This paper addresses polarization detection in Urdu, a low-resource language characterized by complex morphology and insufficient annotated data. We formulate the task as a binary classification problem of social media posts into polarized and non-polarized categories. Our approach is based on Urdu-BERT, a language-specific transformer model combined with language-specific preprocessing, duplicate removal, and data augmentation to mitigate class imbalance and improve generalization. Experimental results show that the fine-tuned Urdu-BERT outperforms TF-IDF-based lexical machine learning baselines and achieves strong performance relative to multilingual transformer baselines. The findings indicate that language-specific pretrained transformers, when combined with appropriate preprocessing and augmentation strategies, provide an effective and generalizable framework for low-resource Urdu polarization detection.