Closure to 'Runoff Estimation by Machine Learning Methods and Application to the Euphrates Basin in Turkey' by Abdullah Gokhan Yilmaz and Nitin Muttil

摘要

The authors wish to thank the discusser for clarifying various issues regarding the work in Yilmaz and Muttil (2014), and they believe that this discussion will contribute to a better understanding of the study. Their responses to the comments of the discusser are as follows: 1. The authors agree with the discusser that machine learning (ML) methods and multiple linear regression (MLR) gave very similar results as mentioned in the article. This was expected because the relationship between flows from the upstream and downstream stations were linear or near-linear, as pointed out by the discusser. This study was undertaken to demonstrate the applicability of ML methods for flow estimation and also for infilling missing data in the records of selected stations in upper Euphrates Basin (UEB) and middle Euphrates Basin, (MEB). It should be highlighted that the capability of the models to simulate peak flows was also investigated to select models to infill the missing data because peak flow values are extremely important for water-engineering design and for flood mitigation measures. Moreover, peak flow magnitude and timing are very important in snow driven basins, because they are a significant indicator of climate change (Yilmaz and Imteaz 2011). Although ML and MLR models gave similar results for flow simulation, ML models outperformed MLR models for peak flow simulation as explained in the "Construction of Missing Data" section on page 1,023 of the article. Therefore, the authors selected ML models [i.e., Adaptive Neuro-Fuzzy Inference System Model 3 (AM3) and Genetic Programming Model 3 (GP3) for Station 2119, and Feedforward Neural Network Model 3 (M3) and AM3 for Station 2174] to infill the missing data. The authors agree with the discusser's comment, but they also believe that the demonstration of the applicability of ML methods for flow estimation and especially for peak flow simulation was an important contribution.