Abstract: A light field （LF） video captures both the temporal evolution and angular information of a scene， enabling advanced immersive applications such as free-viewpoint videos and virtual reality. However， the extremely high dimensionality of LF video data results in massive storage and transmission costs， which severely limit its practical deployment. To address this challenge， sparse-sampling-based reconstruction methods have emerged as effective solutions for low-bitrate LF video compression. By encoding and transmitting only a small number of key views， these methods reconstruct the full LF video at the decoder using deep-learning models， thereby significantly reducing the required bitrate. However， despite their success， most existing sparse-LF-reconstruction methods suffer from a fundamental limitation in that they treat LF videos as collections of independent frames and adopt a frame-by-frame reconstruction strategy. While such approaches effectively exploit the spatial-angular correlations within a single LF frame， they completely overlook the rich temporal information inherently present in LF video sequences. As a result， these methods fail to leverage information from adjacent frames to assist the reconstruction of regions that are poorly constrained by sparse angular sampling in the frame under consideration， leading to a clear performance bottleneck. This limitation becomes particularly severe in scenes with large disparities and frequent occlusions. In such cases， the sparse views within a single frame are often insufficient to infer the content of occluded regions， causing missing structures or texture degradation in the reconstructed LF views. Nevertheless， because of scene or camera motion， these occluded regions are often visible in neighboring frames at the corresponding spatial locations. Ignoring this temporal redundancy prevented the existing methods from fully exploiting the available information for high-quality LF video reconstruction. To overcome these challenges， this study proposes a sparse coding and dense reconstruction framework for LF videos incorporating temporal information， which explicitly incorporates temporal cues into the sparse-reconstruction process. The core component of the proposed framework is an LF video reconstruction network based on temporal enhancement， which is designed to move beyond the traditional independent-frame processing. The proposed network introduces a dedicated temporal-enhancement module that facilitates effective feature interactions across adjacent frames. Specifically， the temporal-enhancement module employs deformable convolutions to adaptively align features extracted from neighboring reference frames with those of the current target frame. This adaptive alignment mechanism allows the network to handle complex motions and large disparity variations， enabling it to selectively retrieve complementary texture details from temporally adjacent frames. By fusing these aligned features， the proposed network compensates for the missing information in the current frame due to sparse angular sampling or occlusions while implicitly enforcing temporal consistency during reconstruction. Extensive experiments conducted on representative LF video datasets demonstrate that the proposed framework consistently outperforms the state-of-the-art sparse-LF-reconstruction methods under various sampling configurations. Notably， the proposed method achieves superior rate-distortion performances without introducing significant additional bitrate overheads. Qualitative results further demonstrate that the proposed temporal-enhancement strategy effectively restores occluded regions and preserves fine texture details， particularly in challenging large-disparity scenes. These results validate the integration of temporal information as an effective and practical solution for improving sparse LF video reconstruction quality.