<![CDATA[Fine particulate matter (PM2.5) pollution has become a serious public health and environmental concern in Indonesia, yet ground-based monitoring stations remain limited and unevenly distributed across the archipelago, restricting comprehensive spatial assessment of pollutant concentrations. This study aimed to develop and comparatively evaluate three deep learning architectures, namely Gated Recurrent Unit (GRU), GRU with Additive Attention (GRU-Attention), and Bidirectional GRU with Additive Attention (BiGRU-Attention), for predicting daily PM2.5 concentrations in Indonesia by integrating Sentinel-5P satellite atmospheric chemistry products and ERA5-Land meteorological reanalysis data. The dataset combined PM2.5 ground-truth measurements from 26 BMKG monitoring stations covering the period 2020–2025, five Sentinel-5P pollutant variables and four ERA5-Land meteorological variables, producing 35,842 cleaned observations and seventeen engineered features. All variables were spatially and temporally aligned, normalized using RobustScaler with log1p target transformation, and reshaped into seven-day sequences using a stratified-station train, validation, and test split with proportions of 70%, 15%, and 15%. The results showed that BiGRU-Attention achieved the best performance with R² of 0.8302, RMSE of 7.4720 µg/m³, MAE of 4.8368 µg/m³, and MAPE of 26.7009%, outperforming GRU-Attention and the baseline GRU. This MAPE is higher than typical single-city PM2.5 models but is consistent with national-scale studies, where low-concentration observations inflate percentage-based errors. The best model was subsequently applied to produce a national daily PM2.5 distribution map, which can help identify regional PM2.5 hotspots, prioritize locations for additional monitoring infrastructure, and inform targeted air quality interventions in regions where ground-based coverage remains sparse.]]>
