Determining flood source areas in watersheds using data-driven models and a geographic information system

Abstract

ABSTRACT This study explores the use of gene expression programming (GEP) and artificial neural networks (ANNs) to estimate flood index values based on the unit flood response (UFR) method in two adjacent watersheds (Ardak and Kardeh) located in northeast Iran. The performances of the studied data-driven models were compared according to certain statistical measures such as Root Mean Square Error (RMSE). The findings indicate that GEP models were more accurate than ANNs (RMSE = 0.0986 vs. 0.1512 for Ardak and RMSE = 0.1024 vs. 0.1112 for Kardeh, respectively). Another advantage of the GEP models was providing an explicit relationship between flood index values and physical attributes. As flood index values derived via the UFR method were close to each other, flood contribution area maps were developed using a geographic information system (GIS) to consider uncertainty. Then, fusion algorithms including ordinary averaging, linear regression, and GEP were applied to develop a flexible regional model.