Prithvi-EO-2.0
A geospatial foundation model for Earth observation jointly developed by NASA and IBM, based on the Transformer architecture. It is self-supervisedly pre-trained on multi-spectral satellite time-series data from the HLS dataset and supports fine-tuning for downstream remote sensing tasks such as crop classification.
Timeline (last 90 days)
A study published on arXiv (arXiv:2610.08810) systematically evaluated the out-of-distribution performance of the Prithvi-EO-2.0 model across 37 events in 12 countries on three continents.
The model's average overall accuracy (OA) dropped from 0.65 in the United States to 0.40 in Europe.
The study found that cross-continental accuracy degradation largely does not stem from spatial transfer itself, but from phenological misalignment.
When the observation window is not aligned with the local crop phenology, model accuracy collapses while deterministic confidence remains high—meaning the model makes errors confidently.
The Expected Calibration Error (ECE) increased in seven out of eight event pairs, and in affected scenarios, 12%–51% of the area was confidently mislabeled by the model with near-zero accuracy.
Consolidating 13 classes into 10 (merging the classes most confused by the model) improved the average OA by 8.4 percentage points.
The model maintains accuracy within observation windows of 45 to 90 days, peaking at approximately 75 days; when the window is compressed to below 30 days, accuracy is approximately 0.11 lower than the plateau.
Both adjustments—merging confused classes and calibrating the observation window—can recover accuracy without retraining the model.
All Facts (8)
A study published on arXiv (arXiv:2610.08810) systematically evaluated the out-of-distribution performance of the Prithvi-EO-2.0 model across 37 events in 12 countries on three continents.
50%UnverifiedThe model's average overall accuracy (OA) dropped from 0.65 in the United States to 0.40 in Europe.
50%UnverifiedThe study found that cross-continental accuracy degradation largely does not stem from spatial transfer itself, but from phenological misalignment.
50%UnverifiedWhen the observation window is not aligned with the local crop phenology, model accuracy collapses while deterministic confidence remains high—meaning the model makes errors confidently.
50%UnverifiedThe Expected Calibration Error (ECE) increased in seven out of eight event pairs, and in affected scenarios, 12%–51% of the area was confidently mislabeled by the model with near-zero accuracy.
50%UnverifiedConsolidating 13 classes into 10 (merging the classes most confused by the model) improved the average OA by 8.4 percentage points.
50%UnverifiedThe model maintains accuracy within observation windows of 45 to 90 days, peaking at approximately 75 days; when the window is compressed to below 30 days, accuracy is approximately 0.11 lower than the plateau.
50%UnverifiedBoth adjustments—merging confused classes and calibrating the observation window—can recover accuracy without retraining the model.
50%