Keynote lecture at the international workshop Machine Learning for Earth Observation MACLEAN 2020: Sparse coding and deep learning in the analysis of hyperspectral images in remote sensing, (slides, video)
Abstract: Hyperspectral imaging is now established as one of the key technologies in Earth observation. While offering rich spectral information in hundreds of spectral bands, hyperspectral images remain to pose challenges for processing due to their huge dimensionality and lack of sufficient training data to match it. In this talk we address mainly clustering and classification of large-scale remotely sensed hyperspectral images, from perspectives of sparse coding and deep learning. Recent results in subspace clustering of very large scale data will be presented and discussed, including clustering of multi-source data. We also address limitations of current deep learning models based on convolutional neural networks (CNN) in hyperspectral image analysis and we discuss recent models based on group CNN designs that require less training data and generalize better to different data sets.
