The Landsat record represents an amazing resource for discovering land cover changes and monitoring the Earth’s surface. But making the most use of the available data, including extending views back into the early 1970s using MSS, requires robust identification of cloud, cloud-shadow, snow, ice, water, and other obstructions between the land surface and satellite sensors. In the current products provided by USGS, the obstructions that are masked, and the reliability of those masks, differ between sensors. As such, products derived from imagery where humans do not manually check the quality of this screening -- which is necessarily the case for modern, automated approaches to global, long-term monitoring -- are contaminated by missed obstructions, leading to incorrect classifications and estimates of land cover, disturbance rates, biomass, etc.
We developed a suite of convolutional neural network (CNN)-driven models to identify clouds, cloud-shadows, water, snow/ice, and flooding in single-date Landsat imagery from MSS, TM, ETM+, and OLI sensors. CNNs are deep-learning algorithms underpinning many of the recent advances in computer vision that utilize both spectral and spatial information in an image, used here to generate a semantic segmentation. Our training dataset was constructed by hand classifying pixels in 300 images. Each image covers a 30 km x 30 km area and are stratified by ecoregion and sensor. This labor-intensive process resulted in a uniquely high-quality dataset needed for the creation of a high-quality model.
Results/Conclusions
Our CNN models generate masks for cloud, cloud-shadow and other obstructions in Landsat imagery at the same accuracy as human classifiers. These masks approach 95% accuracy, with most error due to requiring hard classifications of transitional regions (e.g. clouds often gradually thin at boundaries). Comparing to the quality masks distributed with USGS products, our CNN masks perform consistently better, missing about half as many clouded pixels and an eighth the cloud-shadows in Landsat 8 OLI imagery. The CNN approach, however, is significantly more computationally intensive than the algorithms currently used by USGS. As such, this approach is best suited for platforms where masks can be stored and distributed after processing, such as Google Earth Engine, NASA NEX, or the USGS archive.