Contemporary and historical detection of small lakes using super resolution Landsat imagery: promise and peril

Landsat is the longest-running environmental satellite program and has been used for surface water mapping since its launch in 1972. However, its sustained 30 m resolution since 1982 prohibits the detection of small water bodies, which are globally far more prevalent than large. Remote sensing image resolution is increasingly being enhanced through single image super resolution (SR), a machine learning task typically performed by neural networks. Here, we show that a 10x SR model (Enhanced Super Resolution Generative Adversarial Network, or ESRGAN) trained entirely with Planet SmallSat imagery (3 m resolution) improves the detection of small and sub-pixel lakes in Landsat imagery (30 m) and produces images (3 m resolution) with preserved radiometric properties. We test the utility of these Landsat SR images for small lake detection by applying a simple water classification to SR and original Landsat imagery and comparing their lake counts, sizes, and locations with independent, high-resolution water maps made from coincident airborne camera imagery. SR images appear realistic and have fewer missed detections (type II error) compared to low resolution (LR), but exhibit errors in lake location and shape, and yield increasing false detections (type I error) with decreasing lake size. Even so, lakes between similar to 500 and similar to 10,000 m(2) in area are better detected with SR than with native-resolution Landsat 8 imagery. SR transformation achieves an F-1 score for water detection of 0.75 compared to 0.73 from native resolution Landsat. We conclude that SR enhancement improves the detection of small lakes sized several Landsat pixels or less, with a minimum mapping unit (MMU) of similar to 2/3 of a Landsat pixel - a significant improvement from previous studies. We also apply the SR model to a historical Landsat 5 image and find similar performance gains, using an independent 1985 air photo map of 242 small Alaskan lakes. This demonstration of retroactively generated 3 m imagery dating to 1985 has exciting applications beyond water detection and paves the way for further SR land cover classification and small object detection from the historical Landsat archive. However, we caution that the approach presented is suitable for landscape-scale inventories of lake counts and lake size distributions, but not for specific geolocational positions of individual lakes. Much work remains to be done surrounding technical and ethical guidelines for the creation, use, and dissemination of SR satellite imagery. [GRAPHICS] .

Automated summary

Landsat has been used for surface water mapping since 1972, but its 30 m resolution limits the detection of small water bodies. In this study, a 10x super resolution model trained with high-resolution imagery improved the detection of small lakes in Landsat imagery and produced images with preserved radiometric properties. The super resolution transformation achieved a higher F-1 score for water detection compared to native-resolution Landsat.