An eye in the sky: a use-case for evaluating super resolution

Abstract

The aim of this research was to execute a proof of concept on the added value of deep learning methodologies as part of remote sensing analysis. This was done in collaboration with the Dutch Ministry of Defence to improve the knowledge on this within the geo-domain. Deep learning constituted two different applications as part of this research; super resolution, which is to increase the spatial resolution beyond it is original limit (Yue et al., 2022) and feature extraction. The latter is also interesting as applying a geo-analysis task on super resolution data can prove to be a suitable methodology to evaluate the result and also attributes to the increase scientific interest in deep learning within the field of remote sensing (Yang & Newsam, 2013).

For super resolution models with varying amounts of input data have been tested and the metrical evaluation showed no significant issues although the models could be further optimised. Augmenting data to increase the usability of a dataset proved promising in performance but not conclusive in its added value to modelling super resolution. Visually the super resolution models showed more detail in comparison to a Sentinel 2 image of the same area but their differences in metrics did not result in apparent visual differences between models.

Feature extraction showed that all super resolution models outperformed a Sentinel 2 based extraction model in metrics. In comparison to the ground truth road network the model proved difficult and below expectation.

The conclusion is therefore that super resolution and deep learning based analysis methodologies can be of added value for remote sensing analysis and usable in an accessible and application-oriented manner. On an absolute scale however both the evaluation metrics and evaluation analysis in the form of road extraction showed that it could definitely benefit from further optimization to improve both performance and generalizability of the models.

The discussion touched upon several aspects of the research that could attribute to this, including other types of satellite data, open-source modelling software and alternative analysis tasks using super resolution data.