A hybrid method for small river-water extraction using TM images, covering Hengshan County in Shaanxi Province and acquired in August 20, 2010 and August 2, 1986, is proposed. After the pretreatment of the original image data, WFS feature space is built. Then, WFS is segmented to remove the influence of background spectral interference by aid of overlay analysis with thematic maps, such as present land use map, topographic map and drainage map. Next, the multispectral images containing the preliminary water distribution information are processed with LBV transformation and object oriented segmentation. Further, the precise extraction of river water can be achieved by using SVM supervised classification and mathematical morphology open close operator. Finally, water dynamic analysis is accomplished by adopting the precise water change information acquired from the above results. Results show that using the method provided can get precise water distribution information in Hengshan County, especially can improve the identification accuracy for small river. The map accuracy of water extraction results in 1986 and 2010 are 0.921 and 0.875, respectively, and the user’s accuracy are 0.913 and 0.862, respectively. The hierarchical extraction method proposed is feasible and reliable for small river-water extraction, can reduce the error of loess hilly and gully region identification, significantly.