Uncertainty Quantification for Earth Observation Foundation Models

The recent rise of foundation models in Earth Observation (EO) has reshaped how remote sensing tasks are approached, particularly by allowing strong downstream performance with comparatively limited labeled data. These models have reported impressive results in applications such as land cover classification and semantic segmentation. However, performance gains alone do not resolve a central concern: whether the resulting predictions can be trusted. In practical EO scenarios—including disaster response and environmental monitoring—miscalibrated confidence estimates may lead to incorrect decisions even when overall accuracy appears high.

Motivated by this gap between accuracy and reliability, this study focuses on the uncertainty calibration behaviour of fine-tuned EO foundation models. Using TorchGeo for consistent data handling and the Lightning-UQ-Box framework for uncertainty quantification, we construct an evaluation pipeline that contrasts Vision Transformer–based pretrained models with conventional convolutional neural networks trained from scratch. Experiments are conducted across both image classification tasks (e.g., EuroSAT) and dense prediction settings such as semantic segmentation.

Rather than assuming superior representations automatically yield better-calibrated predictions, we explicitly examine how calibration properties change after fine-tuning large pretrained models. In addition, we evaluate a spectrum of uncertainty quantification approaches, from lightweight post-hoc methods like temperature scaling to more computationally demanding techniques, including Monte Carlo Dropout, deep ensembles, and Laplace approximation. Calibration quality is assessed using expected calibration error and reliability diagrams, alongside predictive accuracy.

By analysing the trade-offs between computational cost, accuracy, and calibration, this work provides practical insight into which UQ strategies are most effective for EO foundation models. Our findings aim to support the deployment of remote sensing systems in operational settings where reliable uncertainty estimates are as critical as raw predictive performance.