A comparison of performance measures of three machine learning algorithms for flood susceptibility mapping of river Silabati (tropical river, India)

Abstract

Flood is the most common phenomenon causing extensive disruption to the environment, socio-economy, infrastructure and many other aspects of human life. Flood susceptibility mapping (FSM) is a crucial step for policymakers in disaster management. However, in the present study, we applied three ensemble machine learning models, namely, Random Forest (RF), Naive Bayes (NB), and Extreme Gradient Boosting (XGB) for FSM of Silabati river (tropical river, India). A total of 500 historical flood points and field observations with considering set of twelve flood influencing factors (rainfall, elevation, slope, curvature, stream power index (SPI), Sediment Transport Index (STI), Terrain ruggedness index (TRI), topographic wetness index (TWI), clay content in soil (SC), distance from the river (DFR), drainage density (DD), and land use and land cover (LULC) for generating the training and validation datasets. To investigate and perceive the flood vulnerability of the study basin, five factors, such as elevation, DD, rainfall, DFR and SC turn out to be the most dominating factors out of the adopted twelve factors considered for the present study in all models. It is found that an area of around 36.08% of the total basin comes under the very high to high FSM. The prediction ability and performance efficiency of three models were comparison and validation measures by statistical techniques such as multicollinearity diagnosis test, Kappa index, MAE, (Mean absolute error), RMSE (Root mean square error), Pearson's correlation coefficients and receiver operating characteristic (ROC). The highest ROC was achieved by the RF model (84.7%), followed by the XGB model (83.1%), and NB model (82.1%) respectively. The RF model performs better for FSM then the other models.