Abstract: Numerical simulation of spatial large-scale watershed landscape evolution faces significant challenges due to high computational demands and extensive storage requirements. Surrogate models are widely used to reduce computation time, but their application in addressing storage issues caused by large-scale grids remains limited. This study conducted landscape evolution simulations in the Tarim River Basin, proposing surrogate models that incorporate spatial downscaling through six prominent machine learning methods, and compared the performance of different models. Subsequently, simulations of landscape evolution and river course changes under four future climate scenarios were conducted based on the surrogate models. Results show that the model based on refined Gaussian Support Vector Machine method performs the best, with a root mean square error of 0.086 m and a determination coefficient of 0.954. During the simulation period (2021—2100), elevation increases are mainly distributed at the slope bottoms and near the river channels, while decreases are concentrated in the plains and basin areas; river course changes primarily occurs in the downstream plains and basin areas, with migrations of the main channel of the Tarim River and the Qarqan River exceeding 20 km. The river course changes in the study area are highly sensitive to climate change. This study presents a surrogate model for landscape evolution with spatial downscaling capabilities, addressing the both computational time and storage challenges, thereby enhancing the feasibility of quantitative analyses of large-scale watershed landscape evolution and river course modifications.