Plastic Talk 2026: When AI and Citizen Science Join Forces to Monitor Marine Plastic Debris

On 7-8 October 2026, a joint program on applying artificial intelligence (AI) to marine plastic debris monitoring and management took place in Hanoi and Hai Phong. Day one was the hybrid forum "Plastic Talk", with nearly 20 participants in person and 64 online. On day two, 16 people took part in field training at Do Son and took nearly 500 photos of coastal plastic debris to help train AI models. The program was organised by the National Plastic Action Partnership (NPAP) - UNDP Viet Nam, GreenU and the IndigoWaters Institute (Taiwan), with support from Taiwan's Ocean Affairs Council.

Day 1 (7 October): Plastic Talk in Hanoi

The forum ran from 14:00 to 16:30 in Conference Room C1 at the UN House (304 Kim Ma, Hanoi) and online via Zoom, bilingual English-Vietnamese with interpretation. Participants included representatives of state management agencies, international organisations, research institutes, universities, civil society and the press. Ms. Dang Nguyet Anh, NPAP Manager at UNDP Viet Nam, opened the event and moderated the discussion. She stressed that marine plastic pollution is a transboundary issue that calls not only for policy solutions but also for more effective ways of collecting, analysing and using data. AI and data science are opening new approaches and giving countries in the region a chance to share experience.

Viet Nam: no baseline data yet, and a circular awaiting issuance

Opening the technical sessions, Dr. Pham Van Hieu (Viet Nam Institute of Meteorology, Hydrology, Environment and Marine Sciences, Ministry of Agriculture and Environment) said Viet Nam has no official, state-published figures on marine plastic debris. Studies use different methods, units and sampling approaches, so data are hard to compare, even within the same area. He noted that microplastics may be reported per litre, per cubic metre or per kilogram, and that net mesh size, tow time and the share of samples analysed also vary between teams. As a result, there is no baseline against which to measure the national target of cutting marine plastic waste by 75% by 2030 (Decision 1746).

To unify methods, the Institute and the Ministry are drafting a circular on technical regulations for marine plastic debris surveys. Work began in 2019 and has been tested in several areas, including Cat Ba and Hai Phong. The draft covers procedures for surveying beach litter, floating litter (observed from vessels), microplastics (net sampling), river-mouth debris and seabed debris. According to Dr. Hieu, surveys in northern Cat Ba showed floating debris density rising several-fold after Typhoon Yagi compared with before. The appendices on data standardisation and debris classification are especially important for later AI applications. The draft does not yet include AI; it focuses on creating a consistent input dataset from the central to the local level. An official letter requesting support to issue it was sent to the Department of Sea and Islands in June 2026.

Korea and Taiwan: it starts with people

Dr. Lee Jongmyoung (Chief Science Officer, Our Sea of East Asia Network - OSEAN, South Korea) and Ms. Ning Yen (Founder and CEO, IndigoWaters Institute, Taiwan) shared experience applying AI to marine debris monitoring in their places. The organisers drew one common lesson: the deciding factor is not necessarily technology but people. South Korea has been at it for more than 15 years, starting with citizen scientists; Taiwan built an open data platform where members of the public contribute photos of litter following common guidelines. According to Ms. Nguyen Thi Thu Trang (Director of GreenU), Taiwan alone has collected around 20,000 photos in the past two years.

A representative of the Department of Digital Transformation (Ministry of Agriculture and Environment) asked a direct question about transfer: could a litter-recognition model trained in Korea's temperate waters be used in Viet Nam's tropical monsoon seas, with mangroves and river sediment as in Hai Phong or Can Gio, and how much data is needed to retrain it while keeping accuracy above 85%. A GIZ participant asked how Taiwan's monitoring data has been used by policymakers. Both questions show Viet Nam's shared concern that data must lead to management decisions.

Reading beach photos with commercial AI: what works, what does not

Mr. Bui Le Thanh Khiet (Head of the Plastic Circular Economy Unit, Institute for Circular Economy Development - ICED) presented a study comparing a commercial general-purpose AI model with human labellers on the same set of beach photos collected over 13 months (September 2023 to December 2024). Each photo contained a 1 m2 frame, and items were classified using a European litter category system with more than 100 groups. At the level of whole survey rounds, AI and human results were fairly similar: both showed which beaches were most and least polluted and the trends over time. At the level of individual photo plots, however, the model recorded only about 77% of litter items; it over-counted in low-litter plots, missed items in heavily littered plots, and tended to confuse white foam fragments with oyster shells in oyster-farming areas. The model also could not distinguish plastic types.

