Special session on ''Novelty Detection and Lifelong Learning for Robust Performance Under Ever-Changing Conditions''
Remote Sensing systems are exposed to the constantly changing conditions in the open world. The conditions in which remote sensing systems operate can change frequently and drastically due to differences in acquisition (i.e., different atmospheric effects, acquisition viewpoints, time periods, etc.) and the heterogeneity of the environments under investigation (e.g., planetary exploration, sustainable development, UAV and drone-based data collection). Researchers have been studying challenges involved with novelty detection, identification and adaptation to build more reliable and intelligent remote sensing systems.
Robust performance of automated systems for open world remote sensing hinges on determining when conditions have changed enough to jeopardize accurate performance, and how to identify critical changes so as to appropriately adapt to them. Novelty detection aims to determine when such changes have occurred without knowledge of what the changes might be. Lifelong learning aims to support robust performance under novel conditions by characterizing these changes and continuously adapting without catastrophically forgetting previous knowledge.
A few examples of novelty scenarios in remote sensing we aim to highlight in our session include:
- Onboard analysis of imagery from UAV and drone-based acquisitions
- Novel situations in controlling the UAV to take imagery
- Supporting robust performance by UAVs flying over novel terrains or conditions
- Detecting unexpected, unmodeled changes in satellite imagery in order to discover novel, unexpected objects, events or processes
- Detecting anthropogenic activity in protected zones to support the goals of sustainable development
Some possible related topics:
- Out of distribution detection and domain adaptation
- Detecting the unknown unknown (what we do not know we are missing)
- Challenges related to novelty detection, adaptation and characterization
- Possible algorithmic approaches
- Novelty detection can also provide new significant and actionable information (e.g., new features)
The paper submission deadline is January 17th, 2022. Please do not submit the paper through the IGARSS 2022 general submission website. The papers for this special session are required to be submitted using the dedicated link that will be shared with the authors after a first contact with the session chairs. If you intend to submit a contribution to this session, please write an email to the session chairs and specify the title and abstract of the paper.
Session chairs and contacts
Katarina Doctor (U.S. Naval Research Laboratory, firstname.lastname@example.org)
Gabriele Cavallaro (Forschungszentrum Jülich, email@example.com)