
Thesis subject
MSc thesis topic: Few-shot learning for fine wetland classification: a meta-learning approach
Few-shot learning (FSL) is an emerging machine learning paradigm that enables models to generalize from only a few labeled examples. This is particularly valuable in remote sensing applications like wetland classification, where collecting labeled data across diverse regions and seasons is challenging. By leveraging prior knowledge, FSL combined with meta-learning strategies can teach models how to learn new categories quickly, making it possible to achieve high classification accuracy with minimal supervision. This approach offers a data-efficient pathway for fine-grained wetland mapping, enhancing scalability and adaptability in wetland monitoring systems.
Background
Wetlands are vital ecosystems that support biodiversity, regulate hydrological cycles, and contribute to climate resilience. However, more than half of the world’s wetlands have been lost or degraded since the 20th century, highlighting the urgent need for accurate and timely monitoring.
Remote sensing has become essential for wetland mapping due to its capability to cover large, inaccessible areas. Yet, conventional classification methods, such as random forests and deep neural networks, rely heavily on large labelled datasets—an impractical requirement for wetlands, which are highly heterogeneous and temporally dynamic. These methods often struggle to generalize across different wetland types, regions, and seasonal conditions.
To address these limitations, few-shot learning (FSL) offers a promising solution by allowing models to adapt to new wetland categories with only a few labelled samples. Meta-learning, often referred to as "learning to learn," enhances FSL by training models to quickly adapt to novel tasks based on prior experience across diverse datasets. This combination enables flexible, data-efficient wetland classification across varied environments and timeframes. Recent studies demonstrate that FSL with meta-learning improves classification performance under data-scarce conditions, making it a valuable approach for advancing global wetland monitoring.
Relevance to research/projects at GRS or other groups
With this research you will contribute to an ongoing PhD research project
Objectives and Research questions
- How can meta-learning improve the adaptability of few-shot classification models across different wetland types and regions?
- How robust is the model when applied to unseen wetland datasets with diverse spectral and temporal characteristics?
Requirements
- Genuine interest in this topic
- GRS-34809 Deep Learning in Data Science
Literature and information
(Expected reading list before starting the thesis research)
- Panuntun, I.A., Jamaluddin, I., Chen, Y.-N., Lai, S.-N., Fan, K.-C., 2024. LinkNet-Spectral-Spatial-Temporal Transformer Based on Few-Shot Learning for Mangrove Loss Detection with Small Dataset. Remote Sens. 16, 1078.
- Rußwurm, M., Wang, S., Kellenberger, B., Roscher, R., Tuia, D., 2024. Meta-learning to address diverse Earth observation problems across resolutions. Commun. Earth Environ. 5, 1–14.
- Tuia, D., Schindler, K., Demir, B., Zhu, X.X., Kochupillai, M., Džeroski, S., van Rijn, J.N., Hoos, H.H., Del Frate, F., Datcu, M., Markl, V., Le Saux, B., Schneider, R., Camps-Valls, G., 2024. Artificial Intelligence to Advance Earth Observation: A review of models, recent trends, and pathways forward. IEEE Geosci. Remote Sens. Mag. 2–25.
- Zhang, X., Liu, L., Zhao, T., Wang, J., Liu, W., Chen, X., 2024. Global annual wetland dataset at 30 m with a fine classification system from 2000 to 2022. Sci. Data 11, 310.
Theme(s): Modelling & visualisation; Integrated Land Monitoring