Projected changes in Arctic river streamflow and shifting climatic drivers: A linear ensemble deep learning approach

Rapid Arctic warming is significantly altering the hydrological regimes of northern high-latitude rivers, with profound impacts for terrestrial water cycles, regional ecosystems, and global climate feedback. In this study, we develop a linear ensemble deep learning approach (LEDLA) that integrates multiple distinct deep learning architectures through linear ensemble meta-models, leveraging their complementary strengths to quantify the river discharge using hydrometeorological variables and project six major Arctic rivers’ streamflow under four Shared Socio-economic Pathway (SSP) scenarios. In the testing dataset, LEDLA obtains 0.5%-10.6% improvement in NSE across various Arctic watershed streamflow predictions, demonstrating that integrating complementary deep learning architectures enhances prediction accuracy and robustness. Projections for 2080-2100 indicate annual streamflow increases of 6.1%-15.3% across six major Arctic watersheds under SSP126-SSP585 scenarios relative to the 2000-2020 baseline. The Kolyma River basin, which is fully covered by the continuous permafrost, is projected to experience the most substantial streamflow increases (32.2%-47.3%). Our results reveal complex permafrost-streamflow dynamics: initially, permafrost degradation suppresses streamflow sensitivity to warming (as observed in the Kolyma River basin, with 100% permafrost covered) while enhancing precipitation-driven responses (as in the Lena River basin, 95% permafrost covered). However, under continued warming and reduced permafrost extent, rising evapotranspiration becomes the dominant control, leading to streamflow declines, such as a 2.4%-8.4% reduction projected for the Ob River (39% permafrost covered) by 2050-2100 compared to 2023-2050. Winter streamflow is projected to rise 1.7-8.7 times relative to historical means, rising by 27%-393% under SSP126-SSP585, with diminished seasonality and muted spring freshet peaks by the end of the century. Streamflow changes are most pronounced under SSP585, driven by accelerated warming and intensifying freeze-thaw transitions. These findings underscore the critical role of permafrost-hydrology interactions in modulating Arctic river discharge and highlight the need to incorporate such dynamics into Earth system models and adaptive water resource planning under rapid climate change.