Plate I Where the smoke starts

The haze has lifted. Can we get ahead of the next one?

Singapore has had a breather. Research at NUS and beyond suggests how computing could help prevent fires and protect people when smoke returns.

October's fires brought haze back to Singapore. On 9 October, the National Environment Agency reported air quality in the Unhealthy range as smoke from Sumatra reached the island.1

To give Singapore more warning, forecasters need a clear picture of fires across the region. Satellite observations can leave gaps when clouds obscure the ground, as NEA reported on 22 October.2

Looking ahead

A drone could inspect suspected fires where the satellite view is unclear. With limited flight time, it would have to choose which sites to visit first.

More than a decade ago, NUS Computing researchers Zhan Wei Lim, David Hsu and Wee Sun Lee studied a general form of this problem in robotics. Their adaptive informative path-planning algorithm uses earlier observations to choose where a robot should measure next, balancing the information it expects to gain against travel cost.3

A more recent approach, CAtNIPP, uses reinforcement learning to choose routes that reduce uncertainty about an area within a travel budget. NUS Mechanical Engineering's Guillaume Sartoretti and colleagues published the work in 2023.4

Looking ahead

A haze-monitoring drone could use such methods to revise its route as it checks suspected fires. Ground crews could then use the findings to decide which sites need a visit.

Fire records also help identify where to concentrate prevention efforts. They point to peatlands, where thick layers of partly decomposed plants have accumulated. In the severe dry year of 2015, peatlands covered 12% of the area studied across Peninsular Malaysia, Sumatra and Borneo, but accounted for 53% of fire detections. NUS researchers Chenghua Shi, Jukka Miettinen and Soo Chin Liew found far more fire activity in deforested peatlands than in intact peat swamp forest.5

Looking ahead

Make prevention worth doing

That contrast makes land use central to prevention. A landholder preparing a plot needs a workable alternative to burning, such as access to affordable land-clearing machinery.

Looking ahead

Governments and companies that benefit from less haze could contribute to a prevention fund. Communities could apply for machinery or repairs to drainage controls. Contributors may favour different projects, so the rules for spending their money would matter.

Looking ahead

Mechanism design studies how rules affect people's choices, including their willingness to contribute. NUS Computing's Warut Suksompong and colleagues studied how to allocate voluntary contributions to shared projects. In their model, each contributor's money goes only to projects they find acceptable, and the allocation rule gives them an incentive to participate.6 A haze fund could draw on this work to select projects that contributors are willing to support together.

Looking ahead

Once money is available, the terms offered to landholders can affect what it achieves. In a Brazilian Amazon trial, flexible conservation contracts allowed landholders to keep part of their payment when clearing stayed below specified limits. During the peak deforestation period, these contracts protected more forest than contracts that cancelled the payment after any clearing.7

Looking ahead

A fire-prevention programme could test an offer of land-clearing machinery alongside payments for completed work. Scheduling software could allocate the shared equipment before clearing begins. Records of its use and subsequent fires would help assess whether the offer changed how people prepared their land.

Before the smoke

Those arrangements need to be in place before the land becomes dangerously dry. On peatlands, falling water levels expose more peat to drying. Monitoring those levels could give agencies a better idea of when fire risk is rising.

Researchers in Riau tested this idea by combining a model of water movement through soil with an artificial neural network, which learns patterns from data. The water model estimated soil moisture and water-table depth. The neural network used these estimates alongside weather data to predict fire activity. In tests on years excluded from training, this combination predicted fires better than a version using weather alone.8

For a view of the coming months, members of the same team developed ProbFire. Its neural-network classifier, called a multilayer perceptron, learns from fire records, climate information and maps of forest and peatland. It then uses seasonal forecasts to estimate the probability of fire activity.9

In tests using past records, it outperformed predictions based on the region's typical climate about two months ahead in south Sumatra and south Kalimantan, and up to four months ahead in south Papua. In west Kalimantan, seasonal forecasts added no improvement.9 The useful warning period therefore depends on the region. Agencies could use these results to decide where seasonal forecasts offer enough notice to organise prevention work.

Looking ahead

For that notice to be useful, the funding has to be ready too. Mozambique's Ready, Set & Go! drought system links seasonal forecasts to agreed actions and funds set aside in advance.10 A fire-prevention programme could use a similar arrangement, releasing funds for canal repairs and equipment bookings when an early warning arrives, then deploying additional crews if the risk rises.

Crews also need local measurements to decide where work is most urgent. Indonesia's peatland agency has a separate programme that connects satellite estimates of soil moisture with measurements on the ground. In its 2022 account, the UN Food and Agriculture Organization describes farmers in Central Kalimantan checking equipment and reporting groundwater levels. When the land becomes too dry, they can warn neighbours and help rewet it.11

Following the smoke to Singapore

When fires do break out, attention turns to the wind. Forecasters need to work out where it will carry the smoke and when that smoke might reach Singapore.

