This work focuses on the comparison of multi-source and multi-scale Burned Area (BA) products derived from Earth Observation (EO) data and, in particular, from satellite images provided by the Copernicus missions. The analyses and comparisons between different BA products have been carried out over Area Of Interest (AOIs) for the African continent and the year 2019. A supervised algorithm for Burned Area (BA) classification based on Random Forest (RF) was implemented and applied to the Sentinel-2 data; the aim is to map burned areas from the time series of each AOIs. The result of this operation, carried out in the Google Earth Engine working environment, are layers containing the perimeters of the burned areas (BA) over the S2 tiles (100 km x 100 km) overlapping the AOIs. Then, BA products are validated in QGIS: BA maps from Sentinel-2, are compared with the burned area perimeters derived from Planet data (which have a better resolution and are considered as ground truth). Omission and commission error metrics and the Dice coefficient were calculated. Error sources were investigated and analysed. Furthermore, an analysis of the effect of the resolution of the input images on the accuracy of BA maps was carried out through the analysis of the Pareto Boundary that was carried out in R environment. Sentinel-2 fire perimeters were also compared with BA maps from the FireCCI51 products, to complete the multiscale comparison. The FireCCI51 has lower spatial resolution (250m) but it delivers multi annual global BA maps. Lower accuracy mapping is therefore expected compared to higher resolution S2 fire perimeters. Results show that a satisfactory accuracy is achieved at each site: Sentinel-2 BA maps have omission error in the range [1,60-20,97 %] and commission error in the range [5,93-19,85 %]. Greater accuracy is obtained over larger burned areas patches and with regular boundaries; in these cases, classification errors due to different resolution of the input image data are less influential. The supervised RF algorithm however provides more accurate than an automatic algorithm since it is trained over each site.
Il lavoro di questa tesi si concentra sul confronto di prodotti multi-sorgente e multi-scala di aree bruciate (BA) derivati da dati di Osservazione della Terra (EO) e, in particolare, da immagini satellitari fornite dalle missioni Copernicus. Le analisi e i confronti tra i diversi prodotti BA sono stati effettuati su aree di interesse (Areas Of Interest, AOIs) in Africa e per l'anno 2019. Un algoritmo supervisionato per la classificazione delle aree bruciate (BA) basato su Random Forest (RF) è stato implementato e applicato ai dati Sentinel-2 per mappare le aree bruciate delle serie temporali di ogni AOIs. Il risultato della classificazione, effettuata nell’ambiente di lavoro Google Earth Engine, sono i perimetri delle aree bruciate (BA) sui tile (100kmx100km) S2 che si sovrappongono alle AOI. Successivamente, i prodotti BA sono validati in QGIS: I risultati della classificazione di Sentinel-2 sono stati confrontati con i perimetri delle aree bruciate derivati dai dati Planet (che hanno una migliore risoluzione e sono considerati come verità a terra). Sono state calcolate le metriche degli errori di omissione e commissione e il coefficiente di Dice, studiando e analizzando le fonti di errore. Inoltre, è stata effettuata un'analisi dell'effetto della risoluzione delle 'immagini di input sull'accuratezza delle mappe BA attraverso l’analisi ed il calcolo della frontiera di Pareto (Pareto Boundary) in ambiente R Studio. I perimetri degli incendi Sentinel-2 sono poi stati confrontati con le mappe BA dei prodotti FireCCI51, per completare il confronto multi-scala. Il dato FireCCI51 fornisce mappe BA globali e pluriennali (2001-2020) con risoluzione di 250 m. Ci si aspetta, quindi, una mappatura di minore accuratezza rispetto ai perimetri di incendio S2 che ha una più alta risoluzione (10/20 m). I risultati mostrano che è stata raggiunta un’accuratezza sufficiente in ogni sito: le mappe BA ottenute da Sentinel-2 hanno errori di omissione e commissione, rispettivamente, nel range [1,60-20,97 %] e [5,93-19,85 %]. L’accuratezza maggiore è stata osservata per aree bruciate più grandi e con confini regolari; in questi casi, l'errore di classificazione dovuto alla diversa risoluzione dei dati dell'immagine di input è meno influente. L'algoritmo supervisionato RF fornisce comunque un'accuratezza maggiore di un algoritmo automatico, poiché è stato addestrato su ogni sito.
Multi-scale analysis and comparison of Burned Area products derived from Earth Observation (EO) data over the African continent
PASTURENSI, MARTINA
2021/2022
Abstract
This work focuses on the comparison of multi-source and multi-scale Burned Area (BA) products derived from Earth Observation (EO) data and, in particular, from satellite images provided by the Copernicus missions. The analyses and comparisons between different BA products have been carried out over Area Of Interest (AOIs) for the African continent and the year 2019. A supervised algorithm for Burned Area (BA) classification based on Random Forest (RF) was implemented and applied to the Sentinel-2 data; the aim is to map burned areas from the time series of each AOIs. The result of this operation, carried out in the Google Earth Engine working environment, are layers containing the perimeters of the burned areas (BA) over the S2 tiles (100 km x 100 km) overlapping the AOIs. Then, BA products are validated in QGIS: BA maps from Sentinel-2, are compared with the burned area perimeters derived from Planet data (which have a better resolution and are considered as ground truth). Omission and commission error metrics and the Dice coefficient were calculated. Error sources were investigated and analysed. Furthermore, an analysis of the effect of the resolution of the input images on the accuracy of BA maps was carried out through the analysis of the Pareto Boundary that was carried out in R environment. Sentinel-2 fire perimeters were also compared with BA maps from the FireCCI51 products, to complete the multiscale comparison. The FireCCI51 has lower spatial resolution (250m) but it delivers multi annual global BA maps. Lower accuracy mapping is therefore expected compared to higher resolution S2 fire perimeters. Results show that a satisfactory accuracy is achieved at each site: Sentinel-2 BA maps have omission error in the range [1,60-20,97 %] and commission error in the range [5,93-19,85 %]. Greater accuracy is obtained over larger burned areas patches and with regular boundaries; in these cases, classification errors due to different resolution of the input image data are less influential. The supervised RF algorithm however provides more accurate than an automatic algorithm since it is trained over each site.| File | Dimensione | Formato | |
|---|---|---|---|
|
TESI_PASTURENSI_10502931.pdf
Open Access dal 01/04/2023
Dimensione
4.68 MB
Formato
Adobe PDF
|
4.68 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/10589/186141