Ecological correlates of blue whale movement behavior and its predictability in the California Current Ecosystem during the summer-fall feeding season

Table 5 Confusion matrix for the binary conversion of the likelihood of ARS estimated by the spatial coordinates NPMR model for the building (n = 1444) and validation sets (n = 364), using the cutoff value that maximized the true skill statistic (TSS_max). FPR is the false positive rate, FNR is the false negative rate, TPR is the true positive rate, and TNR is the true negative rate. The second part of the table reports a set of performance metrics for this binary conversion, including prevalence, accuracy, precision, the area under the receiver operating characteristic curve (AUC, range: 0 to 1 with larger numbers indicating a better fit), the root-mean square error (RMSE, range: 0 to infinity with smaller numbers indicating a better fit), and the Brier score (range: 0 to 1 with lower scores indicating a better calibration of the predictions)

Confusion matrix:
		Predictions		Classification error
		Absence	Presence
Observations in the building set	Absence	113	45	0.28 (FPR)
	Presence	404	696	0.37 (FNR)
Observations in the validation set	Absence	83	39	0.32 (FPR)
	Presence	59	146	0.29 (FNR)
Performance metrics:
	Building set		Validation set
TSS_max	0.35		0.40
Cutoff	0.90		0.65
Observed prevalence^a	0.87		0.63
Predicted prevalence^a	0.59		0.57
TNR (1-FPR)	0.71		0.68
TPR (1-FNR)	0.63		0.71
Accuracy^b	0.64		0.70
Precision^c	0.94		0.79
AUC	0.71		0.72
RMSE	0.32		0.45
Brier score^d	0.10		0.20

^aPrevalence is estimated as: presences/total
^bAccuracy is estimated as: (true positives + true negatives)/(obs. Presences + obs. absences)
^cPrecision is estimated as: true positives/(true positives + false positives)
^dThe Brier score is computed as the mean of the squared residuals

ISSN: 2051-3933