Label-Free Monitoring in Geospatial Foundation Models with Feature-Aware Conformal Prediction

A model trained on satellite data from one region often does not generalize to another region, and the only reliable way to know is to collect new labels there, which is expensive. We study whether the representation space of a frozen geospatial foundation model can provide a label-free signal of how well a downstream model will perform in a new area. We start from split conformal prediction, whose prediction-set size is commonly used as a confidence signal, and show that for a frozen encoder with a linear head this signal is structurally limited: it is computed from the logits alone, and the linear head discards most of the representation, so a large class of representation shifts leaves the confidence unchanged (Kernel Blindness). We propose a simple correction, feature-density-aware conformal prediction (FD-ACP), which rescales the conformal score by how unfamiliar the representation is with respect to the training representations, measured before the downstream head. The correction keeps the standard coverage guarantee and needs no labels at deployment. On EuroSAT, with a frozen DINO ViT encoder and a ten-class classification head, we validate the signal by its correlation with actual accuracy across geographic regions. Under two geographic shift protocols, accuracy falls from 0.836 to 0.281; standard conformal confidence barely follows this drop (r=0.03 and 0.11), while FD-ACP tracks it closely (r=0.72 and 0.69).