DECODING DYNAMIC LANDSLIDE HAZARD ASSESSMENT FOR A MASSIVE REFUGEE CAMP (KTP) IN BANGLADESH

Landslides wreak havoc worldwide by disrupting human ecology. Practical landslide hazard information is crucial to saving lives and assets. Although dynamic hazard assessment provides more realistic insight than static assessments, such characterizations are infrequent because of the lack of multi-temporal landslide inventories. Dynamic hazard assessment is particularly important in a setting like the Kutupalong Rohingya Refugee Camp (KTP), where the landscape and environment have undergone drastic changes in a short time span to accommodate a massive human influx. Here, we attempt to decode dynamic landslide hazard processes for KTP by employing a Generalized Additive Model (GAM). The GAM is a semi-parametric data-driven model retrieves the optimal polynomial fit between landslide (dependent variables) and conditioning factors (independent variables). We successfully use GAM by defining linear (continuous) variables like slope, aspect, etc., and non-linear (categorical) variables such as soil types, landcover change & NDVI change category. To account for the temporal changes, we consider the period of time immediately following the refugee influx (2018 and prior) and after refugee settlements are well established (2021 and prior). We have produced multi-temporal inventories for these two scenarios and used distance from the road, land use change, and NDVI change as proxies for anthropogenic landscape modification. All other factors are treated as predisposing static factors. In this way, we collectively improve existing knowledge in several ways: 1) we performed slope unit-based landslide hazard assessment whereas most of the traditional methods are grid-based approach, 2) our dynamic time-lapse assessment of landslide hazards constrains the temporal change of landslide susceptibility of these slope units, and 3) our implemented slope-unit based GAM approach performs better than standard machine learning (ML) algorithms, i.e., Random Forest (RF), Support Vector Machine (SVM), Linear Discriminate Analysis (LDA), Extreme Gradient Boosting (XGBoost), and Artificial Neural Network (ANN). We validated our findings in three different means: 1) area under the curve (ROC-AUC) approach, 2) correlation between observed and estimated (model-predicted) landslide area, and 3) comparison between GAM-derived and ML outputs.