A lightweight transfer learning framework with dynamic data augmentation for few-shot remote sensing image classification

Artificial intelligence has become increasingly important for Earth observation and remote sensing image analysis; however, deploying deep learning models under limited labeled data and computational resources remains a significant challenge. This study proposes a lightweight transfer learning framework with dynamic data augmentation for few-shot remote sensing image classification. Built upon the EfficientNet-Lite architecture, the proposed framework integrates dynamic data augmentation, synthetic data generation using text-to-image generative models, transfer learning, and ensemble prediction to improve classification performance while preserving computational efficiency. The framework was evaluated on four benchmark remote sensing datasets (OPS-SAT, UC Merced Land Use, WHU-RS19, and RSSCN7). Experimental results demonstrate a 5.33% absolute increase in classification accuracy and a 43% reduction in the loss score compared with a recent EfficientNet-Lite-based baseline. The proposed framework achieves an effective balance between predictive performance, computational efficiency, and model compactness, making it suitable for remote sensing applications operating under limited computational resources and scarce labeled data. The proposed framework is particularly applicable to Earth observation tasks such as environmental monitoring, land-use and land-cover mapping, disaster response, agricultural observation, and onboard satellite image analysis, where both computational efficiency and reliable classification performance are essential. These findings demonstrate that combining lightweight architectures with transfer learning and advanced data augmentation strategies provides an effective and practical framework for few-shot remote sensing image classification and facilitates the deployment of deep learning models in real-world Earth observation applications.