Decoding dynamic landslide hazard processes for a massive refugee camp in Bangladesh

• A dynamic landslide hazard assessment for Kutupalong Rohingya Refugee Camp (KTP). • Slope units (SUs)-based landslide hazard assessment, unlike grid-based approaches. • Comparison of Generalized Additive Model with standard machine learning algorithms. • SUs with increased, decreased, and unchanged landslide susceptibility are identified. • SUs are flagged, which require immediate preventive actions by local authorities. Landslides disrupt human ecology worldwide, making practical hazard assessments essential for saving lives and assets. Dynamic assessments, which capture temporal changes in susceptibility, remain rare due to limited multi-temporal landslide inventories. This study addresses this gap by implementing a slope unit (SU)-based dynamic landslide hazard assessment for the Kutupalong Rohingya Refugee Camp (KTP), a region undergoing rapid environmental and landscape changes due to a massive human influx. Here, we decode dynamic landslide hazard processes by employing a Generalized Additive Model (GAM), a flexible statistical method, to explore how landslides (dependent variables) are linked to independent variables like continuous factors (e.g., slope, soil depth, rainfall) and categorical factors (e.g., soil type, landcover change category). We produced multi-temporal inventories to represent conditions before (2018 and prior) and after (2021 and prior) the establishment of refugee settlements. Anthropogenic modifications, such as distance from roads, land-use/NDVI changes, and rainfall, are treated as dynamic factors, while other factors are considered static predisposing conditions. Our GAM approach performs better than standard machine learning (ML) techniques (e.g., Random Forest, Support Vector Machine, Neural Networks), achieving an overall ROC-AUC of 0.84 and a mean cross-validated AUC of 0.81, compared to AUC (0.64-0.74) for ML models. We also performed uncertainty quantification and repeated random simulations (Monte Carlo simulations) to identify slope units with increased, decreased, or unchanged susceptibility. Priority SUs requiring immediate risk reduction measures are flagged, offering actionable insights for local authorities. Our research findings advance landslide hazard assessments by integrating time-varying dynamic processes with a slope units-based approach and facilitating risk mitigation at KTP.