Predicting animal behaviour using deep learning: GPS data alone accurately predict diving in seabirds. Issue 3 (20th November 2017)

• Record Type: Journal Article
• Authors: Browning, Ella; Bolton, Mark; Owen, Ellie; Shoji, Akiko; Guilford, Tim; Freeman, Robin
• Editors: McPherson, Jana
• Is Part Of: Methods in ecology and evolution. Volume 9:Issue 3(2018)
• Page Start: 681
• Page End: 692
• Publication Date: 2017-11-20
• Subjects: common guillemot -- European shag -- foraging -- machine learning -- prediction -- razorbill -- time–depth recorder; Ecology -- Periodicals; Evolution -- Periodicals; 577
• DOI: 10.1111/2041-210X.12926
• Languages: English
• ISSNs: 2041-210X

Abstract

Abstract: To prevent further global declines in biodiversity, identifying and understanding key habitats is crucial for successful conservation strategies. For example, globally, seabird populations are under threat and animal movement data can identify key at‐sea areas and provide valuable information on the state of marine ecosystems. To date, in order to locate these areas, studies have used global positioning system (GPS) to record position and are sometimes combined with time–depth recorder (TDR) devices to identify diving activity associated with foraging, a crucial aspect of at‐sea behaviour. However, the use of additional devices such as TDRs can be expensive, logistically difficult and may adversely affect the animal. Alternatively, behaviours may be resolved from measurements derived from the movement data alone. However, this behavioural analysis frequently lacks validation data for locations predicted as foraging (or other behaviours). Here, we address these issues using a combined GPS and TDR dataset from 108 individuals by training deep learning models to predict diving in European shags, common guillemots and razorbills. We validate our predictions using withheld data, producing quantitative assessment of predictive accuracy. The variables used to train these models are those recorded solely by the GPS device: variation in longitude and latitude, altitude and coverage ratio (proportion of possible fixes acquired within a set window of time). Different combinations of these variables were used to explore the qualities of different models, with the optimum models for all species predicting non‐diving and diving behaviour correctly over 94% and 80% of the time, respectively. We also demonstrate the superior predictive ability of these supervised deep learning models over other commonly used behavioural prediction methods such as hidden Markov models. Mapping these predictions provides useful insights into the foraging activity of a range of seabird species, highlighting important at sea locations. These models have the potential to be used to analyse historic GPS datasets and further our understanding of how environmental changes have affected these seabirds over time.