Thesis subject
MSc thesis topic: Few-Shot Learning across Remote Sensing Problems
Learning from few training samples is highly useful as annotating images is costly and time intensive. This Thesis will explore different few-shot approaches through pre-trained classification models. It will involve developing and adapting a benchmark dataset involving a variety of remote sensing problems and compare different ways pre-trained models can be employed for few-shot classification.
In the attached figure, you can see a meta-learning model called METEOR (Rußwurm et al., 2024) that was trained on land cover classification tasks. We found that it worked well on a variety of different remote sensing problems. Your Thesis will build upon this and other works to explore few-shot learning in the remote sensing domain and potentially implemented and develop a new few-shot methodology.
Relevance to research/projects at GRS or other groups
It involves several remote sensing applications like marine debris detection, urban scene classification, or detection of deforestation.
Objectives and Research questions
- What datasets are most representative of remote sensing few-shot learning and should be evaluated in a Benchmark
- How to harmonize different resolutions, spectral, bands and label spaces from different downstream tasks?
- Which few-shot models are preferred for individual applications and across applications?
Requirements
- Deep Learning course (required)
- Remote Sensing (optional)
Literature and information
- Rußwurm, M., Wang, S., Korner, M., & Lobell, D. (2020). Meta-learning for few-shot land cover classification. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition workshops (pp. 200-201).
- Qiu, C., Zhang, X., Tong, X., Guan, N., Yi, X., Yang, K., ... & Yu, A. (2024). Few-shot remote sensing image scene classification: Recent advances, new baselines, and future trends. ISPRS Journal of Photogrammetry and Remote Sensing, 209, 368-382.
Theme(s): Modelling & visualisation