Deep learning for semantic segmentation of remote sensing images with rich spectral content

摘要

With the rapid development of Remote Sensing acquisition techniques, there isa need to scale and improve processing tools to cope with the observed increaseof both data volume and richness. Among popular techniques in remote sensing,Deep Learning gains increasing interest but depends on the quality of thetraining data. Therefore, this paper presents recent Deep Learning approachesfor fine or coarse land cover semantic segmentation estimation. Various 2Darchitectures are tested and a new 3D model is introduced in order to jointlyprocess the spatial and spectral dimensions of the data. Such a set of networksenables the comparison of the different spectral fusion schemes. Besides, wealso assess the use of a " noisy ground truth " (i.e. outdated and low spatialresolution labels) for training and testing the networks.