A Hybrid Geostatistical and Deep Learning Framework for Urban Pollutant Concentration Prediction from Sparse Data

Urban air quality management faces substantial challenges in rapidly growing cities where monitoring infrastructure remains severely limited. This paper presents a novel hybrid framework that integrates advanced geostatistical spatial interpolation techniques with state-of-the-art deep learning temporal forecasting models to predict PM2.5and PM10concentrations from sparse sensor networks. Addressing the critical problem of limited monitoring stations in densely populated metropolitan cities like Kolkata, India, where a fixed number of stationary sensors cover the entire metropolitan area, the proposed approach evaluates statistical models such as Universal Kriging, Inverse Distance Weighting (IDW), Radial Basis Function, Thiessen Polygon, and Kernel Density Estimates for spatial estimation in combination with deep learning models including N-BEATS, LSTM, GRU, and RNN for temporal prediction. Among these, the IDW technique demonstrates superior accuracy for spatial interpolation, while the N-BEATS model exhibits the most reliable temporal forecasting capability. The integration of these two methods within a unified spatio-temporal framework enables precise and continuous air quality estimation across unmonitored urban locations. This hybrid approach proves highly advantageous for pollution prediction in areas with sparse sensor networks, supporting efficient environmental management through minimal hardware deployment. Furthermore, the study highlights the potential for deploying such AI/ML systems in edge computing environments, demonstrating their scalability, adaptability, and viability for real-world implementation under resource-constrained conditions.