Remote sensing images have been applied to many aspects in Earth observation work. However, tons of optical remote sensing images are abandoned due to the information loss caused by the clouds and damage of sensing instruments. Recently, many deep learning methods have been proposed to reconstruct the missing information of remote sensing images but they will be non-effective when it comes to the condition where there is no training dataset. In this paper, we propose an unsupervised method which can reconstruct single remote sensing image without training datasets in a deep neural network. The main idea is to process a reference image of the corrupted image with a deep self-regression network and extract the internal map, which possesses the same spatial information as the reference image. The residual information of the corrupted image is used to constrain the spectral authority of internal map to obtain the reconstruction results. We apply the proposed method in three conditions: 1) dead pixel reconstruction, 2) multitemporal reconstruction and 3) heterogeneous data reconstruction. We conduct simulation experiments and real data experiments in three conditions to confirm the superiority of our methods. The results show that the proposed method outperforms some state-of-the-art algorithms.

遥感图像已应用于地球观测工作的许多方面。然而，由于云层造成的信息丢失和传感仪器的损坏，大量的光学遥感图像被遗弃。近年来，人们提出了许多深度学习方法来重建遥感图像的缺失信息，但在没有训练数据集的情况下，它们将是无效的。在本文中，我们提出了一种无监督方法，该方法无需在深度神经网络中训练数据集即可重建单个遥感图像。主要思想是用深度自回归网络处理损坏图像的参考图像并提取内部图，其具有与参考图像相同的空间信息。利用损坏图像的残差信息约束内部图谱的权威性，得到重建结果。我们在三种情况下应用所提出的方法：1) 死像素重建，2) 多时态重建和 3) 异构数据重建。我们在三种条件下进行模拟实验和真实数据实验，以证实我们方法的优越性。结果表明，所提出的方法优于一些最先进的算法。我们在三种条件下进行模拟实验和真实数据实验，以证实我们方法的优越性。结果表明，所提出的方法优于一些最先进的算法。我们在三种条件下进行模拟实验和真实数据实验，以证实我们方法的优越性。结果表明，所提出的方法优于一些最先进的算法。