2023-09-11

To: Pomona

How: By bike, foot, raft, foot, boat, car, foot, bike. Thanks to Nick Key as Adventure Kayak and Cruse for the boat ride home.

Overview

Location	Solo_Male	Solo_Female	Duets	Individual	False_Positives
C05	220	118	41	420	13
D03	301	3	22	348	10
F10	197	47	42	328	29
G05	238	7	11	267	24
D09	184	24	19	246	23
M04	182	21	18	239	149
NB14	163	17	14	208	75
S23	151	12	12	187	21
N20	84	35	11	141	260
H04	102	4	6	118	39
S13T	80	18	10	118	28
V04	43	12	9	73	8
CB11	38	16	3	60	21
H09	47	1	0	48	31
N08M	26	0	4	34	5
A11	20	1	2	25	34
W04	14	4	1	20	12
W20	12	5	0	17	11
J06	5	4	0	9	9
TOTAL	2107	349	225	2906	802

Calls per Hour

location	male	female	duet	individual
C05	0.4357	0.2654	0.0684	0.7011
F10	0.5107	0.1902	0.0897	0.7009
S23	0.6015	0.0886	0.0443	0.6901
G05	0.5533	0.04	0.0244	0.5933
D03	0.5447	0.0422	0.0371	0.5869
S13T	0.3435	0.1069	0.0382	0.4504
D09	0.3389	0.0718	0.0317	0.4107
M04	0.3339	0.0651	0.0301	0.399
NB14	0.2985	0.0523	0.0236	0.3508
V04	0.2185	0.0882	0.0378	0.3067
N20	0.1602	0.0776	0.0185	0.2378
CB11	0.1475	0.0683	0.0108	0.2158
H04	0.1803	0.0167	0.01	0.197
H09	0.1709	0.0036	0.0	0.1745
N08M	0.124	0.0165	0.0165	0.1405
A11	0.0791	0.0108	0.0072	0.0899
W04	0.0647	0.0216	0.0043	0.0863
W20	0.0456	0.019	0.0	0.0646
J06	0.0184	0.0147	0.0	0.0331
TOTAL	0.2721	0.0663	0.0259	0.3384

Changes

Female Calls per Hour

Male Calls per Hour

2023-09-11

I was on Pomona betweeen then 10th and 13th August to:

service 10 permanent kiwi micro moths.
service 10 temporary kiwi moths.
service 5 moths recording 1 minute in every 30. Mainly to check they are functioning, check the batteries are lasting, and bring the sd cards home so I can begin labelling.

The weather was quite dodgy and I needed to call Nick Key for a ride out due to lumpy water. Thanks a million Nick.

Active Moth Locations

Micro Moths: C05, D03, D09, F10, G05, H04, M04, N14, N20, NB14
Kiwi Moths: A11, H09, J06, S23, W20, CB11, S13T, W04, V05, N08M (Was August only, now September too)
Pomona 1/30 Moths: C03, G05, H01, J11, P09
Rona 1/30 Moths: 287, 343, 620, 635, 778.1 (not checked)

Unfortunately the SD card from N14 managed to unplug itself at some point after it wrote its config file to the card, so I have no data from N14 this time.

Calls detected in this batch of data (including the 5 24/7 moths recording 1 minute in 30):

231 (213 last time) Duets (given that 10 extra moths were active, this suggests the 10 permanent micro moths cover most breeding pairs)
350 (249 last time) Solo Female Calls
2156 (1432 last time) Solo Male Calls
885 (222 last time) False Positives

The increase in false positives is due to the return of frogs to N20, and very active kaka.

The C05 kiwi seem extra busy, maybe they are incubating an egg.

