Risk-driven composition decoupling analysis for urban flooding prediction in high-density urban areas using Bayesian-Optimized LightGBM

Abstract

With catastrophic climate change and accelerated urbanization, urban flooding has emerged as the most influential hazard over last few decades. Therefore, a systematic study on the assessment of urban flooding vulnerability and evaluation of multidimensional relationship between flooding indicators and inundation depth is imperatively needed. Machine learning methods have been proven to be extremely effective in predicting urban flooding susceptivity based on a multivariate data-driven approach. In the present study, a cascade modeling chain was explored consisting of integrated Light Gradient Boosting Machine (LightGBM) and decoupling analysis of risk-driven composition, based on a multi-factor database consisting of hydro-meteorological, underlying-surfaces, and building configurations indicators. LightGBM was verified to be reliable and robust for urban flooding vulnerability assessment. Taking Shenzhen as a case study, the results indicated that rainfall volume (TOTAL_R), rainfall duration (LTIME), percentage of impervious surface (PIS), building congestion degree (BCD) and density of buildings (DB) were mainly responsible for the increase in flooding risk, while percentage of water coverage (PW) was highly efficient in flooding mitigation. Areas with high flooding risks are concentrated in older metropolitan areas when the rainfall volume surpassed 125 mm or the rainfall duration was longer than 55h. The urban flooding vulnerability was significantly increased when the PIS, BCD, and DB were greater than 14%, 0.58, and 16 n/ha, respectively. It was recommended that adaptation strategies should be implemented in high-density areas to increase urban resilience to extreme rainfall events. The findings of this study can serve as a scientific basis and technical support for sustainable urban stormwater management.