Flood susceptibility mapping with machine learning, multi-criteria decision analysis and ensemble using Dempster Shafer Theory

Abstract

Listen

Translate article

Floods are one of the most widespread natural hazards occurring across the globe. The main objective of this study was to produce flood susceptibility maps for the province of Salzburg, Austria, using two multi-criteria decision analysis (MCDA) models including analytical hierarchical process (AHP) and analytical network process (ANP) and two machine learning (ML) models including random forest (RF) and support vector machine (SVM). Additionally, we compare which of the MCDA and ML models are better suited for flood susceptibility and evaluate the use of Dempster Shafer Theory (DST) for optimising the resulting flood susceptibility maps based on eleven flood conditioning factors: elevation, slope, aspect, topographic wetness index (TWI), stream power index (SPI), normalised difference vegetation index (NDVI), geology, rainfall, land cover, distance to roads and distance to drainage. The accuracy evaluation of the flood susceptibility maps through the AUC (area under the receiver operating characteristic curve) method along with the relative flood density (R-Index) shows that RF (AUC = 87.8%) and SVM (AUC = 87%) outperform the ANP (AUC = 86.6%) and AHP (AUC = 85.9%) models. Therefore, the predictive performance of ML models was slightly better than the MCDA models. The DST could further increase the accuracy of both ML models (AUC = 88.3%) and MCDA models (AUC = 87.3%). However, the best accuracy (AUC = 89.3%) is reached through an ensemble of all four models.