Machine learning-based streamflow projections in the upper indus basin under CMIP6 climate scenarios

Pakistan's water security relies heavily on the Upper Indus basin (UIB), which provides nearly half of the country's surface water essential for agriculture, domestic use, and hydropower generation. Therefore, accurate future streamflow projections of this region are crucial for effective water resources management. This study used machine learning algorithms, artificial neural networks (ANNs) and convolutional neural networks (CNNs), to downscale ten Coupled model intercomparison project phase 6 Global Circulation Models (GCMs) climatic data at a regional scale. This research also developed ANNs and CNNs algorithms to project streamflow in the Upper Indus Basin (UIB) using downscaled climate data under SSP245 and SSP585 scenarios for the period of 2026–2055 and 2056–2085, validated against observed streamflow (1985–2015). The results indicated that ANNs and CNNs showed strong performance during testing and training for the downscaling of climatic data and based on (Coefficient of determination) R² and (Kling–Gupta Efficiency) KGE values, five best performing downscaled GCMs (i.e., MIROC6, CNRM-CM6-1, CNRM-ESM2-1, CanESM5, IPSL-CM6A-LR) were selected for streamflow projection. For streamflow projection, CNNs model significantly outperformed than ANNs model during both training and testing phases, achieving mean squared error (0.0302) as the loss function and mean absolute error (0.0032) as the evaluation metric. Future streamflow projections, based on downscaled climatic data from five GCMs using the CNNs algorithm, showed an increasing behavior for the periods 2026–2055 and 2056–2085 compared to the 1985–2014 baseline, with a more intense increase projected under the SSP245 scenario than under SSP585. Among the five GCMs analyzed, CNRM-ESM2-1 projected the most intense increase in streamflow, with rises of 84 % and 97 % for the periods 2026–2055 and 2056–2085 under SSP245, and 59 % and 91 % under SSP585 respectively. This improved understanding of future streamflow will guide adaptive management for hydropower and water resources in UIB.