Deforestation can cause multifarious and profound consequences to the environment and species such as soil degradation, loss of variability of species and global warming. Google Earth Engine (GEE) as an efficient analytical tool for land cover mapping has been introduced in this study. The central theme is to compute and monitor the forest change from 2000 to 2019 for our AOI located in Brazilian Amazon. Before starting the analysis in Amazon, a trial study was conducted in an area about 154.68 km2 in North Italy due to the windstorm happened from 27 to 30 October, in north-east Italy which caused enormous forest loss in the affected areas. The area size and evident land cover change make it suitable to be a test site of efficiency and accuracy of GEE. The pre- and post- land cover maps were created to verify forest and non-forest areas with data accessed from Landsat 8 and Sentinel-2 on GEE cloud-computing platform. The accuracies of the outcomes were evaluated by confusion matrices and kappa index and we have obtained overall accuracy more than 0.9 and kappa more 0.85 for all the maps. The forest loss generated by the event as 18.80 km2 and 25.03 km2 was computed by subtraction between the pre- and post-event classifications from Landsat data and Sentinel-2 data separately. Two different approaches have been proposed to resample the high-resolution image from 1m to 10m to make it comparable with the classification results. One is based on the ‘Prevalence’ rule and the other based on the ‘Percentage’ rule. When the performance of GEE as a tool was proved and the model was tuned for the case in Italy, the main case study which covered a large scale from both spatial and temporal aspects has been analysed. The purpose is to track the deforestation pattern in an area (about 49449.53 km2) in Brazilian Amazon located province Para rainforest from 2000 to 2019. The mapping was done at five-yearly intervals, with a focus on the dry season periods in the Amazon for less cloud coverage in image composites. The land cover transitions in these 20 years were mapped with Landsat 5, 7 and 8 data. The validation procedures were done manually with 533 validation points generated for each class with reference to a high-resolution panchromatic image obtained from CBERS (China-Brazil satellite programme) on the platform of CollectEarth. NDVI datasets were used to supplement the photo interpretation and to reduce the uncertainties during the validation phase. Overall accuracy and Cohen’s Kappa were computed to indicate the performance level of the classifications. The forest loss for about each five years was computed using the derived from GEE land cover maps. The obtained forest loss for each time interval is: 2000 - 2006: 5081.90 km2, 2006 - 2010: 1942.71 km2, 2010 - 2015: 1779.41 km2, 2015 - 2019: 2569.81 km2. Consequently, further discussion of the trends for forest loss and gain were obtained for each five-year timestep and they were compared with the data recorded by Instituto Nacional de Pesquisas Espaciais (INPE). The forest change prediction for the future 9 years from 2019 has been conducted with a plugin called MOLUSCE in QGIS in the end.
La deforestazione può causare molteplici e profonde conseguenze sull'ambiente e sulle specie come il degrado del suolo, la perdita di variabilità delle specie e il riscaldamento globale. Google Earth Engine (GEE) come strumento analitico efficiente per la mappatura della copertura del suolo è stato introdotto in questo studio. Il tema centrale è calcolare e monitorare il cambio di foresta dal 2000 al 2019 per il nostro AOI situato nell'Amazzonia brasiliana. Prima di iniziare l'analisi in Amazon, è stato condotto uno studio di prova in un'area di circa 154,68 km2 nel Nord Italia a causa della tempesta di vento avvenuta dal 27 al 30 ottobre, nel nord-est dell'Italia, causando un'enorme perdita di foreste nelle aree colpite. Le dimensioni dell'area e l'evidente cambiamento della copertura del suolo lo rendono adatto per essere un sito di test di efficienza e precisione di GEE. Le mappe di copertura pre e post-terra sono state create per verificare le aree forestali e non forestali con accesso ai dati da Landsat 8 e Sentinel-2 sulla piattaforma di cloud computing GEE. Le accuratezze dei risultati sono state valutate da matrici di confusione e indice kappa e abbiamo ottenuto una precisione complessiva superiore a 0,9 e kappa più 0,85 per tutte le mappe. La perdita di foresta generata dall'evento come 18,80 km2 e 25,03 km2 è stata calcolata per sottrazione tra le classificazioni pre e post evento dai dati Landsat e Sentinel-2 separatamente. Sono stati proposti due approcci diversi per ricampionare l'immagine ad alta risoluzione da 1m a 10m per renderla comparabile con i risultati della classificazione. Uno si basa sulla regola "Prevalenza" e l'altro sulla regola "Percentuale". Quando le prestazioni di GEE come strumento sono state dimostrate e il modello è stato messo a punto per il caso in Italia, è stato analizzato il caso di studio principale che ha coperto su larga scala sia gli aspetti spaziali che temporali. Lo scopo è quello di tracciare il modello di