River surfaces serve as a type of river boundary condition that can be extensively observed and automatically extracted using remote sensing observation. Accurate river surfaces provides vital data for the study of river research, such as research on watershed-scale river evolution, total water resources assessment, and river carbon dioxide emission estimation. However, extracting extremely small river surfaces from satellite imagery remains a challenging issue, which easily leads to missing river information. In this research, we proposed GRF-ANN method for the extraction of river surfaces based on deep learning, in order to tackle the challenge of extracting river surfaces during dry season periods. This study focused on the Wuding River Basin, a primary tributary of the Yellow River Basin, with dry climate and seasonal rivers, and utilizes multi-source high-resolution remote sensing data to extract entire river surfaces of the whole river basin. The extraction results show that the Kappa coefficient of GRF-ANN is 0.89, with a river extraction accuracy of approximately 92%. Compared to existing research, this method achieved a 25.6% improvement in accuracy, showed a favorable extraction results. Meanwhile, the study significantly enhanced the extraction efficiency by establishing a CPU-GPU intelligent acceleration algorithm and optimizing the storage structure, resulting in 5 times increases in computational speed. The study provided methods and data supports for the extraction of river geometric information such as cross section morphology, and the research of seasonal river morphology and spatial distribution in mountainous areas.