Onboard conformal prediction for domain shift in Earth observation

Deep learning (DL) models are increasingly deployed in onboard Earth observation (EO) platforms, supporting a wide range of applications such as land-use classification, wildfire detection, maritime surveillance and atmospheric observation. In these settings, reliable uncertainty quantification (UQ) is essential, as predictions must remain trustworthy despite distributional shifts induced by noise, compression and varying environmental conditions. While conformal prediction provides distribution-free, finite-sample theoretical guarantees, its validity breaks down under distribution shift. In this paper, we study weighted conformal prediction for EO applications, where covariate shift is addressed by reweighting calibration samples using estimated density ratios.