A data-based landslide susceptibility map of Africa

Abstract Our understanding of the spatial patterns of landslides in Africa is limited with available landslide studies typically focusing on only one or a few study areas. Moreover, Africa is clearly underrepresented in terms of available landslide inventories. This study aims to produce a first continent-wide landslide susceptibility map for Africa, calibrated with a well-distributed landslide dataset. We reviewed the literature on landslides in Africa and compiled all available landslide inventories (ca. 10,800 landslides), supplemented by additional landslide mapping using Google Earth imagery in underrepresented regions (ca. 7250 landslides). This resulted in a dataset of approximately 18,050 landslides. Various environmental variables were investigated for their significance in explaining the observed spatial patterns of landslides. To account for potential mapping biases in the dataset, we used Monte Carlo simulations that selected different subsets of mapped landslides to test the significance of the considered environmental variables. Based on these analyses, we constructed two landslide susceptibility maps for Africa: one for all landslide types and one excluding the known rockfalls. In both maps, topography is by far the most significant variable. We evaluated the performance of the fitted multiple logistic regression models using independent subsets of landslides, selected from the total dataset. Overall, both maps perform very well in predicting intra-continental patterns of landslides in Africa and explain about 80% of the observed variance in landslide occurrence. To further test the robustness and sensitivity to mapping biases, we also modelled landslide susceptibility while excluding regions with arid climates, as landslides in these environments are expected to be better preserved over time and therefore likely relatively overrepresented. Despite this potential bias, the effect on the landslide susceptibility model is limited. Based on the constructed database and our analyses we further discuss potential research gaps for landslide prediction in Africa and at continental scales. For example, analysis of the African countries' mean landslide susceptibility shows a lack of landslide research in various countries prone to landsliding (e.g.: Guinea, Gabon, Lesotho, Madagascar). Apart from the intrinsic value of this landslide susceptibility map as a natural hazard risk management tool, the map and compiled database are highly promising for other applications. For example, we explored the potential significance of landslides as a geomorphic process by confronting our landslide susceptibility map with an available database of measured catchment sediment yield for 500 rivers in Africa. Overall, a significant positive, but relatively weak relation between landslide susceptibility and sediment yield is observed.