Shifting species ranges and changing phenology: A new approach to mining social media for ecosystems observations

Geoscientists & ecologists are increasingly using social media to solicit 'citizen scientists' to participate in the data collection process. However, social media users are also a largely untapped resource of spontaneous, unsolicited observations of the natural world. Of particular interest are observations of species phenology & range to better develop a predictive understanding of how ecosystems are affected by a changing climate and human-mediated influences. Social media users' observations include information on phenological & biological phenomena such as flowers blooming, native & invasive species sightings, unusual behaviors, animal tracks, droppings, damage, feeding, nesting, etc. Our AGU2011 pilot study on the North American armadillo suggests that useful observational data can be extracted from Twitter to map current species ranges to compare with past ranges. We have expanded that work by mining Twitter for a number of North American species and ecosystem observations to determine usefulness for environmental applications such as: 1) supplementing existing databases, 2) identifying outlier phenomena, 3) guiding additional crowd-sourced studies and data collection efforts, 4) recruiting citizen scientists, 5) gauging sentiment about the observations and 6) informing ecosystems policy-making and education. We present the results for our evaluation of a representative sample from a list of 200+ species for which we've collected data since August 2011. Our results include frequency of reports and sightings by day, week and month, where the number of observations range from a few per month to ten or more per day. We discuss challenges, best practices and tools for distilling information from crowd-sourced observations gathered via Twitter in the form of 140-character 'tweets'. For example, geolocation is a critical issue. Despite the prevalence of smart phones, specific latitudinal and longitudinal coordinates are included in fewer than 10% of the observations. This number can be substantially increased at both local & regional scales by using user profile and contextual geolocation algorithms. We identify potential outlier observations, map ranges, and evaluate the usefulness of citizen sentiment conveyed in the observations as a potential metric for policy makers and managers. Based on these results we draw conclusions on best applications for use of crowd-sourced social media observations: Identifying outliers, front-tracking, guiding traditional data collection efforts and informing policy- and decision-makers about citizen sentiment toward resources.