The research, led by Bojana Petrović of the Vinča Institute of Nuclear Sciences at the University of Belgrade, together with colleagues from Spain’s Institute of Environmental Assessment and Water Research and the Norwegian Institute for Air Research, was conducted at an urban background site called Ada Marina, located on Ada Ciganlija, Belgrade’s largest recreational area along the Sava River. From June 2023 to May 2024, the team collected daily filter samples using three co-located low-volume samplers, yielding 84 representative PM10 samples for analysis. What set this campaign, dubbed WeBaSOOP, apart from earlier work in the region was its unprecedented chemical scope: rather than measuring only elements or ions, the researchers quantified 34 species, including organic and elemental carbon, 27 elements, four major ions, and — for the first time in Serbia and the wider Western Balkans — a suite of specific organic tracer molecules.

These tracers are the forensic key to the study. Levoglucosan, mannosan and galactosan are anhydrosugars released when cellulose burns, making them unambiguous fingerprints of biomass combustion. Sugars and sugar alcohols such as glucose, mannitol and arabitol mark primary biological aerosol particles — fungal spores, pollen, plant fragments and microbial debris. Two-methyltetrols, meanwhile, are oxidation products of isoprene, a volatile organic compound emitted in abundance by vegetation, and serve as markers of biogenic secondary organic aerosol. Feeding this rich chemical inventory into the United States Environmental Protection Agency’s Positive Matrix Factorization model, version 5.0, the team resolved the PM10 mass into seven factors, each with a distinct chemical profile, seasonal rhythm and wind-direction signature.

The dominant factor was biomass burning, contributing 21 percent of PM10 mass annually and a striking 34 percent during the heating season from October to April. Its profile was anchored by 80 percent of total galactosan, 79 percent of mannosan and 76 percent of levoglucosan, alongside roughly half of the potassium and chloride and a third of the organic carbon. Diagnostic ratios — an organic carbon to elemental carbon ratio of 9.4 and an organic carbon to levoglucosan ratio of 8.3 — matched values reported for wood smoke across Europe. The seasonal peak aligns with a well-known reality in Belgrade: an estimated 300,000 households are not connected to the district heating network and rely largely on wood as their heating fuel. During cold spells, temperature inversions trap these emissions in a stable near-surface layer, allowing pollutants to accumulate unchecked.

Intriguingly, the study also caught biomass burning outside the official heating season. On 2 October, before heating officially begins on 15 October and with ambient temperatures around 15 degrees Celsius, concentrations of the anhydrosugar tracers spiked sharply. The ratio of levoglucosan to mannosan, which distinguishes softwood burning (ratios below about 4) from hardwood burning (around 14 to 15), averaged 8.9 for the burning factor overall but reached 18.9 in summer versus 8.0 in winter. Satellite observations confirmed active fires in Vojvodina on that October day, pointing to agricultural waste burning — a phenomenon documented in Serbia every October since 2000 — while summer contributions may also reflect charcoal combustion at the restaurants dotting the Ada Ciganlija lakeside near the monitoring site.

The second-largest source, at 18 percent, was ammonium sulphate, a secondary aerosol formed when sulphur dioxide is oxidised in the atmosphere and neutralised by ammonia. This factor carried 55 percent of total ammonium and 54 percent of sulphate, with the coexistence of ammonium bisulphate and ammonium sulphate indicating only partial neutralisation of sulphuric acid. Its high organic-to-elemental carbon ratio of 9.1 confirmed the aged, regional character of these particles. Wind analyses pointed to influences from the south and southeast, consistent with transport from the coal-fired Nikola Tesla and Kolubara thermal power plants located 25 and 50 kilometres from the city, and the Kostolac plant 60 kilometres to the east. Notably, the 2024 launch of flue gas desulphurisation at Nikola Tesla A may already be reshaping this contribution.

Mineral dust ranked third at 17 percent, characterised by crustal elements — aluminium, titanium, strontium, magnesium and calcium — and showing influence from the southwest. The campaign captured a Saharan dust outbreak on 31 March 2024, confirmed by sodium-to-aluminium and potassium-to-aluminium ratios closely matching literature values for Saharan material, and corroborated by FLEXPART footprint modelling that traced air masses back across the Sahara. But dust was not purely natural: preparatory construction for the future Belgrade metro depot at Makiš, roughly seven kilometres upwind, left its imprint on the factor’s wind polar plots, blending local construction dust with long-range transported soil.

Perhaps the most novel finding concerns the biological fraction. Two factors of biogenic origin — primary biological aerosol particles and biogenic secondary organic aerosol from isoprene — together accounted for 20 percent of PM10 mass, each contributing 10 percent. The primary biological factor, traced by sugars and sugar alcohols, peaked in the warm months and showed influence from the vegetation of Ada Ciganlija and the Košutnjak forest three kilometres away. Using the mannitol-to-arabitol ratio of 1.0, consistent with fungal spore signatures across Europe, the team estimated that 21 to 44 percent of the organic carbon in this factor derives from fungal spores. The isoprene-derived secondary factor, marked by 2-methylerythritol and 2-methylthreitol, appeared exclusively during the growing season and peaked in summer, its co-occurrence with sulphate species suggesting acid-catalysed aqueous-phase formation — a mechanism first proposed in a landmark 2004 Science paper.

