Multi-Scale Augmentation for Enhanced Long-Tailed Remote Sensing Object Classification

Long-tailed distribution presents a significant challenge in remote sensing image classification, often resulting in poor recognition of tail classes and diminished overall classification performance. While diffusion models have shown potential in addressing data imbalance, existing methods struggle with inter-class quantity imbalances and intra-class scale variations in remote sensing data. To address these issues, we propose a multi-scale data augmentation approach that utilizes pre-trained diffusion models to generate realistic remote sensing images. By implementing a multi-scale uniform sampling strategy, we construct a fine-tuning dataset with balanced quantity and scale distributions. Fine-tuning the pre-trained model on this dataset facilitates the generation of tail class samples with diverse scales, effectively alleviating long-tailed distribution challenges. Evaluated on the DOTA, DIOR, and FGSC-23 datasets, our method achieves Top-1 accuracy improvements, with further gains when integrated with model-level approaches.