A data science challenge for converting airborne remote sensing data into ecological information
Sergio Marconi, Sarah J. Graves, Dihong Gong, Shahriari Nia Morteza, Marion Le Bras, Bonnie J. Dorr, Peter C. Fontana, Justin Gearhart, Craig S. Greenberg, Dave J. Harris, Sugumar A. Kumar, Agarwal Nishant, Joshi Prarabdh, Sandeep U. Rege, Stephanie A. Bohlman, Ethan P. White, Daisy Z. Wang
In recent years ecology has reached the point where a data science competition could be very productive. Large amounts of open data are increasingly available and areas of shared interest around which to center competitions are increasingly prominent. The University of Florida ran a competition to help improve three tasks that are central to converting remote sensing images into the kinds of information that would traditionally collected by ecologists in the field: 1) crown segmentation, for identifying the location and size of individual trees; 2) alignment to match ground truth data on trees with remote sensing; and 3) species classification to identify trees to species. Most papers in ecology do not compare their methods and different papers focus on different data sets. In an effort to overcome some of these barriers to comparing methods and determining the degree of generalizability, over the last two years the National Institute of Standards and Technology (NIST) has embarked on an effort to examine data science problems and to build out the NIST IAD Data Science Evaluation Series (DSE) for cross-domain evaluation of algorithms. NIST has devised a general evaluation paradigm to address data science problems that span diverse disciplines, domains, problems, and tasks. Plant identification is one such task that has been examined in the context of the DSE, which serves as a framework for evaluation-driven research, supporting evaluation for data-science solutions in the ecology domain as well as in other domains for which analogous tasks have been devised.