Graph-Based Propagation for Multispectral Remote Sensing Image Completion

Image completion refers to the problem of recovering the missing, corrupted or obscured entries in image data. In this paper, we consider the problem in the remote sensing domain, where regions of an image are missing due to difficulties such as cloud cover, sensor failures or partial sensor coverage. Where previous work in this field generally falls into the category of low-rank completion methods, we propose a novel graph-based diffusion approach to the problem. The method, referred to as GraphProp, propagates observed entries around a graph-based representation of the image region in order to recover the missing entries. The graph-based diffusion approach to completion is to the best of our knowledge a novel method for remote sensing image completion. Using real-world multispectral image data acquired from the Landsat 7 platform, we validate our approach using experiments which synthetically obscure image sections. In these tests, we benchmark against alternative image completion approaches and demonstrate the superior reconstruction performance of our method versus the state of the art. Code which implements the method has been made publicly available at https://github.com/iainrolland/GraphProp.