Back to AI News
arxiv

Matched-Pair Study Finds Earth-Observation Pretraining Is Not Uniformly Better

A new arXiv study compares five matched pairs of Earth-observation and generic vision encoders across 14 frozen models and five RGB scene-classification datasets. EO pretraining significantly helped SatDINO, Scale-MAE and RemoteCLIP, but significantly hurt DINOv3-SAT and SatMAE on linear probing, with DINOv3-SAT dropping by up to 22.5 points. Two MAEs pretrained on the same fMoW-RGB imagery produced opposite outcomes, and gains or losses were largest with 1-5% labels. The authors conclude that satellite pretraining is not a reliable selection criterion and that candidate checkpoints should be validated on target data.

AI Brief

It gives AI practitioners a concrete, controlled warning that Earth-observation foundation models can underperform generic encoders depending on the checkpoint and task.

arXiv:2611.13849v1 Announce Type: cross Abstract: Geospatial foundation models are pretrained on satellite and aerial imagery, but whether such an encoder beats a strong generic one for few-label land-use/land-cover mapping remains unclear, as prior comparisons rarely isolate the pretraining data. We study five matched pairs of Earth-observation (EO) and generic encoders sharing objective and architecture (DINO/SatDINO, DINOv3/DINOv3-SAT, MAE/SatMAE, MAE/Scale-MAE, CLIP/RemoteCLIP) within a benchmark of 14 frozen encoders on five RGB scene-classification datasets (0.06-30 m GSD). With native preprocessing, we evaluate k-NN, linear and attentive probes and image retrieval over six low-label splits, using paired per-image significance tests. EO pretraining is not uniformly helpful: on most datasets and metrics, it significantly improves three pairs (SatDINO, Scale-MAE, RemoteCLIP, the latter evaluated only on unseen datasets) and significantly hurts two (linear probe: DINOv3-SAT by up to 22.5 points, SatMAE by up to 8.8), with opposite outcomes even for two MAEs pretrained on the same fMoW-RGB imagery. These effects are robust to patch-token pooling and attentive probing, concentrate on small man-made objects, and are largest with 1-5% labels. Across encoders, the silhouette score of frozen embeddings predicts k-NN accuracy and retrieval mAP almost perfectly (Spearman $\rho \geq 0.96$ on four of five datasets). Pretraining on satellite imagery is thus not a reliable selection criterion; candidate checkpoints must be validated on target data.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!