The core of the method rests on two numbers extracted from satellite imagery. The first is aerosol optical depth, or AOD, a measure of how much sunlight particles in the atmosphere block as it travels toward the sensor. The second is the Ångström Exponent, or AE, a spectral fingerprint that distinguishes particle sizes. When the Ångström Exponent climbs above 1.2, the atmosphere is dominated by fine-mode particles, the submicron specks characteristic of combustion processes such as coal burning, biomass fires, traffic, and industrial emissions. By requiring both a high AE and an AOD above 0.4, the researchers could flag days when the air near a power plant was genuinely loaded with fine combustion aerosols rather than coarse dust or sea salt.

The dataset behind the study is enormous. The team assembled 41,119 daily observations spanning 2015 through 2025 across thirteen major coal-fired power plants on Java, drawing on Himawari-8 aerosol products distributed through the Japan Aerospace Exploration Agency’s P-tree system. Because coal plants operate continuously rather than episodically, their persistent emissions create a recognizable local aerosol signature that the satellite can track over years. Across all sites, 8.52 percent of days met the criteria for a fine-mode combustion event, but that average conceals enormous spatial variability that turned out to be one of the study’s most important findings.

Nowhere is that variability sharper than in the contrast between Suralaya and Pacitan. Suralaya, a sprawling power complex in Banten province, recorded the highest mean aerosol optical depth of any site at 1.01, along with very high pollution days on 51.38 percent of the observations, meaning that for roughly half the decade the air above it was severely loaded with particles. Pacitan, by comparison, showed the cleanest conditions of the thirteen sites with a mean AOD of just 0.46. Two plants on the same island, burning the same fuel for the same grid, yet producing radically different atmospheric footprints, a difference the authors attribute to local meteorology, topography, and dispersion conditions rather than to any single operational factor.

With the historical record established, the researchers trained three machine learning models to classify days as combustion aerosol events or non-events: a random forest, the gradient-boosted XGBoost algorithm, and a long short-term memory network designed to capture temporal sequences. The results were almost unnervingly good. Random forest and XGBoost both achieved perfect scores across every metric the team measured, with accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve all reaching 1.000. The LSTM performed slightly below that ceiling but still delivered strong classification. Perfect scores on held-out data suggest the event signal is exceptionally well separated in the feature space, though the authors treat the models as classification tools within a larger predictive framework rather than as standalone oracles.

The genuinely novel contribution lies in how the team projected the future. Standard Monte Carlo simulation, the workhorse of probabilistic forecasting, typically resamples from historical distributions while assuming events are independent of one another. The researchers instead built a modified Monte Carlo approach that explicitly incorporates seasonal patterns, temporal autocorrelation, and extreme events, recognizing that a polluted day is more likely to follow another polluted day and that monsoon-driven seasonality shapes the entire annual cycle of aerosol loading. When the two approaches were run side by side to generate projections from 2026 to 2060, the modified method predicted 9.3 percent more aerosol event days than the standard version, a difference the authors report as statistically significant at p less than 0.001.

Even more consequential than the mean difference is what the modified approach revealed about uncertainty. The conventional Monte Carlo produced confidence intervals spanning 172 days across the projection period, while the modified version produced intervals of 212 days, and average standard deviations ballooned from 43.1 days to 82.7 days. Wider intervals are not a weakness; they are an honest accounting. The authors argue that conventional methods generate falsely precise projections by ignoring natural variability, effectively promising a certainty the atmosphere does not offer. For policymakers deciding where to spend limited emission-control budgets, knowing the plausible range of outcomes matters as much as knowing the average.

The spatial heterogeneity of the projections is where the study becomes genuinely actionable. Pelabuhan Ratu, on the southwest coast of West Java, shows the largest projected increase in fine-mode combustion aerosol events at 62.3 percent, a trajectory that would transform the site’s pollution profile over the coming decades. Cilacap, in Central Java, moves in the opposite direction with a substantial projected decrease of 38.1 percent. These divergent futures mean that a single island-wide pollution strategy would misallocate resources in both directions, over-investing where conditions are already improving and under-investing where they are deteriorating fastest. The framework, in effect, hands regulators a ranked priority list grounded in probabilistic evidence.

The broader context sharpens the stakes. Coal remains the backbone of Indonesian electricity generation, and national planning documents from the Ministry of Energy and Mineral Resources and the state utility PLN chart the sector’s expansion through the 2020s and beyond. Fine particulate matter from coal combustion is linked to cardiovascular and respiratory disease, and recent research on Indian coal plants has estimated substantial reductions in premature mortality from pollution controls, a parallel the Indonesian authors cite. Satellite-based monitoring offers a way to hold that trade-off in view continuously, since ground-based air quality networks in the region remain sparse relative to the scale of the problem. Himawari-8’s geostationary vantage, scanning the full disk every ten minutes, provides coverage that no surface network can match.

What the Telkom University team has demonstrated is a template that could travel well beyond Java. The ingredients, a geostationary aerosol record, a size-sensitive filtering criterion, machine learning classifiers, and a variability-aware Monte Carlo engine, are available for any region where persistent emission sources meet satellite coverage. The study’s own numbers carry the clearest message: nearly a decade of daily observations, thirteen power plants, and a projection horizon of thirty-five years all converge on the conclusion that aerosol futures are local, seasonal, and deeply uneven. Treating them as such, the authors suggest, is the difference between emission-control strategies that look good on paper and ones that actually clean the air where people breathe it.

Subject of Research: Predictive modelling of fine-mode combustion aerosol events near coal-fired power plants in Java using satellite data and machine learning

Article Title: Predictive modelling of fine-mode combustion aerosol events near Indonesian coal-fired power plants using AE-filtered Himawari-8 satellite data

Article References: Agustia, S., Raharjo, J., Wijayanto, I., & Karna, N. (2026). Predictive modelling of fine-mode combustion aerosol events near Indonesian coal-fired power plants using AE-filtered Himawari-8 satellite data. Environmental Monitoring and Assessment, 198(10), Article 1074. https://doi.org/10.1007/s10661-026-15804-1

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15804-1

Keywords: Himawari-8, aerosol optical depth, Ångström Exponent, coal-fired power plants, fine-mode aerosols, machine learning, random forest, XGBoost, LSTM, Monte Carlo simulation, air pollution, Indonesia