Deep learning models for satellite image classification often rely on large-scale annotated datasets, which are expensive to collect and rarely available across diverse geographies and spatial resolutions. In this work, we present a label-efficient and resolution-agnostic training framework based on contrastive self-supervised learning (SSL) for satellite image classification. Using SimCLR as the backbone, we pretrain a ResNet encoder on unlabeled RGB data from Sentinel-2 (EuroSAT) and UC Merced datasets, followed by supervised fine-tuning on varying fractions (10%, 20%, 50%) of labeled data. Our method consistently recovers over 93% of the fully supervised accuracy using only 10% labels, and outperforms supervised baselines by up to 13% points. To assess generalization, we conduct cross-resolution experiments (10 m vs. 0.3 m) where the pretrained encoder transfers across domains without fine-tuning. Results show robust domain adaptation, with SSL models maintaining over 82% accuracy in transfer settings. These findings establish contrastive SSL as a scalable, reproducible, and cost-effective paradigm for geospatial AI, particularly in low-label and multi-resolution remote sensing scenarios. Future work will explore extensions to multispectral and temporal modalities, as well as transformer-based architectures.