Fauna identification service

Acoustics

Increasingly, technology is improving the options for fauna survey, but this comes with a need for advanced tools to analyse the vast amounts of data that can be collected. This is especially apparent in the world of eco-acoustics. NSW DPIRD is using artificial intelligence to develop tools for identifying key forest species from acoustic recordings.

Currently, open-source software AviaNZ (Marsland et al. 2019) is being used for this purpose.

Recognisers for different species have been developed. Read the Fauna Call Recogniser Report (PDF, 633.74 KB), the final report prepared for the NSW NRC outlining early progress and preliminary (draft) recognisers.

Examples of Dingo, Yellow-bellied Glider and Koala calls can be viewed here (displayed on Kaleidoscope software).

Download species recognisers

Koala browsing a blackbutt Species recognisers were developed in AviaNZ using CNNs (AI algorithms). The final algorithm for each species aims to provide a balance between maximising true positives and minimising false positives.

It is important to note that all output from these recognisers must be validated manually as results from true positives will be mixed with false positives. Note that recognisers are continually being improved, so this website should be checked regularly for updates.

Like all survey methods, perfect detection is not possible and detection probability ideally should be modelled for different species to identify how much survey effort is required for 95% confidence of detection (see Gonsalves et al. 2024).

View a video showing the use of AviaNZ, especially in relation to call validation.

Note: The equations for calculating Recall and Precision are supplied here.

Click below to download the species recogniser(s) of interest, which can be run in AviaNZ software and view their associated metadata (reporting the species test results).

These recognisers are © NSW Department of Primary Industries and Regional Development. They must not be adapted, remixed, transformed or built upon for any purpose or in any way without written consent from DPIRD. The user must attribute the 'NSW Department of Primary Industries and Regional Development' or 'NSW DPIRD'.

By downloading these recognisers you agree to abide by the Creative Commons Attribution-NonCommercial 4.0 International licence (CC BY-NC 4.0).

Barking Owl

Barking Owl metadata information

Species	Barking Owl (Ninox connivens)
Software	AviaNZ, Version 3.2.3
Recogniser name	Barking_Owl_GV_040524 (ZIP, 4168.79 KB) - download recogniser files here
Training call count	6,010 (10,000 with augmentation)
Call description	Two-note “woof-woof” with the first note higher pitched than the second. The duration between the two notes is shorter compared to a boobook call.
Calls sourced from	NSW - Pilliga, North Coast, Sydney Northern Beaches
Test files	Event and segment-level= 267 files (100 x 30sec files containing the target species and 167 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Grey-headed Flying Fox, Insects, Yellow-bellied Glider, Ambient noise, Boobook

Test results

Event-level: recogniser performance at level of every event (hit); Segment-level and real-world: 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)	656	327	894	1,119	0.67	0.42	0.52	0.56	0.59
Segment-level (no. of 30 sec blocks)	89	11	3	164	0.89	0.97	0.93	0.98	0.95
No. of hits	33	1	2	324	0.97	0.94	0.95	0.99	0.99

Example of true positive call in a 30 second spectrogram image

Glossy Black Cockatoo

Glossy Black Cockatoo metadata information

Species	Glossy Black Cockatoo (Calyptorhynchus lathami)
Software	AviaNZ, Version 3.2.3
Recogniser name	Glossy_Black_Cockatoo_GV_190324 (ZIP, 4169.48 KB) - download recogniser files here
Training call count	2,418 (10,000 with augmentation)
Call description	Creaky and wavering squawks. Quieter and less harsh than other cockatoos.
Calls sourced from	NSW - North Coast, Pilliga, South Coast, Xeno-Canto (web source)
Test files	Event and segment-level= 190 files (100 x 30sec files containing the target species and 90 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Chorus calls, Grey-headed Flying Fox, Insects, Sulphur-crested Cockatoo, Galah

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	431	432	367	1,770	0.50	0.54	0.52	0.83	0.73
Segment-level (no. of 30 sec blocks)	73	27	3	87	0.73	0.96	0.83	0.97	0.84
No. of hits	36	14	7	302	0.73	0.84	0.78	0.98	0.94

Example of true positive call in a 30 second spectrogram image

Greater Sooty Owl

Greater Sooty Owl metadata information

Species	Greater Sooty Owl (Tyto tenebricosa)
Software	AviaNZ, Version 3.2.3
Recogniser name	Greater_Sooty_Owl_ROCNN_20-45-54 (ZIP, 4165.35 KB) - download recogniser files here
Training call count	219 (1,500 with augmentation)
Call description	Loud descending screech, also referred to as the ‘falling bomb whistle’.
Calls sourced from	NSW - North Coast
Test files	Event and segment-level= 225 files (33 x 30sec files containing the target species and 192 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Diurnal birds

