Authors: Ethan D. Kyzivat*, Brown University, Ekaterina M.D. Lezine, Brown University, Laurence C. Smith, Brown University
Topics: Water Resources and Hydrology, Remote Sensing, Land Use and Land Cover Change
Keywords: Deep learning, machine learning, landcover classification, surface water mapping, hydrology, Arctic, Canada, GAN
Session Type: Virtual Paper
Start / End Time: 8:00 AM / 9:15 AM
Room: Virtual 49
Presentation File: No File Uploaded
The Canadian Shield (CS), a Precambrian geologic formation underlaying most of central and eastern Canada, is the most lake-rich region on Earth. Recent studies using high-resolution CubeSat satellite imagery have revealed its surface water hydrology to be surprisingly dynamic at fine spatial scales. However, the high-resolution satellite data necessary to observe these scales has only been readily available in recent years. Here, we propose a method for resampling 30m resolution satellite imagery to 3m using a neural network and show that it is possible to accurately detect water from the high-resolution output.
We use a generative adversarial network (GAN) to perform this resampling through a process known as super resolution (SR). We degrade high-resolution Planet CubeSat images of the CS, then resample the coarsened imagery back to its native resolution using our model and traditional cubic resampling (CR). To test the accuracy of the GAN-generated SR imagery, we apply a water classification and find that SR outperforms CR, as measured by the Cohen kappa coefficient, for shoreline regions with substantial differences between the two methods. This work opens up the possibility of retroactive application of SR to older, coarser-resolution satellite datasets to infer historical changes in fine-scale surface water dynamics.