Early Earth observation data is available from former spy images taken by the CIA within the “Keyhole“ programme in the 1960s and 1970s. These so called “Corona“ images are available at a high spatial resolution but provide only panchromatic (grey-scale) information. Additionally, the images were taken with different camera systems and under different observation geometries, which causes various radiometric properties and diverse visual appearance of the images. However, the Corona images are very valuable for the assessment of historical changes in land cover and land use.
Given the limited radiometric information in those images, most methods for land cover classification make use of spatial information such as image texture (Deshpande et al., 2021). Additionally, one study has shown that an artificial colouring of the grey-scale images strongly improves the visual appearance and interpretability and can result in higher accuracy in land cover classification (Agapiou, 2021). A deep learning based method called DeOldify based on generative adversarial networks allows colourising historical Corona imagery (see illustration). However, in order to obtain a natural-looking appearance of the coloured image, a parameter needs to be determined that is affected by the grey-scale range of the original image. In order to compare and visualize images from several acquisitions, there is the need to match the grey scale information across images and to determine a parametrisation for the colouring that is consistent. Those needs should be addressed in this research.
We will provide you a big collection of georeferenced Corona imagery with derived land cover maps. First, different (histogram) matching methods should be tested and compared to achieve a grey-scale information that is consistent across images. Secondly, an approach should be developed based on DeOldify to obtain a consistent natural-like appearance of artificially coloured images.
Third, the obtained results should be visualised in an interactive web visualization based on a web mapping or tile service, allowing others to view data in a web browser or GIS client of their choice. In addition, there should be a possibility to obtain the underlying data of the service as a custom extract. The web viewer should offer an interface to create the selection for the extract. For this part of the task, there should be a draft of the aimed application and an investigation of possible software for the browser frontend and the backend software, which serves data. Finally, there should be some short evaluation of the created application or the colouring result of the Corona image data.

Staff working in this domain (potential supervisors): Matthias Forkel, Mathias Gröbe, Christopher Marrs, Lucas Kugler