Boosting algorithms for projecting streamflow in the Lower Godavari Basin for different climate change scenarios

Abstract

Abstract The present study investigates the ability of five boosting algorithms, namely Adaptive Boosting (AdaBoost), Categorical Boosting (CatBoost), Light Gradient Boosting (LGBoost), Natural Gradient Boosting (NGBoost), and eXtreme Gradient Boosting (XGBoost) for simulating streamflow in the Lower Godavari Basin, India. Monthly rainfall, temperatures, and streamflow from 1982 to 2020 were used for training and testing. Kling Gupta Efficiency (KGE) was deployed to assess the ability of the boosting algorithms. It was observed that all the boosting algorithms had shown good simulating ability, having KGE values of AdaBoost (0.87, 0.85), CatBoost (0.90, 0.78), LGBoost (0.95, 0.93), NGBoost (0.95, 0.95), and XGBoost (0.91, 0.90), respectively, in training and testing. Thus, all the algorithms were used for projecting streamflow in a climate change perspective for the short-term projections (2025–2050) and long-term projections (2051–2075) for four Shared Socioeconomic Pathways (SSPs). The highest streamflow for all four SSPs in the case of NGBoost is more than the historical scenario (9382 m3/s), whereas vice-versa for the remaining four. The effect of ensembling the outputs of five algorithms is also studied and compared with that of individual algorithms.