Multi-Horizon Air Quality Index Prediction incorporating Spatial Hotspot Analysis and Meteorological Drivers

Abstract

Urban air pollution poses significant environmental and public health challenges in rapidly urbanizing Indian cities. This study presents a GIS-integrated MLframework for multi-horizon (1-24 hours) forecasting of the Air Quality Index (AQI) in Nashik using multi-pollutant time-series data. Air quality observations from four Maharashtra Pollution Control Board (MPCB) monitoring stations at Gangapur Road, Hirawadi, MIDC Ambad and Pandav Nagari were analysed for representative seasonal months (January, April, July and October) of 2025. The dataset included PM2.5, PM10, NO, NO2, NOx, NH3, SO2, CO, wind speed and wind direction. Seasonal spatial distribution maps were generated using GIS software to examine pollutant variability across industrial, residential and traffic-dominated zones. Winter exhibited the highest particulate concentrations, with mean PM2.5 levels approaching 85-90 µg/m3 in industrial regions, while monsoon values declined to approximately 30-45 µg/m3 , reflecting nearly 45% seasonal reduction due to rainfall-induced washout. For temporal forecasting, Long Short-Term Memory (LSTM), XGBoost and RF models were implemented for 1-24 hour ahead AQI prediction. LSTM demonstrated superior long-horizon stability, achieving an R2 of 0.90 for 1-hour forecasts, maintaining R2 above 0.80 for 24-hour predictions, while XGBoost performed competitively for short-term horizons. Feature importance analysis identified PM2.5, NO2 and wind speed as dominant predictors influencing AQI dynamics. The integrated GIS-ML framework effectively captures spatial heterogeneity and temporal variability of urban air pollution and provides a scalable approach for real-time AQI early warning systems in medium-sized cities.