CONVOLUTIONAL NEURAL NETWORK-BASED PREDICTIONS OF POTENTIAL FLASH FLOOD HOTSPOTS IN SINGAPORE: INSIGHTS AND STRATEGIC INTERVENTIONS

Abstract

Amid increasing urbanization, changing climate, and limited storm water infrastructure, urban flooding is a global issue, and Singapore is no exception. Traditional identification of flood-prone areas in Singapore has relied on historical flash flood data. However, by applying the booming influx of big data across various domains, including geography, weather, and DEM data, and using the deep learning model, Convolutional Neural Network (CNN), this research proposes a method that can accurately and effectively predict flash flood spots in an urban environment. Specifically, datasets including elevation, slope, aspect, rainfall, canals, drainage, and land use are fed into the CNN model to predict the locations of flash floods. The model, with a testing accuracy of 0.962, generates a comprehensive flash flood assessment map identifying high-risk areas in Singapore. Contrary to the current flood-prone area identification, which classifies only 0.79%of the country as susceptible to flash floods based on historical events, our CNN model-based assessment indicates that 11.4% of the country is at high risk. These newly identified zones are predominantly located along the coastline and in low-lying watershed outlets. Additionally, we propose corresponding storm water infrastructure enhancements to mitigate flash flooding in these locations.