Speaker Bio & Abstract
Photogrammetry and GeoInformation
Leibniz University of Hannover
AbstractIn the last few years deep learning approaches, mainly based on convolutional neural networks (CNN) have had a stunning success for image analysis tasks like object classification, localisation and detection, and have reached similar results in related fields (speech recognition, natural language processing etc.). For example, so called “super human performance” has been reached in the classification and detection challenge ImageNet Large Scale Visual Recognition Challenge (ILSVRC) for everyday images depicting individual objects and terrestrial image sequences often acquired for assessing autonomous driving algorithms.
When considering the domain of remote sensing, a number of open questions remain. These include the transferability of the available approaches and pre-trained networks to for aerial and satellite images, for multi-temporal and multi-/hyperspectral or radar image analysis and for other data types such as 3D point clouds. Another issue is the need for a large amount of training data which may not always be available, and if so, may contain gross errors.
This presentation will assess the state-of-the-art in deep learning for aerial and satellite image analysis and will point out the potential and limitations of this promising technique as well as some directions for future research.