225 / 2024-02-28 11:42:11

Data-driven modelling for urban flooding forecasting in Birmingham

PCA(principal component analysis),urban flood forecasting,logistic regression,face recognition,data-driven model

Compared with the physical model, the data-driven model can save the computational complexity of urban flood prediction. The study is an attempt to seek a relatively optimal data-driven model that could be used to set up for a flood forecasting system for the Birmingham city (UK), and identify critical variables to support the forecasting.



For this, the study area was divided into a regular 1 km2 grid, which matches the radar grid. Flood records from Birmingham City Council are analysed , for the period 1998-2016; rainfall data from rain gauges (2004-2017) and radar (2005-2018) are organised and integrated to determine the storm events corresponding to the flood records. And the rainfall characteristics are calculated for each event. The response variable is the number of flood records, per grid, per storm event, which was expressed by the proportion it accounted for the total number of flood records during an event. And the predictors are four rainfall characteristics: rainfall accumulation, peak rainfall rate, maximum return period and its corresponding duration.



Simple linear regression analysis is made between rainfall characteristics and flood data, but weak correlation is found. Further, based on logistic regression, rainfall accumulation is found to be the most critical predictor, while peak rainfall rate is proved to be a poor predictor. Principal component analysis (PCA) is conducted on rainfall accumulation maps, which identifies that similar rainfall accumulation maps would lead to similar flooding occurrence places. This provides an idea for urban flooding forecasting in the area where the correlations between rainfall and flood data are weak.