Je suis
Citoyen / Grand public
Étudiant / Futur étudiant
Partenaire public
Enseignant / Elève

Large scale land cover mapping from very high resolution (VHR) satellite images with convolutional neural network


Campus Paris-Rive-Gauche


Séminaires Planétologie et Sciences Spatiales

522, bât. Lamarck

Tristan Postadjian


Land cover mapping is the task of assigning each pixel of an overhead image a label, given a set of classes. This can be done manually, but coming at high costs, especially if one considers large geographic regions. Existing land cover maps are not satisfactory due to (i) a poor updateness rate (ii) a low spatial resolution (preventing from monitoring urban evolution) (iii) an inappropriate set of classes. Considering an automatic generation of such maps would allow annual updates and task-specific classes; specifically, SPOT 6&7 satellites provide us with a full France coverage at a resolution of 1.5m, in four spectral bands (RGB-IR) every year. This work investigates the potential of recent machine learning methods, namely the deep convolutional neural networks, in order to classify the aforementioned images.