Access to accurate and detailed land cover (LC) maps is crucial for many applications such as environmental monitoring, land and
agricultural policy, population estimation, urban climate modelling, etc. In the remote sensing (RS) community, the approach
combining object-based image analysis (OBIA) and traditional supervised machine learning (ML) classifiers such as Random Forest (RF) has
been the state of the art (SoA), for many years, for detailed LC
mapping from very-high resolution remote sensing (VHRRS) data.
However, during the last decade, deep learning (DL) approaches and more specifically deep convolutional neural networks (CNN)
facilitated a breakthrough in the field of computer vision (CV) by beating traditional ML approaches. CNNs also proved their ability
to overperform former approaches in the field of RS and is now the new SoA.
Most of the cutting-edge methodological development on
CNN occurs in the field of CV and is then applied to the specific characteristics of RS images. However, the technological
transfer from the field of CV to RS is not straightforward and many challenges still exist. The goal of this project is to provide
solutions to alleviate one of the major issues faced by the RS community when applying cutting-edge DL approaches on VHRRS data for LC
mapping,
namely, the difficulty to get access to a large amount of labelled training data.