Open-Earth-Monitor Global Workshop 2026

Deep learning represents a powerful tool to interpret Earth observation data at large geographic scales. However, in cases where abundant reference data is not available and cannot easily be collected, new approaches are needed to benefit from this technology. Several Earth observation tasks, especially in environmental remote sensing, remain challenging due to the limited number of samples and the geographic and temporal bias in the reference data. Furthermore, mapping biophysical variables from single sensor inputs often leads to high ambiguities. Multimodal models pretrained in a self-supervised fashion promise to overcome such challenges.

In this talk, I will first present our recent research project MMEarth-Bench, a multimodal benchmark dataset for environmental remote sensing. I will discuss our evaluation of existing pretrained models and present our test-time adaptation approach that adapts any model at test time using multimodal data to construct adaptation signals. Lastly, I will present SuperF, an approach for multi-image super-resolution. This test-time optimization approach based on implicit neural representations makes use of repeated observations with sub-pixel shifts and does not require any high-resolution training data. Under a static scene assumption, it can be applied to super-resolve e.g. Sentinel-2 time series for any place on Earth.

• Personal website: https://langnico.github.io/
• MMEarth-Bench project: https://mmearth-bench.com/
• SuperF project: https://sjyhne.github.io/superf/