Oregon State University



Event Details

MS Final Examination – Liqiang He

Friday, June 9, 2017 1:00 PM - 3:00 PM

Species Distribution Modeling of Citizen Science Data as a Classification Problem with Class-Conditional Label Noise
Species distribution models (SDM), which quantify the correlation between the distribution of a species and environmental factors, are increasingly used to map and monitor animal and plant distributions in the context of awareness of environmental change and its ecological consequence. For perfect data, this is a straightforward classification problem from environmental features to presence or absence labels. But for imperfect data, such as the citizen science data from eBird, in which volunteers report locations where they observed or failed to observe sets of species, mistakes will cause label noise. In this case, both the class features and the observation features would be sources of false positive noise and false negative noise. However, few common modeling approaches for this task address these sources of noise explicitly. In this work, I explore the idea of treating this problem as a classification problem with class-conditional label noise. By lever aging additional information about observation features, this model outperforms other candidates significantly when sufficient data is available. I describe the conditions under which the parameters of my proposed model are identifiable, explore the impact of model misspecification, and apply this model to simulated data and real data from the eBird citizen science project.

Major Advisor: Rebecca Hutchinson
Committee: Thomas Dietterich
Committee: Amir Nayyeri
GCR: John Bolte

Kelley Engineering Center (campus map)
Calvin Hughes
1 541 737 3168
Calvin.Hughes at oregonstate.edu
Sch Elect Engr/Comp Sci
This event appears on the following calendars: