Air pollution monitoring remains constrained by the sparse spatial distribution of ground-based monitoring stations and by the challenge of translating satellite-derived atmospheric observations into accurate estimates of near-surface pollutant concentrations. This thesis presents a scalable geoinformatics and environmental machine learning framework for the integration of heterogeneous atmospheric datasets, combining European Environment Agency (EEA) ground measurements, ERA5 meteorological reanalysis products, Sentinel-5P TROPOMI atmospheric observations, and Copernicus Atmosphere Monitoring Service (CAMS) variables. The proposed framework was developed at the European scale to support harmonized metadata reconstruction, pollutant-specific filtering, spatial co-location, temporal synchronization, and reproducible environmental data fusion. The predictive capabilities of the framework were subsequently evaluated through a localized case study in the Milan Metropolitan Area, focusing on daily surface-level SO2 estimation. Multiple modeling paradigms were investigated, including classical machine learning algorithms, ensemble learning approaches, AutoML forecasting systems, and Time Series Foundation Models (TSFMs), such as Chronos and TimesFM. The results demonstrate that forecasting based exclusively on historical SO2 observations is insufficient for accurate prediction, confirming that surface-level SO2 variability is strongly influenced by meteorological forcing, atmospheric transport processes, boundary-layer dynamics, and regional background atmospheric conditions. Classical machine learning performance improved progressively through enhanced feature engineering, chronological validation strategies, autoregressive lag structures, rolling-window statistics, and cyclical temporal encodings. The highest predictive performance was achieved by Chronos-2 using environmental covariates and Low-Rank Adaptation (LoRA) fine-tuning, highlighting the potential of pretrained temporal foundation models when adapted to atmospheric forecasting applications. The final covariate-aware forecasting results should be interpreted as a retrospective upper-bound performance scenario because observed meteorological and atmospheric variables were available throughout the evaluation period. Operational deployment would require the integration of forecasted covariates obtained from numerical weather prediction and atmospheric composition models. Overall, the findings demonstrate that scalable air-quality estimation benefits substantially from the integration of ground observations, Earth Observation products, meteorological reanalysis, physically informed feature engineering, and advanced temporal forecasting architectures.
Il monitoraggio dell’inquinamento atmosferico è ancora limitato dalla distribuzione spazialmente disomogenea delle stazioni di monitoraggio a terra e dalla difficoltà di trasformare le osservazioni satellitari dell’atmosfera in stime affidabili delle concentrazioni di inquinanti al livello del suolo. Questa tesi propone un framework scalabile di geoinformatica e machine learning ambientale per l’integrazione di dataset atmosferici eterogenei, combinando misure a terra dell’European Environment Agency (EEA), prodotti di rianalisi meteorologica ERA5, osservazioni satellitari Sentinel-5P TROPOMI e variabili del Copernicus Atmosphere Monitoring Service (CAMS). Il framework è stato sviluppato a scala europea per supportare la ricostruzione armonizzata dei metadati, il filtraggio degli inquinanti, la co-localizzazione spaziale, la sincronizzazione temporale e la fusione riproducibile di dati ambientali provenienti da fonti indipendenti. Le sue capacità predittive sono state successivamente valutate attraverso un caso di studio nell’Area Metropolitana di Milano, focalizzato sulla stima giornaliera delle concentrazioni superficiali di SO2. Sono stati confrontati diversi paradigmi di modellazione, tra cui algoritmi classici di machine learning, approcci ensemble, configurazioni AutoML e Time Series Foundation Models (TSFM), quali Chronos e TimesFM. I risultati dimostrano che la previsione basata esclusivamente sulla serie storica di SO2 non è sufficiente per descrivere accuratamente la variabilità delle concentrazioni al suolo. Le prestazioni ottenute confermano infatti che la dinamica della SO2 è fortemente influenzata dalle condizioni meteorologiche, dai processi di trasporto atmosferico, dalla variabilità dello strato limite planetario e dalle condizioni atmosferiche di fondo a scala regionale. Le prestazioni dei modelli