Enhancing temporal series of Sentinel-2 and Sentinel-3 data products: from classical regression to deep learning approach
Ver/ Abrir
Metadatos
Mostrar el registro completo del ítemcomunitat-uji-handle:10234/158176
comunitat-uji-handle2:10234/71345
comunitat-uji-handle3:10234/141145
comunitat-uji-handle4:
TFG-TFMMetadatos
Título
Enhancing temporal series of Sentinel-2 and Sentinel-3 data products: from classical regression to deep learning approachAutoría
Tutor/Supervisor; Universidad.Departamento
Pla, Filiberto; Fernández Beltrán, Rubén; Departament de Llenguatges i Sistemes InformàticsFecha de publicación
2021-03-05Editor
Universitat Jaume IResumen
The free and open availability of satellite images covering global extent in recent days provides many novel opportunities for global monitoring of the earth’s surface. Sentinel-2 (S2) and Sentinel-3 (S3) satellite ... [+]
The free and open availability of satellite images covering global extent in recent days provides many novel opportunities for global monitoring of the earth’s surface. Sentinel-2 (S2) and Sentinel-3 (S3) satellite missions capture mid to high resolution imagery with frequent revisit and show data synergy as they both focus on land and ocean observational needs. Specifically, the high temporal resolution of S3 (1-2 day revisit) presents potential in filling the data gaps in S2 (5 day revisit) vegetation products. In this scenario, this study assesses the feasibility of using Sentinel-3 images for Sentinel-2 vegetation products estimation using machine learning (ML) and deep learning (DL) approaches. This study employs four state of the art ML regression algorithms, linear regression, ridge regression, Support Vector Regression (SVR) and Random Forest Regression (RFR) and two DL network architectures with different depth and complexities, Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN) to predict the S2 NDVI and SAVI maps from the S3 spectral bands information. A paired S2/S3 dataset is prepared for the study area covering one S2 tile in Extremadura, Spain. The results demonstrate that all the DL architectures except pixel-wise MLP outperformed the ML models with the 3D CNN performing the best. The best performing 3D CNN architecture obtained remarkable mean squared error (MSE) of 0.00198 for NDVI and 0.00282 for SAVI while the best performing ML algorithms were patch-wise RFR with MSE of 0.0035 in case of NDVI and patchwise SVR with MSE of 0.00586 for SAVI. The models and the dataset prepared for this study will be useful for further research that focus on capitalizing the free and open availability of Sentinel-2 and Sentinel-3 imagery as well as new and advanced technologies to provide better vegetation monitoring capabilities for our planet. [-]
Palabras clave / Materias
Màster Universitari Erasmus Mundus en Tecnologia Geoespacial | Erasmus Mundus University Master's Degree in Geospatial Technologies | Máster Universitario Erasmus Mundus en Tecnología Geoespacial | Sentinel-2 | Sentinel-3 | inter-sensor data products estimation | machine learning | deep learning | convolutional neural networks
Descripción
Treball de Final de Màster Universitari Erasmus Mundus en Tecnologia Geoespacial (Pla de 2013). Codi: SIW013. Curs acadèmic 2020-2021
Tipo de documento
info:eu-repo/semantics/masterThesisDerechos de acceso
info:eu-repo/semantics/openAccess