In the context of increasing climate-related risks and urbanisation, decisionmakers require robust, scalable, and interpretable tools to support urban resilience and sustainable planning. Environmental hazards such as air pollution, urban heat, and flooding pose serious threats to public health and infrastructure, particularly in urban centres. To address this need, a general machine learning (ML) framework was developed to model and map the susceptibility and intensity of urban environmental hazards. The framework is based on an extensive literature review and adopts a holistic approach, encompassing data collection from multiple sources, preprocessing and harmonisation, ML modelling, interpretation, and practical applications. The proposed framework is showcased by modelling three key environmental hazards in the city of Milan, Italy: air pollution, Urban Heat Island (UHI), and floods. A robust technical workflow was developed for each hazard, including the integration of diverse data sources (e.g., in-situ stations, satellite imagery, and raster datasets), implementation of state-of-the-art ML models such as Artificial Neural Networks (ANN), XGBoost, Random Forest (RF), and Support Vector Machines (SVM), and the production of susceptibility and intensity maps. Custom software tools were also developed to support the modelling of multiple hazards based on multiple data sources and the testing of multiple models. Air pollution susceptibility was modelled as the likelihood of surpassing a pollution concentration limit based on governmental directives. The specific modelled pollutants were PM2.5, PM10, SO2, O3, and NO2, based on pollutant concentration measurements from a network of in-situ stations, maintained by ARPA Lombardia. Conditioning factors included land cover, digital terrain model, distance to water, distance to main roads, and meteorological variables (e.g., temperature, humidity, radiation, precipitation, and wind). Monthly susceptibility maps were generated, based on the average meteorological conditions considering the seasonal behaviour of the multiple pollutants’ concentration and sparsity of the target data. Additionally, the results were interpreted using Explainable Artificial Intelligence (XAI) techniques. UHI susceptibility was modelled using Land Surface Temperature (LST) derived from Landsat-8, referred to as Surface UHI (SUHI). SUHI was modelled as the likelihood of a location surpassing a reference LST, defined as the average temperature in forest and agricultural areas. SUHI was modelled during summer using an ANN, conditioned on LST, land cover, building height, population density, and derived spectral indices. The SUHI susceptibility map was produced based on average conditions from recent years. Furthermore, SUHI intensity was also modelled, defined as the LST difference between a location and the reference LST. The same conditioning factors were used, excluding LST, to enable prediction of the expected LST difference independently of temporal LST fluctuations. Flood susceptibility was modelled using historical flood occurrences provided by the local authority, Regione Lombardia. These occurrences consisted of polygons delimiting the extent of floods in specific years. The flood conditioning factors included topographic (digital elevation model, slope, aspect), geological, and hydrological variables (distance to water, stream power index, topographic wetness index), as well as land cover and others. Multiple models were tested to evaluate their accuracy, including RF, ANN, SVM, and Logistic Regression. The flood susceptibility map was produced using the best-performing model, RF. Additionally, a further application of the models was explored. In the context of Nature-based Solutions (NbS) and their use for the potential mitigation of environmental hazards, multiple studies have measured the cooling effects of urban vegetation, e.g., green roofs and parks. Therefore, the SUHI intensity model was leveraged to measure the cooling effect of simulated land cover changes introducing urban vegetation, i.e., simulating a vegetation pixel by modifying the land cover class and related spectral indices. Two cases were tested to showcase the model’s usefulness for urban planning: (i) the addition of green roofs on residential buildings in a neighbourhood, and (ii) the expansion of existing urban vegetation cores. These scenarios resulted in an average cooling effect of 5 K and 1 K, respectively. In conclusion, the proposed ML framework provides a robust and scalable approach to modelling and mapping urban environmental hazards, integrating diverse data sources, advanced ML techniques, and XAI for enhanced interpretability. By successfully applying this framework to air pollution, SUHI, and flood susceptibility in Milan, it demonstrates its adaptability to different hazard types and its capacity to generate actionable insights. Furthermore, the extension of the framework to assess the impact of NbS highlights its potential for supporting sustainable urban planning and climate resilience strategies. Future research could expand its applicability to other cities and hazards, refine predictive capabilities through additional data sources, and further explore NbS interventions for hazard mitigation.
