About the workshop

This virtual workshop is a continuation of the NOAA series of workshops on “Leveraging AI in Environmental Sciences.” The third event continues the successes of previous workshops and encourages participation by scientists, program managers, and leaders from the public, academic and private sectors who work in AI and environmental sciences. The theme for this year’s workshop is “Transforming Weather, Climate Services, and Blue Economy with Artificial Intelligence.” This year’s workshop is led by the NOAA Center for Artificial Intelligence (NCAI), a program in the formulation stage.

All time listed on this page is Mountain Time (UTC-6) by default. You can change to your local time zone on the right side of this page. Questions regarding the workshop can be addressed to ai.workshop@noaa.gov. View the participant handbook for how to navigate the workshop technology platforms.

Registration for the workshop is closed and this page is viewable to registered participants only. Please do not share your access code publicly.

Suzanne Romain

Pacific States Marine Fisheries Commission

Challenges associated with integration of machine vision algorithms into camera based catch accounting programs.

The Electronic Monitoring Innovation (EMI) group has been developing generalized Machine Vision (MV) models for fish identification and accounting since 2014. Image collections from Alaska were utilized to train models and included fish coming aboard longliners and sliding through enclosures called camera chutes. Model backbones were intended to be utilized with less training when applied in new regions. While initial model performance was promising, it degraded when applied to new vessels, areas, and camera/lighting conditions. EMI’s experience across years of fisheries image collection has highlighted several sources of such problems likely to be common in other applications of MV in fisheries monitoring. These challenges are endemic to fisheries monitoring tasks and are not addressed by basic applications of AI, which are primarily developed on large, clean, and balanced datasets that do not reflect some features common in fisheries monitoring imagery. When applied to real world data, models need to be calibrated for varying image characteristics, a process known as Domain Adaptation by developers. Our UW partners have developed analytical tools to address this, and are implementing them into basic applications. The EMI group has explored Domain Adaptation methods that address unbalanced data and background changes simultaneously, with or without human review. While Domain Adaptation applications can certainly improve classifications even without human review, they need further development, especially to account for Label Shift (different species compositions), long tail distributions, and novel species detection.

My Speakers Sessions

8:30am MDT

Virtual Poster Walk - Part V QiqochatJordan Watson • Yoonjin Lee • Erin Moreland • Ashley Woods • Ali Muhamed Ali • Sanjeev Koppal • Peter Tueller • Suzanne Romain • Laura McGee • KUNAL BHAVESH MEHTA • Liese Siemann • Yawen Zhang • Wu-Jung Lee • William Tarpeh • Shlomi Ziskin Ziv • Marina Chaji • Taylor Lee • Earth Science Information Partners

3rd NOAA Workshop on Leveraging AI in Environmental Sciences

Suzanne Romain

My Speakers Sessions

8:30am MDT

Recently Active Attendees