Synthesising aerial imagery with DCGAN
Note: This is a work-in-progress; I am currently adding to this post on a weekly basis.
The uses of Generative Adversarial Networks are now almost ubiquitous across the deep learning landscape. Their usefulness for data augmentation is an active are of research. Though they may successfully learn the ‘blob-like’ features of faces, such as noses and eyes, it is potentially much harder for GANs to replicate fine linear features such as those present in aerial imagery of peatland. I have given this a go; replicating papers and other blogs examining whether GANs can be used to generate realistic synthetics of aerial or satellite imagery.
Understanding the data
Data sizes
Normalisation
Checkpointing and trying to avoid silly mistakes
Resources
Generating Synthetic Multispectral Satellite Imagery from Sentinel-2