Mr. Khiet therefore recommended treating commercial AI as a first screening step for citizen science programs, not a replacement for field surveys. A sub-sample checked by a second human labeller is needed to calibrate errors, and model results should be kept separate from human results. Asked whether the method could be included in the circular, he said this would be very difficult because uploading data to commercial platforms raises data security concerns. A participant also noted that vendors may quietly change their algorithms, so a result that is right today may not hold next year. The method is therefore suited to research or NGO surveys, not as a basis for penalties or for identifying sources of discharge.

Panel discussion: what does AI data need to carry legal weight?

The panel was moderated by Ms. Dang Nguyet Anh, with Dr. Pham Van Hieu, Dr. Lee Jongmyoung, Ms. Ning Yen and Ms. Nguyen Thi Thu Trang. Asked what data regulators would recognise, Dr. Hieu set out four requirements: clear provenance (time, place and survey coordinates); reference data from conventional field surveys; labelled sample sets to check completeness and correctness; and data governance to keep data consistent across regions. He also stressed that AI is only a supporting tool: to show a 75% reduction you first need to know the baseline, and managers will not rely on a single tool.

The representative of the Department of Digital Transformation added that the Law on Artificial Intelligence, in force since 1 March 2026, sets responsibilities and transparency requirements for data used in AI, and could therefore be the basis for subordinate regulations guiding the use of AI in plastic monitoring. He also asked for three conditions to be clarified so that the circular is workable: technical rules on data format and security, a data reporting system, and funding, staff and infrastructure at local level.

On the community side, Ms. Nguyen Thi Thu Trang said the simplest way for people to take part is to photograph litter with their phones, following a consistent photo method with clear guidance. If even a small share of millions of people contributed, an open-source system would hold a very large amount of data. But the relationship must run both ways: contributors need useful information in return, such as a litter hotspot map for choosing where to hold a weekend activity, rather than waiting a year for a research report, by which time the hotspots have shifted with the seasons.

Three core actions

At the close, the organisers summarised three directions for action. First, standardise data: Viet Nam needs to develop and issue a set of technical standards for plastic monitoring data soon, as a legal basis for policymaking as AI tools become more common. Second, build an open platform as Taiwan has done, cooperating with international platforms such as MDImageNet and learning from Korea and Taiwan, to create a shared training data repository that saves resources. Third, empower communities: develop friendly, accessible AI tools to mobilise citizen science networks and local initiatives.

A representative of the Department of Environment (Ministry of Agriculture and Environment) said at the end of the session that the department is revising the circular guiding the Law on Environmental Protection, is interested in data collection methods that can be turned into policy, and welcomes comments and cooperation from experts. The organisers noted that this is also the program's goal: getting research results and monitoring methods in front of regulators.

Day 2 (8 October): field day at Do Son, nearly 500 photos for AI

On the morning of 8 October, 16 people, made up of the coordination team, technical partners, selected registrants and volunteers, left Hanoi and took the expressway to the Do Son tourist area in Hai Phong. The first training session was held right by the sea: Ms. Ning Yen explained the MDImageNet photography protocol (camera angle, lighting, how to separate single items from groups of litter, avoiding duplicates), while Ms. Nguyen Thi Thu Trang interpreted and coordinated. The group split into three small teams, each assigned equipment, measuring tools and phones with location tracking.

Two field sessions, morning and afternoon, followed the beach sections of Zones 1, 2 and 3 and places where litter accumulates with the tide. Teams photographed many kinds of plastic litter: fishing gear, EPS foam buoys, single-use plastic fragments, household packaging and tourism waste. The afternoon session added accumulation points shaped by currents, shipping and tourism, to make sure the photos were sharp and diverse enough to reflect the particular conditions of Northern Viet Nam's coast. In total, the group took nearly 500 photos.

At the end of the day, the teams reviewed some representative photos in a debrief. Ms. Ning Yen assessed photo quality against MDImageNet's AI standards, gave feedback on framing, lighting and litter categorisation, and agreed with the team on how to compile, filter and upload the photos to the system after the trip. These are among the first datasets in Viet Nam collected by a community group using a consistent photo protocol, paving the way for larger collection drives.

GreenU thanks all speakers, participants, partners and volunteers who joined us over the two days.