Enhancing Flood Susceptibility Mapping Through Machine Learning and Multi-Criteria Decision-Making Techniques

Abstract

Flooding has historically caused massive social, economic and environmental damage affecting lives and livelihood. There is a concerted effort to study the phenomenon of flooding, hazard assessment, mitigation and prevention of flood and associated risks. In the planning phase, Flood Susceptibility Mapping (FSM) to identify vulnerable areas are useful to  provide critical data for preparedness, risk management, and sustainable land use planning. Most flood susceptibility maps are generated using multi-criteria Decision models (MCDM) and Machine Learning models. The reliability and accuracy of Flood Susceptibility Maps are increased by systematic evaluation of multiple flood related factors and their interrelationship which are more easily analyzed by machine learning models. Therefore, the aim of the study is to  develop FSM for Toorsa River and the Pasakha River. using the two multi-criteria Decision models, the Shannon Entropy model and the Criteria Importance Through Intercriteria Correlation  method. Further FSM were also developed using two machine learning algorithms of Random Forest  and Support Vector Machine.  Factors such as elevation, slope, soil type, and land use/land cover, rainfall, and Topographic Wetness Index were assigned weights based on the two MCDM techniques to develop the maps. Similarly, the same factors were used for training ML models and validating their performance in flood-prone area classification. The results of the study shows that Area Under Curve scores were high (0.98) in both the study area using Random Forest while the lowest score of 0.21 was obtained for Pasakha river using the Shannon Entropy Model. In general the Machine Learning models are found to be more accurate, which may be attributed to its ability to interpret data in a non-linear manner unlike the MCDM methods. The final Flood Susceptibility Maps of the study area were produced based on Random Forest models, as it provided the most accurate results.