To address limitations in existing deep unfolding methods regarding stage-wise feature modeling, we propose a novel Unfolding network based on Dual Priors (UDP) for hyperspectral image reconstruction. By interpreting the reconstruction task as a Bayesian optimization problem, UDP takes advantage of deep neural networks to achieve Proximal Gradient Descent (PGD) unfolding. More specifically, our UDP framework simultaneously integrates spectral and spatial priors through a hybrid design: Spectral-wise Dual Attention Block (SDAB) captures spectral dependencies via self-attention and cross-stage attention mechanisms, while Convolutional Attention Block (CAB) enhances spatial modeling capability through a Transformer-inspired structure and multi-scale fusion. Experimental results on both simulated and real degraded datasets demonstrate that UDP achieves state-of-the-art performance in PSNR and SSIM, especially under complex spectral conditions. Furthermore, UDP exhibits superior reconstruction quality with lower model complexity, showcasing potential in real-world remote sensing and spectral imaging applications.