A Positive-Unlabeled Learning Algorithm for Urban Flood Susceptibility Modeling

Flood susceptibility modeling helps understand the relationship between influencing factors and occurrence of urban flooding and further provides spatial distribution of flood risk, which is critical for flood-risk reduction. Machine learning methods have been widely applied in flood susceptibility modeling, but traditional supervised learning requires both positive (flood) and negative (non-flood) samples in model training. Historical flood inventory data usually contain positive-only data, whereas negative data selected from areas without flood records are prone to be contaminated by positive data, which is referred to as case-control sampling with contaminated controls. In order to address this problem, we propose to apply a novel positive-unlabeled learning algorithm, namely positive and background learning with constraints (PBLC), in flood susceptibility modeling. PBLC trains a binary classifier from case-control positive and unlabeled samples without requiring truly labeled negative data. With historical records of flood locations and environmental covariates, including elevation, slope, aspect, plan curvature, profile curvature, slope length factor, stream power index, topographic position index, topographic wetness index, distance to rivers, distance to roads, land use, normalized difference vegetation index, and precipitation, we compared the performances of the traditional artificial neural network (ANN) and the novel PBLC in flood susceptibility modeling in the city of Guangzhou, China. Experimental results show that PBLC can produce more calibrated probabilistic prediction, more accurate binary prediction, and more reliable susceptibility mapping of urban flooding than traditional ANN, indicating that PBLC is effective in addressing the problem of case-control sampling with contaminated controls and it can be successfully applied in urban flood susceptibility mapping.