Unconventional applications for geo-spatial deep learning
Synopsis: Deep learning methods have emerged and proved very successful in providing meaningful insights from big archives of Earth Observation data, leading to breakthroughs e.g. in land cover / use classification, time series analysis, change detection or data fusion. However, most deep learning-based methods developed for remote sensing are supervised, and require a significant amount of labelled training data that is tedious and expensive to collect, or not available in certain areas. Furthermore, supervised methods do not effectively tap the potential of massive amounts of unlabelled Earth observation data, which can be used to learn effective representations to describe the world.
For these reasons, we propose a session about Unsupervised and weakly supervised deep learning for Earth Observation, emphasizing the development and application of novel methods such as (but not limited to):
- Unsupervised learning
- Transfer learning
- Semi-supervised learning
- Weakly supervised learning
- Meta learning
- Few-shot learning
In this thematic session, we welcome new and ambitious talks in various application domains of photogrammetry and remote sensing, that can demonstrate how to leverage unlabelled or weakly labelled EO data.
The session is open to all authors.