The remaining sources were a mixed traffic-and-industry factor at 15 percent, blending exhaust and brake-wear tracers such as iron, antimony, copper and tin with industrial markers including arsenic, cadmium, lead and vanadium, and a smaller 9 percent factor combining aged sea salt transported from the Atlantic with road salting and local combustion emissions. Mass reconstruction independently confirmed the source apportionment: the carbonaceous fraction dominated PM10 composition at 54 percent annually, followed by secondary inorganic aerosol at 19 percent and mineral dust at 13 percent. The average PM10 concentration over the campaign was 23.9 micrograms per cubic metre — compliant with the current EU limit of 40 but exceeding both the 20 microgram limit that the new EU directive will impose from 2030 and the World Health Organization’s 2021 guideline of 15 micrograms. Winter concentrations of 28 micrograms far outstripped summer’s 19, and the daily limit was breached four times, peaking at 92 micrograms on 19 December.

The implications reach well beyond Belgrade. Because roughly a fifth of the particulate mass originates from biogenic sources that no emission control can touch, and mineral dust adds another irreducible background of around 4 micrograms per cubic metre, meeting future air quality standards will demand disproportionately deep cuts in controllable anthropogenic emissions. The study’s authors argue that targeted mitigation of residential wood burning — the largest single controllable source — offers the clearest path forward, while the identification of biogenic factors for the first time in the region provides a scientific basis for the kind of natural-source accounting that emerging legislation may eventually permit. As Europe tightens its air quality rules and the Balkans remain one of the continent’s pollution hotspots, this Belgrade campaign demonstrates how comprehensive chemical forensics can turn a hazy urban sky into a readable ledger of exactly what society burns, builds, drives and grows.

Subject of Research: Source apportionment and chemical characterization of PM10 particulate matter in Belgrade, Serbia

Article Title: Chemical characterization and source apportionment of PM10 in Belgrade, Serbia: influence of local and regional anthropogenic and natural sources

Article References: Petrović, B., Alastuey, A., Yttri, K. E., Pandolfi, M., Jovanović, M., Radović, B., Kovačević, R., Stojanović, D. B., Davidović, M., Platt, S. M., Bartonova, A., & Jovašević-Stojanović, M. (2026). Chemical characterization and source apportionment of PM 10 in Belgrade, Serbia: influence of local and regional anthropogenic and natural sources. Atmospheric Chemistry and Physics, 26(19), 14165-14184. https://doi.org/10.5194/acp-26-14165-2026

Image Credits: AI Generated

DOI: 10.5194/acp-26-14165-2026

Keywords: PM10, air pollution, source apportionment, biomass burning, Belgrade, positive matrix factorization, organic tracers, secondary organic aerosol, mineral dust, biogenic aerosols, heating season, Western Balkans

The study, led by Yumiao Li and colleagues at Southwest Jiaotong University in Chengdu, together with researchers at the Tianfu Yongxing Laboratory, deployed continuous monitoring of particulate matter and a suite of meteorological parameters in a plateau village for a full annual cycle. The team recorded annual mean concentrations of 24.3 micrograms per cubic meter for PM10, the coarse particle fraction, and 18.5 micrograms per cubic meter for PM2.5, the finer fraction that penetrates deepest into human lungs. While those annual averages may appear moderate by the standards of heavily polluted megacities, the seasonal and diurnal structure behind them tells a far more troubling story about when and how villagers are actually exposed.

The monthly pattern was unambiguous. December registered the highest monthly mean concentrations of both PM10 and PM2.5, while July recorded the lowest. That winter maximum aligns squarely with the heating season, when yak dung stoves burn for extended hours to keep homes warm in conditions where temperatures plunge far below freezing and conventional fuels such as coal or natural gas are scarce or unavailable. During the non-heating months, by contrast, stove use drops and atmospheric conditions improve, allowing particulate levels to fall to their annual minimum. The contrast between heating and non-heating season concentration peaks was pronounced, underscoring that the seasonal rhythm of pollution on the plateau is dictated less by industry or traffic than by the domestic hearth.

The diurnal cycles revealed an equally distinctive fingerprint. Concentrations of particulate matter rose sharply during two windows: 8:00 to 10:00 in the morning and 18:00 to 22:00 in the evening. These are precisely the periods when villagers light up their stoves, first to prepare breakfast and later to cook the evening meal and heat the house against the falling nighttime temperatures. The pattern is a classic signature of residential solid-fuel combustion, and it mirrors findings from other rural regions of China where household coal and biomass burning dominate ambient pollution. On the plateau, however, the fuel is yak dung, and its incomplete combustion in traditional stoves releases a substantial load of particulate matter that escapes indoors and accumulates in the village air.

