Transfer Learning for Remote Sensing
|A GRSL special stream – 2018
With the development of modern satellite sensors, remote sensing data have been explosively increasing. To make full use of these newly collected data, remote sensing image analysis is facing new challenges. Traditional methods would be labor intensive and ineffective, while machine learning (ML) based methods have attracted great attention as they are proved to be more effective. However, such conventional ML frameworks highly rely on the labeled data, which is not an easy task especially for some large data sets. Revealing the relationship among datasets to make the best usage of prior knowledge of pre-existing data would be a promising research topic as a consequence. If these labeled samples could be reused, many efforts could be saved. Unfortunately, directly applying the ML models trained on existing data to new data is unrealistic due to their spectra divergence. Transfer learning is thus thought to have great potential to solve this problem, which employs knowledge from one domain to another new yet related domain by constructing a common subspace, on which two domains are with the same data representation.
This special stream aims to invite original contributions reporting the latest progress in transfer learning based methods to solve remote sensing problems. The motivation is to generate more effective cross-domain machine learning and image understanding methods to deal with the ever-increasing remote sensing data.
The topics of interest include, but are not limited to:
- Theories for domain adaptation and generalization
- Transfer learning in remote sensing
- Multi-task learning in remote sensing
- Multi-view learning in remote sensing
- Multi-label learning for remote sensing
- Domain generalization algorithms for visual problems
- Deep representation learning for domain adaptation and generalization
Submission open: 1 July 2018
Submission close: 30 November 2018