Deep Learning of Geospatial Patterns for Remote Sensing Image Downscaling

Remote sensing image downscaling can be regarded as an inverse problem that each coarse pixel is corresponding to infinite combinations of fine pixels. With the increasing demanding of fine imagery in various fields, many researchers have devoted themselves to the studies of efficient downscaling algorithms in the past few decades. Geostatistical methods, particularly the kriging family of methods, have been extensively used in remote sensing image downscaling to account for the geospatial patterns of imagery. These geostatistical methods enjoy the flexibility of integrating point spread function and the possibility of preserving spectral information of original imagery [e.g., area-to-point kriging (ATPK)]. As one of state-of-the-art geostatistical methods, ATPK, which relies on two-point statistics, is incapable of modeling complex geospatial patterns in heterogeneous area. In the past few years, deep learning-based algorithms have shown great potential in learning complex spatial patters for various computer vision applications, and many researchers have successfully adopted these methods in different remote sensing applications including image downscaling. The application of deep learning in remote sensing, however, is largely limited by the requirement of massive training datasets which are extremely difficult to come by. Recently, an unsupervised deep learning method, namely Deep Image Prior (DIP), was proposed for single image super-resolution without any training data. In this study, we explore the performance of DIP in modeling complex geospatial patterns for remote sensing image downscaling and integrate it into a regression framework to include the statistical relationships with available ancillary datasets. In this thesis, the methodological details of DIP and the DIP-based regression framework are first discussed. We then apply the deep learning-based framework into real case studies (with Landsat and MODIS) and highlight the advantages with a performance comparison with ATPK-based framework.