Assessing the impact of training samples overlap and density in random forest for landslide susceptibility mapping: Implications for degraded land management in Bandung Regency, Indonesia

Landslide susceptibility mapping is essential for disaster mitigation and land management in degraded mountainous regions. Machine learning algorithms, particularly Random Forest (RF), have been increasingly applied due to their robustness in handling complex, non-linear relationships. However, classification performance is often affected by the quality of training samples, especially when landslide and non-landslide points exhibit spatial overlap. This study investigated how varying densities of fully overlapping samples influence RF performance in Bandung Regency, West Java, Indonesia, an area characterised by steep slopes, rapid land-use change, and post-mining degradation. Balanced datasets ranging from 50 to 700 samples per class were evaluated with hyperparameter tuning. The highest validation accuracy (89%) was achieved with 500 samples at a max_depth of 2, while training accuracy was approximately 10% lower, indicating the algorithm’s difficulty in separating overlapping classes. A more stable trade-off was obtained with 300 samples and a max_depth of 4, suggesting that moderate densities enhance generalisation. To translate these findings into practice, we propose an ensemble zoning and uncertainty mapping framework that integrates multiple model outputs to identify consensus zones for slope stabilisation, vegetation restoration, and adaptive spatial planning. This approach improves the reliability of susceptibility maps and provides actionable insights for managing degraded and landslide-prone landscapes.