Abstract:
This study quantifies the transboundary contribution of fine particulate matter (PM2.5) in Colombo
and Kandy, Sri Lanka (2021–2023), originating from other countries in the Indian subcontinent,
and evaluates a machine learning (ML) approach to predict these episodes. Transboundary
pollution days were identified using ground-based PM2.5 and meteorological data, particularly
when winds from 315◦–45◦ (NW–N–NE) transported pollutants from the continent. Temporal
patterns, meteorological influences, and pollutant persistence were examined. Observed data
revealed frequent pollution episodes associated with northeasterly winds, indicating long-range
transport from external sources. Random Forest regression models performed strongly, achieving
R2 = 0.91 (RMSE = 2.60 μg/m3) models performed strongly, achieving R2 = 0.91 (RMSE
= 2.60μg/m3 for Colombo. Classification models to detect high-pollution days (≥50 μg/m3)
achieved 100% accuracy and ROC AUC = 1.00 at both sites. These results demonstrate the high
predictability of pollution episodes when combining local meteorological factors with persistence
features. The findings underscore the importance of transboundary air pollution control
and highlight the potential of machine learning–based forecasting to strengthen early warning
and public health response in Sri Lanka. The strongest predictors were 7-day and 30-day PM2.5
averages and wind components, indicating that both historical pollution and regional airflows
strongly influence concentrations. However, the models struggled to detect extreme events due
to the limited number of high-pollution days (22 in Colombo and only one in Kandy). The
findings demonstrate that transboundary pollution substantially impacts PM2.5 in Colombo during
the Northeast Monsoon and inter-monsoon periods. This approach highlights the potential
of simple ML systems to forecast transboundary episodes, support targeted warnings, promote
regional cooperation, and integrate cross-border pollution signals into air-quality alert systems.
