Deep learning method requires a substantial amount of labeled data to achieve the state-of-the-art performance. However, annotating a large volume of data is often costly and impractical. Active Learning is a approach that reduces labeling costs by intelligently selecting and annotating the most crucial data points, which benefits from the integration of uncertainty and diversity as key criteria for sampling. Existing uncertainty-based methods often fall short in capturing the distinct sources of uncertainty, resulting in a diminished quality of uncertainty estimation. Additionally, clustering is commonly used to ensure diversity, which requires multiple iterations but overlooks the global correlations present throughout the entire unlabeled dataset. As a result, a well-designed ad hoc combination is essential to balance uncertainty and diversity. To address above limitations, we propose Evidential Uncertainty-aware Determinantal Point Process active learning method. Specifically, we employ the theory of Subjective Logic to measure multifaceted uncertainty including vacuity and dissonance. On these grounds, we first focus on the samples with high dissonance and employ Determinantal Point Process to select the samples considering both vacuity and diversity. The proposed method explores and exploits the information associated with the latent feature space near the decision boundary to select the most valuable samples. The experimental results on various real-world datasets reveal the superiority of our method.