Optimization of state-of-the-art fuzzy-metaheuristic ANFIS-based machine learning models for flood susceptibility prediction mapping in the Middle Ganga Plain, India
Aman Arora,Alireza Arabameri,Manish Pandey,Masood A Siddiqui,U K Shukla,Dieu Tien Bui,Varun Narayan Mishra,Anshuman Bhardwaj
DOI: https://doi.org/10.1016/j.scitotenv.2020.141565
2021-01-01
Abstract:This study is an attempt to quantitatively test and compare novel advanced-machine learning algorithms in terms of their performance in achieving the goal of predicting flood susceptible areas in a low altitudinal range, sub-tropical floodplain environmental setting, like that prevailing in the Middle Ganga Plain (MGP), India. This part of the Ganga floodplain region, which under the influence of undergoing active tectonic regime related subsidence, is the hotbed of annual flood disaster. This makes the region one of the best natural laboratories to test the flood susceptibility models for establishing a universalization of such models in low relief highly flood prone areas. Based on highly sophisticated flood inventory archived for this region, and 12 flood conditioning factors viz. annual rainfall, soil type, stream density, distance from stream, distance from road, Topographic Wetness Index (TWI), altitude, slope aspect, slope, curvature, land use/land cover, and geomorphology, an advanced novel hybrid model Adaptive Neuro Fuzzy Inference System (ANFIS), and three metaheuristic models-based ensembles with ANFIS namely ANFIS-GA (Genetic Algorithm), ANFIS-DE (Differential Evolution), and ANFIS-PSO (Particle Swarm Optimization), have been applied for zonation of the flood susceptible areas. The flood inventory dataset, prepared by collected flood samples, were apportioned into 70:30 classes to prepare training and validation datasets. One independent validation method, the Area-Under Receiver Operating Characteristic (AUROC) Curve, and other 11 cut-off-dependent model evaluation metrices have helped to conclude that the ANIFS-GA has outperformed other three models with highest success rate AUC = 0.922 and prediction rate AUC = 0.924. The accuracy was also found to be highest for ANFIS-GA during training (0.886) & validation (0.883). Better performance of ANIFS-GA than the individual models as well as some ensemble models suggests and warrants further study in this topoclimatic environment using other classes of susceptibility models. This will further help establishing a benchmark model with capability of highest accuracy and sensitivity performance in the similar topographic and climatic setting taking assumption of the quality of input parameters as constant.