Abstract: Accurately identifying water features is a crucial technical means for analyzing the spatiotemporal changes of surface water.In response to the problem of low accuracy in various long-term water extraction methods, we utilized Landsat remote sensing imagery to select 5 484 scenes of usable imagery from the Yellow River source regions spanning from 1986 to 2022.Two methods, Convolutional Neural Networks(CNN)combined with Object-based Image Analysis(OBIA)and Multi-Index Water Detection Rules(MIWDR),were utilized to extract surface water features in the Yellow River source area.The accuracy of the two methods was compared and analyzed.Subsequently, the spatiotemporal characteristics of water features in the Yellow River source area from 1986 to 2022 were explored, and correlation analysis was conducted to investigate the main climatic factors.The results revealed that:(1) CNN-OBIA achieved an overall accuracy of 96.78% and a Kappa coefficient of 0.93,while MIWDR achieved an overall accuracy of 94.28% and a Kappa coefficient of 0.88.Overall, CNN-OBIA exhibited higher extraction accuracy than the MIWDR method.CNN-OBIA results better preserved the integrity of water boundaries, effectively removed mountain shadows, and improved the accuracy of extracting smaller rivers.(2) Total water area of the study area showed a decreasing trend in 1986～2001,followed by an increasing trend in 2001～2022.(3) Correlation analysis indicated a significant positive correlation between precipitation, temperature, and changes in water area.