In urban analytical studies, verifying the adaptability of analytical methodologies is essential to ensure their applicability across varying urban environments. This thesis systematically examines well-established analytical techniques, evaluating their potential for generalization within distinct urban contexts, with a focused investigation on the specific conditions presented by Milan’s urban landscape. The study provides a critical assessment of these methodologies, identifying limitations and enhancing our understanding of Milan’s unique environmental attributes. The research objectives encompass the adaptation and validation of machine learning methodologies within Milan’s air quality context, alongside an exhaustive analysis of feature importance to determine the relative impacts of meteorological and traffic variables on pollution concentrations. The analytical approach, executed via the FastAI framework integrated with the TabNet architecture, refines established methods by employing open-source data meticulously curated for this investigation. A dual modeling approach is adopted, including both multi-target and single-target frameworks, aimed at maximizing predictive performance. The findings reveal that the single-target configuration achieves superior precision in predicting particulate concentrations relative to the multi-target counterpart. Results underscore the prominence of meteorological variables—such as humidity, dew point, temperature, and atmospheric pressure—as primary contributors to air quality predictions, with traffic-related metrics showing a lesser impact. These insights highlight the necessity of high-resolution meteorological data integration with traffic metrics for enhancing urban environmental monitoring systems. While challenges such as spatial limitations and seasonal variability of sensor networks persist, this study establishes foundational work for expanding urban air quality models to various environmental contexts.
Nell'ambito degli studi di analisi dei contesti urbani, la verifica dell'adattabilità delle metodologie analitiche è fondamentale per garantirne l'applicabilità in contesti urbani diversi. Questa tesi esamina sistematicamente le tecniche analitiche consolidate, valutandone il potenziale di generalizzazione in contesti urbani distinti, con un'indagine specifica sulle condizioni proprie del paesaggio urbano di Milano. Lo studio offre una valutazione critica di queste metodologie, identificandone le limitazioni e approfondendo la comprensione delle caratteristiche ambientali uniche di Milano. Gli obiettivi della ricerca includono l'adattamento e la validazione di metodologie di machine learning nel contesto della qualità dell'aria di Milano, insieme a un'analisi esaustiva dell'importanza delle feature per determinare l'impatto relativo delle variabili meteorologiche e del traffico sulle concentrazioni di inquinanti. L'approccio analitico, realizzato tramite il framework FastAI integrato con l'architettura TabNet, affina i metodi consolidati utilizzando dati open-source accuratamente selezionati per questa indagine. È stato adottato un approccio a doppia modellazione, che comprende framework multi-target e single-target, mirato a massimizzare le prestazioni predittive. I risultati rivelano che la configurazione single-target raggiunge una precisione superiore nella previsione delle concentrazioni di particolato rispetto al modello multi-target. I risultati evidenziano la rilevanza delle variabili meteorologiche—come umidità, punto di rugiada, temperatura e pressione atmosferica—come principali contributori nelle previsioni della qualità dell'aria, mentre le metriche legate al traffico mostrano un impatto minore. Questi risultati sottolineano la necessità di integrare dati meteorologici ad alta risoluzione con metriche del traffico per migliorare i sistemi di monitoraggio ambientale urbano. Sebbene permangano sfide come le limitazioni spaziali e la variabilità stagionale delle reti di sensori, questo studio getta le basi per l'espansione dei modelli di qualità dell'aria urbana a vari contesti ambientali.
Leveraging deep learning and feature importance methods to study the influence of traffic and weather on air quality in Milan
GIOVIA, GIUSEPPE
2023/2024
Abstract
In urban analytical studies, verifying the adaptability of analytical methodologies is essential to ensure their applicability across varying urban environments. This thesis systematically examines well-established analytical techniques, evaluating their potential for generalization within distinct urban contexts, with a focused investigation on the specific conditions presented by Milan’s urban landscape. The study provides a critical assessment of these methodologies, identifying limitations and enhancing our understanding of Milan’s unique environmental attributes. The research objectives encompass the adaptation and validation of machine learning methodologies within Milan’s air quality context, alongside an exhaustive analysis of feature importance to determine the relative impacts of meteorological and traffic variables on pollution concentrations. The analytical approach, executed via the FastAI framework integrated with the TabNet architecture, refines established methods by employing open-source data meticulously curated for this investigation. A dual modeling approach is adopted, including both multi-target and single-target frameworks, aimed at maximizing predictive performance. The findings reveal that the single-target configuration achieves superior precision in predicting particulate concentrations relative to the multi-target counterpart. Results underscore the prominence of meteorological variables—such as humidity, dew point, temperature, and atmospheric pressure—as primary contributors to air quality predictions, with traffic-related metrics showing a lesser impact. These insights highlight the necessity of high-resolution meteorological data integration with traffic metrics for enhancing urban environmental monitoring systems. While challenges such as spatial limitations and seasonal variability of sensor networks persist, this study establishes foundational work for expanding urban air quality models to various environmental contexts.| File | Dimensione | Formato | |
|---|---|---|---|
|
2024_12_Giovia_Executive Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
658.25 kB
Formato
Adobe PDF
|
658.25 kB | Adobe PDF | Visualizza/Apri |
|
2024_12_Giovia_Tesi.pdf
accessibile in internet per tutti
Descrizione: Tesi
Dimensione
8.46 MB
Formato
Adobe PDF
|
8.46 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/10589/231009