Research Projects

Flash floods, characterized by their sudden onset, high intensity, and short duration (typically under six hours), cause significant human and economic losses globally. In India, between 1978 and 2006, there were 225 documented flash flood events leading to 4,756 fatalities. The increasing frequency and severity of such events due to climate change highlight the urgent need for reliable, high-resolution, real-time flash flood forecasting systems. Current operational flash flood forecasting systems depend largely on ground-based gauge observations, which, while accurate at point locations, lack the spatial and temporal coverage necessary for detecting and predicting highly localized flash flood events. Moreover, majority of such approaches are deterministic, or based on statistical approaches, thus limiting their ability to accurately forecast flash floods. Satellite-based Precipitation Estimates (SPEs) offer a potential solution by providing near real-time rainfall data at global scales. However, these datasets suffer from biases and often lack the spatial resolution required to detect fine-scale three-dimensional (3D) storm structures, which often trigger flash floods. This project aims to address these critical limitations by developing an advanced flash flood forecasting framework using a combination of satellite data, numerical weather prediction (NWP) models, deep learning techniques, and distributed hydrological modelling. The approach will begin by evaluating the 3D spatiotemporal features of historical extreme rainfall events, such as peak intensity, storm core location, spatial spread, etc., and thereafter comparing them with those obtained from SPEs. A novel hybrid deep learning model, combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, will be developed to correct the 3D biases and spatially downscale the SPEs. The corrected SPE will then be fused with short-term NWP nowcasts using a Bayesian Model Averaging approach. This fusion will ensure that both the spatial memory and temporal evolution of rainfall events are preserved and better captured. The combined rainfall product will serve as input to a distributed hydrological model—LISFLOOD and LISFLOOD-FP—which will simulate runoff, river flow, inundation extent, and flood depths in near real time. The Brahmani–Baitarani basin in eastern India has been selected as the pilot area due to its high susceptibility to flash floods caused by intense rainfall, complex topography, and frequent cyclonic activities. By integrating satellite and model forecasts through ML-based bias correction and dynamic data fusion, the system will offer improved accuracy in both timing and magnitude of flash flood predictions. The outcomes will be especially valuable for safeguarding lives and infrastructure in vulnerable regions, and the methodology can be scaled to other high-risk basins across India.