Stochastic LSTM network for flood hazard map forecast

Abstract

Flood hazard has gained importance worldwide in recent decades owing to its impact on human lives, socio-economic, and infrastructure and is one of the most destructive in natural disasters. Consequently, several flood prediction models are developed either using physical modeling, statistical approaches, or data-driven machine learning (ML) approaches. While developing physical models would provide the desired insights, it is often challenging to obtain extensive information regarding the river basin and hydrological parameters and requires expertise. Additionally, often, they fail to predict short-term accurately. ML approaches have the advantage of mimicking any complex function when trained with sufficient data, and the data requirements drastically reduce with the incorporation of appropriate physical constraints. Hence, over the years, ML approaches contributed significantly to the advancement of flood forecasting and reducing hazards.   A recent study estimated €581 million in damage annually to the EU railways. In regards to the French railway network it is 28,000 km long and, therefore, exposed to various natural hazards. There is, therefore, a dire need to estimate the flood hazard to railway infrastructure accurately. Initially, data is curated about the frequency of occurrence and the flood intensity that can impact railways. From the collected data, feature engineering is performed to understand parameters (e.g., corresponding to hydrological, geology, and topography) that cause significant hazards. The obtained parameters are used to develop an LSTM-based forecasting model for short-term and long-term predictions and consequently develop hazard maps along the regions of railway infrastructure that are significantly prone to risk. Finally, we conclude the work with a few mitigation strategies to reduce the risk.