Scholarlink

<p>Frequent outbreaks of <i>Ulva prolifera</i> in China’ s Yellow Sea—characterized by strong seasonality, large scale, and long drift distances—have become a major marine ecological disaster. Satellite remote sensing integrated with deep learning enables rapid and accurate <i>U. prolifera</i> monitoring. However, surface-aggregated <i>U. prolifera</i> exhibits variable morphology and density, which imposes high requirements on the robustness and generalization of deep learning-based segmentation networks. In this study, we propose a multi-attention and edge deep supervision neural network (MA-EDSNN) model based on Gaofen-1 (GF-1) data to achieve high-precision <i>U. prolifera</i> semantic segmentation (especially for sparse floating regions and edge details). The model integrates a self-designed scale-aware channel attention (SACA) module with attention gate (AG) and convolutional block attention module (CBAM) in each decoder block to achieve multi-scale feature fusion. It further constructs a novel Hybrid Edge-supervised Segmentation loss (HES loss) to optimize boundary and small-object segmentation, and incorporates a high-frequency enhancement module (HFEM) and Swin Transformer into the encoder to enhance edge perception and contextual understanding. Applications of the model in the southern Yellow Sea demonstrate that MA-EDSNN outperforms its pre-improvement version (Accuracy=99.08%, mIoU=91.77%, F1=95.4%) in extracting fine-scale <i>U. prolifera</i> patches, sparse <i>U. prolifera</i>, and their edge details; it also outperforms methods like index-threshold method, random forest, and U-Net across multiregional test sets, verifying its generalization and robustness, and providing an efficient solution for marine ecological environment 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

A multi-attention and edge deep supervision neural network for capturing floating Ulva prolifera using GF-1 satellite imagery

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Dingfeng Yu,
  • Huiping Li,
  • Chen Chen,
  • Qianguo Xing,
  • Bangyi Tao,
  • Xiangfeng Kong,
  • Shunqi Pan

摘要

Frequent outbreaks of Ulva prolifera in China’ s Yellow Sea—characterized by strong seasonality, large scale, and long drift distances—have become a major marine ecological disaster. Satellite remote sensing integrated with deep learning enables rapid and accurate U. prolifera monitoring. However, surface-aggregated U. prolifera exhibits variable morphology and density, which imposes high requirements on the robustness and generalization of deep learning-based segmentation networks. In this study, we propose a multi-attention and edge deep supervision neural network (MA-EDSNN) model based on Gaofen-1 (GF-1) data to achieve high-precision U. prolifera semantic segmentation (especially for sparse floating regions and edge details). The model integrates a self-designed scale-aware channel attention (SACA) module with attention gate (AG) and convolutional block attention module (CBAM) in each decoder block to achieve multi-scale feature fusion. It further constructs a novel Hybrid Edge-supervised Segmentation loss (HES loss) to optimize boundary and small-object segmentation, and incorporates a high-frequency enhancement module (HFEM) and Swin Transformer into the encoder to enhance edge perception and contextual understanding. Applications of the model in the southern Yellow Sea demonstrate that MA-EDSNN outperforms its pre-improvement version (Accuracy=99.08%, mIoU=91.77%, F1=95.4%) in extracting fine-scale U. prolifera patches, sparse U. prolifera, and their edge details; it also outperforms methods like index-threshold method, random forest, and U-Net across multiregional test sets, verifying its generalization and robustness, and providing an efficient solution for marine ecological environment monitoring.