Evaluating machine learning models for air quality error mapping in Kraków, Poland

Accurate air quality prediction is essential for sustainable urban development. This study evaluates the performance of machine learning models, including DLinear and XGBoost, in comparison with the traditional Autoregressive Integrated Moving Average (ARIMA) method for predicting fine particulate matter (PM_2.5) concentrations in Kraków, Poland. A dense network of low-cost sensors was used to generate high-resolution spatial and temporal data. Prediction errors were analysed using the Getis-Ord Gi* spatial statistics method during both extreme pollution events and low pollution days. The results indicate that DLinear achieved the lowest root mean square error (RMSE = 3.8 µg/m³), followed by XGBoost (RMSE = 6.7 µg/m³) and ARIMA (RMSE = 9.2 µg/m³). The spatial distribution of errors highlights the influence of environmental factors, such as humidity and proximity to water bodies, on model accuracy. These findings show the limitations of current prediction models and emphasize the need for spatially adaptive approaches to improve pollution.