Marc Rußwurm, M.Sc

The Earth is dominated by physical spatiotemporal processes that are monitored by an increasing number of space-borne sensors. Optical satellites measure a rich spectral signature on a spatial pixel-grid at regular time intervals. The acquired data is highly dimensional in terms of color, space and time. This makes the whole of Earth observation data impossible to visualize for human eyes, but highly expressive for data-driven learning. Also, processes on the Earth follow complex spatio-physical relations of which only the spectral change on the surface can be monitored. The large number of hidden variables of these physical models can be approximated using neural networks while a direct functional formulation is often intractable. Finally, data-driven models allow us to efficiently utilize the increasing quantity of gathered Earth observation data for research and applications at global scale.

The success of deep machine learning in combination with today's vast availability of data has spawned new research fields and applications used in daily life. Open source code distribution, public benchmark datasets, and new GPU-driven parallelization technology enable weekly innovation and fast turn-over of applications. The model-driven design of hand-crafted functional characteristics using domain-specific expert-knowledge has been superseded by learning these features directly from the data. This new level of abstraction tasks domain-experts to determine the optimal network-design instead of directly the set of optimal features for specific tasks. This level of abstraction also allows the cross-application of methods from seemingly unrelated research fields, such as natural language processing and Earth observation, wherever a similar format of data can be applied.

I started my Ph.D. studies following up my Master in the Earth science related subject of Geodesy with focus on remote sensing and Earth observation. I closely worked with my current Ph.D. supervisor already in my studies. His expertise in computer vision and machine learning guided me to methodical research that could be published while finalizing my studies. With this background, my research applications are space related, while the methods are on broad interdisciplinary applicability.

Forschung

Schwerpunkte:

  • Computer Vision
  • Earth Observation
  • Temporal Modelling
  • Recurrent Neural Networks
  • Multitemporal Land Cover Classification
  • Change Detection

Projekte:

Publikationen

  • Rußwurm, Marc; Lefèvre, Sébastien; Courty, Nicolas; Emonet, Rémi; Körner, Marco; Tavenard, Romain: End-to-end Learning for Early Classification of Time Series. arXiv preprint arXiv:1901.10681, 2019 more…
  • Rußwurm, Marc; Körner, Marco: Multi-Temporal Land Cover Classification with Sequential Recurrent Encoders. ISPRS International Journal of Geo-Information 7 (4), 2018, 129 more…
  • Rußwurm, Marc; Körner, Marco: Convolutional LSTMs for Cloud-Robust Segmentation of Remote Sensing Imagery. Proceedings of the Conference on Neural Information Processing Systems Workshops (NeurIPSW), 2018 more…
  • Rußwurm, Marc; Körner, Marco: Multitemporal Crop Identification from Medium-Resolution Multi-Spectral Satellite Images based on Long Short-Term Memory Neural Networks. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences (ISPRS) XLII-1/W1, 2017, 551--558 more…
  • Rußwurm, Marc; Körner, Marco: Temporal Vegetation Modelling using Long Short-Term Memory Networks for Crop Identification from Medium-Resolution Multi-Spectral Satellite Images. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017 more…
  • Rußwurm, Marc; Körner, Marco: Temporal Vegetation Modelling using Long Short-Term Memory Networks for Crop Identification from Medium-Resolution Multi-Spectral Satellite Images. (other entry) more…