Nowadays, climate emergencies attracted wide attention from the international community. Forest fires can be considered one of the most diffuse disasters affecting our planet due to their natural or human-related causes. In order to manage such critical events, an effective strategy must be defined through a deep analysis of the phenomenon. As known, Satellites’ data offer the exceptional opportunity to investigate these dangerous events’ consequences. For this purpose, it is presented here the development of a standalone tool for burnt area mapping after a fire event. After an extensive analysis of the relevant techniques and algorithms existing in the literature, a new methodology is proposed. For what concerns the involved data, Sentinel-1 and Sentinel-2 products have been selected by exploiting both multispectral and Synthetic Aperture Radar (SAR) sensors. Due to their spatial and temporal resolution and free availability, Sentinels’ data allowed to build a diversified fire events dataset. To transform those data into valuable information for burnt area mapping, a Machine Learning approach was adopted. Indeed, through the implementation of an Artificial Neural Network, it was possible to manage the vast amount of data to train models capable of providing a burnt area map. In particular, three models were defined to exploit optical, SAR, and optical-SAR fusion data. As expected, the results showed remarkable performances of the optical-based model while the SAR-based model has proved to be less efficient. Nevertheless, the last one can be considered a valuable alternative to the optical-based model since the SAR sensor can penetrate clouds and fire smoke to acquire the scene. Besides, the third model implies the improvement of the SAR-based training dataset due to the integration of optical-derived information. As expected, better accuracies were achieved if compared to the exclusive use of the SAR-based model. Eventually, the implemented tool was designed to make the user experience easy and intuitive. All the preprocessing steps, the proposed models, and the standalone tool were implemented, leveraging the Python programming language’s versatility.
Al giorno d'oggi, l'emergenza climatica ha attirato l'attenzione della comunità internazionale. Gli incendi boschivi, considerati tra i disastri ambientali più diffusi, colpiscono il nostro pianeta sia per cause naturali che antropologiche. Eventi di questo genere rendono quindi necessaria la definizione di una strategia di contrasto efficace, definita attraverso un'analisi accurata del fenomeno. Come è noto, i dati forniti dai Satelliti in orbita offrono l'incredibile opportunità di analizzare le conseguenze di questi pericolosi eventi. A questo proposito, viene qui presentato lo sviluppo di uno strumento "stand-alone" volto all'estrazione dell'area bruciata prodotta da un incendio boschivo. Dopo un'analisi approfondita delle tecniche e degli algoritmi relativi a questa tematica, viene inoltre proposta una nuova metodologia. In riferimento ai dati utilizzati, si è scelto di sfruttare il potenziale di sensori multispettrali e radar mediante l'impiego di prodotti forniti dalle missioni Sentinel-1 e Sentinel-2. La loro alta risoluzione sia spaziale che temporale e la loro facile reperibilità hanno reso possibile la costruzione di un dataset di eventi incendiari decisamente diversificato. Con riguardo alla trasformazione di questi dati in informazioni pertinenti all'estrazione delle aree incendiate, è stato adottato un algoritmo di Machine Learning. Pertanto, l'implementazione di una rete neurale artificiale ha reso possibile la costruzione di modelli capaci di stimare l'area incendiata conseguente all'incendio. In particolare, sono stati prodotti tre modelli, per processare rispettivamente i dati ottici, SAR e quelli ottenuti dalla fusione di entrambi. Mentre i risultati ottenuti dal modello basato su dati ottici sono ottimi, il modello basato sui dati SAR è parso meno performante, ma comunque abbastanza efficace da costituire una valida alternativa al primo. Inoltre, l'utilizzo sinergico dei due sensori consente di ottenere risultati migliori di quelli che si avrebbero col solo utilizzo dei dati SAR. Infine, il linguaggio di programmazione Pyhton ha permesso l'implementazione di un'applicazione autonoma che rende semplice ed intuitiva l'esperienza dell'utente.
Forest fires from space : burnt area mapping using multispectral and SAR sensors data and a machine learning approach
PAPALE, LORENZO GIULIANO
2019/2020
Abstract
Nowadays, climate emergencies attracted wide attention from the international community. Forest fires can be considered one of the most diffuse disasters affecting our planet due to their natural or human-related causes. In order to manage such critical events, an effective strategy must be defined through a deep analysis of the phenomenon. As known, Satellites’ data offer the exceptional opportunity to investigate these dangerous events’ consequences. For this purpose, it is presented here the development of a standalone tool for burnt area mapping after a fire event. After an extensive analysis of the relevant techniques and algorithms existing in the literature, a new methodology is proposed. For what concerns the involved data, Sentinel-1 and Sentinel-2 products have been selected by exploiting both multispectral and Synthetic Aperture Radar (SAR) sensors. Due to their spatial and temporal resolution and free availability, Sentinels’ data allowed to build a diversified fire events dataset. To transform those data into valuable information for burnt area mapping, a Machine Learning approach was adopted. Indeed, through the implementation of an Artificial Neural Network, it was possible to manage the vast amount of data to train models capable of providing a burnt area map. In particular, three models were defined to exploit optical, SAR, and optical-SAR fusion data. As expected, the results showed remarkable performances of the optical-based model while the SAR-based model has proved to be less efficient. Nevertheless, the last one can be considered a valuable alternative to the optical-based model since the SAR sensor can penetrate clouds and fire smoke to acquire the scene. Besides, the third model implies the improvement of the SAR-based training dataset due to the integration of optical-derived information. As expected, better accuracies were achieved if compared to the exclusive use of the SAR-based model. Eventually, the implemented tool was designed to make the user experience easy and intuitive. All the preprocessing steps, the proposed models, and the standalone tool were implemented, leveraging the Python programming language’s versatility.| File | Dimensione | Formato | |
|---|---|---|---|
|
2021_04_Papale.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
42.37 MB
Formato
Adobe PDF
|
42.37 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/10589/174072