<![CDATA[This study investigates long-tailed satellite cloud image classification by comparing CNN and Vision Transformers (ViT) built upon vision–language foundation models. A large-scale satellite cloud dataset with 11 highly imbalanced classes, including a dominant non-phenomenon category, is used to represent realistic atmospheric variability. The data are split using stratified sampling, standardized to a fixed resolution, and used to fine-tune CLIP-based backbones from RemoteCLIP and GeoRSCLIP through parameter-efficient adaptation. Several loss functions Cross Entropy, Logit Adjustment, Focal, Class-Balanced, and label-distribution–aware variants are evaluated, along with experiments examining majority-class removal and adapter bottleneck adjustments. Initial results show that Logit Adjustment causes majority-class collapse under default settings. After optimization, ViT-based models consistently outperform CNN models, achieving higher accuracy and more balanced macro-level performance. Class-Balanced loss emerges as the most effective objective, offering a strong trade-off between overall accuracy and per-class fairness. Increasing the adapter bottleneck dimension further boosts ViT performance, enabling the best configuration to match or exceed prior benchmarks while improving minority-class recognition. The final optimized model is deployed in a web-based prediction system, demonstrating the practical potential of foundation-model approaches for satellite-driven weather analysis.]]>
