Classifying imagery based on identifiable pixels—labeling pixels of a certain color “water” and others as “land”, for example—is known as image segmentation, a common method of processing natural science data into meaningful units. Manual image segmentation is time-consuming, though. What if artificial intelligence could be trained to do it instead?
Teaching Software to Track Coastal Change
USGS scientists studying coastal change are increasingly using remote sensing techniques to collect data—from satellites, time-series cameras, drones, and other automated imaging methods. While remote sensing allows scientists to capture coastal data of unprecedented quality and quantity, these data must first be processed before they are useful—they must be sorted in an intelligible, reproducible, and standardized way.
Classifying imagery based on identifiable pixels—labeling pixels of a certain color “water” and others as “land”, for example—is known as image segmentation, a common method of processing natural science data into meaningful units. Manual image segmentation is time-consuming, though. What if artificial intelligence could be trained to do it instead?
To find out, USGS researchers and partners created a software program that uses artificial intelligence, guided by a human user, or a “human-in–the-loop", to perform fast and accurate image segmentation of coastal change data.
The program, called Doodler, can be “trained” by a human user to identify pixels as distinct data types. The user labels a few pixels, and the program then learns how to identify the rest. The researchers trained Doodler on a variety of different data set scenarios, including flood detection in post-hurricane aerial imagery, coastal evolution in satellite imagery, and benthic physical habitat mapping in sidescan sonar data.
“Our project Remote Sensing Coastal Change is moving more and more into exploring the many ways that remote sensing can capture data,” said Dan Buscombe, a USGS contractor at the Pacific Coastal and Marine Science Center who led the development of Doodler. “Many of these methods are citizen-science based: We have CoastSnap that gives you shoreline location, the King Tides Project that gives you flooding, and Sandsnap that gives you grain size. Doodler joins this suite of tools at our disposal. All these observations can combine to make this really big data set that is accessible to people working on coast change topics, either within USGS or outside of it.”
The results are published in the study “Human-in-the-Loop Segmentation of Earth Surface Imagery”.