Flood Risk Assessment in Peninsular India: Integrating Data, Multivariate Statistics, and Machine Learning

Abstract

Floods are among the most significant natural hazards in India; however, their drivers, spatial coherence, and potential impacts remain poorly understood, particularly in data-scarce catchments. This thesis addresses these challenges by integrating newly developed hydrological data, multivariate risk assessment, spatial flood analysis, and machine learning-based modelling to improve the characterization of flood potential across Peninsular India. The study develops CAMELS-IND —a large-sample hydrometeorological dataset encompassing 472 catchments in Peninsular India. The dataset comprises 41 years of catchment-averaged meteorological time series, observed streamflow for a subset of catchments, and over 200 catchment attributes derived from remote sensing, reanalysis, and ground-based sources. CAMELS-IND adheres to international standards, facilitating regional and global hydrological studies and providing the foundational data infrastructure for subsequent analyses. In the next chapter, the thesis investigates the compound nature of flood events by examining the joint influence of rainfall, soil moisture, and storm surge in two catchments: one in peninsular India and the other in the United Kingdom. Using copula-based joint distribution functions, the study reveals that these flood drivers frequently co-occur and exert interdependent effects, leading to significant deviations in return period estimates when compared to traditional univariate methods. The findings underscore the importance of adopting a multivariate framework in flood risk assessment. Subsequently, the spatial organization and synchronization of floods across Peninsular India are analysed using long-term streamflow records. The results indicate increasing regional coherence in flood occurrence, often extending beyond individual river basin boundaries. Five distinct clusters of catchments with synchronous flood behaviour are identified using F-madogram-based dissimilarity measures and partitioning algorithms. These clusters highlight the limitations of basin-centric flood assessments and suggest the need for a regional perspective in flood hazard mapping and management. Finally, a novel hybrid modelling framework is proposed to estimate seasonal quick flow—an indicator of flood generating surface runoff—using climate-based water availability and a machine learning-derived Baseflow Index (BFI). The framework demonstrates robust performance across diverse hydroclimatic regimes and enables spatially distributed flood sensitivity and flashiness assessment in both gauged and ungauged catchments. Together, the contributions of this thesis advance the understanding of flood-generating processes, enhance data availability, and provide practical tools for regional flood risk assessment. The findings have implications for hydrologic modelling, infrastructure planning, and climate-resilient water resources management in data-limited and hydrologically diverse regions.