This paper proposes an extensible ground weapon system classification model based on a dataset with 11 categories obtained through web crawling and publicly available operation and training videos. A base model is trained using this dataset to extract discriminative features, similar to conventional supervised learning. By applying few-shot learning and triplet loss, an extended model is trained based on the previous training. The evaluation shows the base model achieved 93.35% to 97.3% accuracy for 5 categories and 93.54% to 99% accuracy for 11 categories. The extended model, achieved 72.18% to 75.54% accuracy when trained on 5 categories and tested on 11. These experiments demonstrate the effectiveness of few shot learning and triplet loss in creating an extended model capable of predicting new, untrained categories even with limited data.