Since the last Pomona trip:

I recieved a batch of very high quality data from Haast, there were a lot of calls (Duet 1231, Female 1183, Male 3369) and a lot of false positives (8752) due to all the new sounds. When I get a chance I will do an analysis to compare Haast Kiwi at Haast to Haast Kiwi on Pomona
I did some work on training a new classifier in 100% Julia (rather than Python/Opensoundscape). It took some work getting the pipeline right. My model is trained on colour images instead of the B&W used by opensoundscape.
using my clip png images (around 48500, duet, female, male, not_kiwi) I got 78% accuracy training a Resnet18 from scratch after many many epochs (170).
using a pretrained Resnet model I managed around 87% accuracy in 30 epochs or so.
using a pretrained Resnet with data augumentation modelled on opensoundscape (random levels of gaussian noise, random frequency and time masks) we got over 90% on 18 epochs, extremely efficient, but 48000 images is not enough for a top class model, it learns the training set by heart and results on the test set top out too quickly.

Epoch: 24
accuracy = 0.9067
│    4×4 Named Matrix{Int64}
│    targets ╲ predicts │    1     2     3     4
│    ───────────────────┼───────────────────────
│    1                  │  380     8    61    11
│    2                  │    3   396    45    18
│    3                  │   28    20  2264    81
└    4                  │    8    24   147  1370

I used this model this morning to sort the trip detections into duet (1), female (2), male (3), not_kiwi (4). It performs extremely well in practice and saved me hours if not days. It is really good at catching most of the false positives, and makes very few miss-classifications. I now have more data to add to this data set, the model can only improve.
Before I left for Pomona I constructed a new big dataset (for my main model kiwi/not_kiwi) out of all kiwi detections 5 500 000 images, 5 second clips, with 9% kiwi. After its first epoch it was close to 97% accuracy. It died on the 12th epoch (my fault) with 98% accuracy. The good news is it was a long way from topping out, accuracy on a 5% sample of the training set was the same as the test set. False negatives below are an overestimate, due to the generous segmentation of my auto segmented data. I have not done any hand labelling, impossible at this scale. It may benefit from less aggressive data augumentation next time.

accuracy = 0.9804
┌ Info: eval
│   v_cmatrix =
│    2×2 Named Matrix{Int64}
│    targets ╲ predicts │      1       2
│    ───────────────────┼───────────────
│    1                  │  21621    4731
└    2                  │    669  247923

I now need a Something detector, to help with the new daytime data labelling. The next step is to label the existing nights dataset with other bird species, and to add daytime data. The next models will be Kiwi, Morepork, Kaka ..... + Something (anything) + Nothing.
The goal is to catch new species in Something, so they can be labelled without wading through hours of audio. My experience tells me labelling need not be perfect, quantity goes a long way.
I will be attending the online Audio Moth conference again soon. Unfortunately timing is unfortunate, 2 am start, but they have a speaker this year on using acoustic indicies to estimate biodiversity, which is the plan for the 1 min in 30 audio sampling on Pomona, Rona, at the Bowen, and hopefully still the Hollyford (when I have spare moths), as well as the bird call monitoring.
The battery voltage in the 5 24/7 moths serviced dropped from 4.9V to 4.3V over the 3 months they were out, recording 1 minute in 30. Since they do not cut out until 3.1V and they were all still dry and functioining, it looks like they will last a full year in the field.
Female kiwi were recorded at all 5 24/7 moths and duets at all except H01. Most calls were detected at C03, G05 next, around 7 calls each at J11, P09, H01. Total of 56 calls in 3 months recording 1 in 30 at 5 locations.

This is still work in progress, too busy:

My next exploratory analysis, coming soon, will be around the number of male calls recorded vs female.

There are a few possible explanations:

- Male kiwi just call more
- Male calls carry further, therefore I am getting the same call on multiple recorders
- Could my model be missing female calls?
- Am I misclassifying some male calls as female?

As part of this I also plan to compare Avianz female detections (including some human searching at the time) with my unassisted opensounscape detections.

Some Audio:

Apologies, I can't describe each link. Luck of the draw. Most are closer duets or female kiwi. There are some calls from new locations here.

C05