Abstract
Landslide susceptibility mapping (LSM) is vital for disaster mitigation and early warning systems. However, traditional machine learning approaches often suffer from poor "non-landslide" sample selection and the limited accuracy of individual classifiers. This study addresses these issues by introducing a constrained sampling strategy using Logistic Regression (LR) and applying ensemble learning techniques—Bagging and Boosting—to Classification and Regression Trees (CART) and Multi-Layer Perceptrons (MLP). Focusing on the Zigui-Badong section of China’s Three Gorges Reservoir Area (TGRA), the research identifies elevation and distance to rivers as primary causal factors. The results demonstrate that the LR-MLP-Boosting model achieved the highest predictive accuracy (AUC = 0.986), significantly outperforming individual classifiers and traditional random sampling methods.
Introduction
Predicting landslide occurrence is a complex challenge involving nonlinear relationships between terrain, geology, and environmental triggers. While machine learning (ML) has become a standard tool for LSM, the quality of the training data—specifically the selection of "non-landslide" (negative) samples—is often overlooked. Randomly selecting negative samples from any non-landslide area can introduce noise, as these areas might possess the geological conditions for failure but haven't yet experienced a landslide. This study proposes a refined workflow to improve both sample quality and model performance.
Methods and Materials
Study Area
The research focused on the Zigui-Badong section of the TGRA, a region characterized by intense human engineering and fluctuating reservoir water levels, which contribute to high landslide frequency. A total of 179 landslides were identified for the study.
Data Preparation
Twelve influencing factors were selected, including altitude, slope, lithology, and proximity to rivers and faults. The dataset was split into 70% for training and 30% for validation.
Modeling Workflow
- Constrained Sampling: A preliminary Logistic Regression (LR) model was used to map susceptibility. Non-landslide samples were then exclusively selected from "Very Low" susceptibility zones to ensure high-quality negative training data.
- Base Classifiers: Two distinct ML models were employed: CART (decision tree) and MLP (neural network).
- Ensemble Techniques: To enhance these base learners, the researchers applied Bagging (parallel processing to reduce variance) and Boosting (sequential processing to reduce bias).
Results and Discussion
Factor Importance
Quantitative analysis using Information Value (IV) revealed that elevation (specifically <240 m) and distance to rivers (<300 m) were the most significant drivers of landslide development in the study area. This underscores the impact of reservoir operations on slope stability.
Model Performance
The integration of constrained sampling and ensemble learning yielded superior results across all metrics. The Area Under the Curve (AUC) values for the primary models were:
- LR-MLP-Boosting: 0.986 (Highest performance)
- LR-CART-Boosting: 0.981
- LR-MLP-Bagging: 0.978
- LR-CART-Bagging: 0.973
In comparison, models using traditional random sampling (NO-constraint) performed significantly worse. For instance, the basic LR-MLP model achieved an AUC of approximately 0.901, while the ensemble version reached 0.986—an improvement of roughly 8.5%.
Conclusion
The combination of ensemble learning and constrained sample selection effectively addresses the "garbage in, garbage out" problem in landslide susceptibility mapping. By using a simple model (LR) to define a "safe zone" for negative samples and advanced ensemble techniques (Boosting/Bagging) to handle complex terrain data, researchers can produce highly accurate maps. These high-fidelity maps are essential for monitoring and developing early warning systems in landslide-prone regions like the Three Gorges Reservoir.
Summary updated