Balancing Data Imbalance in Remote Sensing Image Road Segmentation With Gradient-Guided Multi-Level Data Augmentation

In remote sensing image segmentation, data imbalance remains a critical challenge, particularly in datasets from networks and natural scenes, increasing task complexity. Recent studies have explored adversarial training to address this issue, but these methods rely heavily on labeled data, introducing annotation bias and efficiency concerns. To overcome these limitations, this paper proposes a gradient-guided image-level perturbation strategy. Specifically, we design a multi-level, multi-angle data augmentation framework that integrates patch-level perturbation with gradient-based adaptive region selection, enhancing the model's recognition of rare samples while preserving contextual integrity. Experimental results demonstrate that our method surpasses traditional segmentation algorithms and data-balancing strategies, achieving an F1 score of 82.67% on the Massachusetts Roads dataset and exhibiting strong generalization on the LoveDA dataset. This research provides valuable insights for advancing remote sensing image analysis.