A novel methodology for Explainable Artificial Intelligence integrated with geostatistics for air pollution control and environmental management
Mateusz Zaręba, Tomasz Danek
Abstract
This study focuses on understanding air pollution dynamics through the integration of Explainable Artificial Intelligence (XAI) and spatio-temporal geostatistics for air pollution control. It combines the classic approach with the simultaneous evaluation of multiple machine learning (ML) models to enhance the interpretability of model decisions. Data from a dense network of 52 IoT (internet of things) sensors deployed across an urban municipality and surrounding areas were analyzed to identify key predictors of air pollution trends. XAI is employed to interpret complex relationships between pollution levels, meteorological conditions, human activity patterns, and geographic factors. Spatio-temporal geostatistics serve as the foundation for interpreting XAI in both spatial and temporal contexts. July exhibited the least spatial variability in PM 2.5 concentrations, with the hour of the day being the key predictor (13.04% of predictive importance), while December showed the highest spatial variability, with atmospheric pressure as the dominant factor (13.84% of predictive importance). Precipitation was the least influential predictor in both months (3.00 %). Four and nine distinct clusters with significant spatial variability in predictor importance were identified. Analysis of transition matrices revealed both stable and dynamic clusters, highlighting the complex nature of PM 2.5 emissions as they vary between warm and cold seasons, characteristic of moderate climate zones. This methodology enhances the understanding of air pollution dynamics and provides a transparent framework for urban pollution control and management decision-making. The findings contribute to the development of smart urban management strategies, supporting sustainable city planning and pollution mitigation efforts, while advocating for the right of citizens to a cleaner environment. • Geostatistics and AI integration enhances air pollution control understanding. • XAI reveals significant seasonal shifts in PM 2.5 clusters and key pollution drivers. • Predictor importance heatmaps showed clear spatio-temporal dependencies. • XAI and geostatistical insights empower spatially targeted air pollution control. • GeoAI framework supports human-centered AI focused on environmental insights.