A Head-to-Head Study of Ensemble and Deep Learning Algorithms for Flood Damage Prediction in Japan

Japan experiences frequent and severe flooding, yet accurate flood damage estimation remains a significant challenge. This paper examines the efficacy of contemporary machine learning methodologies in forecasting flood damage, juxtaposing them with conventional techniques through the analysis of 28 years of flood event data gathered from diverse Japanese water systems between 1993 and 2020. Specifically, five different regression models were considered: Random Forest, XGBoost, Support Vector Regression (SVR), a deep neural network (DNN), and linear regression as a basic standard. Using log-transformed flood damage per capita as the target variable, a hybrid feature selection method was applied that pinpointed population, catchment area, year, terrain slope, river code, and short-window antecedent rainfall as the most effective predictors. XGBoost and Random Forest did much better than SVR, DNN, and linear regression, which shows that tree-based ensembles work well with structured tabular data. These models can be used for operational risk assessment and resource allocation in hazard zones.