Determination of flood probability and prioritization of sub-watersheds: A comparison of game theory to machine learning

Floods often significantly impact human lives, properties, and activities. Prioritizing areas in a region for mitigation based on flood probability is essential for reducing losses. In this study, two game theory (GT) algorithms – Borda and Condorcet – were used to determine the areas in the Tajan watershed, Iran that were most likely to flood, and two machine learning models – random forest (RF), and artificial neural network (ANN) – were used to model flood probability (the probability of flooding). Twelve independent variables (slope, aspect, elevation, topographic position index (TPI), topographic wetness index (TWI), terrain ruggedness index (TRI), land use, soil, lithology, rainfall, drainage density, and distance to river) and 263 locations of flooding were used to model and prepare flood-probability maps. The RF model was more accurate (AUC = 0.949) than the ANN model (AUC = 0.888). Frequency ratio (FR) was calculated for all factors to determine which had the most influence on flood probability. The values of twelve factors that affect flood probability were estimated for each sub-watershed. Then, game-theory algorithms were used to prioritize sub-watersheds in terms of flood probability. A pairwise comparison matrix revealed that the sub-watersheds most likely to flood. The Condorcet algorithm selected sub-watersheds 1, 2, 4, 5, and 11 and the Borda algorithm selected sub-watersheds 2, 4, 5, 20 and 11. Both models predicted that most of the watershed has very low flood probability and a very small portion has a high probability for flooding. The quantitative analysis and characterization of the watersheds from the perspective of flood hazard can support decision making, planning, and investment in mitigation measures.