Orcasound at ONC/Meridan 2019 workshops

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

Documentation at Orcasound's orcadata repository Github wiki
Storage in Orcasound's acoustic sandbox on S3
Annotation tools: Audacity and Pod.Cast

"Dory" (Erika Pelaez)

Orcasound hackathaon participant, 2018-2019

Semi-supervised(-ish) learning approach:

Watkin's library: (Watkins library: orca class & other MarMam class)
Librosa: pre-processing and feature selection
Simple random forest model (non-NN, 128-bin vector from MelSpectrgram, 10-fold cross validation)
SciKit learn (precision, recall, F1)
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