A data science challenge for converting airborne remote sensing data into ecological information

• DOI: 10.7287/peerj.preprints.26966v1
• Publication Date: 2018-05-29T12:51:47Z
• Authors: Sergio Marconi; Sarah J. Graves; Dihong Gong; Morteza Shahriari...; Marion Le Bras; Bonnie J. Dorr; Peter Fontana; Justin Gearhart; Craig Greenberg; Dave J. Harris; Sugumar Arvind Kumar; Agarwal Nishant; Joshi Prarabdh; Sundeep U. Rege; Stephanie Ann Boh...; Ethan P. White; Daisy Zhe Wang
• Keywords: Tree (set theory); Hindsight bias; Identification

Ecology has reached the point where data science competitions, in which multiple groups solve same problem using by different methods, will be productive for advancing quantitative methods tasks such as species identification from remote sensing images. We ran a competition to help improve three that are central converting images into information on individual trees: 1) crown segmentation, identifying location and size of trees; 2) alignment, match ground truthed trees with sensing; 3) classification trees. Six teams (composed 16 participants) submitted predictions one or more tasks. The segmentation task proved most challenging, highest-performing algorithm yielding only 34% overlap between remotely sensed crowns However, algorithms performed better larger For alignment task, an based minimizing difference, terms both position tree size, yielded perfect alignment. In hindsight, this was over simplified including targeted instead all possible crowns. Several well classification, correctly classifying 92% individuals performing common rare species. Comparisons results across provided number insights improving overall accuracy extracting ecological sensing. Our experience suggests kind can benefit development ecology biology broadly.