NUS's Centre for Remote Imaging, Sensing and Processing, or CRISP, publishes satellite maps of fires and airborne particles.12 It also produces an experimental 48-hour forecast of fine-particle pollution, known as PM2.5. The model combines weather forecasts with estimates of emissions, including smoke from fires, to calculate how pollution may spread across the region.13

Following the smoke across a map is only part of that forecast. On 22 October, NEA reported smoke high above Singapore, with little reaching the air near the ground.2 To estimate what people will breathe, forecasters also need to know how high the smoke is.

During Singapore's October 2010 haze, NUS researcher Santo Salinas and colleagues combined measurements of sunlight passing through the atmosphere with laser readings at different heights and ground-level PM2.5 measurements. Together, these observations let them examine the smoke above Singapore alongside the air near the ground.14

When should a warning trigger action?

A forecast can appear accurate and still give poor warnings. Imagine 1,000 days, with severe haze on 50 of them. A system that predicts no severe haze every day would be right 95% of the time, yet miss every severe episode. Counting missed episodes and false alarms tells us more about whether a warning is useful.

The risk threshold determines when a forecast triggers a warning. Lower it and more severe days are caught, at the cost of more false alarms. The choice depends on what happens after the warning. ProbFire's authors compared the cost of acting with the losses preventive action could avoid. Their analysis estimated that different users would benefit from different thresholds.9

A fire-prevention agency would therefore need to choose a threshold that fits the work it can carry out and afford. In Singapore, official health advice should guide changes to school and workplace activities.1 Advance notice gives them time to prepare suitable rooms and equipment.

Looking ahead

A plan for the day

Moving a lesson indoors means finding a suitable room, arranging staff and checking that the air there is cleaner. A school may also need to decide where to put a limited number of air cleaners.

Looking ahead

These choices form a constrained optimisation problem. A planning tool could compare timetables using the forecast and indoor particle measurements. It would seek lower exposure with manageable disruption, subject to room capacity and staff availability. Teachers could review the suggested timetable and change it as conditions develop.

Looking ahead

NUS research gives us ways to measure whether such changes help. The MyPEQ project, led by Rajasekhar Balasubramanian, used portable sensors to study pollution exposure during daily activities, indoors and outdoors.15 In a separate study, Kwok Wai Tham and colleagues tested a fan filter unit that reduced particle levels in a classroom during haze.16 Room measurements could show whether moving a class or changing its filtration gives pupils cleaner air.

Looking ahead

Employers could use similar planning tools to adjust outdoor work schedules. Community centres could use them to arrange access to suitable indoor rooms. The forecast, room measurements and available space would all inform the plan.

Before the next episode

To prevent future fires, restoration teams work to keep peatlands wet. Blocking drainage canals helps retain water in the ground. A 2026 study found fewer fire hotspots in restored Indonesian peatlands than in comparable unrestored areas during 2019. The difference was not detectable in 2023, when fire activity was much lower.17

Looking ahead

With a limited budget, a restoration team must choose which canals to repair. A computer model could use canal layouts, water levels and rainfall to estimate how each repair would affect water retention through a dry spell. Planners could then compare the area kept wet with the cost and effects on neighbouring farmland before sending crews out.

NUS Computing's Sanka Rasnayaka and colleagues developed DUALFloodGNN, a graph neural network for modelling floodwater. It treats an area as connected locations, predicting how much water each holds and how water flows between them. Training penalises predictions that break physical rules about water balance. The team tested it on simulated floods in an Australian river catchment.18

Looking ahead

A peatland model could borrow this approach, combining learning from observations with the physics of water movement through the ground. Planners could use it to compare canal-repair plans under different rainfall scenarios, then check and update its predictions with field measurements.

Looking ahead

Once the season ends, fewer fires would be encouraging, but wetter weather might explain the change. To assess the prevention work, agencies would need to compare sites while accounting for local weather, land conditions and the repairs or support each area received.

Looking ahead

The records needed for that comparison may belong to several agencies, villages and companies. NUS Computing's Tze-Yun Leong and colleagues have developed methods for federated causal inference, which estimate the effect of an intervention using data held by separate organisations without exchanging their raw records.19 A study could use these methods to compare repaired sites with suitable unrepaired sites and estimate whether the repairs changed fire risk, and by how much. Agencies could then use the findings to decide which work to fund in the next dry season.

Use the clear spell to prepare

Bookmark NEA's haze service. Check the one-hour PM2.5 readings and accompanying advice before outdoor activities. Use the 24-hour PSI forecast to plan for the next day, and follow the guidance for your health and activity level.2021 The regional outlook tracks dry weather, fires and winds that could carry smoke. Check the update time when reading or sharing a forecast.21

Schools and workplaces can prepare now. Decide who will monitor advisories and which activities can move indoors. Check that suitable rooms and air-cleaning equipment are ready. NEA's response guidance covers indoor arrangements for schools and adjustments to outdoor work.1

If you are considering an AI warning service, ask how much notice it gives and how often it misses episodes or raises false alarms in your region. A planning tool should also show how its proposals fit the staff, equipment and budget you have.

For agencies funding prevention, the clear spell is time to test improvements before the next dry season. Trials could compare computer-guided inspections with existing routes, or use water models to choose canal repairs. Record the work, its cost and what happens afterwards, alongside weather and land conditions. Use those results to decide which measures deserve wider funding.