deforestazione in un'area (circa 49449,53 km2) nella foresta pluviale Para situata nella regione brasiliana dell'Amazzonia dal 2000 al 2019. La mappatura è stata effettuata a intervalli di cinque anni, con particolare attenzione ai periodi di stagione secca in Amazzonia per minore copertura del cloud nei compositi di immagini. Le transizioni di copertura del suolo in questi 20 anni sono state mappate con i dati Landsat 5, 7 e 8. Le procedure di validazione sono state eseguite manualmente con 533 punti di validazione generati per ogni classe con riferimento a un'immagine pancromatica ad alta risoluzione ottenuta da CBERS (programma satellitare Cina-Brasile) sulla piattaforma di CollectEarth. I set di dati NDVI sono stati utilizzati per integrare l'interpretazione della foto e per ridurre le incertezze durante la fase di validazione. La precisione generale e il Kappa di Cohen sono stati calcolati per indicare il livello di prestazione delle classificazioni. La perdita di foresta per circa ogni cinque anni è stata calcolata usando il derivato dalle mappe di copertura del suolo GEE. La perdita di foresta ottenuta per ogni intervallo di tempo è: 2000-2006: 5081,90 km2, 2006-2010: 1942,71 km2, 2010-2015: 1779,41 km2, 2015-2019: 2569,81 km2. Di conseguenza, sono state ottenute ulteriori discussioni sulle tendenze per la perdita e il guadagno delle foreste per ogni periodo di cinque anni e sono state confrontate con i dati registrati dall'Instituto Nacional de Pesquisas Espaciais (INPE). La previsione del cambio di foresta per i futuri 9 anni a partire dal 2019 è stata condotta con un plug-in chiamato MOLUSCE in QGIS alla fine.
Monitoring forest change using multi-temporal remote sensing classification on Google Earth engine
SUN, YARU
2019/2020
Abstract
Deforestation can cause multifarious and profound consequences to the environment and species such as soil degradation, loss of variability of species and global warming. Google Earth Engine (GEE) as an efficient analytical tool for land cover mapping has been introduced in this study. The central theme is to compute and monitor the forest change from 2000 to 2019 for our AOI located in Brazilian Amazon. Before starting the analysis in Amazon, a trial study was conducted in an area about 154.68 km2 in North Italy due to the windstorm happened from 27 to 30 October, in north-east Italy which caused enormous forest loss in the affected areas. The area size and evident land cover change make it suitable to be a test site of efficiency and accuracy of GEE. The pre- and post- land cover maps were created to verify forest and non-forest areas with data accessed from Landsat 8 and Sentinel-2 on GEE cloud-computing platform. The accuracies of the outcomes were evaluated by confusion matrices and kappa index and we have obtained overall accuracy more than 0.9 and kappa more 0.85 for all the maps. The forest loss generated by the event as 18.80 km2 and 25.03 km2 was computed by subtraction between the pre- and post-event classifications from Landsat data and Sentinel-2 data separately. Two different approaches have been proposed to resample the high-resolution image from 1m to 10m to make it comparable with the classification results. One is based on the ‘Prevalence’ rule and the other based on the ‘Percentage’ rule. When the performance of GEE as a tool was proved and the model was tuned for the case in Italy, the main case study which covered a large scale from both spatial and temporal aspects has been analysed. The purpose is to track the deforestation pattern in an area (about 49449.53 km2) in Brazilian Amazon located province Para rainforest from 2000 to 2019. The mapping was done at five-yearly intervals, with a focus on the dry season periods in the Amazon for less cloud coverage in image composites. The land cover transitions in these 20 years were mapped with Landsat 5, 7 and 8 data. The validation procedures were done manually with 533 validation points generated for each class with reference to a high-resolution panchromatic image obtained from CBERS (China-Brazil satellite programme) on the platform of CollectEarth. NDVI datasets were used to supplement the photo interpretation and to reduce the uncertainties during the validation phase. Overall accuracy and Cohen’s Kappa were computed to indicate the performance level of the classifications. The forest loss for about each five years was computed using the derived from GEE land cover maps. The obtained forest loss for each time interval is: 2000 - 2006: 5081.90 km2, 2006 - 2010: 1942.71 km2, 2010 - 2015: 1779.41 km2, 2015 - 2019: 2569.81 km2. Consequently, further discussion of the trends for forest loss and gain were obtained for each five-year timestep and they were compared with the data recorded by Instituto Nacional de Pesquisas Espaciais (INPE). The forest change prediction for the future 9 years from 2019 has been conducted with a plugin called MOLUSCE in QGIS in the end.| File | Dimensione | Formato | |
|---|---|---|---|
|
Sun Yaru Thesis Final.pdf
accessibile in internet per tutti
Dimensione
81.43 MB
Formato
Adobe PDF
|
81.43 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/10589/153687