A machine learning framework for runoff simulation in small catchments: Integrating atmosphere–ocean–land large-scale indices and catchment variables

Study region Cherchen River catchment. Study focus This study presents a machine learning framework that integrates Extra Trees with Shapley Additive Explanations for variable selection, and adopts an Extended Long Short-Term Memory network as the core model for monthly runoff simulation. The framework considers multiple variables including atmosphere–ocean–land large-scale indices and catchment variables, and effectively identifies the top 10 key variables for runoff simulation from 756 variables. This study focuses on developing a runoff simulation framework for small catchments that generalizes across diverse hydrological conditions, while remaining effective under high dimensional inputs and limited local hydrological data. New hydrologic Insights In the Cherchen River catchment, annual runoff exhibited a significant increasing trend of 7.29 m³ s⁻¹ a⁻¹ from 1991 to 2021. Runoff variation is strongly influenced by local meteorological and hydrothermal conditions, such as wind speed and air temperature, and is also significantly associated with sea surface temperature variability in the Indian Ocean Warm Pool. In addition, large-scale circulation patterns, including the East Asian trough, the Northern Hemisphere polar vortex, and the mid-tropospheric geopotential height field over the Tibetan Plateau, play a secondary role relative to these factors. The framework proposed in this study achieves robust performance on the annual (MAE = 1.27 ×10 7 m 3 , RMSE = 1.34 ×10 7 m 3 , and R 2 = 0.99) and monthly (MAE = 7.03 ×10 6 m 3 , RMSE = 8.83 ×10 6 m3, and R 2 = 0.95) scales.