Estimating elements susceptible to urban flooding using multisource data and machine learning

The accuracy of flood susceptibility prediction (FSP) could be affected by inadequate representation of flood conditioning factors (FCFs) and the approaches used to identify the most relevant FCFs. This study analyzed twenty-eight FCFs derived from open-access earth observation datasets to develop FSP model for a highly urbanized Akaki catchment, which hosts and surrounds the capital city of Ethiopia, Addis Ababa. In the study, relevant FCFs were first identified using different collinearity-based and model-integrated feature selection methods, and sequentially introduced into a machine learning model. Simulated FSPs were compared against a reference flood inventory dataset to determine the most effective selection method. Findings show that: (i) using extreme rainfall indices improved the accuracy of FSP, (ii) Mean Decrease Impurity (MDI) was found to be the most effective feature selection method, (iii) geomorphological and physiographic FCFs showed the highest and the lowest predictive power, respectively, and (iv) the quantile method outperformed other approaches in classifying the flood susceptibility map. Findings indicate that an area of 217 km 2 , 43000 buildings, 163 km of paved roads and 0.54 million inhabitants are highly susceptible to flooding in the catchment. In particular, Addis Ababa contains almost 75 % of the estimated susceptible elements in only one-third of the catchment area. The results of this study provide valuable insights for urban planning and flood management, helping to reduce the socio-economic impacts of flooding and enhance urban resilience.