针对视频卫星超时相数据高时间分辨率的特点，单帧超分辨率重建算法无法充分利用其信息，提出一种基于自注意力的超时相遥感图像超分辨率重建方法。对超时相序列帧先依据帧速率划分为多个时间组，利用3D卷积对空间及时间同时建模的特点提取不同时间组下的时空信息，之后通过改进的广泛自注意力残差块增加注意力计算范围并完成高动态映射，在实现超时相序列帧建模的同时提取丰富的空间细节信息，最后将多个时间组的特征融合，送入亚像素卷积提升分辨率完成重建。所提算法的优点在于，对超时相数据的多时间组处理充分利用了其丰富的时空信息，改进的自注意力块可在无配准情况下完成序列帧建模并提高空间细节信息的提取能力。在高分四号数据集上的实验表明，算法的主观视觉效果及客观评价指标均优于对比算法，在高分四号数据2倍上采样重建时，较双三次插值算法在PSNR值上提升了2.49dB以上，且在4倍重建时依然有很强的重建性能，较双三次插值算法有巨大提升。实验结果表明，该方法拥有较好的超分辨率重建效果，有利于超时相数据在各领域的应用。

Due to the high temporal resolution of hyper-temporal data of video satellite, the single-frame super-resolution reconstruction algorithm cannot make full use of its information. A hyper-temporal remote sensing image super-resolution reconstruction method based on self-attention is proposed. The hyper-temporal sequence frame is first divided into multiple temporal groups according to the frame rate, and the spatial and temporal simultaneous modeling characteristics of 3D convolution are used to extract spatio-temporal information in different time dimensions, and then through the improved wide self-attention residual block increase the range of attention calculation and complete high dynamic mapping, extract rich spatial detail information while realizing hyper-temporal sequence frame modeling. Finally, the features of multiple time groups are fused and sent to subpixel convolution to improve the resolution for reconstruction. The advantages of the proposed algorithm are that the multi-time group processing of hyper-temporal data makes full use of its rich spatio-temporal information, and the improved self-attention block can complete sequence frame modeling without registration and improve the ability of extracting spatial details. Experiments on the GF-4 data set show that the subjective visual effects and objective evaluation indicators of the algorithm are better than those of the comparison algorithm. Compared with the bicubic interpolation algorithm, the PSNR value of GF-4 data is improved by more than 2.49dB when the GF-4 data is 2 times reconstruction and the reconstruction performance is still very strong when the GF-4 data is reconstructed at 4 times, which is a huge improvement compared with the bicubic interpolation algorithm. Experimental results show that this method has a good super-resolution reconstruction effect and is beneficial to the application of hyper-temporal data in various fields.