Johnson (2026) Calibrated spatial uncertainty for Earth observation foundation models via Matérn-motivated latent stochastic regularization

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-08-14
• Authors: Daniel Johnson
• DOI: 10.1016/j.rse.2026.115610

Research Groups

Department of Geography, School of Liberal Arts, Indiana University Indianapolis, Indianapolis, IN, USA
The Polis Center, Indiana University Indianapolis, Indianapolis, IN, USA

Short Summary

This study introduces a Matérn-motivated latent stochastic regularization layer for frozen foundation model embeddings to quantify spatial uncertainty and improve prediction interval calibration in Earth observation foundation models.

Objective

• Investigate the effect of incorporating spatial statistical structure into deep learning uncertainty methods for Earth observation foundation models.

Study Configuration

• Spatial Scale: Continental-scale satellite imagery, with a focus on 10km × 10km training patches.
• Temporal Scale: Not specified.

Methodology and Data

• Models used: OlmoEarth embeddings, Matérn-motivated latent stochastic regularization layer.
• Data sources: Satellite imagery, land surface temperature (LST), normalized difference vegetation index (NDVI).

Main Results

• The proposed layer achieves 90% prediction interval coverage within 0.8 percentage points of nominal for both NDVI and LST.
• Under spatial block holdout, coverage degrades by 3.9 percentage points for NDVI and 7.7 for LST compared to baselines.

Contributions

• Original contribution: introducing a Matérn-motivated latent stochastic regularization layer for frozen foundation model embeddings to quantify spatial uncertainty and improve prediction interval calibration.
• The proposed method retains a higher prediction interval coverage probability at the 90% nominal level (PICP@90) than matched MLP and CNN baselines.

Funding

• This research was supported by [no specific funding mentioned in the paper].

Citation

@article{Johnson2026Calibrated,
  author = {Johnson, Daniel},
  title = {Calibrated spatial uncertainty for Earth observation foundation models via Matérn-motivated latent stochastic regularization},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115610},
  url = {https://doi.org/10.1016/j.rse.2026.115610}
}