Conventional methods including threshold method, spectral structure method, normalized difference water index and classification method are used to delineate and extract water body from TM/ETM+ data collected by Landsat with medium spatial resolution. However, none of them take the impact of mixed pixel into account. Mixed pixels, unavoidable on remote sensing image, generally result in low precision of pixel-based information extraction, which makes it hard to meet the requirements of highly accurate applications. In this article, the water area of the East Lake is calculated based on linear spectral unmixing method from Landsat TM image obtained in 2006. The result proves to be more accurate compared with those by pixel-based methods. Furthermore, linear spectral unmixing method for mixed pixel is exerted on varied Landsat TM/ETM+ imageries to monitor the shrink of lake area in Wuhan from 1995 to 2006. The monitoring results demonstrate that lake areas in Wuhan decreased continuously during this period, and the Index of Lake Loss Intensity (ILLI) was unevenly distributed across time and space.