Self-Supervised Contrastive Learning for Remote Sensing Image Classification Without Labels

Labeled remote sensing imagery is expensive to acquire at scale, creating a critical bottleneck for supervised deep learning approaches to land use and land cover classification. We propose GeoContrast, a self-supervised contrastive learning framework tailored to the unique characteristics of remote sensing data, including multi-temporal consistency, multi-scale spatial hierarchy, and spectral diversity. GeoContrast introduces geographically-informed augmentations—temporal perturbation, scale-consistent cropping, and spectral mixing—that align with domain semantics rather than arbitrary transformations. Pre-trained on 2.1M unlabeled Sentinel-2 images, GeoContrast achieves 89.7% overall accuracy on BigEarthNet with only 1% labeled data, surpassing fully supervised baselines trained on 100% labels.