One of the study’s most striking technical findings concerns the composition of the particle load. In every month, the ratio of PM2.5 to PM10 exceeded 69 percent, ranging from 68.6 to 76.2 percent. In other words, fine particles made up the dominant share of the total particulate burden year-round. This matters enormously for health, because PM2.5 particles are small enough to bypass the body’s upper respiratory defenses, lodge deep in the alveoli of the lungs, and even enter the bloodstream. A long body of epidemiological research, including the foundational work of C. Arden Pope and Douglas Dockery, has linked fine particulate exposure to cardiovascular disease, respiratory illness, and premature mortality. A pollution profile dominated by the fine fraction therefore carries disproportionate health implications, even at concentrations that might seem tolerable if judged on coarse particle numbers alone.

To understand why concentrations rise and fall as they do, the researchers turned to the meteorology of the plateau itself. Using Spearman correlation analysis alongside machine learning models, they found that PM2.5 concentrations were negatively correlated with solar radiation, temperature, and wind speed, and positively correlated with relative humidity. The physical logic is intuitive. Stronger winds disperse and dilute locally emitted particles; higher temperatures and stronger solar radiation enhance vertical mixing of the boundary layer, lifting pollutants away from the surface; and calm, cold, humid conditions trap particles near the ground where people live and breathe. On the Qinghai-Tibet Plateau, where winter brings both intense stove use and stagnant, cold air, these factors conspire to produce the December maximum the monitoring recorded.

The machine learning component of the study was designed as a rigorous head-to-head comparison. The team evaluated three widely used ensemble algorithms: Random Forest, Extreme Gradient Boosting, known as XGBoost, and Light Gradient Boosting Machine, known as LightGBM. All three are tree-based methods capable of capturing nonlinear relationships between meteorological drivers and pollutant concentrations, but they differ in how they build and combine their constituent trees. Random Forest grows many decision trees on bootstrapped samples of the data and averages their predictions, while XGBoost and LightGBM build trees sequentially, with each new tree correcting the errors of its predecessors. Across three evaluation metrics, Random Forest outperformed the two boosting approaches, proving the more reliable tool both for identifying the meteorological factors that influence PM2.5 and for predicting concentrations.

The predictive power of the trained model allowed the researchers to extend their analysis beyond the single monitored village. Applying the framework to villages in four cities across the Qinghai-Tibet Plateau, they predicted outdoor PM2.5 concentrations during periods of yak dung stove operation. The predictions exhibited similar variation trends across all four locations, suggesting that the pollution dynamics documented in the monitored village are not a local anomaly but a regional pattern rooted in shared fuel practices and a shared climate. That generalizability is what elevates the study from a single-site case report to a template for understanding air quality across the entire plateau, home to millions of people whose exposure has historically been invisible to national monitoring networks concentrated in eastern cities.

The findings arrive at a moment of growing recognition that household solid-fuel burning is a major and underappreciated source of ambient air pollution, not merely an indoor air problem. Previous research, including work published in the Proceedings of the National Academy of Sciences, has argued that Chinese household emissions constitute a substantial share of the country’s particulate burden, and field campaigns measuring traditional biomass cookstoves in Tibet and South Asia have documented their high emission factors. The new study adds a crucial temporal dimension to that literature, showing exactly when in the day and the year plateau villages are most affected, and demonstrating how machine learning can translate sparse field measurements into actionable exposure estimates for communities that lack permanent monitoring infrastructure.

The authors argue that their results can inform future strategies for improving outdoor air quality on the plateau and support assessments of the environmental and health benefits those strategies would deliver. Cleaner-burning stove designs, improved combustion efficiency, fuel alternatives, and ventilation interventions could all, in principle, blunt the morning and evening peaks and shrink the winter maximum. Because the study establishes a quantitative baseline and a validated predictive model, any such intervention could be evaluated against measured and modeled expectations rather than guesswork. For a region where the energy transition must balance cultural tradition, fuel scarcity, and the imperative of public health, that kind of evidence is the essential first step. The image that emerges from the data is vivid: twice a day, as stoves are lit across plateau villages, a fine-particle plume rises into cold, still air, and only when the sun climbs and the wind picks up does the atmosphere begin to clear.

Subject of Research: Temporal variation of outdoor particulate matter pollution from yak dung combustion in Qinghai-Tibet Plateau villages and its correlation with meteorological factors

Article Title: Temporal variation of outdoor particulate matter concentration in a village of the Qinghai-Tibet Plateau and its correlation with meteorological factors: based on long-term field measurement and machine learning approach

Article References: Li, Y., Chen, J., Wu, D., Ma, R., Yi, Y., Yu, T., & Deng, M. (2026). Temporal variation of outdoor particulate matter concentration in a village of the Qinghai-Tibet Plateau and its correlation with meteorological factors: based on long-term field measurement and machine learning approach. Air Quality, Atmosphere & Health, 19(10), Article 211. https://doi.org/10.1007/s11869-026-02094-2

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02094-2

Keywords: Qinghai-Tibet Plateau, yak dung stove, particulate matter, PM2.5, air quality, machine learning, Random Forest, meteorology, heating season, household air pollution, village air monitoring, XGBoost