Projecting future climate extremes in the glacier-fed upper indus basin using machine learning based downscaling of CMIP6 GCMs
摘要
Climate change threatens Pakistan’s Upper Indus Basin, where future projections are hindered by uncertainties and limited reliable data in this glacier-fed region. This study employs two machine learning (ML) algorithms, Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs) to downscale climatic data of ten Coupled Model Intercomparison Project phase 6 (CMIP6) global circulation models (GCMs) for improved regional applicability. ML algorithms were trained and tested using GCMs and observed historical data (1985–2014). The future downscaled data was used to assess future climate projections at twelve UIB meteorological stations for two periods: near-term (2026–2055) and long-term (2056–2085) under two Shared Socioeconomic Pathways SSP245 and SSP585. The Modified Mann-Kendall (MMK) test was used for the seeking of trends, while RClimdex was applied to calculate twelve precipitation and temperature extreme climate indices (ECIs). For CNNs, the values of R² (0.93–0.71), KGE (0.93–0.77), RMSE (3.00–5.90), and PBIAS (–3.90–1.00%) indicated slightly better performance than ANNs for downscaling. Based on R² and KGE, the top five CMIP6 GCMs (MIROC6, MRI-CM6-1, CNRM-ESM-2, CanESM5, and IPSL-CM6A-LR) were selected for future climate projections. Trend analysis indicates robust warming signals in minimum and maximum temperatures across all downscaled CMIP6 GCMs by CNNs while precipitation trends remain uncertain and inconsistent among models and scenarios. The CNNs-downscaled five CMIP6 GCMs projected large variability in the annual mean precipitation and temperature over the UIB. Under SSP245, annual precipitation change ranged from + 89.27% to − 81.50% in the near-term (NT) and from + 74.24% to − 50.74% in the long-term (LT), while under SSP585 the range was + 86.00% to − 81.44% (NT) and + 86.00% to − 77.63% (LT). Projected mean Tmax under SSP245 varied between + 3.79 °C and − 5.89 °C (NT) and + 3.97 °C to − 5.49 °C (LT), whereas under SSP585 it ranged from + 6.41 °C to − 2.43 °C (NT) and + 3.41 °C to − 3.81 °C (LT). Similarly, mean Tmin under SSP245 varied between + 2.37 °C and − 5.04 °C (NT) and + 1.91 °C to − 5.96 °C (LT), while under SSP585 it ranged from + 5.47 °C to − 3.95 °C (NT) and + 2.71 °C to − 6.56 °C (LT). ECI of precipitation (Sdii, rx1day, R10mm, R 25 mm, R95p, R99p, Prcptot) and temperature ((Su25, Tr20, Txx, Txn, Tnx, Tnn) revealed pronounced warming trends, particularly under the higher-emission SSP585 scenario. Both models consistently predicted temperature increases, with CNNs capturing more nuanced spatial variations and extreme event patterns. The research underscores the critical need for comprehensive climate adaptation strategies in the UIB, emphasizing the potential for significant climatic transformations in the coming decades.