**3. Methods**

The architecture of our proposed GAN-GL model for the segmentation of glacial lakes is shown in Figure 3. In GAN-GL, we incorporated a water attention module and image segmentation module into the generator. The discriminator was designed based on ResNet-152 to encode the lake area as vectors and determine their categories. Given a remotely sensed image input, the generator attempts to produce glacial lake masks. Then, the generated masks and true labeled masks are both fed into the discriminator for training until they can correctly predict whether the input data are generated or real. In the following subsections, we describe each process in more detail.

**Figure 3.** Architecture of the proposed GAN-GL model, which mainly consists of three parts—A water attention module and an image segmentation module in the generator, and the ResNet-152-based discriminator.

#### *3.1. Generator*