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	83	20	32	6,569	0.81	0.72	0.76	1.00	0.99
Segment-level (no. of 30 sec blocks)	31	2	1	191	0.94	0.97	0.95	1.00	0.99
No. of hits	1	9	0	350	0.10	1.00	0.18	1.00	0.98

Example of true positive call in a 30 second spectrogram image

Grey-headed Flying Fox

Grey-headed Flying Fox metadata information

Species	Grey-headed Flying Fox (Pteropus poliocephalus)
Software	AviaNZ, Version 3.2.3
Recogniser name	Grey_Headed_Flying_Fox_GV_040824 (ZIP, 4184 KB) - download recogniser files here
Training call count	14,237 calls (32,000 with augmentation)
Call description	High-pitched, single-note shriek; cackling and squabble calls.
Calls sourced from	NSW - North Coast, Sydney Northern Beaches
Test files	Event and segment-level= 264 files (96 x 30sec files containing the target species and 168 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Masked Owl, Sooty Owl, Bird chorus, Yellow-bellied Glider, Kookaburra, Insects, Day birds

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	557	351	105	1,987	0.61	0.84	0.71	0.95	0.85
Segment-level (no. of 30 sec blocks)	81	15	5	163	0.85	0.92	0.88	0.97	0.92
No. of hits	103	38	6	213	0.84	0.94	0.82	0.97	0.88

Example of true positive call in a 30 second spectrogram image

Grey-headed Flying Fox spectrogram image

Koala

Koala metadata information

Species	Koala (Phascolarctos cinereus)
Software	AviaNZ, Version 3.2.3
Recogniser name	Koala_CNN_LG_071223 (ZIP, 4178.59 KB) - download recogniser files here
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
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	Trucks, Trains, Kookaburras

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	1,146	904	30	58,482	0.56	0.97	0.71	1.00	0.98
Segment-level (no. of 30 sec blocks)	38	17	2	1,940	0.69	0.95	0.80	1.00	0.99
No. of hits	39	21	2	3,701.4	0.65	0.95	0.77	1.00	0.99

Example of true positive call in a 30 second spectrogram image

Release notes for previous versions

Koala_CNN_LG_010822 – released September 2022

Improved recall and precision compared to DPI_Male_Koala_V3_CNN15_10-02-23

DPI_Male_Koala_V3_CNN15_10-02-23 – released June 2022

Initial release. Please contact author if you would like access to this release

Masked Owl / Tyto Owl

This Masked Owl recoginser is trained on Masked Owl calls but other Tyto Owls also screech and will be recognised, thus requiring expert validation.

Masked Owl metadata information

Species	Masked Owl (Tyto novaehollandiae)
Software	AviaNZ, Version 3.2.3
Recogniser name	Masked_Owl_GV_210424 (ZIP, 4165.52 KB) - download recogniser files here
Training call count	466 (5,000 with augmentation)
Call description	Loud, rasping shriek.
Calls sourced from	NSW - North Coast, South Coast, North Sydney
Test files	Event and segment-level= 277 files (100 x 30sec files containing the target species and 177 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Insects, Sirens. Difficult to distinguish between other Tyto sp. and may need expert opinion.

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	240	49	49	2,653	0.83	0.83	0.83	0.98	0.97
Segment-level (no. of 30 sec blocks)	81	19	2	175	0.81	0.98	0.87	0.99	0.92
No. of hits	16	2	0	342	0.89	1.00	0.94	2.0	0.99

Example of true positive call in a 30 second spectrogram image

Powerful Owl

Powerful Owl metadata information

Species	Powerful Owl (Ninox strenua)
Software	AviaNZ, Version 3.2.3
Recogniser name	Powerful_Owl_CNN_LG_061222 (ZIP, 4162.76 KB) - download recogniser files here
Training call count	2,702 (5,000 with augmentation)
Call description	Two-note “woo-hoo” with the second note often higher pitched that the first. Male calls are typically lower pitched and females higher, but there can be overlap.
Calls sourced from	NSW - North Coast
Test files	Event and segment-level= 229 files (100 x 30sec files containing the target species and 129 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Barking Owl, Frogs, Kookaburra, Insects, Siren

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	709	115	874	1,302	0.86	0.45	0.59	0.60	0.67
Segment-level (no. of 30 sec blocks)	94	6	9	120	0.94	0.91	0.93	0.93	0.93
No. of hits	55	32	0	273	0.63	1.00	0.77	1.00	0.91

Example of true positive call in a 30 second spectrogram image

Southern Boobook

Southern Boobook metadata information

Species	Southern Boobook (Ninox novaeseelandiae)
Software	AviaNZ, Version 3.2.3
Recogniser name	Southern_Boobook_GV_190623 (ZIP, 4174.85 KB) - download recogniser files here
Training call count	4,057 (10,000 with augmentation)
Call description	Double-note hoot with the first note higher pitched than the second.
Calls sourced from	NSW - North Coast, South Coast, Southern Highlands, Gunnedah
Test files	Event and segment-level= 200 files (100 x 30sec files containing the target species and 100 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	There were no recorded false positive detections during testing, however, when this recogniser was used in various datasets, it has detected Cattle, Lyrebirds, Dogs barking, Bandicoot

