A Comparative Study of Heterogeneous Ensemble-Learning Techniques for Landslide Susceptibility Mapping
- 1. AComparativeStudyof HeterogeneousEnsemble-Learning Techniquesfor LandslideSusceptibilityMapping DSA 554 3.0 - Spatio-Temporal Data Analysis Erandika Lakmali
- 2. Introduction o Research Paper Title: A Comparative Study of Heterogeneous Ensemble-Learning Techniques for Landslide Susceptibility Mapping o Authors: Zhice Fang, Yi Wang, Ling Peng & Haoyuan Hong o Journal: International Journal of Geographical Information Science (Q1) o Problem Statement: • Landslides cause massive economic losses and casualties worldwide • Traditional prediction models struggle to generalize across different regions and terrain types • Need for integrating deep learning and machine learning methods for a robust landslide susceptibility model
- 3. Introduction (cont.) o Research Goal: • Develop a robust landslide susceptibility mapping (LSM) model using ensemble-learning techniques • Compare four ensemble methods: Stacking, Blending, Simple Averaging (SA), Weighted Averaging (WA) o Research Objective: • Investigate the effectiveness of heterogeneous ensemble learning techniques in predicting landslide susceptibility using spatial datasets o Significance: • Landslides cause severe economic and human losses globally • Accurate susceptibility mapping can aid in disaster risk assessment and mitigation
- 4. StudyArea &Data The study follows a structured ensemble-learning framework for landslide susceptibility mapping. o Study Area: Yanshan County, Jiangxi Province, China • Covers 2178 km² with varying elevations (-27m to 2144m) • Terrain includes mountains, hills, valleys, and plains
- 5. StudyArea &Data (cont.) o Data Used: • 16 Landslide Conditioning Factors, e.g., - Topographical: Slope, Aspect, Curvature, Elevation - Hydrological: Distance to Rivers, SPI, TWI - Geological: Lithology, Distance to Faults - Climatic: Annual Rainfall - Human Influence: Distance to Roads, Land Use • 380 Landslide Historical Locations (extracted from field studies and satellite data) • Split data into 70% training and 30% testing • Data Sources: NASA SRTM, Landsat 7 ETM+, China Geological Survey, Jiangxi Meteorological Bureau
- 6. Methodology Overview o Feature Selection Techniques Used: 1. Multicollinearity Analysis (Variance Inflation Factor, VIF) • Removed highly correlated features (VIF > 10) 2. Relief-F Algorithm • Identified most influential factors for landslides
- 7. Methodology Overview (cont.) o Heterogeneous Ensemble Learning Methods Used: • Base Classifiers: - Deep Learning: Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) - Machine Learning: Support Vector Machines (SVM), Logistic Regression (LR) • Four Ensemble Methods Implemented: - Stacking: Uses meta-classifiers to combine model predictions - Blending: Similar to stacking but prevents information leakage - Simple Averaging (SA): Computes an average probability of base models - Weighted Averaging (WA): Assigns weights based on AUC scores
- 9. Methodology Overview (cont.) o Metrics Used for Performance Evaluation of Models: • Overall Accuracy (OA) • Precision • Recall • F1-Score • Matthews Correlation Coefficient (MCC) • Index of Agreement (IOA)
- 10. Methodology Overview (cont.) o Metrics Used for Performance Evaluation of Models: • Area Under Receiver Operating Characteristic (ROC) Curve (AUC)
- 11. Results (cont.)
- 12. Results o Generated Landslide Susceptibility Maps: • This shows a comparison of different model outputs and how each predicts landslide-prone areas • Classified into five risk categories: Very High, High, Moderate, Low, Very Low • Higher landslide occurrence in steep terrain with high rainfall and loose lithology
- 13. Results (cont.)
- 14. Results (cont.) o Performance Evaluation: • Blending achieved the highest AUC (0.858), followed by Stacking (0.853) • Blending & Stacking ensembles outperformed individual base classifiers • CNN had the highest AUC among base classifiers (0.843), followed by SVM (0.836) and RNN (0.815) • Logistic Regression (LR) performed the worst among the base classifiers with an AUC of 0.774 • Weighted Averaging (WA) and Simple Averaging (SA) ensembles had AUC values of 0.849, slightly lower than Blending and Stacking • CNN was found to be a critical component for Stacking and WA ensembles, whereas SVM played a major role in Blending’s success
- 15. Replicationof Results o Objective: • Reproduce the key findings of the research paper using the provided dataset • Evaluate the performance of ensemble models in landslide susceptibility prediction.
