Landsat at 3 m resolution?!? Applying a super resolution model to surface water detection

Keywords: Super resolution, SR, Landsat, Planet labs, geoAI, lake-size distribution, small object detection, upscaling, downscaling, ESRGAN
Abstract Type: Paper Abstract

Abstract

Landsat is the longest-running environmental satellite program. However, its sustained 30 m resolution since 1982 prohibits the detection of prevalent small water bodies. Single image super resolution (SISR)—the enhancement of pixel resolution of an image—offers a neural network-based method to reconstruct fine-scale information from coarser satellite images. Here, we obtain 289 Planet Dove-R (PS2.SD) scenes from a globally-inclusive domain and create high-resolution and low-resolution image pairs for a range of hydrologic environments worldwide. The low-resolution Planet images, derived from bicubic resampling, are used to train a SISR GAN model (Enhanced Super Resolution GAN, ESRGAN) that resamples them to a 10x-finer resolution (i.e. from 30 m to 3 m). Here, we demonstrate that this model, trained entirely on Planet data, can be used to resample 30 m Landsat 5 and 8 optical bands to native 3 m Planet resolution. We then show that this high-resolution output can be used to detect sub-pixel water bodies unresolvable by Landsat, with a minimum mapping unit (MMU) of ~ 2/3 of a Landsat pixel for small water bodies that balances its efficacy at detecting lakes against its false positive rate. Much work remains to be done surrounding technical and ethical guidelines for the creation, use, and dissemination of SR satellite imagery. This ability to generate realistic, high-resolution imagery from Landsat data dating to 1985 has exciting applications beyond water detection and paves the way for high-resolution land cover classification and object detection from the Landsat archive.