Scholarlink

<p>Accurate contrail segmentation is crucial for climate modeling and aviation emission monitoring. Existing methods are constrained by low-resolution satellite imagery and the absence of comprehensive ground-level contrail datasets. This study introduces GACCIS, the first high-quality ground-level contrail segmentation dataset, and proposes MDH-YOLO, a lightweight instance segmentation model. Built upon the YOLOv11 framework, MDH-YOLO incorporates the Multi-Scale Edge Information Selection (MSEIS) module with adaptive pooling and the Dual-Domain Attention Mechanism (DSM) to enhance edge extraction. The Dynamic Re-parameterized Feature Pyramid Network (DRGFPN) refines high-frequency details through dynamic sampling and adaptive kernel reconfiguration. The Hybrid Group-Decoupled Segmentation Head (HGD-SegHead) reduces the model’s parameter count to 2.56M. Experimental results demonstrate that MDH-YOLO achieves 87.9% mAP (Mask) at 188.68 FPS on the GACCIS dataset, outperforming YOLOv11n-seg by 3% in mAP while maintaining real-time performance. This work provides an accurate and efficient solution for real-time contrail monitoring.</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

MDH-YOLO: A Novel Algorithm for Aircraft Contrail Image Segmentation

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Jianhui Xi,
  • Zhaojie Wang

摘要

Accurate contrail segmentation is crucial for climate modeling and aviation emission monitoring. Existing methods are constrained by low-resolution satellite imagery and the absence of comprehensive ground-level contrail datasets. This study introduces GACCIS, the first high-quality ground-level contrail segmentation dataset, and proposes MDH-YOLO, a lightweight instance segmentation model. Built upon the YOLOv11 framework, MDH-YOLO incorporates the Multi-Scale Edge Information Selection (MSEIS) module with adaptive pooling and the Dual-Domain Attention Mechanism (DSM) to enhance edge extraction. The Dynamic Re-parameterized Feature Pyramid Network (DRGFPN) refines high-frequency details through dynamic sampling and adaptive kernel reconfiguration. The Hybrid Group-Decoupled Segmentation Head (HGD-SegHead) reduces the model’s parameter count to 2.56M. Experimental results demonstrate that MDH-YOLO achieves 87.9% mAP (Mask) at 188.68 FPS on the GACCIS dataset, outperforming YOLOv11n-seg by 3% in mAP while maintaining real-time performance. This work provides an accurate and efficient solution for real-time contrail monitoring.