbor method to ensure consistent spatial reference; S103. Landslide Sample Construction: Positive samples are generated with the centroid of the landslide surface as the center, and an equal number of negative samples are randomly generated 1 km away, for a total of 645 positive samples and 1290 negative samples, forming an equilibrium sample set N = 2P, where P is the number of landslide points.

[0009] S2. Data Preprocessing and Standardization

[0010] S201. Handling missing values: Continuous variables are imputed using multiple imputation, and categorical variables are imputed using the mode. S202. Outlier Removal: Outliers are removed using the 3σ criterion for continuous variables, with a removal rate of <2%; S203. Z-score standardization: For continuous variables, standardize according to the formula... x′ = (x - μ) / σ Standardization is performed, where μ is the mean and σ is the standard deviation; categorical variables are coded using One-Hot encoding.

[0011] S3. Generation of metaprobabilistic features from heterogeneous basis learners

[0012] S301. Model Configuration: RandomForest: n_estimators=800, max_depth=None; XGBoost: n_estimators=800, max_depth=6, learning_rate=0.05; LightGBM: n_estimators=800, max_depth=6, learning_rate=0.05; S302. Temporal Crossover Framework: A 5-fold temporal crossover is adopted, with each fold divided according to the year to ensure that the training set year is earlier than the validation set and to avoid information leakage; S303. Meta-probability output: For the i-th sample, the k-th base learner outputs probability pik, forming the meta-feature matrix. Pmeta = [p RF , p XGB , p LGB ] ∈ n×3 .

[0013] S4. Dual-channel deep cross-fusion

[0014] S401. Channel Separation: Original High-Dimensional Factor X orig ∈ n×m (m=33) Enter the "spatial channel", and the meta-feature Pmeta enters the "semantic channel"; S402. Sub-network structure: Spatial channel: Dense(128)→BN→ReLU→Dropout(0.3)→Dense(64)→BN→ReLU→Dropout(0.3); Semantic channels: Dense(16) → BN → ReLU → Dropout(0.2); S403. Adaptive Gated Fusion: G = σ(Wg·[H orig H meta ] + b g ) Hfuse = G ⊙ H orig + (1-G) ⊙ H meta Where σ is the Sigmoid function, W g b g Here are the learnable parameters, and ⊙ represents the element-wise product; S404. Output after fusion: Hfuse is processed through Dense(32) → BN → ReLU → Dropout(0.3) → Dense(1) → Sigmoid to obtain the probability of landslide occurrence. = σ(W out ·H fuse + b out ).

[0015] S5. Model Training and Optimization

[0016] S501. Loss Function: Binary Cross-Entropy

[0017] L = ; S502. Optimizer: Adam, initial learning rate 1e-3, batch size 256; S503. Early Stop Mechanism: Monitor and verify loss, patience=10, restore_best_weights=True; S504. Regularization: L2 coefficient 1e-4, Dropout ratio 0.2–0.3, to prevent overfitting.

[0018] S6. Accuracy Verification and Regional Mapping

[0019] S601. Evaluation Indicators: AUC=

[0020] AP=

[0021] F1 = 2·Precision·Recall / (Precision + Recall); S602. Result Output: Divide the test set probability into five levels: extremely high, high, medium, low, and extremely low according to the natural breakpoint method, and generate a 30 m resolution susceptibility raster map; Attached Figure Description

[0022] Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a schematic diagram of the ROC curve; Figure 3 This is a schematic diagram of the PR curve; Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0026] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0027] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature being directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0029] Example

[0030] S1. Collection of multi-source heterogeneous geospatial data and unification of spatiotemporal benchmarks

[0031] S101. Data Scope Definition: The research unit is Yongding District, Wulingyuan District, Cili County, and Sangzhi County of Zhangjiajie City. Data for 10 years, from 2013 to 2023, will be collected, including: 30 m DEM (ALOS PALSAR) → Elevation, Slope, Aspect 1:200,000 geological map (Hunan Provincial Geological Survey Institute) → Stratigraphy and Lithology (Vector) 30 m Landsat-8 NDVI (Annual Average Maximum) 1 km TRMM precipitation data (annual cumulative, bilinear interpolation to 30 m) 30 m CLCD Land Use (2020 Edition) 1:200,000 faults, roads, and rivers (OSM) → Euclidean distance grid Historical landslide record (remote sensing interpretation, 645 points / areas in total) S102. Spatiotemporal reference unification: The CGCS2000 coordinate system is adopted, and all grids are resampled to 30 m × 30 m using the nearest neighbor method to ensure consistent spatial reference; S103. Landslide Sample Construction: Positive samples are generated with the centroid of the landslide surface as the center, and an equal number of negative samples are randomly generated 1 km away, for a total of 645 positive samples and 1290 negative samples, forming an equilibrium sample set N = 2P, where P is the number of landslide points.

[0032] S2. Data Preprocessing and Standardization

[0033] S201. Handling missing values: Continuous variables are imputed using multiple imputation, and categorical variables are imputed using the mode. S202. Outlier Removal: Outliers are removed using the 3σ criterion for continuous variables, with a removal rate of <2%; S203. Z-score standardization: For continuous variables, standardize according to the formula... x′ = (x - μ) / σ Standardization is performed, where μ is the mean and σ is the standard deviation; categorical variables are coded using One-Hot encoding.

