Unravelling flood complexity: statistical and neural network approaches for Cauvery River Basin, India

Floods are widespread natural calamities with substantial socio-economic effects that demand adequate management measures and forecasts. In this study, the most popular traditional statistical distribution techniques, Gumbel, Log Pearson-III (LP-III), Log-Normal (LN) and three soft computing techniques such as Artificial Neural Networks (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Adaptive Neuro-Fuzzy Inference System-Firefly Algorithm (ANFIS-FFA) were examined for their potential to predict floods. These approaches were employed for modelling yearly maximum discharge at M.H. Halli, T. Narasipur, Kollegal, Biligundulu, Urachikottai, Kodumudi, and Musiri gauging stations of the Cauvery River using 40 years (1980 to 2019) data. Two statistical constraints, the wilton index (WI) and the root mean square error (RMSE), are employed to determine the performance of the proposed hybrid model. The results showed that, for M.H. Halli, Biligundulu, Urachikottai, Kodumudi, and Musiri gauging stations, ANFIS-FFA gave the highest WI values as 0.9141, 0.9636, 0.9205, 0.9373, and 0.8939, whereas and for T. Narasipur, and Kollegal gauging stations the ANN gave the highest WI value 0.8524 and 0.9440, respectively, during the testing phase. Futhermore, the coefficient of determination (R2) for ANFIS-FFA at M.H. Halli, Biligundulu, Urachikottai, Kodumudi, and Musiri gauging stations were 0.9140, 0.9636, 0.9205, 0.9378, and 0.8999, respectively and for ANN at T. Narasipur, Kollegal gauging stations were of R2 as 0.8574 and 0.9440, respectively. Based on these results, it can be concluded that the soft computing techniques (ANFIS-FFA and ANN) outperformed the statistical techniques.