classici di machine learning sono migliorate progressivamente grazie all’introduzione di tecniche avanzate di feature engineering, validazione cronologica, variabili lag, statistiche mobili e codifiche cicliche delle componenti temporali. Le migliori prestazioni predittive sono state ottenute dal modello Chronos-2 mediante l’integrazione di covariate ambientali e l’applicazione di strategie di fine-tuning basate su Low-Rank Adaptation (LoRA), evidenziando il potenziale dei modelli fondazionali temporali preaddestrati quando adattati a problemi di previsione della qualità dell’aria. I risultati finali ottenuti mediante l’utilizzo delle covariate ambientali devono tuttavia essere interpretati come uno scenario retrospettivo di limite superiore, poiché durante la fase di valutazione erano disponibili osservazioni meteorologiche e atmosferiche reali. Un’applicazione operativa del framework richiederebbe invece l’impiego di covariate previste da modelli numerici di previsione meteorologica e da sistemi di previsione della composizione atmosferica. Nel complesso, questa ricerca dimostra come la stima scalabile della qualità dell’aria possa beneficiare significativamente dell’integrazione tra osservazioni a terra, prodotti di Earth Observation, dati di rianalisi meteorologica, tecniche di feature engineering fisicamente motivate e avanzate architetture di modellazione temporale.
Geoinformatics-based environmental data fusion and machine learning for surface-level air pollution estimation: a european-scale framework with a Milan case study
SAUD MINO, CLAUDIA ISABELA
2025/2026
Abstract
Air pollution monitoring remains constrained by the sparse spatial distribution of ground-based monitoring stations and by the challenge of translating satellite-derived atmospheric observations into accurate estimates of near-surface pollutant concentrations. This thesis presents a scalable geoinformatics and environmental machine learning framework for the integration of heterogeneous atmospheric datasets, combining European Environment Agency (EEA) ground measurements, ERA5 meteorological reanalysis products, Sentinel-5P TROPOMI atmospheric observations, and Copernicus Atmosphere Monitoring Service (CAMS) variables. The proposed framework was developed at the European scale to support harmonized metadata reconstruction, pollutant-specific filtering, spatial co-location, temporal synchronization, and reproducible environmental data fusion. The predictive capabilities of the framework were subsequently evaluated through a localized case study in the Milan Metropolitan Area, focusing on daily surface-level SO2 estimation. Multiple modeling paradigms were investigated, including classical machine learning algorithms, ensemble learning approaches, AutoML forecasting systems, and Time Series Foundation Models (TSFMs), such as Chronos and TimesFM. The results demonstrate that forecasting based exclusively on historical SO2 observations is insufficient for accurate prediction, confirming that surface-level SO2 variability is strongly influenced by meteorological forcing, atmospheric transport processes, boundary-layer dynamics, and regional background atmospheric conditions. Classical machine learning performance improved progressively through enhanced feature engineering, chronological validation strategies, autoregressive lag structures, rolling-window statistics, and cyclical temporal encodings. The highest predictive performance was achieved by Chronos-2 using environmental covariates and Low-Rank Adaptation (LoRA) fine-tuning, highlighting the potential of pretrained temporal foundation models when adapted to atmospheric forecasting applications. The final covariate-aware forecasting results should be interpreted as a retrospective upper-bound performance scenario because observed meteorological and atmospheric variables were available throughout the evaluation period. Operational deployment would require the integration of forecasted covariates obtained from numerical weather prediction and atmospheric composition models. Overall, the findings demonstrate that scalable air-quality estimation benefits substantially from the integration of ground observations, Earth Observation products, meteorological reanalysis, physically informed feature engineering, and advanced temporal forecasting architectures.| File | Dimensione | Formato | |
|---|---|---|---|
|
Saud_Mino_Thesis.pdf
accessibile in internet per tutti
Descrizione: Thesis PDF
Dimensione
17.29 MB
Formato
Adobe PDF
|
17.29 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260711