*4.1. Implementation Details and Evaluation Metrics*

Segmentation experiments were conducted using Tensorflow 1.14 on the Python 3.7 platform. The GAN-GL dataset was split into 70% for training and 30% for validation. In the training stage, ResNet-152 in the discriminator was pre-trained on ImageNet. The training of the model was configured with a batch size of 1 for 100 epochs, and the optimizer used was AdamOptimizer, with a learning rate of 0.0001. To quantitatively evaluate the glacial lake mapping accuracy, the number of glacial lake pixels was counted using the predicted mask and the true labeled mask, and five performance indicators, Precision (P), Recall (R), Overall Accuracy (OA), F1 Score (F1), and Intersection over Union (IoU) were used. The corresponding formulations are as follows:

P = all correctly predicted water pixels/all predicted pixels;

R = all correctly predicted water pixels/all water pixels;

OA = all correctly predicted pixels/all pixels;

F1 = 2 × P × R/(P + R);

IoU = (predicted water pixels ∩ true water pixels)/(predicted water pixels ∪ true water pixels).
