Novel machine learning methods for learning models of bird distribution and migration from citizen science data
Professor Tom Diettrich (Oregon State University)
NICTA SML SEMINARDATE: 2012-05-15
TIME: 11:15:00 - 12:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
Many migratory bird species are in decline. Many factors are believed to contribute to this including air and water pollution, habitat fragmentation, and habitat destruction. Improving our scientific understanding of these factors and predicting the consequences of future environmental changes (e.g., climate warming) requires developing much more accurate models of bird migration and distribution. However, like many ecological phenomena, bird migration occurs over continental and planetary scales, which makes it very difficult for scientific research teams to study.
The Cornell Lab of Ornithology leads a citizen science project, eBird (ebird.org), in which bird watchers can go birding and then upload a checklist of the birds that they saw (and did not see) and the time and place. eBird is receiving more than 1M data points per month. An interesting challenge for machine learning is to develop methods that can learn accurate models of distribution and migration from this noisy, incomplete, and spatially-biased data. This talk will discuss ongoing work addressing three challenges of this data: (a) birds may be present at a site but fail to be detected by an observer, (b) birds are not directly observed during migration but only after they land at a particular site, and (c) bird watchers tend to observe near home, so the ebird observations are very highly biased.
To address these problems, we are developing three machine learning techniques: (a) non-parametric latent variable models, (b) collective graphical models, and (c) non-parametric density ratio methods for covariate shift adjustment.
