Metadata information for koala

This metadata information for koala (Phascolarctos cinereus) outlines tests of the recogniser accuracy.

Species Koala (Phascolarctos cinereus)
Software AviaNZ, Version 3.1
Recogniser name Koala_CNN_LG_010822 (ZIP, 4170.99 KB) - download recogniser files here
Author(s) Leroy Gonsalves, NSW DPI Forest Science
Contact information leroy.gonsalves@dpi.nsw.gov.au or brad.law@dpi.nsw.gov.au
Training call count 2,738 male bellows (10,000 with augmentation)
Call description Males produce a low frequency bellow with inhalation and exhalation
Calls sourced from NSW - North Coast, Southern Highlands, South Coast and Gunnedah
Test files 19 x 1-hr recordings and 1 x 0.5-hr recording from between dusk and dawn were used for testing. This included a mix of files generated using Songmeter SM4 and Audiomoth. Files contained the target species and a range of negative examples incl. anthropogenic, biophonic and geophonic noise.
False positives Highway and vehicle noise, planes, kookaburras, sirens, boobooks, dogs barking, frogs, insects, yellow-bellied gliders. N.B. list not exhaustive. Diurnal birds at dusk and dawn can generate large numbers of false positives.

Table 1. Test results.

Event-level: recogniser performance at level of every event (hit); Segment-level: recogniser performance at detecting an event anywhere within a 30sec file; TP = true positives; FN = false negatives; FP = false positives; TN = true negatives.

Test TP FN FP TN Recall Precision F1 Specificity Accuracy
Event-level (secs) 2050 604 292 58220 0.77 0.88 0.82 1.00 0.99
Segment-level (no. of 30 sec blocks) 55 15 19 1923 0.79 0.74 0.76 0.99 0.98
No. of hits 60 15 18 3727* 0.80 0.77 0.78 1.00 0.99

*TN for no. of hits assumes average hit is 15.8 secs in duration.


The equations for calculating Recall and Precision are supplied below.

Equations for calculating Recall and Precision


Example of False Negative (faint)


Example of True Positive