Assessing the utility and limitations of accelerometers and machine learning approaches in classifying behaviour during lactation in a phocid seal

Table 2 Confusion matrix of behaviour classified from random forests

	Rest	Alert	Presenting/Nursing	Locomotion	Comfort Movement	Flippering pup	Tot. pred.
Rest	*23,825*	5445	834	113	1091	271	31,579
Alert	6994	*19,308*	897	741	1639	433	30,012
Presenting/Nursing	276	293	*2361*	13	184	148	3275
Locomotion	0	1	1	25	0	0	27
Comfort Movement	10	9	4	0	46	3	72
Flippering pup	58	45	117	2	44	*587*	853
Tot. obs.	31,163	25,101	4214	894	3004	1442	65,818

Confusion matrix of classified discrete feature windows (each 1-s wide) showing the predicted behaviours (rows; originating from random forest classification) against the observed reference behaviours (columns; originating from the focal video data) for accelerometers mounted on the head in 2015 (50 Hz). Behaviours included Rest, Alert, Presenting/Nursing, Locomotion, Comfort Movement, and Flippering pup. Values across the diagonal (bold italic) represent those behaviours that were correctly identified, true positives, while those in italics represent false positives (rows) and false negatives (columns)

ISSN: 2050-3385