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

Table 4 Summary of feature variables extracted from acceleration data

Feature code	Description	References
stX, stY, stZ	Static acceleration (body posture) in each separate axis	[17, 54,55,56]
dyX, dyY, dyZ	Dynamic acceleration in each separate axis	[23, 54, 55, 57]
PBDAx, PBDAy, PBDAz	Partial dynamic body acceleration (absolute acceleration in each axis)	[40, 58, 59]
ODBA	Overall dynamic body acceleration	[60, 61]
VeDBA, VeDBAs	Vectorial dynamic body acceleration, smoothed
ratioX, ratioY, ratioZ	Ratio of VeDBA to PDBA	[25]
jerkX, jerkY, jerkZ	Jerk, derivative of acceleration, in each separate axis	[18]
jerkN	Norm of jerk in all axes
Pitch, Roll	Pitch and roll in radians	[54]
PSD1x, PSD1y, PSD1z	Primary dominant power spectrum density in each axis	[25, 62]
PSD2x, PSD2y, PSD2z	Secondary dominant power spectrum density in each axis
Freq1x, Freq1y, Freq1z	Frequency corresponding to the primary dominant power spectrum density in each axis
Freq2x, Freq2y, Freq2z	Frequency corresponding to the secondary dominant power spectrum density in each axis

Summary of feature variables used to classify behaviour through machine learning methods. All feature variables, with the exception of those relating to power spectrum density and frequency (e.g. PSD1x, Freq1x), were summarized according to their mean value over 1-s windows of time (50 samples at 50 Hz, 25 at 25 Hz). Power spectrum and frequency elements (PSD1x–Freq2z) were derived over a 3-s moving window (1-s overlap on either side) to minimize spectral leakage

ISSN: 2050-3385