Abstract: Natural resource monitoring and water resource management boost the demand for high-precision and high-frequency river monitoring. However, conventional semantic segmentation models based on deep learning are constrained by high requirements for samples and limited generalization capabilities. Meanwhile, the Segment Anything Model (SAM), despite strong generalization and zero-shot learning (ZSL) capability, faces challenges in automatically identifying land cover types. To address these issues, this study proposes a coarse-to-fine intelligent river information extraction framework oriented to remote sensing images by integrating classical deep learning models with SAM. First, coarse river information extraction is conducted by employing a classical semantic segmentation model that integrates atrous spatial pyramid pooling (ASPP) and convolutional block attention module (CBAM). Then, a method is proposed to automatically acquire positive and negative sample points of SAM based on conventional image processing algorithms, such as Canny edge detection and flood fill. This method enables the fully automatic acquisition of high-precision sparse prompt points. The results indicate that compared to classical semantic segmentation models, the coarse-to-fine river extraction framework proposed in this study achieved an increase of 7.4% in the mean Intersection over Union (mIoU). Under the same basic SAM framework, the proposed method for automatically acquiring positive and negative sample points increased the mIoU of river information extraction by 5.7% compared to the random extraction of prompt points. This method can fully leverage the advantages of conventional and large models, enabling high-precision, fully automatic, and intelligent river information extraction with few samples and low costs. This study will provide technical support for applying the results of intelligent 0river information extraction to engineering practices.