Spatial Loss Function for Super-Resolution of Geoscientific Data

Abstract

Super Resolution is a method for artificially increasing the imaging system's resolution by post processing without having to collect new datasets. It is mostly developed and used in computer graphics by the computer science community for image and video enhancement due to its capacity to add spatial variations in the data and perform better than conventional interpolation methods such as bicubic interpolation been used extensively in the geoscience community. After the advancement of deep learning-based super-resolution methods in the 2010s, it has shown great potential for use in data scare regions where high-resolution geoscientific data are not available, and collection of such data is also not possible due to financial and technical reasons. Even though Super-Resolution is used geospatial data, the loss functions to optimize those models are mostly used as it is from computer vision, and they do not account for the spatial relationship between neighbouring pixels. For example, in the case of elevation data, the relative elevation between two pixels (slope) and their direction (aspect) is more important compared to absolute elevation in the case of geoscientific modelling. However, those relations which must be valid in the ground observation are not well considered in Super-Resolution of Geospatial data.

Our research aims to develop a loss function that can respect the spatial relationship between pixel and its neighbours, representing the ground reality. In contrast to existing methods, these new loss functions are better in generating the geospatial datasets that can be further used in geoscientific analysis and modelling approaches rather than mere visualization. The developed loss function is tested with multiple super-resolution models, both generative adversarial networks and the end-to-end models trained with geoscientific data. Our research shows that using spatial relation aware loss function and the super-resolution model can better reconstruct the ground reality even though training them is more complex than using simple mean squared error loss functions. The use of such a novel loss function can also generate better terrain in the case of Digital Elevation Models, which can be observed in the slope and aspect of such datasets.