Auteurs
E. Colin Koeniguer, G. Le Besnerais, A. Chan Hon Tong, B. Le Saux, A. Boulch, P. Trouvé, R. Caye Daudt, N. Audebert, G. Brigot, P. Godet, B. Le Teurnier, M. Carvalho, J. Castillo-Navarro (ONERA)
The purpose of this article is to take stock of the progress made in remote sensing thanks to the recent development of deep learning techniques. This assessment is made by means of a systematic presentation of the various activities carried out at ONERA in remote sensing imagery using deep learning methods. It covers a large part of the observation problems: filtering, object detection, land-use classification, change detection, and biomass estimation. In light of these activities, we highlight the practical challenges of deep learning, mainly physical feature definition and training database construction. Some directions for future research are also given, such as the development and use of dedicated remote sensing platforms, hybrid supervised/unsupervised strategies, and the further exploitation of multimodal/multitemporal data.