*4.2. Ablation Study*

To investigate the effectiveness of each module in GAN-GL and its influence on the final glacial lake mapping results, an ablation study was performed. Several specific combinations of individual modules are as follows:


The results of the ablation study are shown in Table 3 and are based on the three GAN-GL datasets. The water attention module combined with the GAN-based structure (Attn + ISeg + ResNet-152) obtained the highest values of Precision (93.34%), Recall (92.01%), F1 score (92.17%), and IoU (86.34%).


**Table 3.** Experimental results of ablation study for the three glacial lake subsets.

Note: -<sup>1</sup> ISeg. - Attn + ISeg. - ISeg + ResNet-50. - ISeg + ResNet-101. - ISeg + ResNet-152. - Attn + ISeg + ResNet-50. - Attn + ISeg + ResNet-101. -Attn + ISeg + ResNet-152.

Comparison for attention module: Because the water attention mechanism enables the model to focus on the identification of lake pixels, the water attention module markedly improves the segmentation performance of the glacial lakes (with an increase of 2~3% in accuracy).

Comparison for ResNet backbone: We tested the effects of different ResNet backbones in the discriminator, including ResNet-50, ResNet-101, and ResNet-152. Table 3 shows that the deeper the layers of the ResNet backbone, the better its performance. This can be explained by the fact that ResNet-152 records more details about glacial lakes by using deeper convolution layers compared to ResNet-101 and ResNet-50. This facilitates the accurate extraction of the complex edges of glacial lakes.

Comparison for the discriminator: Clear improvements were observed in the evaluation results when the discriminator was used (e.g., the ISeg and ISeg + ResNet backbone, the Attn + ISeg and Attn + ISeg + ResNet backbone). This is because the discriminator can guide the generator to learn the real distribution of the data.

Furthermore, it should be noted that accuracies were the highest for the densitycropped dataset, which contains sufficient glacial lake information in each patch to improve the training level of the model. This shows that the density of glacial lakes in the training data is an easily overlooked but important factor that affects the overall segmentation results.
