TL;DR
This study introduces soil erosion (SE) as a crucial predisposing factor for landslide susceptibility prediction (LSP), demonstrating its significant impact on model accuracy. By comparing machine learning models with and without SE, the research highlights that incorporating SE leads to more reliable predictions, with finer grid resolutions generally yielding better results.
What Problem Does This Paper Address?
Landslides are a major global hazard, causing significant loss of life and economic damage. Accurately predicting where landslides are likely to occur is vital for mitigation efforts. Traditional landslide susceptibility prediction (LSP) models often rely on factors like topography, geology, hydrology, and land cover. However, this study identifies a critical gap: the influence of soil erosion (SE) as a source of landslide material and a reflection of long-term rainfall erosion has been largely overlooked. This omission can lead to less reliable predictions, especially in areas prone to both erosion and landslides. The paper aims to address this by investigating the impact of incorporating SE into LSP models and comparing their performance against traditional models.
Methods
The researchers employed a multi-faceted approach to model landslide susceptibility:
- Data Preparation: They collected landslide inventory data and several predisposing factors, including topography (elevation, slope, aspect, curvature, TWI), land cover (NDVI, NDBI), lithology, and crucially, soil erosion (SE). These factors were processed at two different grid resolutions: 30m and 60m.
- Soil Erosion Modeling: Soil erosion intensity was calculated using the Revised Universal Soil Loss Equation (RUSLE) model, incorporating factors like rainfall erosivity, soil erodibility, slope length and steepness, and cover management.
- Machine Learning Models: Four popular machine learning algorithms were utilized:
- Multilayer Perceptron (MLP): A feed-forward neural network.
- Logistic Regression (LR): A statistical method for binary classification.
- Support Vector Machine (SVM): A powerful classification algorithm.
- C5.0 Decision Tree (C5.0 DT): An improved version of the C4.5 algorithm, known for its efficiency and accuracy.
- Model Comparison: For each machine learning model, two versions were developed: one incorporating the SE factor (SE-MLP, SE-LR, etc.) and another without it (MLP, LR, etc.). These models were trained and tested using a 70/30 split of the data.
- Evaluation: The performance of all models was assessed using the Area Under the Receiver Operating Characteristic Curve (AUC), a standard metric for evaluating classification accuracy. The relative importance of each predisposing factor was also analyzed. Statistical significance of differences between SE-based and single models was tested using the Wilcoxon rank test.
Datasets
- Landslide Inventory: 446 landslide locations in Ningdu County, China, collected from government disaster survey reports and high-resolution image interpretation (1970-2003).
- Predisposing Factors:
- Topography: Derived from ASTER GDEM (30m resolution).
- Land Cover: Landsat-8 Thematic Mapper (TM) images (October 2013) for NDVI and NDBI.
- Geology: 1:100,000 scale geological map of Ningdu County for lithology.
- Hydrology: Daily rainfall data from 7 meteorological stations (1998-2017) for rainfall erosivity.
- Soil Erosion (SE): Calculated using RUSLE.
The study used existing datasets and calculated derived factors, rather than introducing new benchmark datasets.
Key Insights and Ideas
- SE is a Dominant Factor: Soil erosion emerged as the most influential predisposing factor for landslide susceptibility across all models and resolutions. This underscores the importance of considering erosion processes in landslide risk assessment.
- SE Improves Prediction Accuracy: Models that included the SE factor consistently outperformed their counterparts without SE, demonstrating a tangible improvement in predictive power.
- Resolution Matters: Finer grid resolutions (30m) generally led to higher prediction accuracy compared to coarser resolutions (60m), suggesting that detailed spatial data can capture more nuanced landslide-influencing patterns.
- Model Performance Varies: While SE generally improved accuracy, the C5.0 DT and SVM models consistently showed better performance than MLP and LR, indicating their suitability for this type of complex spatial prediction task.
Key Results
- Accuracy Improvement: SE-based models showed a 2-4% increase in AUC under 30m resolution and a 2-5% increase under 60m resolution compared to models without SE.
- Best Performing Models:
- Under 30m resolution, the SE_C5.0 DT model achieved the highest AUC (86.7%).
- Under 60m resolution, the SE_SVM model achieved the highest AUC (82.4%).
- SE's Importance: The SE factor was consistently ranked as the most important predisposing factor in most models, contributing significantly to the prediction accuracy. For instance, in the SE_SVM model, SE accounted for 23% of the contribution rate.
- Grid Resolution Impact: The 30m grid resolution generally resulted in higher overall LSP accuracy than the 60m resolution.
- Statistical Significance: The Wilcoxon rank test confirmed statistically significant differences in landslide susceptibility indices between SE-based and single models, reinforcing the impact of SE.
Limitations and Future Work
- Resolution Impact: While 30m resolution generally performed better, the paper notes that the effect of SE on landslide susceptibility was not "great" under different resolutions, suggesting that while important, its influence might be less sensitive to resolution changes than other factors.
- Model Specificity: The study focused on four specific machine learning models. Future work could explore other advanced algorithms or ensemble methods.
- Dynamic Factors: The current study primarily uses static predisposing factors. Incorporating dynamic factors like rainfall intensity and duration in real-time could further enhance prediction capabilities.
- Generalizability: The findings are specific to Ningdu County. Further research in diverse geological and climatic settings would be valuable to assess the broader applicability of SE as a predisposing factor.
Summary updated