Nel contesto dell’aumento dei rischi legati al cambiamento climatico e della crescent urbanizzazione, i decisori politici e i pianificatori urbani necessitano di strumenti solidi, scalabili e interpretabili per supportare la resilienza urbana e la pianificazione sostenibile. I pericoli ambientali, come l’inquinamento atmosferico, le isole di calore urbane e le inondazioni, rappresentano gravi minacce per la salute pubblica e le infrastrutture, in particolare nei centri urbani. Per rispondere a questa esigenza, è stato sviluppato un framework generale di apprendimento automatico (Machine Learning, ML) per modellare e mappare la suscettibilità e l’intensità dei pericoli ambientali urbani. Il framework, basato su un’ampia revisione della letteratura, adotta un approccio olistico che comprende la raccolta di dati da fonti eterogenee, la loro pre-elaborazione e armonizzazione, la modellazione ML, l’interpretazione e le applicazioni pratiche. Il framework è stato applicato nella città di Milano, Italia, per modellare tre principali pericoli ambientali: l’inquinamento atmosferico, l’isola di calore urbana (Urban Heat Island, UHI) e le inondazioni. Per ciascun pericolo è stato sviluppato un flusso tecnico robusto, che comprende l’integrazione di diverse fonti di dati (ad esempio stazioni in-situ, immagini satellitari e dati raster), l’implementazione di modelli ML all’avanguardia come le Reti Neurali Artificiali (ANN), XGBoost, Random Forest (RF) e Support Vector Machine (SVM), e la produzione di mappe di suscettibilità e intensità. Sono stati inoltre sviluppati strumenti software su misura per supportare la modellazione di più pericoli basata su molteplici fonti di dati e su modelli differenti. La suscettibilità all’inquinamento atmosferico è stata modellata come la probabilità di superare i limiti di concentrazione previsti dalle normative vigenti. I principali inquinanti considerati sono stati PM2.5, PM10, SO2, O3 e NO2, utilizzando misurazioni provenienti da una rete di stazioni in-situ gestite da ARPA Lombardia. I fattori condizionanti includevano l’uso del suolo, il modello digitale del terreno, la distanza da corpi idrici e strade principali, e variabili meteorologiche (es. temperatura, umidità, radiazione, precipitazioni e vento). Le mappe di suscettibilità sono state generate mensilmente, considerando le condizioni meteorologiche medie e la stagionalità delle concentrazioni degli inquinanti, nonché la scarsità dei dati obiettivo. I risultati sono stati interpretati utilizzando tecniche di Intelligenza Artificiale Spiegabile (XAI). La suscettibilità all’UHI è stata modellata utilizzando la Temperatura Superficiale del Suolo (Land Surface Temperature, LST) derivata da Landsat-8, con riferimento alla componente superficiale dell’isola di calore urbana (Surface UHI, SUHI). La SUHI è stata modellata come la probabilità che una località superi una LST di riferimento, definita come la temperatura media nelle aree forestali e agricole. La modellazione, effettuata per il periodo estivo tramite una rete neurale artificiale (ANN), ha considerato come fattori LST, uso del suolo, altezza degli edifici, densità di popolazione e indici spettrali derivati. La mappa di suscettibilità è stata prodotta considerando le condizioni medie degli ultimi anni. È stata inoltre modellata l’intensità della SUHI, definita come la differenza tra la LST di una località e la LST di riferimento. Per questa modellazione sono stati utilizzati gli stessi fattori condizionanti, escludendo però la LST, al fine di consentire la previsione dell’intensità attesa in modo indipendente dalle fluttuazioni temporali della LST. La suscettibilità alle inondazioni è stata modellata utilizzando dati storici forniti da Regione Lombardia, rappresentati da poligoni che delimitano le aree inondate in anni specifici. I fattori condizionanti includevano variabili topografiche (modello digitale del terreno, pendenza, esposizione), geologiche e idrologiche (es. distanza dall’acqua, stream power index, topographic wetness index), oltre all’uso del suolo. Sono stati testati diversi modelli ML per valutarne l’accuratezza, tra cui RF, ANN, SVM e Regressione Logistica. La mappa finale di suscettibilità alle inondazioni è stata ottenuta con il modello RF, risultato il più performante. È stata inoltre esplorata un’ulteriore applicazione dei modelli sviluppati, nel contesto delle Soluzioni Basate sulla Natura (Nature-based Solutions, NbS) e del loro potenziale ruolo nella mitigazione dei pericoli ambientali. Diversi studi hanno dimostrato gli effetti di raffrescamento della vegetazione urbana, come tetti verdi e parchi. Il modello di intensità della SUHI è stato quindi utilizzato per valutare l’effetto di raffrescamento di modifiche simulate della copertura del suolo, introducendo vegetazione urbana, ovvero simulando un pixel vegetato mediante la modifica della classe di uso del suolo e degli indici spettrali associati. Due scenari sono stati testati per mostrare la potenziale utilità del modello nella pianificazione urbana: (i) l’aggiunta di tetti verdi negli edifici residenziali di un quartiere e (ii) l’espansione delle aree verdi esistenti. I due interventi hanno prodotto un effetto medio di raffrescamento pari a 5 K e 1 K rispettivamente. In conclusione, il framework ML proposto offre un approccio solido e scalabile per la modellazione e la mappatura dei pericoli ambientali urbani, integrando fonti di dati eterogenee, tecniche ML avanzate e strumenti XAI per una migliore interpretabilità. L’applicazione a Milano ne dimostra la flessibilità rispetto a differenti tipologie di pericoli e la capacità di generare risultati utili e concreti. Inoltre, l’estensione del framework alla valutazione delle NbS evidenzia il suo potenziale nel supportare strategie di pianificazione urbana sostenibile e resilienza climatica. Futuri sviluppi potranno estenderne l’utilizzo ad altre città e pericoli, migliorare le capacità predittive tramite nuovi dati e approfondire le potenzialità delle NbS per la mitigazione degli impatti ambientali.
