Meta-learning with Meteor

Robin sits down with Marc Rußwurm to unpack METEOR, a meta-learning framework that turns a single Sentinel-2 land-cover data cube into a model that can be fine-tuned on any new remote-sensing task with as few as five labelled examples. Marc explains how meta-learning differs from classic foundation-model pre-training, why he splits the globe into thousands of “micro-tasks,” and how a laptop-sized GPU can train the whole backbone in two days. Real-world demos include rapid deforestation mapping and detecting the Beirut port explosion with only a handful of pixels per class. PROJECT & PAPER LINKS • GitHub repository – https://github.com/marccoru/meteor • Nature Communications Earth & Environment paper – https://www.nature.com/articles/s43247-023-01146-0 • Marc on LinkedIn – https://www.linkedin.com/in/marcrusswurm/ • Episode blog & podcast – https://www.satellite-image-deep-learning.com/p/meta-learning-with-meteor KEY TAKE-AWAYS • Task-centric training: the Earth is treated as thousands of tiny classification problems, teaching the model to generalise. • Tiny data, big results: five annotated images can fine-tune METEOR for deforestation, urban density, or disaster damage. • Sensor flexibility: channel-wise adapters let you swap Sentinel-2 for PlanetScope RGB or Sentinel-1 SAR without retraining from scratch. • Laptop-friendly: full meta-training finishes in ~48 h on a single mid-range GPU; inference is even lighter. • Open & extensible: all code, pre-trained weights, and evaluation scripts are on GitHub — extend the task set or try the model on your own project. Enjoy the episode? Like, Subscribe, and share your questions below! Bio: Marc Rußwurm is Assistant Professor of Machine Learning and Remote Sensing at Wageningen University. His background is in Geodesy and Geoinformation, and he obtained a Ph.D. in Remote Sensing Technology at TU Munich. During his Ph.D., he could visit the European Space Agency and the University of Oxford as a participant in the Frontier Development Lab in 2018, the Obelix Laboratory in Vannes, and the Lobell Lab in Stanford. As a postdoctoral researcher, he joined the Environmental Computational Science and Earth Observation Laboratory at EPFL, Switzerland. His research interests are developing modern machine learning methods for real-world remote sensing problems, such as classifying vegetation from satellite time series and detecting marine debris in the oceans. He is interested in domain shifts and transfer learning problems naturally arising from geographic data.