Data‑Driven Inverse Modelling of Urban Pollutant Emissions Using a ConvLSTM3D Spatiotemporal Network

Accurate quantification of atmospheric pollutant emissions is important for interpreting urban air quality, evaluating mitigation strategies and informing evidence-based decision making for public health protection. Inverse modelling methodologies can provide spatially resolved emission estimates that complement traditional bottom-up inventories and support the analysis and forecasting of urban air pollution episodes, using data from observed concentrations and meteorology. This is particularly relevant in complex urban environments, where heterogeneous sources and meteorological conditions introduce substantial uncertainties in conventional inventories.

This study explores a ConvLSTM3D-based inverse modelling framework that learns spatiotemporal relationships between pollutant concentration fields, meteorological variables and gridded emission fields over an urban domain (Thessaloniki, Greece). Using a preliminary three-month dataset (October–December 2019) of NO₂ and CO concentrations together with wind components and temperature, the convolutional recurrent architecture infers emission patterns directly from sequences of concentration and meteorological inputs, with ongoing work expanding both the feature set, to include turbulent diffusion, humidity, precipitation, pressure and solar radiation, and the temporal record.

Training with early stopping achieved a best validation loss of 0.100 at epoch 22 and a final test loss of 0.115, indicating stable convergence and strong generalization to unseen days without evident overfitting. Predicted emission maps closely reproduce true hotspot structure and spatial gradients for both pollutants, while representative time series capture diurnal peaks and troughs with good phase agreement, quantitatively supporting the qualitative match observed. Unlike traditional bottom-up inventories, which rely on static activity data and coarse spatial disaggregation, and unlike simpler forward-modelling or shallow machine-learning approaches, the proposed framework couples convolutional and recurrent learning to resolve fine-grained spatial structure and temporal dynamics jointly, directly from observations. These results position enriched ConvLSTM3D-based inverse modelling as a promising, scalable alternative for high-resolution urban air-quality applications, warranting systematic benchmarking against alternative architectures in future work.