GeoAI for resilient urban development: a machine learning framework for modelling and mapping environmental hazard susceptibility and simulation of nature-based solutions
PUGLIESE VILORIA, ANGELLY DE JESÚS
2024/2025
Abstract
In the context of increasing climate-related risks and urbanisation, decisionmakers require robust, scalable, and interpretable tools to support urban resilience and sustainable planning. Environmental hazards such as air pollution, urban heat, and flooding pose serious threats to public health and infrastructure, particularly in urban centres. To address this need, a general machine learning (ML) framework was developed to model and map the susceptibility and intensity of urban environmental hazards. The framework is based on an extensive literature review and adopts a holistic approach, encompassing data collection from multiple sources, preprocessing and harmonisation, ML modelling, interpretation, and practical applications. The proposed framework is showcased by modelling three key environmental hazards in the city of Milan, Italy: air pollution, Urban Heat Island (UHI), and floods. A robust technical workflow was developed for each hazard, including the integration of diverse data sources (e.g., in-situ stations, satellite imagery, and raster datasets), implementation of state-of-the-art ML models such as Artificial Neural Networks (ANN), XGBoost, Random Forest (RF), and Support Vector Machines (SVM), and the production of susceptibility and intensity maps. Custom software tools were also developed to support the modelling of multiple hazards based on multiple data sources and the testing of multiple models. Air pollution susceptibility was modelled as the likelihood of surpassing a pollution concentration limit based on governmental directives. The specific modelled pollutants were PM2.5, PM10, SO2, O3, and NO2, based on pollutant concentration measurements from a network of in-situ stations, maintained by ARPA Lombardia. Conditioning factors included land cover, digital terrain model, distance to water, distance to main roads, and meteorological variables (e.g., temperature, humidity, radiation, precipitation, and wind). Monthly susceptibility maps were generated, based on the average meteorological conditions considering the seasonal behaviour of the multiple pollutants’ concentration and sparsity of the target data. Additionally, the results were interpreted using Explainable Artificial Intelligence (XAI) techniques. UHI susceptibility was modelled using Land Surface Temperature (LST) derived from Landsat-8, referred to as Surface UHI (SUHI). SUHI was modelled as the likelihood of a location surpassing a reference LST, defined as the average temperature in forest and agricultural areas. SUHI was modelled during summer using an ANN, conditioned on LST, land cover, building height, population density, and derived spectral indices. The SUHI susceptibility map was produced based on average conditions from recent years. Furthermore, SUHI intensity was also modelled, defined as the LST difference between a location and the reference LST. The same conditioning factors were used, excluding LST, to enable prediction of the expected LST difference independently of temporal LST fluctuations. Flood susceptibility was modelled using historical flood occurrences provided by the local authority, Regione Lombardia. These occurrences consisted of polygons delimiting the extent of floods in specific years. The flood conditioning factors included topographic (digital elevation model, slope, aspect), geological, and hydrological variables (distance to water, stream power index, topographic wetness index), as well as land cover and others. Multiple models were tested to evaluate their accuracy, including RF, ANN, SVM, and Logistic Regression. The flood susceptibility map was produced using the best-performing model, RF. Additionally, a further application of the models was explored. In the context of Nature-based Solutions (NbS) and their use for the potential mitigation of environmental hazards, multiple studies have measured the cooling effects of urban vegetation, e.g., green roofs and parks. Therefore, the SUHI intensity model was leveraged to measure the cooling effect of simulated land cover changes introducing urban vegetation, i.e., simulating a vegetation pixel by modifying the land cover class and related spectral indices. Two cases were tested to showcase the model’s usefulness for urban planning: (i) the addition of green roofs on residential buildings in a neighbourhood, and (ii) the expansion of existing urban vegetation cores. These scenarios resulted in an average cooling effect of 5 K and 1 K, respectively. In conclusion, the proposed ML framework provides a robust and scalable approach to modelling and mapping urban environmental hazards, integrating diverse data sources, advanced ML techniques, and XAI for enhanced interpretability. By successfully applying this framework to air pollution, SUHI, and flood susceptibility in Milan, it demonstrates its adaptability to different hazard types and its capacity to generate actionable insights. Furthermore, the extension of the framework to assess the impact of NbS highlights its potential for supporting sustainable urban planning and climate resilience strategies. Future research could expand its applicability to other cities and hazards, refine predictive capabilities through additional data sources, and further explore NbS interventions for hazard mitigation.File | Dimensione | Formato | |
---|---|---|---|
Pugliese_GeoAI_hazard_susceptibiliy_mapping_revised.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Revised thesis
Dimensione
142.93 MB
Formato
Adobe PDF
|
142.93 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/10589/237717