The quantitative assessment of impact of pumping capacity and LID on urban flood susceptibility based on machine learning

Drainage facilities such as drainage pumping systems and Low Impact Development (LID) practices are effective measures to reduce urban flood risk. The quantitative identification of their influence on the reduction of urban flood susceptibility (UFS) is of great significance in providing scientific references for urban flood control. In this study, we constructed a conceptual method to investigate the spatial variation of UFS based on the machine learning models (i.e., Convolution Neural Network (CNN) and Support Vector Machine (SVM)), which has been tested in Wuhan City of China with good performances. After model evaluation, we have quantitatively studied the impact of two flood mitigation measures (pumping stations and LID practices) on the UFS. In particular, we evaluated the UFS mitigation efficiency of several designed scenarios using different combinations of pump discharges and LID area fractions by comparing them against default scenarios. We found a nonlinear negative response relation between the reduction of UFS with either the increase in pump discharge or LID area fractions. The proportion of the area of highest susceptibility (PAH) decreases as the pumping capacity increases, and when the pumping capacity is 2.5 times the default condition, the PAH reduces to 45% from 73.7% of no pump stations and reaches its minimum value. When the LID layout area is 100% of the whole region, the PAH can reduce to 51% from 67.7% of no LID. The findings can be beneficial for the design of optimal preventative strategy to sufficiently reduces UFS.