EarthJev

A non-generative vision–language model for Sentinel-2: it reads all 12 bands of a patch and, for every typed question, returns a calibrated probability for each declared option in one forward pass, without generating any token.

These are precomputed outputs of the released model on test patches, never seen in training. The patches were chosen by a fixed rule before looking at the outputs, and wrong answers are kept (). Paper: arXiv:2611.02096 · Model: dBenf/EarthJev · Code: CosmoIknosLab/EarthJev.

Accuracy by question type

Accuracy of the released model on all the test questions of each type. Tick : chance level (one over the number of options).

Question types seen in training (CORINE test, BigEarthNet v2.0)

Questions written by other authors, never used for training

Examples

Bars: the model's probabilities. Blue: the model's answer. ✓ bold: the reference answer. The number on each thumbnail counts the correct answers for that patch.