Cote: |
1115 |
Auteur: |
TALON Patrick |
Année: |
Septembre 2018 |
Titre: |
Convolutional neural networks for semantic segmentation of multi-scale remote sensing imagery: comparison of model architectures |
Sous la direction de: |
Dr Christian Kaiser |
Type: |
Mémoire de master en géographie |
Pages: |
63 |
Complément: |
16 pages d'annexes |
Fichier PDF: |
Mémoire [5.7 Mo]
|
Mots-clés: |
Artificial / neural / network / convolutional / remote / sensing / semantic / segmentation |
Résumé: |
This master thesis addresses the question of how convolutional neural networks (CNN) can help to achieve better accuracies for semantic segmentation of remote sensing data, in a view of automated and decision oriented applications. The attempt to answer this question is made by comparing the results achieved by two distinct models inspired by the popular U-Net and SegNet architectures. The experiments will be conducted on two remote sensing datasets that have been captured at different geographic scales. Results show that there is no unique solution for semantic segmentation of remotely sensed data and that different levels of scale would require different architectures of CNN. In order to achieve acceptable level of accuracies for public or private decision oriented applications, CNNs require more imposing labelled data for training and more balance between the land cover classes. |