SatViT-V1 and SatViT-V2 pretrained model weights

Both models were pretrained on an unlabelled dataset of 1.3 million images acquired from Sentinel-1 and Sentinel-2 (stacked along the channel dimension).

BEN-500 dataset (120x120 pixels, 13 optical and 2 SAR channels)

Each split contains 500 images sampled from the BigEarthNet-MM dataset. Each of the 43 classes are represented at least 12 times per split. Images are PyTorch tensors of shape (500, 120, 120, 15) and labels are (500, 43).