- 16. Replicationof Results (cont.) o Steps Followed for Replication (Aligned to the Paper) • Step 1: Data Preprocessing & Feature Engineering - Standardization: Applied StandardScaler to normalize numerical features - Multicollinearity Check: Used Variance Inflation Factor (VIF) to identify and remove redundant features - Feature Selection: Applied Relief-F Algorithm to extract the most relevant features for classification • Step 2: Model Implementation & Hyperparameter Optimization - Implemented Base Classifiers: - Convolutional Neural Network (CNN) - Recurrent Neural Network (RNN) - Support Vector Machine (SVM) - Logistic Regression (LR)
- 17. Replicationof Results (cont.) • Step 2: Model Implementation & Hyperparameter Optimization (cont.) - Ensemble Methods Implemented: - Stacking Ensemble: Used Logistic Regression as the meta-learner - Blending Ensemble: Trained base models on a split dataset, using a meta-model on predictions - Weighted Averaging Ensemble: Assigned dynamic weights to base models based on AUC scores - Simple Averaging Ensemble: Combined model predictions using equal-weight averaging - Hyperparameter Optimization: - CNN & RNN: Optimized layers, activation (ReLu), dropout (0.2), epochs (150) - SVM: Used RBF kernel (C=1000, gamma=0.001) - LR: Applied Ridge regularization (C=0.1)
- 18. Replicationof Results (cont.) • Step 3: Model Training & Evaluation - Trained each model with 70% training data & 30% test data (consistent with the paper) - - Evaluated models using the following metrics: - Overall Accuracy (OA) - Precision - Recall - F1-Score - Matthews Correlation Coefficient (MCC) - Area Under the Curve (AUC) - Index of Agreement (IOA)
- 19. Replicationof Results (cont.) • Step 4: Results Comparison with Original Paper: - Result: Replication successfully matches the original study, validating the findings! 1 2 3 4 5
- 20. Replicationof Results (cont.) • Step 4: Results Comparison with Original Paper (cont.): - Result: Replication successfully matches the original study, validating the findings!
- 21. Replicationof Results (cont.) o Key Findings & Observations: • Stacking performed well nearly matching paper’s accuracy values • SVM achieved nearly the same accuracy as Stacking • LR underperformed, aligning with paper's results that it was the weakest classifier • Indicating that LR was less effective in capturing spatial dependencies in landslide susceptibility • Feature importance selection via Relief-F positively impacted model performance • Blending & Weighted Averaging ensembles had errors due to data format mismatches and probability-based meta- learning issues
- 22. Replicationof Results (cont.) o Conclusion of Replication: • Replication successfully validated the findings of the original study • Findings confirm that ensemble-learning methods outperform (showed better prediction performance) than the base classifiers • Stacking and CNN emerged as the best-performing models in replication, in contrast to the paper’s emphasis on Blending and Stacking • Challenges in replicating Blending and Weighted Averaging approaches indicate areas for further improvement in future replications • Overall, results align closely with the paper, with minor variations due to dataset-specific factors and tuning methodologies
- 23. Limitationsof theStudy o Dataset Limitations: • Geographical Bias: Only focuses on Yanshan County; not generalize to other regions • Temporal Variability Not Considered: Lacks multi-temporal analysis, making it a static rather than dynamic • Imbalanced Dataset: Unequal distribution of landslide vs. non-landslide samples may impact recall & precision • Remote Sensing Data Constraints: Spatial resolution=25m • Limited Feature Diversity: Socio-economic & anthropogenic factors (deforestation, urbanization) are not considered o Methodological Limitations: • Feature Selection Assumptions: Relief-F feature selection assumes independence among features, not true in spatial datasets • Blending ensemble methods can be overfit to training data
- 24. Limitationsof theStudy (cont.) o Methodological Limitations (cont.): • Hyperparameter Sensitivity: Model performance is highly dependent on hyperparameter tuning, and minor variations can lead to different results • Computational Costs: Deep learning models (CNN, RNN) require high computational power compared to traditional ML models (SVM, LR) o Practical & Interpretability Limitations: • Black-Box Nature of Deep Learning Models: CNN and RNN are less interpretable than traditional ML models like SVM and LR • Does not analyze uncertainty in model predictions, crucial for real-world disaster management • Does not validate susceptibility maps with actual post- event landslide occurrences, limiting practical applicability
- 25. Suggestionsfor Improvement o Enhance Data Collection: • Integrate real-time landslide monitoring sensors to improve predictive performance in practical applications • Use higher resolution satellite imagery o Model Enhancements: • Improve Feature Selection Techniques to reduce reliance on feature independence assumption • Train models on multi-region datasets to improve generalizability • Apply Explainable AI (XAI) techniques to enhance model interpretability
- 26. Conclusion& Takeaways o Key Takeaways: • Heterogeneous ensemble learning outperforms traditional models for LSM • Stacking & Blending are the most effective techniques • Deep learning models (CNN, RNN) enhance feature extraction o Final Thought: • Future studies should focus on real-time susceptibility mapping using spatio-temporal data streams