2022-12-06
I finally finished labelling data from my last trip 2 months ago, its been a marathon. Hoping to go to Pomona again next week to collect the latest data.
Notebook at skraak.kiwi
H15, WB09, WD05 no detections. J11 only one detection. I think this is possibly because the Kiwi are less mobile as they may have been incubating eggs. Locations with lots of detections for this period may be the home base of breeding pairs.
New active pair at KS06 New active pair at S13T
F09 is very active and I now have almost as much data from them as I do from C05. The pair at N20 make a lot of noise and so do the frogs, which makes it a challenging location.
D03 male seems to have paired up but not certain, need more data.
Kaka heard at M04 twice, and once at S13T. Please stay kaka.
Next trip I am going to do a big reorganisation, the core moths will remain but there are some new places that need exploring, gaps to fill in.
Last week I attended an online audiomoth conference, and learned some stuff that can help me to move forward.
I have discovered opensoundscape.org. They have written many bioacoustic primitives I can use.
I learned that the audiomoths save the temperature in the metadata of their files. I rewrote my code and now my data includes temperature, battery voltage, time zone, moth ID.
My data now also makes it easier to write labels in a format compatible with Raven Pro, a commercial package from Cornell University. I need my labels in this format to be able to use opensoundscape.
Soon I will take a look at how call behaviour relates to temperature.
Battery voltage and moth ID can help to track down the cause of noise, I will be able to exclude noisy data automatically to some extent, where required. Moths get noisy when the battery is flat, and some are noisier than others.
Time zone info is just a good check to make sure my data is consistent.
My next steps: Move the thousands of 15 minute files with kiwi calls, and their labels into one place. Chop up those files into labelled 5 second wav files. Train a binary classifier with more than 1 million 224px square images, kiwi/noise. Lack of data is not a problem. The problem is now too much data. If it gets it right more than 20% of the time that will be better than AviaNZ, but 90% would be more useful. Train a model to take binary kiwi data and identify Male/Female/M&F, Close/Far so I dont have to do so much labelling, this last run I labelled calls in 1700 files, probably 2500 calls. Many hours of work.
Some random audio