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	985	361	901	753	0.73	0.52	0.61	0.46	0.58
Segment-level (no. of 30 sec blocks)	90	14	0	96	0.87	1.00	0.91	1.00	0.93
No. of hits	26	26	0	308	0.5	1.00	0.67	1.00	0.93

Example of true positive call in a 30 second spectrogram image

Squirrel Glider

Squirrel Glider metadata information

Species	Squirrel Glider (Petaurus norfolcensis)
Software	AviaNZ, Version 3.2.3
Recogniser name	Squirrel_Glider_EK_010823 (ZIP, 4170.1 KB) - download recogniser files here
Training call count	4,015 (10,000 with augmentation)
Call description	Short one-note honk, increasing then decreasing in pitch, with emphasis on the down-sweep. Calls can vary in length though are usually consistent within a single calling bout.
Calls sourced from	NSW - North Coast
Test files	Event and segment-level= 199 files (100 x 30sec files containing the target species and 99 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Frogs, Anthropogenic and Ambient Noise, Dog, Grey-headed Flying Fox, Owlet Nightjar, Sugar Glider, Fox

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	260	866	296	1,572	0.23	0.47	0.31	0.84	0.61
Segment-level (no. of 30 sec blocks)	55	45	12	87	0.55	0.82	0.66	0.88	0.71
No. of hits	41	33	20	266	0.55	0.67	0.61	0.93	0.85

Example of true positive call in a 30 second spectrogram image

Sugar Glider

Sugar Glider metadata information

Species	Sugar Glider (Petaurus breviceps)
Software	AviaNZ, Version 3.2.3
Recogniser name	Sugar_Glider_EK_040623 (ZIP, 4178.94 KB) - download recogniser files here
Training call count	7,912 (10,000 with augmentation)
Call description	Short one-note yip or bark. Will often call repetitively at relatively consistent intervals. 1-4 harmonics are often present.
Calls sourced from	NSW - North Coast, South Coast, Southern Highlands
Test files	Event and segment-level= 247 files (100 x 30sec files containing the target species and 147 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 720 files (30 sec files split from 6 x 1hr continuous recordings).
False positives	Dog, Bandicoot, Barking Owl, Boobook

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	1,581	411	216	792	0.67	0.89	0.95	0.99	0.96
Segment-level (no. of 30 sec blocks)	93	7	2	145	0.93	0.98	0.76	0.79	0.79
No. of hits	68	11	3	638	0.86	0.96	0.91	1.00	0.98

Example of true positive call in a 30 second spectrogram image

Yellow-bellied Glider

Yellow-bellied Glider metadata information

Species	Yellow-bellied Glider (Petarus australis)
Software	AviaNZ, Version 3.2.3
Recogniser name	Yellow_Bellied_Glider_CNN_LG_100423 (ZIP, 4168.65 KB) - download recogniser files here
Training call count	283 (4,000 with augmentation)
Call description	A very low-pitched soft hoot followed by a high-pitched loud shriek and a lower-pitched gurgle. Sometimes calls only consist of the hoot and gurgle.
Calls sourced from	NSW - North Coast
Test files	Event and segment-level= 236 files (96 x 30sec files containing the target species and 140 x 30sec files containing false positive examples incl. anthropogenic, biophonic and geophonic noise). Real-world = 360 files (30 sec files split from 3 x 1hr continuous recordings).
False positives	Fantail Cuckoo, Grey-headed Flying Fox, Kookaburra, Insects, Squirrel Glider, Sugar Glider

Test results

Test	TP	FN	FP	TN	Recall	Precision	F1	Specificity	Accuracy
Event-level (secs)	329	183	180	2,305	0.64	0.65	0.64	0.93	0.88
Segment-level (no. of 30 sec blocks)	73	23	9	131	0.76	0.89	0.82	0.94	0.86
No. of hits	7	14	1	338	0.33	0.88	0.48	1.00	0.96

Example of true positive call in a 30 second spectrogram image

Ultrasonics

Bat taking off DPIRD Forest Science also has a long history of using automated methods to identify the thousands of bat calls that can be recorded with bat detectors. Software called Anascheme (developed by Matt Gibson) has been used for this purpose, in association with identification keys developed by DPIRD for different regions of NSW. More details on the identification keys and the software can be found in this publication. DPIRD can provide a service of using Anascheme to analyse large datasets. New AI models for full spectrum calls are currently being developed.

Fauna identification service

DPIRD researchers have extensive experience in developing recognisers and validating their outputs, and we offer this identification service for larger datasets.

Fees for this service can be obtained by contacting brad.law@dpi.nsw.gov.au.