Asbestos contamination represents a significant health and environmental hazard, raising the concern for accurate asbestos detection and mapping towards its effective mitigation. However, conventional methodologies for asbestos detection, including field inspections and laboratory analyses, are typically labor-intensive, expensive, and frequently constrained by limited spatial coverage. Remote sensing technologies provide cost-effective and scalable alternatives to these practices, facilitating geographic coverage extension and enabling continuous monitoring over time. This thesis explores an automated approach to asbestos detection using deep learning and multispectral satellite imagery. High-resolution images from WorldView-3 and Pléiades Neo satellites are pansharpened and enhanced to create a dataset of complex urban environments. Three convolutional neural networks — EfficientNetB0, ResNet50, and InceptionNetV3 — are adapted to process multispectral inputs, utilizing transfer learning from ImageNet. The findings highlight enhancements in binary classification metrics when employing multispectral configurations as opposed to standard RGB inputs, thus underscoring the importance of integrating supplementary spectral bands to enhance the effective discrimination of asbestos containing materials. Moreover, this thesis performs a qualitative evaluation approach based on exploiting the insights provided by Gradient Class Activation Maps as a powerful tool to understand where networks focus to output predictions. This approach enables both prediction validation and material localization through Weakly Supervised Learning techniques.
La contaminazione da amianto rappresenta un grave rischio per la salute e l’ambiente, sollevando la necessità di sviluppare metodi accurati per il rilevamento e la mappatura dell’amianto, al fine di promuoverne una mitigazione efficace. Tuttavia, le metodologie convenzionali per il rilevamento dell’amianto, come le ispezioni in campo e le analisi di laboratorio, risultano generalmente dispendiose in termini di tempo, costose e spesso limitate a una copertura spaziale ristretta. Le tecnologie di remote sensing offrono alternative scalabili e più economiche a queste pratiche, permettendo un’estensione della copertura geografica e il monitoraggio continuo nel tempo. Questa tesi esplora un approccio automatizzato per il rilevamento dell’amianto utilizzando deep learning e immagini satellitari multispettrali. Immagini ad alta risoluzione provenienti dai satelliti WorldView-3 e Pléiades Neo sono sottoposte a pansharpening e miglioramenti per creare un data set relativo a contesti urbani complessi. Tre reti neurali convoluzionali — EfficientNetB0, ResNet50 e InceptionNetV3 — sono state adattate per processare input multispettrali, sfruttando il transfer learning da ImageNet. I risultati evidenziano miglioramenti nei metriche di classificazione binaria quando si utilizzano configurazioni multispettrali rispetto agli input RGB standard, sottolineando l’importanza dell’integrazione di bande spettrali aggiuntive per migliorare la discriminazione efficace dei materiali contenenti amianto. Inoltre, questa tesi adotta un approccio qualitativo basato sull’analisi delle Gradient Class Activation Maps, uno strumento potente per comprendere dove le reti focalizzano l’attenzione per generare previsioni. Questo approccio consente sia la validazione delle previsioni che la localizzazione dei materiali attraverso tecniche di Weakly Supervised Learning.
Deep learning approaches for asbestos classification via multispectral satellite images
MAZZOLA, DARIO
2023/2024
Abstract
Asbestos contamination represents a significant health and environmental hazard, raising the concern for accurate asbestos detection and mapping towards its effective mitigation. However, conventional methodologies for asbestos detection, including field inspections and laboratory analyses, are typically labor-intensive, expensive, and frequently constrained by limited spatial coverage. Remote sensing technologies provide cost-effective and scalable alternatives to these practices, facilitating geographic coverage extension and enabling continuous monitoring over time. This thesis explores an automated approach to asbestos detection using deep learning and multispectral satellite imagery. High-resolution images from WorldView-3 and Pléiades Neo satellites are pansharpened and enhanced to create a dataset of complex urban environments. Three convolutional neural networks — EfficientNetB0, ResNet50, and InceptionNetV3 — are adapted to process multispectral inputs, utilizing transfer learning from ImageNet. The findings highlight enhancements in binary classification metrics when employing multispectral configurations as opposed to standard RGB inputs, thus underscoring the importance of integrating supplementary spectral bands to enhance the effective discrimination of asbestos containing materials. Moreover, this thesis performs a qualitative evaluation approach based on exploiting the insights provided by Gradient Class Activation Maps as a powerful tool to understand where networks focus to output predictions. This approach enables both prediction validation and material localization through Weakly Supervised Learning techniques.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Mazzola_Thesis_01.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
12.95 MB
Formato
Adobe PDF
|
12.95 MB | Adobe PDF | Visualizza/Apri |
2024_12_Mazzola_Executive_Summary_02.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
3.41 MB
Formato
Adobe PDF
|
3.41 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/10589/230433