As an essential resource in High Mountain Regions (HMR), evaluating the glacial lake dynamics is of great significance to explore the impacts of climate changes and predict the risks of Glacial Lake Outburst Floods (GLOFs). However, complicated and laborious methods for glacial lake mapping are unpractically applied in automatically monitoring glacial lakes at a large-scale region. In this work, we explored the mapping efficiency of glacial lakes by combing Generative Adversarial Networks and multi-level feature pyramid (GANMFP) in HMR. We first sampled the image patches containing glacial lakes from Landsat-8 raw data to evaluate the model performance. Totally, 6583 patches with 256x256x7 pixels are randomly cropped from 62 Landsat-8 images. Then we employed these data to train and test the GAN model, which integrated a multi-level feature fusion module in generator and a Resnet-152 networks in discriminator. From the validation results and result visualization, our method achieved good performances in Precision (88.42), Recall (59.61), and Overall Accuracy (99.28), which shows excellent potential in glacial lake mapping at a large-scale region. © 2021 IEEE.