Detection & classification of SRKW calls

through human & machine learning, in real-time

Val Veirs and Scott Veirs, Beam Reach, SPC

Data Workshop
Organized by ONC/Merdian | 21-22 Nov 2019 | Victoria, BC

Slides at: orcasound.net/talks/ | Gmail: vveirs, sveirs | Slack | Trello

Orcasound 2.0: new locations, node hardware/software, and app in 2019

Vision: Open source software, open data access, real-time engagement of citizen scientists and cloud-computing

Real-time inspiration: ~5 decades of live-stream pioneering by OrcaLive (Paul Spong and Helena Symonds)

Thanks to: 2017 Kickstarter backers, "hall of fame" devs & designers, and the 2018-2019 hackathons at UW, Microsoft, and DemocracyLab

Public launch of Orcasound app in Nov 2018: live.orcasound.net

Beta-testing more interactive UI in fall 2019: beta.orcasound.net

Resources for teaching humans (& machines?)

Outreach/education nodes

  • Seattle Aquarium listening kiosk
  • Port Townsend Marine Science Center listening station built with Killer Whale Tales and NOAA support
  • 2018: Langley Whale Center Ocean Listening Exhibit
  • 2019: Lime Kiln Visitor Center audio/video kiosk
  • 2020: Integration of new live streaming tech and citizen science into extant educational nodes

Growing signal library

Orcasound machine learning efforts

Open training & testing data for SRKW models

"Dory" (Erika Pelaez)

Orcasound hackathaon participant, 2018-2019

Semi-supervised(-ish) learning approach:

  1. Watkin's library: (Watkins library: orca class & other MarMam class)
  2. Librosa: pre-processing and feature selection
  3. Simple random forest model (non-NN, 128-bin vector from MelSpectrgram, 10-fold cross validation)
  4. SciKit learn (precision, recall, F1)
  5. May try PCEN (per channel energy normalization)...

UW ocean acoustic hack days (Nov 2018)

Led by Valentina Staneva and Shima Abadi

Challenge: find SRKW calls in Oregon shelf OOI/RSN hydrophone node (80 m depth) during period when NOAA satellite teg track narrows down when they may have been audible...

DEMO validation tool: WhaleDr

Microsoft Pod.Cast annotation tool

Product of July 2019 4-day Microsoft hackathon

Developed by Akash Mahajan, Prakruti Gogia, & Nithya Govindarajan

DEMO

Microsoft Pod.Cast ML model

VGG-ish, ResCNN, transfer learning (Akash, Prakruiti, Nithya)

  • validation via Pod.Cast UI
  • Round 0: labeled data from Watkin's "killer whales"
  • Round 1: data from Watkin's "killer whale" unlabeled master tapes
  • Round 2: 90 minutes data from Orcasound Lab 5 July 2019
  • Performance based on 30 min 5 July 2019 test set
  • Round 3 (in progress): ~3.5 hrs from Orcasound Lab 27 Sep 2017

Tour: Pod.Cast page of the Orcadata wiki...

UW grad students

Jennifer, John, Wai Sing, Yuhao

Invitation to the Orcasound Slack workspace

Localization of orca calls

Leads to: call source level, lombard effect, & great curiosities about cetacean communication

Orca communication animation

Rare separation of a calf documents call-respons in SRKWs

Red dot is mother/brother location; blue dot is calf. Towed array is in lower right corner.

Orcasound DC (~2002): WhoListener

Admiralty Inlet study (unpublished): of 22 SRKW transits, humans detect 45%, Wholistener 64%, combined 77%. (71/79/93% during local daytime)

During 2009-2012 *many* .jpg spectrograms and .mp3 clips were archived (example raw data directory) & there is a mySQL database with O(1000) expert validations

Orcasound DC (2009): ASA Portland, OR

Orcasound DC: flow chart

Orcasound DC: triggering

Current DC effort: Zorbita

From Visual Basic to QT to the clouds?

Questions?

/