Evaluation of machine learning, information theory and multi-criteria decision analysis methods for flood susceptibility mapping under varying spatial scale of analyses

Abstract

The annual average economic losses due to various natural disasters are increasing exponentially across the globe and have reached a mark of US$239.2 billion per year between the period 2000–2019. In India, due to flood hazards, approximately 6.5 million people were affected and the total economic damages were approximately US$3 billion per year between 1990 and 2019. This study proposes to compare and evaluate the performance of the Support Vector Machine (SVM), Shannon's Entropy (SE), and Analytical Hierarchy Process (AHP) methods in the preparation of flood susceptibility mapping and assess their contextual suitability. The study has been carried out in coastal districts of West Bengal, India namely, South 24 Parganas and East Midnapur. Nine flood conditioning factors were identified and used as input parameters for the study. Each flood susceptibility map was subdivided into four categories where 27.29% of the study area were categorized as highly susceptible, while 39.45% of area were categorized as moderately susceptible to flood hazards. The Receiver Operating Characteristic (ROC) curve and the Seed Cell Area Index (SCAI) were applied for accuracy assessment and validation of the selected models. AUC values for ROC curves showed that the performance of the SVM model is better than that of the AHP and SE methods. Moreover, Hot Spots analyses for identification of high flood susceptible clusters revealed that the SVM model was far superior when compared to the other two models. The factors selected for assessment of flood susceptibility in AHP and SE models could not respond to the changing scale of analyses and hence produced significant erroneous outcomes. Hence, the SVM model evolved as the most versatile method in this comparison.