Improving Land Cover Classification Under Limited Data Using Advanced Augmentation and Transfer Learning

Land cover classification from satellite imagery under limited labeled data remains a critical challenge in operational remote sensing, where annotation costs are prohibitive. This paper presents a systematic label-efficiency study of advanced data augmentation combined with transfer learning for remote sensing (RS) scene classification, evaluated across four public benchmarks: EuroSAT (10 classes, 27,000 images), UC Merced (21 classes, 2,100 images), NWPU-RESISC45 (45 classes, 31,500 images), and AID (28 classes,  8,400 images), totaling  69,000 images across 104 land cover classes. We combine an ImageNet-pretrained ResNet-50 backbone with three augmentation strategies—Mixup, CutMix, and Random Erasing—evaluated at six training data ratios from 1% to 50% with three independent random seeds per experiment. Transfer learning provides the most significant improvement, yielding gains of 10–28% over from-scratch baselines. Augmentation benefit is strongly data-regime dependent: mixing augmentations yield +21–31% gains under extreme label scarcity (1% data) but converge toward baseline at sufficient data (≥20%). The optimal augmentation strategy correlates significantly with dataset-specific inter-class overlap, quantified by the Intra/Inter-Class Distance Ratio (point-biserial r=0.261, p=0.0075, n=104 classes). EuroSAT CutMix at 1% training matches scratch-trained models using 20× more labeled data. Mechanistic evidence is provided via t-SNE visualization and Grad-CAM attention analysis. A practical three-step augmentation selection framework is proposed and validated.