Contemporary foundation models for Earth Observation have uncertain predictive power in extreme environments, where sparse sampling during training leads to poor generalisation. For some downstream tasks, models trained on these backbones fail to beat supervised baselines entirely. This may be from locations that were poorly sampled while training, or in target locations where current environmental conditions are significantly different to training data. What can we do about it?

After ESL 2024, we have a newfound ability to make 3D volumetric reconstructions of global clouds. How can we use and improve our models to support a wide variety of scientific use-cases? From better forecasting of extreme events like heat domes or dust storms, to more discriminative cloud classification or a more nuanced understanding of how deforestation affects cloud cover and type.

Around a quarter of greenhouse gas emissions are from gases other than CO2; this “long tail” includes methane, nitrous oxide and a variety of other gas species that contribute to radiative forcing in the atmosphere. In the STARCOP project, we showed that we can detect methane using multi-spectral sensors using ML on small satellite hardware. Can we detect other gases using visible-wavelength (silicon) sensors? How can we best extract the weak signatures that identify gas plumes?