[0034] S3. Generation of metaprobabilistic features from heterogeneous basis learners

[0035] S301. Model Configuration: RandomForest: n_estimators=800, max_depth=None; XGBoost: n_estimators=800, max_depth=6, learning_rate=0.05; LightGBM: n_estimators=800, max_depth=6, learning_rate=0.05; S302. Temporal Crossover Framework: A 5-fold temporal crossover is adopted, with each fold divided according to the year to ensure that the training set year is earlier than the validation set and to avoid information leakage; S303. Meta-probability output: For the i-th sample, the k-th base learner outputs probability pik, forming the meta-feature matrix. Pmeta = [p RF , p XGB , p LGB ] ∈ n×3 .

[0036] On the Zhangjiajie test set, the performance of the three base learners is as follows: RF: AUC=0.880, AP=0.873 XGB: AUC=0.876, AP=0.860 LGB: AUC=0.879, AP=0.863 Generator feature matrix Pmeta∈ 165×3 (Test set sample size n=165) S4. Dual-channel deep cross-fusion S401. Channel Separation: Original High-Dimensional Factor X orig ∈ n×m (m=33) Enter the "spatial channel", and the meta-feature Pmeta enters the "semantic channel"; S402. Sub-network structure: Spatial channel: Dense(128)→BN→ReLU→Dropout(0.3)→Dense(64)→BN→ReLU→Dropout(0.3); Semantic channels: Dense(16) → BN → ReLU → Dropout(0.2); S403. Adaptive Gated Fusion: G = σ(Wg·[H orig H meta ] + b g ) Hfuse = G ⊙ H orig + (1-G) ⊙ H meta Where σ is the Sigmoid function, W g b g Here are the learnable parameters, and ⊙ represents the element-wise product; S404. Output after fusion: Hfuse is processed through Dense(32) → BN → ReLU → Dropout(0.3) → Dense(1) → Sigmoid to obtain the probability of landslide occurrence. = σ(W out ·H fuse + b out ).

[0037] S5. Model Training and Optimization

[0038] S501. Loss Function: Binary Cross-Entropy

[0039] L = ; S502. Optimizer: Adam, initial learning rate 1e-3, batch size 256; S503. Early Stop Mechanism: Monitor and verify loss, patience=10, restore_best_weights=True; S504. Regularization: L2 coefficient 1e-4, Dropout ratio 0.2–0.3, to prevent overfitting.

[0040] S6. Accuracy Verification and Regional Mapping

[0041] S601. Evaluation Indicators: AUC=

[0042] AP=

[0043] F1 = 2·Precision·Recall / (Precision + Recall); The model performance was calculated using an independent test set, and the results are as follows: Figure 2 (ROC curve) and Figure 3 As shown in the PR curve: Coupled dual-depth model (this invention): AUC = 0.890, AP = 0.862, F1 = 0.843, Precision = 0.845, Recall = 0.841; It significantly outperforms the single model in the ROC space, demonstrating a stronger ability to distinguish between positive and negative samples.

[0044] Contrast base learner: RandomForest: AUC = 0.880, AP = 0.873 XGBoost: AUC = 0.876, AP = 0.860 LightGBM: AUC = 0.879, AP = 0.863 While maintaining a similar AP, this invention improves AUC by approximately 1.1–1.4 percentage points, indicating that it has better control over false positives and is suitable for unbalanced landslide sample scenarios.

[0045] S602. Output Results: Divide the test set probabilities into five levels—extremely high, high, medium, low, and extremely low—using the natural breakpoint method, and generate a 30 m resolution susceptibility raster map.

Claims

1. A landslide susceptibility assessment method based on spatiotemporal enhancement and heterogeneous integration, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous geospatial data of the study area, including elevation, slope, aspect, stratigraphy, land use, annual cumulative rainfall, distance from faults / roads / rivers, normalized difference vegetation index (NDVI), and historical landslide records, and unify all data to the same spatiotemporal reference and raster resolution. S2. The data obtained in S1 is cleaned, missing values ​​are imputed, outliers are removed, and Z-score is standardized. A balanced sample is constructed using the strategy of "landslide point + random non-landslide point 1 km away". S3. Select three heterogeneous base learners, Random Forest, XGBoost and LightGBM, and output leak-free meta-probability features under a 5-fold temporal cross framework to form a second-level meta-feature set; S4. Separate the original standardized features and the meta-feature set obtained in S3 according to the dual-path input architecture of "spatial channel-semantic channel", and connect them to two independent multilayer perceptron sub-networks respectively. Cross-fusion is completed through adaptive gating weights to obtain the fused feature vector. S5. Using the fused feature vector obtained in S4 as input, the model is trained by passing it through a fully connected layer and a Sigmoid activation layer to output the probability of landslide occurrence. The Adam optimizer and early stopping mechanism are used to complete the model training. S6. Calculate AUC, AP, F1, and Recall indices based on the test set, plot ROC and PR curves, reclassify the probability results to generate a landslide susceptibility zoning map, and complete the accuracy verification and regional application.

2. The method according to claim 1, characterized in that: The hyperparameters of the heterogeneous base learners in S3 are determined by a combination of grid search and Bayesian optimization to ensure the diversity and complementarity of meta-probabilistic features.

3. The method according to claim 1 or 2, characterized in that: In S4, each of the dual-channel subnetworks introduces batch normalization and Dropout layers, and the gate weights are automatically learned through one-dimensional convolution, avoiding the subjectivity of manual weighting.

4. The method according to any one of claims 1-3, characterized in that: In S5, L2 regularization is introduced before the output of the fully connected layer to prevent overfitting and improve the model's generalization ability.

5. The method according to any one of claims 1-4, characterized in that: The susceptibility zoning in S6 adopts a five-level classification using the natural breakpoint method, and is compared with the frequency of field verification points to ensure that the results are consistent with the actual disaster distribution.

Images