AI for Earth’s Land Cover Mapping

Опубликовано 17 июля 2019, 1:18

High-resolution land cover mapping is a process of assigning land use labels, such as “impervious surface,” or “tree canopy” to each pixel in high resolution (&lt1m) aerial or satellite imagery. Such maps are an essential component in environmental science, agriculture, forestry, urban development, the insurance and banking industries, and for demography in developing countries.

Traditional Geographic Information System (GIS) workflows start with a variety of segmentation and classification algorithms and involve specialized human labor that refines these initial maps. Open street maps and other crowdsourcing platforms, on the other hand, utilize work of thousands of volunteers through mapping runs in which land cover maps are created from scratch. Both approaches are labor-intensive and this can dissuade agencies and researchers from pursuing some studies, especially the ones that track the land use longitudinally at high temporal rates. And so, with the optimism over modern deep learning and reinforcment/active learning came renewed interest within ML and computer vision communities in these applications, which involve terabytes of mostly unlabeled or weakly/indirectly labeled data. I will talk about several approaches to rapid land cover mapping we have investigated on our road to creating the first ever 1m resolution land cover map of the entire USA. These approaches involve novel deep learning models that can super-resolve low res labels, active learning, machine teaching and hybrid intelligence. I will also discuss exciting opportunities both to develop new machine learning models or hybrid systems, and to make practical impact.

See more at microsoft.com/en-us/research/v...