Deep Learning for Satellite Imagery: Semantic Segmentation, Non-Rigid Alignment, and Self-Denoising
Guillaume Charpiat (Équipe TAU, INRIA Saclay / LRI - Université Paris-Sud)
Neural networks have been producing impressive results in computer vision these last years, in image classification or
segmentation in particular. To be transferred to remote sensing, this tool needs adaptation to its specifics: large
images, many small objects per image, keeping high-resolution output, unreliable ground truth (usually
mis-registered). We will review the work done in our group for remote sensing semantic segmentation, explaining the
evolution of our neural net architecture design to face these challenges, and finally training a network to register
binary cadaster maps to RGB images while detecting new buildings if any, in a multi-scale approach. We will show in
particular that it is possible to train on noisy datasets, and to make predictions at an accuracy much better than the
variance of the original noise. To explain this phenomenon, we build theoretical tools to express input similarity from
the neural network point of view, and use them to quantify data redundancy and associated expected denoising effects.
If time permits, we might also present work on hurricane track forecast from reanalysis data (2-3D coverage of the
Earth's surface with temperature/pressure/etc. fields) using deep learning.
After a PhD thesis at ENS on shape statistics for image segmentation, and a year in Bernhard Schölkopf's team at MPI
Tübingen on kernel methods for medical imaging, Guillaume Charpiat joined INRIA Sophia-Antipolis to work on computer
vision, and later INRIA Saclay to work on machine learning. Lately, he has been focusing on deep learning, with in
particular remote sensing imagery as an application field.