Because of sensor malfunction and poor atmospheric conditions, remotebrsensing imagery usually suffers from missing information. Missing information inbrremote sensing imagery reduces the usage rate and hinders the follow-upbrinterpretations, making the reconstruction of them a critical problem. In recent years,brthe success of deep learning methods in computer vision and image processingbrshowed us a new way to solve such a problem. We propose this research to study thebrapplication of deep learning methods on the reconstruction of missing information inbrremote sensing imagery. In this research, Fully Convolutional Networks (FCN) isbradopted to reconstruct missing information in small regions with spatial and spectralbrreference data while Generative Adversarial Networks (GAN) is adopted tobrreconstruct missing information in large regions with multi-temporal and multi-sourcebrreference data. In the FCN model, equal numbers of convolution layers andbrdeconvolution layers are used to establish an end-to-end mapping from input to outputbrin the GAN model, a pair of Discriminator and Generator is trained simultaneouslybrwith remaining information in the target image and their corresponding reference data.brWe also proposed a universal deep model framework to explore the generalization ofbrdeep learning to other models. By defining the data flow between input feature mapsbrand output feature maps, we can build deeper structures with any model with multiplebrkernels. In the proposed research, we expect to achieve better experimental resultsbrthan any currently existing methods using the same type of reference data. On thebrother hand, we also expect to establish a non-neural-network deep model, whichbrachieves equal results to deep neural networks but takes less time to train. (keywords: remote sensing, deep learning, missing information, image reconstruction)

Runtime: 2018 - 2020