Exploring interactions between socioeconomic context and natural hazards on human population displacement

Abstract

Climate change is leading to more extreme weather hazards, forcing human populations to be displaced. We employ explainable machine learning techniques to model and understand internal displacement flows and patterns from observational data alone. For this purpose, a large, harmonized, global database of disaster-induced movements in the presence of floods, storms, and landslides during 2016-2021 is presented. We account for environmental, societal, and economic factors to predict the number of displaced persons per event in the affected regions. Here we show that displacements can be primarily attributed to the combination of poor household conditions and intense precipitation, as revealed through the interpretation of the trained models using both Shapley values and causality-based methods. We hence provide empirical evidence that differential or uneven vulnerability exists and provide a means for its quantification, which could help advance evidence-based mitigation and adaptation planning efforts.

Fig. 1. Spatial distribution of newly displaced people (NDP) per sudden-onset disaster over selected countries for years 2016–2021.

A Colors represent the sum of NDP per country registered in the years under consideration; pie charts indicate the event counts and percentages with respect to the global number of events. B The total number of NDP per continent and hazard type occurred in the period of interest.

Fig. 2. Map of the areas impacted by sudden-onset hazards.

Polygons are given at the administrative level 1 or 2 or by a combination of the two, depending on the disaster-affected area. Color represents the total sum of newly displaced people (NDP) produced by hazards in each polygon in years 2016–2021.

Fig. 3. Performance of the trained Random Forest models.

A Predictions versus true values in logarithmic scale were obtained by averaging over all test batches in the bootstrapping. The color levels show the density of points. The Pearson correlation is 0.57. B Distribution of the R² on the hold-out set for all the bootstrapping iterations. The blue density is obtained with all the covariates, while the red one is obtained by excluding the two weather variables, namely maximum precipitation accumulation and maximum 10 m wind speed.

Fig. 4. Relation between the input features and their importance scores averaged over the different test batch configurations in the bootstrapping.

The horizontal axis represents the normalized feature values. At the same time, the color scale is given by the mean product between the weights of the linear model (A) and the feature value or the mean Shapley value per event for the Gradient Boosting Machines (B) and Random Forests (C). The covariates are displayed in decreasing order of importance.

Fig. 5. Scatter plots of Shapley values versus precipitation (A) and area (B), and box plots of the treatment effects obtained with causal forests (C, D).

In the upper plots (A, B), the color scale is given by the value of the Absolute Wealth Index (AWI). The blue and red curves are smoothed averages of the Shapley values for instances having AWI < 650 US dollars and AWI > 650 US dollars, respectively. In the bottom graphs (C, D), we show the distribution (median and spread) of the causal relationship between the target (i.e., NDP) and each of the covariates considered one by one as treatments.