Spatio-Temporal Predictive Modelling of PM2.5 in Karachi from Multi-Source Satellite, Meteorological and Ground Observations (2019–2023)

Karachi's ambient PM2.5 concentrations average approximately 12.3 times the World Health Organization 2021 annual guideline of 5 μg m⁻³ on the 2019–2023 dataset analysed in this study, placing it among the most polluted megacities in South Asia. Systematic, data-driven characterisation of the city's pollution sources and temporal structure remains limited in the peer-reviewed literature. We address this gap by constructing a five-year daily spatio-temporal dataset from eight ground monitoring stations fused with multi-source satellite retrievals (Sentinel-5P trace gases, MODIS MAIAC aerosol optical depth, ERA5 meteorological reanalysis, VIIRS nighttime lights) and a strict ground-truth provenance chain that combines OpenAQ per-station observations, a US Consulate reference sensor, and MERRA-2 citywide surface PM2.5. The final modelling set contains 14,400 station-day observations across 13 feature columns. We trained five traditional machine-learning models and a unidirectional causal LSTM with self-attention on a 2019–2022 training set, evaluated on a held-out 2023 test set. Random Forest achieved the best predictive performance (RMSE = 16.30 μg m⁻³, MAE = 10.63 μg m⁻³, R² = 0.612). The LSTM produced a negative R² of −0.104, interpreted as evidence of a fundamental data-density mismatch between complex sequence architectures and eight monitoring stations. SHAP analysis identified pm25_lag1 and Optical_Depth_055 as dominant predictors. Moran's I on RF residuals is −0.19 (p = 0.28, not significant), confirming no detectable spatial autocorrelation remains after the RF absorbs satellite and meteorological features. The full preprocessing pipeline, processed dataset, trained model weights, and an interactive 3D digital twin are released as an open-source repository reusable for other data-sparse South Asian cities. Code and data: https://github.com/SidhartSami/karachi-spatio-temporal-airq