An Adaptive Transformer Framework Integrating Scarcity-Guided Augmentation and Reliability Aware Domain Alignment for Few-Shot Satellite Image Classification

Abstract

Satellite image classification is essential for land-use analysis, geographic monitoring, and environmental assessment; however, geographic domain shifts, class imbalance, and limited labeled target-domain samples remain significant challenges. Existing deep learning approaches often rely on abundant labeled data and may experience reduced classification performance when transferred across geographic regions, while conventional augmentation and domain adaptation methods may inadequately address class scarcity and synthetic-sample reliability. This study proposes an adaptive Transformer framework integrating scarcity-guided augmentation and reliability-aware domain alignment for few-shot satellite image classification. The framework uses the Domain Shift across Geographic Regions (DSGR) dataset with corresponding fMoW satellite imagery, covering 41 land-use classes across six geographic regions. Implementation is planned in Python using PyTorch, with a pretrained Swin Transformer-Tiny backbone. The proposed method estimates class scarcity, allocates adaptive augmentation, generates class-conditioned synthetic images using a generative adversarial network, and filters synthetic samples according to class consistency. Scarcity-Conditioned Token Adaptation (SCTA) refines token representations, while adaptive local–global feature fusion captures spatial details and contextual information. Reliability-aware class-conditional prototype alignment reduces geographic feature discrepancies, and class balanced focal loss addresses imbalanced classification. Evaluation will consider 1-, 5-, 10-, and 20-shot target support settings using accuracy, precision, recall, Macro-F1, and balanced accuracy. Preliminary illustrative results indicate 95.0% accuracy, 92.4% Macro-F1, 92.8% balanced accuracy, and 90.6% minority-class recall. Compared with the illustrative Inception-v3 accuracy of 92.0%, the proposed accuracy represents a 3.0 percentage-point improvement, equivalent to approximately 3.26% relative improvement. The framework aims to improve reliable cross-region classification under data scarcity, with potential applications in geographically diverse satellite-image analysis.