Flood susceptibility mapping at Ningdu catchment, China using bivariate and data mining techniques
Book Chapter
Khosravi, K, Melesse, AM, Shahabi, H et al. (2019). Flood susceptibility mapping at Ningdu catchment, China using bivariate and data mining techniques
. 419-434. 10.1016/B978-0-12-815998-9.00033-6
Khosravi, K, Melesse, AM, Shahabi, H et al. (2019). Flood susceptibility mapping at Ningdu catchment, China using bivariate and data mining techniques
. 419-434. 10.1016/B978-0-12-815998-9.00033-6
A flood susceptibility map is the first and the most important action in flood mitigation and risk assessment. The main goal of the this study is to prepare flood susceptibility maps for Ningdu catchment in China using frequency ratio (FR) as bivariate model and Logistic Model Tree (LMT) and Random Forest (RF) as data mining techniques as well as their comparison. At first, flood inventory mapping was constructed using 156 flood historical locations which finally divided randomly into two parts by a ratio of 70:30 for training (111 locations) and testing (48 locations) or model building and validation, respectively. Information gain ratio (IGR) and multicollinearity diagnosis tests were then applied to select conditioning factors. The results of the IGR showed that the altitude is the most effective factor contributing to flood occurrences at the Ningdu catchment while curvature factor was removed from modeling due to null predictive capability. Three flood susceptibility maps were prepared using the three aforementioned models and validated using some criteria including receiver operating characteristic (ROC), area under the ROC curve (AUC), kappa, root mean square error (RMSE), and mean absolute error (MAE). The results showed that LMT model has the highest prediction power (AUC=0.982), followed by RF (AUC=0.978) and FR (AUC=0.969). Results also indicated that according to the LMT model, about 21% of the study area which accounts for 836km2 has been covered by high and very high flood susceptibility classes.
