Sample Augmentation and Balance Approach for Improving Classification Performance With High-Resolution Remote Sensed Image

High-resolution remote sensing images can provide detailed information for land cover classification. However, the performance of a classifier depends on the quality and quantity of training samples. Additionally, manually labeling a sufficient number of samples is often time-consuming and labor-intensive. To this end, this study proposes a novel sample augmentation and balance approach (SABA) to improve remote sensing classification accuracy. First, potential candidate samples are identified around existing labeled samples using radial direction lines and gradient constraints for sample augmentation. Then, a refinement strategy based on pixel-to-pixel (P2P) and region-to-region (R2R) rules is applied to refine the candidate samples, effectively removing noisy and mislabeled instances to improve sample quality. Finally, the proposed sample augmentation and refinement strategy method is embedded into an iterative framework to enhance and balance the training samples for each class. Experimental results on two datasets demonstrated that the proposed SABA framework outperforms five state-of-the-art methods, particularly under small-sample and class-imbalanced conditions. The improvement achieved by our proposed approach is about 0.79% $\sim ~13.57$ % in terms of the overall accuracy. The code of the proposed approach is available at https://github.com/ImgSciGroup/2025-GRSL-SABA