Analysis of Rainfall induced Landslide Susceptibility Using Deep Forest Model for Decision Boundary Interpretation

The frequency and magnitude of landslide damage have increased due to the impact of heavy rainfall, which has been exacerbated by climate change. Consequently, the importance of landslide susceptibility analysis for identifying high-risk areas is being further emphasized. Previous susceptibility studies have utilized various data-driven analyses, including machine learning and deep learning, to understand the complex nonlinear relationships among landslide-influencing factors. In particular, ensemble techniques have been shown to enhance overall performance and stability by combining the prediction results of individual models. However, previous bagging and boosting-based ensemble techniques have primarily focused on improving average classification performance. Further examination is necessary to assess the stability and interpretability of decision boundaries under varying threshold values and the distribution characteristics of prediction probabilities. This is especially challenging in landslide datasets with significant class imbalance, where pixels in the boundary region can exhibit highly sensitive prediction changes depending on threshold settings.

To address these limitations, this study employed the gcForest (multi-grained cascade forest) model, also known as Deep Forest. gcForest is a deep learning alternative that utilizes a cascade structure, comprising multiple layers of random forests. Each layer receives both original features and class probability outputs from the preceding layer. This structure facilitates the incremental updating of probability information for samples near decision boundaries, enabling iterative reclassification. This structure is distinct from existing ensemble techniques in that it enables stepwise improvement of decision boundaries for samples with high prediction uncertainty. This is in contrast to the existing ensemble techniques that determine predictions at a single stage. In order to make a comparison with existing ensemble techniques, this study has set bagging-based random forest and boosting-based XGBoost as the base model of deep forest.

The proposed analysis approaches were applied to Pohang City, Gyeongsangbuk-do, South Korea, where a large-scale landslide occurred in 1998. The analysis results demonstrated that the gcForest-based model exhibited enhanced prediction performance (gcForest_RF AUC = 91.62%, gcForest_XGBoost AUC = 91.40%) in comparison to the prevailing ensemble methods, random forest and XGBoost. Specifically, the XGBoost-based gcForest model demonstrated enhanced accuracy, improving from 0.797 to 0.814, and an elevated f1-score from 0.789 to 0.814 when compared to the prevailing XGBoost model. These results indicate that gcForest's stepwise improvement structure contributes to enhanced performance in classifying uncertain samples near decision boundaries, thereby enabling more stable landslide susceptibility prediction.

 

Acknowledgement

This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (RS-2025-00515970).