Novel view synthesis under sparse-view conditions remains a critical challenge for real-world scene reconstruction. Conventional Neural Radiance Fields demonstrate limited capability in exploiting priors from observed regions when input views are restricted, resulting in suboptimal reconstruction quality. Although recent approaches have attempted to incorporate uncertainty modeling, they typically focus on a single type of uncertainty, failing to capture the comprehensive uncertainty landscape. To address these limitations, we present ENeRF, a novel approach that jointly models both aleatoric and epistemic uncertainty through evidential learning. Our proposed method employs a Normal-Inverse Gamma distribution to represent color and their associated uncertainties. Additionally, an uncertainty-aware adaptive sampling strategy is introduced, which dynamically allocates computational resources to regions with high uncertainty, particularly focusing on areas with limited observational coverage. Experimental results show that the proposed method achieves notable improvements in image reconstruction quality and uncertainty estimation accuracy under extremely sparse input conditions, establishing a new paradigm for 3D reconstruction in resource-constrained environments.