Unlabeled to Accurate: Self-Supervised Learning for Land Use Classification in Sentinel-2 Imagery

Self-supervised learning (SSL) provides a compelling alternative to traditional supervised methods by leveraging large volumes of unlabeled data to learn generalizable feature representations. In this paper, we evaluate three state-of-the-art SSL frameworks—SimCLR, DINO-MC, and LightlyTrain—for land use classification using Sentinel-2 satellite RGB imagery from the EuroSAT dataset. While a fully supervised ResNet-18 baseline achieves 84.6% accuracy, our SSL models demonstrate superior performance, especially in low-label regimes. The DINO-MC model, utilizing a Wide ResNet backbone and pretrained on the SeCo-100K corpus, achieves a state-of-the-art accuracy of 95.7%. SimCLR and LightlyTrain also outperform the baseline, with LightlyTrain offering 88.7% accuracy at significantly reduced computational cost. These results highlight the effectiveness of SSL in remote sensing applications where labeled data is scarce but unlabeled imagery is abundant. Our study provides practical insights into accuracy-compute trade-offs and label efficiency, supporting broader adoption of SSL in Earth observation tasks.