Abstract
Air pollution in Brazilian agricultural–urban interfaces is strongly driven by land-use change and seasonal biomass burning. In southeastern Brazil’s Atlantic Forest biome, these processes interact with local meteorology to produce complex air-quality dynamics. This study provides one of the first long-term (2017–2025), observation-based assessments of air pollution variability and machine-learning diagnostic prediction of particulate matter in an agricultural–urban interface of the Atlantic Forest biome affected by recurrent biomass burning. In situ measurements of carbon monoxide (CO), nitrogen oxides (NOx), ozone (O₃), fine particulate matter 2.5 μm (PM2.5) and coarse particulate matter 10 μm
(PM10) from the local air-quality monitoring station revealed pronounced seasonality, with particulate concentrations peaking during the winter (dry season) under intense solar radiation, low humidity, strong biomass burning, and weak winds that favor pollutant accumulation. Exceedances of World Health Organization (WHO) air-quality guidelines occurred on more than 30% of monitored days for PM2.5, with 2024 being the most polluted year. Multiple statistical and machine learning models, including multiple linear regression, generalized additive models (GAM), random forest (RF), extreme gradient boosting (XGBoost), deep neural network (DNN), and long-short term memory (LSTM), were evaluated for same-day particulate matter (PM) prediction. Multiple linear regression, GAM, and tree-based models captured the mean behavior of PM2.5 and PM10 but systematically underestimated high-concentration regimes, while the LSTM model did not improve the representation of daily PM variability. Incorporating feature-engineered temporal predictors improved performance, particularly for XGBoost, which identified CO, NOx, relative humidity, and dew-point temperature as dominant predictors. The feature-engineered DNN, obtained with time-lagged features, achieved the best results (PM2.5 and PM10: R2 = 0.96; and MAE = 2.02 and 5.17 μg m⁻³, respectively), better reproducing both low- and high-concentration regimes. A simple persistence baseline was used as a reference benchmark. RF, XGBoost, and DNN consistently outperformed it, whereas the LSTM did not outperform the persistence baseline, even with feature engineering, highlighting the limitations of recurrent architectures for this dataset and modeling configuration. These results show that particulate matter variability in this agricultural-urban interface is controlled by the combined influence of combustion tracers, atmospheric moisture, and temporal persistence. They also highlight the value of combining long-term observations with diagnostic machine-learning benchmarks to support PM estimation, gap filling, and the future development of site-specific air-quality assessment frameworks in biomass-burning regions.