A validated approach for supervised dive classification in diving vertebrates

Supervised dive classification is a commonly used technique for categorising time-depth profiles of diving vertebrates. Such analyses permit the description and quantification of dive behaviour and foraging tactics, and highlight feeding locations. Ideally, classification functions should be validat...

Full description

Bibliographic Details
Published in:Journal of Experimental Marine Biology and Ecology
Main Authors: Thums, M., Bradshaw, C., Hindell, M.
Format: Article in Journal/Newspaper
Language:English
Published: Elsevier Science BV 2008
Subjects:
Online Access:http://hdl.handle.net/2440/48230
https://doi.org/10.1016/j.jembe.2008.06.024
Description
Summary:Supervised dive classification is a commonly used technique for categorising time-depth profiles of diving vertebrates. Such analyses permit the description and quantification of dive behaviour and foraging tactics, and highlight feeding locations. Ideally, classification functions should be validated, and this is commonly done visually. Visual classification is subjective, but it is currently one of the few measures available for validation. We develop several approaches to validate a supervised dive classification: (1) two people visually assigning dives and developing a dataset where both agree, and (2) use of dives from southern elephant seals identified as "drift dives" with characteristic velocity signatures. We classified the dives of three seals from their post-moult foraging trips and estimated the error associated with visual classification. We found classification error (disagreement) between classifiers up to 57%. We created a training dataset based on dives with agreement and applied this to a relatively new, automated classification method - the random forests (RF) algorithm. A supervised function developed using this algorithm estimated a classification error of 5% on elephant seal dives; classification error on underrepresented dive classes ranged from 2 - 12%. Testing this classification function on independent data produced a low error (1.6%). RF function errors were lower than for visual classification, and errors were similar to or better than those estimated using discriminant functions. Swim velocity parameters were the most important predictors, but their absence did not reduce the random forests function's effectiveness by much. Our results suggest that there is a temporal shift in diving behaviour as seals become more buoyant. We compared the temporal patterns in drift rate from the drift dives classified using the RF function with the drift dives validated via characteristic velocity signatures. This indicated that the classifications produced by RF function are valid even though some ...