Over the past half-century, global warming and humidification have led to an accelerated rate of glacier melting in China, highlighting the increasing importance of monitoring glacier distribution. However, current automated glacier extraction methods have significant limitations, such as boundary fragmentation, omission of glaciers in shaded mountain areas, and misclassification in cloud-covered regions. To address these issues, this study selected Menyuan County in Qinghai Province as the experimental area. Sentinel-2 imagery and DEM data were utilized, applying object-oriented automatic classification technology in combination with the C5.0 decision tree model to develop a multi-feature glacier extraction rule set and a neighborhood feature rule set. An improved two-stage object-oriented glacier extraction method was subsequently proposed. The findings revealed that glaciers exhibited distinct response patterns across various features, including spectral mean, spectral standard deviation, NDSI (normalized difference snow index), DEM (digital elevation model), adjacency, and slope orientation. A two-stage glacier extraction method effectively enabled automatic glacier extraction. It also significantly enhanced the recognition accuracy in cloud-covered and shaded mountain regions, achieving an overall glacier recognition accuracy of 98.50%.