Abstract: Glacier remote-sensing segmentation suffers from high inter-class spectral similarity and strong texture homogeneity, which makes high-quality annotation costly and limits the size of labeled datasets. To address this, we proposed LmgFormer, a global multi-scale lightweight glacier segmentation algorithm based on adversarial learning. SegFormer was adopted as the backbone of the generator to extract features and produce pseudo feature maps. The atrous spatial pyramid–object level context (ASP-OLC) module was designed to jointly model multi-scale features with atrous convolutions and object-level contextual information, while reducing the overall number of parameters and computations. The global attention mechanism (GAM) was introduced to further extract and fuse global channel-spatial interaction features. The fully convolutional discriminator (FCD) replaced the conventional classification discriminator, enhancing feature discriminability and model robustness through adversarial learning on real and generated segmentation results. The conditional random field (CRF) was finally employed to remove isolated artifacts in the segmentation maps. Experimental results show that LmgFormer achieves 94.2% accuracy on glacier segmentation with only 8.123 M parameters.