Abstract

Flooding has been a significant problem over the past century in the United States (US), causing growing threats to human lives and socioeconomic damage. In the state of Kansas, since 1996, more than 1,500 flood events were recorded, resulting in an economic loss of between US$2b and US$5b. Many factors influence flood-susceptibility at a local scale. It may be helpful and timely to improve community resilience to flood disasters in Kansas. Our initial step was to assess factors that trigger flooding using Machine Learning (ML). Six ML algorithms: 1) Logistic Regression (LR); 2) Random Forest (RF); 3) Support Vector Machine (SVM); 4) K-nearest neighbor (KNN); 5) Adaptive Boosting (Ada Boost); 6) Extreme Gradient Boosting (XG boost) were used to evaluate their ability to classify locations in terms of floodsusceptibility. The learning data for these ML algorithms comprised a geo-spatial database of twelve floodsusceptibility factors from 1,528 flood inventories since 1996. The susceptibility factors comprised: rainfall, elevation, slope, aspect, flow direction, flow accumulation, Topographic Wetness Index (TWI), distance from the nearest stream, evapotranspiration, land cover, land surface temperature, and hydrographic soil type. The ML algorithms were compared, and the best algorithm was selected to estimate floodsusceptibility for each location in the geodatabase resulting in a flood-susceptibility map. A sensitivity analysis of floodsusceptibility factors indicated that the intensity or magnitude of the rainfall, land cover and soil type were the most significant factors for Kansas during this period.