Scholarlink

<p>Semantic segmentation in remote sensing is frequently hindered by severe class imbalance, where over-represented land-cover categories overshadow critical minority ones. Conventional mitigation techniques, including cost-sensitive loss functions and stochastic data augmentation, often fail to generate sufficient diversity or preserve the structural and contextual integrity required for high-resolution imagery. On the other hand, generative approaches may be very data demanding, or become prone to creating semantic contradictions, thereby introducing significant noise into the training signal. To address this problem, we present dynamic imbalance-aware oversampling (DIAO) and its refined variant, DIAO-CP, specifically designed to overcome these challenges in pixel-level classification, specific to semantic segmentation tasks. DIAO replaces random sampling with an iterative density-driven scoring function that identifies images to be resampled based on their potential to minimize global distribution entropy. DIAO-CP extends this logic by integrating an object-centric synthesis loop that ensures context-sensitive placement, preventing semantic collisions by utilizing a background selection filter. These mechanisms ensure that synthesized samples maintain both statistical equilibrium and geographic plausibility. The proposed methods were evaluated on relevant semantic segmentation benchmarks, specifically the Chesapeake Conservancy and OpenEarthMap (OEM). Results indicate that DIAO-CP successfully optimized the information structure of the datasets, reducing the Kullback–Leibler divergence from an initial 0.3380 to 0.1924 in the OEM benchmark. In terms of predictive performance, our solutions yielded an average improvement of 10.5 and 11.8 percentage points (p.p.), respectively, in Macro F1 and mIoU compared to baseline model without using DIAO for all regions. As such, the DIAO strategies demonstrated a superior ability to recover rare semantic entities for minority classes. These findings confirm the effectiveness of our approach in enhancing the discriminative power for imbalanced classes in diverse geographic scales within semantic segmentation workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing minority class recovery in high resolution land cover mapping with dynamic imbalance aware oversampling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月9日

查看全文
  • Gabriel Amarante,
  • Matheus T. P. Souza,
  • Gabriel Lima,
  • Willian Barreiros Jr.,
  • Renato Ferreira,
  • George Teodoro

摘要

Semantic segmentation in remote sensing is frequently hindered by severe class imbalance, where over-represented land-cover categories overshadow critical minority ones. Conventional mitigation techniques, including cost-sensitive loss functions and stochastic data augmentation, often fail to generate sufficient diversity or preserve the structural and contextual integrity required for high-resolution imagery. On the other hand, generative approaches may be very data demanding, or become prone to creating semantic contradictions, thereby introducing significant noise into the training signal. To address this problem, we present dynamic imbalance-aware oversampling (DIAO) and its refined variant, DIAO-CP, specifically designed to overcome these challenges in pixel-level classification, specific to semantic segmentation tasks. DIAO replaces random sampling with an iterative density-driven scoring function that identifies images to be resampled based on their potential to minimize global distribution entropy. DIAO-CP extends this logic by integrating an object-centric synthesis loop that ensures context-sensitive placement, preventing semantic collisions by utilizing a background selection filter. These mechanisms ensure that synthesized samples maintain both statistical equilibrium and geographic plausibility. The proposed methods were evaluated on relevant semantic segmentation benchmarks, specifically the Chesapeake Conservancy and OpenEarthMap (OEM). Results indicate that DIAO-CP successfully optimized the information structure of the datasets, reducing the Kullback–Leibler divergence from an initial 0.3380 to 0.1924 in the OEM benchmark. In terms of predictive performance, our solutions yielded an average improvement of 10.5 and 11.8 percentage points (p.p.), respectively, in Macro F1 and mIoU compared to baseline model without using DIAO for all regions. As such, the DIAO strategies demonstrated a superior ability to recover rare semantic entities for minority classes. These findings confirm the effectiveness of our approach in enhancing the discriminative power for imbalanced classes in diverse geographic scales within semantic segmentation workflows.