Using Social Media Imagery to Nowcast Air Pollution


Commuters pictured wearing masks while approaching the MRT station at Bundaran HI

Air pollution has emerged as a growing health issue across Asia and the Pacific and affects the lives of millions of citizens. With this concern in mind, Pulse Lab Jakarta has been investigating how to deploy a machine learning model it developed to nowcast air quality using deep learning. Building on exploratory research conducted in 2018 and in line with our mission to build collaboration and exchange expertise and technical skills with the private sector, we applied to the Computer Vision for Global Challenges (CV4GC) initiative and were delighted that our proposal was selected as one of the final 17 challenge winners. During the workshop, we presented our nowcasting air pollution model and received expert feedback from the computer vision community on how to refine its development.

The Computer Vision for Global Challenges initiative is focused on bringing the computer vision community closer to socially impactful tasks, data sets as well as applications, which can yield positive results for worldwide impact. Our data scientist attended the first workshop that was held in Long Beach, California in conjunction with the IEEE Conference on Computer Vision and Pattern Recognition last month. Special mention to the organisers who put a significant amount of effort into carefully selecting challenge proposals that covered a wide range of development issues, and sought to bridge the gap between the computer vision research and international development communities.

As a data innovation lab that is based in Indonesia and operates across the region on impactful projects using artificial intelligence methods and human centered design, the overall purpose of this initiative aligns with our research principles. From designing the scope of a research project in the early phase to developing and testing prototypes later on, it is important for us to always ensure and design with the users in mind, so that whatever insights uncovered can be useful for actionable, evidence-based decision and policy making.

The World Health Organisation reports that millions of people die from outdoor air pollution, and those deaths are overwhelmingly concentrated in the largest cities of the developing world. We have seen through our experimental process in AI for good that the opportunities for scaling up prototypes and ideas will only come when we collaborate in extensive public-private partnerships in order to ensure that we get the best technical expertise; get access to cutting edge research; and also ensure that the domain expertise for solving development problems is equally superb.

Why is air quality important?

As mentioned above, we (along with millions of other citizens) have become increasingly concerned about the rising air pollution across Southeast Asian cities. Air pollution is now one of the leading environmental risks impacting sustainable health, development and economic growth. Mitigation of air pollutants is critical to safeguard current development goals and to achieve further progress towards the sustainable development goals (SDGs), particularly in terms of health (SDG3), urbanisation (SDG11) and climate action (SDG13).

In 2016 PLJ created Haze Gazer, a crisis analysis and visualisation tool to track and manage the impact of fire and haze events. Haze Gazer combines open, government and social media data to provide a near real-time view of haze and hotspots. Citizen generated data especially have the potential to provide the context needed to really understand the effects of the problem on the local community. In 2018, PLJ conducted an exploratory analysis to estimate air quality using a deep learning model that fuses social media images, satellite and ground sensor information. The approach yielded promising results, but requires further exploration to validate the model as a suitable replacement for or adjunct to traditional air quality monitoring approaches, such as the use of satellite and official weather station data.

Why do we use a different method this time?

Pulse Lab Jakarta’s poster presentation at the Computer Vision for Global Challenges conference
  • Official air quality information is sparse and expensive, particularly in the developing world.
  • People take photos every day and several of them have a view of the sky.
  • Using deep learning, we can train a machine learning algorithm to:
  1. Recognise which photos are indoor / outdoor,
  2. Recognise which outdoor photos have a view of the sky,
  3. Extract a ‘haze feature’ using existing dehazing algorithms designed to make photos look clearer, and
  4. Use that haze feature in a neural net that automatically predicts air quality down to the village level.

The beginning of endless opportunities

Facebook AI and the University of Edinburgh were the main drivers behind having the global impact challenges convened at this highly acknowledged computer vision conference. Mentors were drawn from leading Silicon Valley firms, and Pulse Lab Jakarta’s data scientist was fortunate to be paired up with the machine learning lead at CrowdAI who was fantastic in giving essential technical advice on how to tailor the nowcasting air pollution project and providing introductions regarding key experts in this area who can contribute to shaping our research going forward.

We came away from the conference excited at the level of enthusiasm we observed from the world’s leading computer vision experts who lent their technical skills and advisory support to address some of the world’s most serious challenges. With our niche access to social media companies, our AI expertise in house and prime location within the social media capitals of the world, we are uniquely positioned to embark on identifying how social media imagery can better showcase air pollution around the region in real-time rather than waiting for the effects to become more obvious through medical interventions. We also have a state-of-the-art visualisation platform that presents data regarding natural disasters from several unconventional data sources in a simple, intuitive and visually-appealing way that can be repurposed for visualising this data.

We recognise that air pollution is a global challenge that calls for cross-sector collaboration and citizen engagement and we’re keen to collaborate with other entities working to address this issue in Indonesia and across the region. If you’re interested in knowing more about our approach, please get in touch with us.

Pulse Lab Jakarta is grateful for the generous support from the Government of Australia.



UN Global Pulse Asia Pacific
United Nations Global Pulse Asia Pacific

UN Global Pulse Asia Pacific is a regional hub that aims to drive data innovation and sustainable development to ensure that no one is left behind.