Application of Deep Learning and Rain‐On‐Grid Hydrodynamic Modeling for Catchment‐Scale Flood Forecasting in Guwahati, India

ABSTRACT This study integrates deep learning‐based real‐time rainfall forecasting with hydrodynamic flood modeling to provide near real‐time predictions of urban flood inundation. The research focuses on urban Guwahati, a rapidly urbanizing city in Northeast India prone to recurrent pluvial flooding. A novel hybrid model, Multivariate Singular Spectrum Analysis integrated with Dual Attention‐Long Short‐Term Memory (MSSA‐DA‐LSTM), was developed using hourly meteorological data (rainfall, temperature, pressure, humidity, and wind speed) obtained from the Indian Monsoon Data Assimilation and Analysis (IMDAA) dataset (2015–2022, 12 km resolution). A multistep ahead forecast is achieved using a direct forecast approach, demonstrating high predictive skill ( R 2 > 0.7) for lead times of +1 to +2 h. The forecasted rainfall serves as an input to the HEC‐RAS Rain‐on‐Grid (RoG) model to simulate urban flood inundation and depth variations. The proposed MSSA‐DA‐LSTM model demonstrated significant improvements over both standalone LSTM and MSSA‐LSTM approaches, reducing Root Mean Square Error (RMSE) by 72.41% and Symmetric Mean Absolute Percentage Error (SMAPE) by 74.04%. Improved forecast accuracy was primarily attributed to the dual‐attention mechanism, selectively emphasizing critical temporal and meteorological features, and a customized weighted loss function specifically capturing peak rainfall. The forecast driven HEC‐RAS RoG model simulated urban flood inundation depths ranging from 0.1 to 1 m during the June 2022 flood event, with deeper inundation (up to 3 m) observed in natural depressions and low‐lying urban areas. The simulated flood extent was rigorously validated using Sentinel‐1 SAR imagery, achieving a Hit Rate (HR) of 0.78 and Critical Success Index (CSI) of 0.62 at a 0.05 m threshold. A return period‐based flood hazard assessment identified severe flood risks in the study area. Overall, the proposed framework provides a practical and adaptable solution for urban flood forecasting, specifically in areas with limited data.