This article demonstrates training self-supervised DINO vision transformers on satellite imagery using Amazon SageMaker and the BigEarthNet-S2 dataset.

• Self-supervised learning reduces annotation costs for satellite/aerial image analysis tasks
• DINO algorithm trains student/teacher networks on unlabeled overhead imagery without labels
• Custom PyTorch dataset class loads BigEarthNet-S2 multispectral images from S3
• SageMaker distributed data parallel library enables multi-GPU training on large instances
• Pre-trained DINO model transfers to downstream land cover classification task
• BigEarthNet-S2 pre-training achieved 6.7% higher average precision than ImageNet pre-training
• Training took approximately 11 hours on ml.p3.16xlarge instance with 83% GPU utilization

The solution provides a template for training self-supervised vision transformers on large-scale unlabeled satellite imagery, improving model performance on downstream tasks compared to generic ImageNet pre-training.