SAT-493M is a large-scale pretraining corpus of commercial Maxar satellite imagery used by the DINOv3 project. It contains approximately 493 million RGB, ortho-rectified image chips (tiles) at 512×512 pixels, sampled at ~0.6 meter ground sampling distance. The collection was assembled from Maxar high-resolution optical imagery and was used to pre-train DINOv3 satellite models (e.g., ViT-7B and distilled variants) to produce high-quality dense features for remote-sensing / overhead-vision tasks. The dataset is a proprietary/commercial compilation of Maxar imagery (not published as an open Hugging Face dataset) and is provided to the DINOv3 team under Maxar licensing; access and redistribution are therefore restricted. Primary sources: DINOv3 paper (arXiv:2508.10104) and the DINOv3 repository / model cards which describe models pretrained on “SAT-493M.”
No results indexed yet — be the first to submit a score.
Submit a checkpoint and a reproduction script. We will run it, publish the score, and — if it takes the top — annotate the step on the progress chart with your name.