ANN-Based Climate Downscaling for Future Precipitation Projections in the Mi Oya Basin Using CMIP6 Data

This study investigates the use of artificial neural networks (ANNs) and a unidirectional Long Short-Term Memory (LSTM) model for statistical downscaling of precipitation projections from the CanESM5 climate model in the IPCC Sixth Assessment Report (AR6), focusing on the Mi Oya River Basin, Sri Lanka. Climate downscaling addresses the mismatch between the coarse spatial resolution of General Circulation Models (GCMs) and the finer scales needed for localized climate impact assessments. Despite the global uptake of CMIP6 data sets, their application for statistical downscaling in Sri Lanka remains limited. CanESM5 was selected from a multi-model CMIP6 ensemble, based on correlation coefficient, root-mean-square error, and standard deviation metrics for the baseline period (1980–2014). Leveraging the capacity of LSTMs to capture complex, non-linear, and sequential relationships, a unidirectional many-to-one LSTM architecture was employed to align with real-time forecasting constraints using only historical inputs. The model achieved strong performance in baseline testing, with Nash–Sutcliffe Efficiency (NSE) ≈ 0.43 and low normalized RMSE, demonstrating its suitability for projection purposes. Downscaled future precipitation (2020–2100) under Shared Socioeconomic Pathway (SSP) scenarios indicates increasing rainfall trends, with SSP5-8.5 showing the largest changes, including moderate increases in seasonal rainfall (typically within 10–40% based on median values) and enhanced variability towards the end of the century. These findings demonstrate the potential of machine learning–based downscaling to enhance localized climate projections and support decision-making in water resource management, disaster preparedness, and climate-resilient agricultural planning in vulnerable tropical basins.