Table 5.

Performance metrics for the best severity CNN model found by the optimiser for different sizes of the STFT.

FE domain	Model name	Accuracy	Macro precision	Macro recall	Macro F1	Weighted F1	# Training samples
STFT	CNN–1	0.689	0.709	0.665	0.676	0.684	9094
	CNN–1/8	0.780	0.785	0.771	0.776	0.780	75,443
	CNN -1 (32 windows)	0.615	0.629	0.591	0.599	0.609	226

Size is represented by the number by which the STFT window length is multiplied (e.g. 1/4 represents a quarter of the normal 10-wavelength-long window).