Abstract: Across the globe, there exist numerous data-scarce regions characterized by incomplete ground-based hydrological monitoring networks and insufficient observational data. Traditional hydrological models which rely heavily on high-quality long-term in-situ measurements are difficult to apply effectively in such areas, posing significant challenges to runoff simulation and water resource management. Meanwhile, advancements in remote sensing technology and deep learning methodologies have offered novel solutions to addressing this predicament. This study focuses on the Tentek River Basin in southeastern Kazakhstan as the research area. A CNN-LSTM runoff simulation model driven by multi-source remote sensing data was constructed, with dozens of types of remote sensing and reanalysis datasets incorporated into the model. Daily-scale runoff simulation was thereby realized in the basin. The results demonstrate that the proposed model can still achieve relatively accurate runoff simulation in the absence of support from ground-based observational data. Specifically, the model yielded the optimal performance when utilizing a data combination comprising GPM (Global Precipitation Measurement) precipitation data, near-surface air temperature, deep soil moisture content, Normalized Difference Vegetation Index (NDVI), and snow depth, with the Nash-Sutcliffe Efficiency (NSE) up to 0.869. This study verifies the feasibility and practical values of remote sensing data in hydrological modeling for data-scarce regions, and provides a viable approach for runoff simulation in such areas.