A fusion-based framework for daily flood forecasting in multiple-step-ahead and near-future under climate change scenarios: a case study of the Kan River, Iran

This study proposes a novel fusion framework for flood forecasting based on machine-learning, statistical, and geostatistical models for daily multiple-step-ahead and near-future under climate change scenarios. An efficient machine-learning model with three remote-sensing precipitation products, including ERA5, CHIRPS, and PERSIANN-CDR, was applied to gap-fill data. Four individual machine-learning models, including Random Forest, Multiple-Layer Perceptron, Support Vector Machine, and Extreme Learning Machine, were developed twelve days ahead of streamflow modeling. Then, three fusion models, including Random Forest, Bayesian Model Averaging, and Bayesian Maximum Entropy, were applied to combine the outputs of individual machine-learning models. The proposed framework was also implemented to downscale the precipitation variables of three general climate models (GCMs) under SSP5-8.5 and SSP1-2.6 scenarios. The application of this approach is investigated on the Kan River, Iran. The results indicated that individual models illustrated weak performance, especially in far-step-ahead flood forecasting, so it is necessary to utilize a fusion technique to improve the results. The RF model indicated high efficiency in the fusion step compared to other fusion-based models. This technique also demonstrated effective proficiency in downscaling daily precipitation data of GCMs. Finally, the flood forecasting model was developed based on the fusion framework in the near future (2020–2040) by using precipitation data from two scenarios. We conclude that flood events based on SSP5-8.5 and SSP1-2.6 will increase in the future in our case study. Also, the frequency evaluation shows that floods under SSP1-2.6 will occur about 10% more than SSP5-8.5 in the Kan River basin from 2020 to 2040.