| 张安健, 侯世闯, 吴淑雷, 陈焕东.基于多特征融合和注意力机制的红树林遥感图像语义分割网络[J].海南师范大学学报自科版,2026,(2):223-235 |
| 基于多特征融合和注意力机制的红树林遥感图像语义分割网络 |
| Semantic Segmentation Network of Mangrove Remote Sensing Image Based on Multi-feature Fusion and Attention Mechanism |
| |
| DOI：10.12051/j.issn.1674-4942.2024.04.013 |
| 中文关键词: 多特征融合 注意力机制 图像分割 T-Net SET-Unet |
| 英文关键词: multi-feature fusion attention mechanism image segmentation T-Net SET-Unet |
| 基金项目:国家自然科学基金项目（61966013） |
|
| 摘要点击次数: 194 |
| 全文下载次数: 7 |
| 中文摘要: |
| 红树林是一种对全球生态平衡和生物多样性具有重要意义的生态敏感区域。由于复杂的地形、变化的光照条件和多样的植被，红树林区域的遥感图像分割面临许多挑战。使用传统的图像分割方法对于红树林遥感图像分割小区域和边缘区域的精准度较差。本文提出了一种基于深度学习的红树林遥感图像分割算法：结合多特征融合和注意力机制的红树林遥感图像语义分割网络。首先，在U-Net结构的基础上加以改进，结合Darknet-53模型结构特征，使用了三重解码特征，构建三重U-Net网络结构的T-UNet网络（Triple Unet），提升了小区域的分割精度。其次，在 T-Unet基础上通过引入SE-Net和CBAM注意力模块，构建SET-Unet（Squeeze Excitation Triple Unet）网络结构，进一步提高了边界区域的分割效果。实验结果显示，T-Unet模型在小区域分割的场景中相比标准U-Net模型效果有显著提升；SET-Unet模型在边界区域的召回率显著优于对比模型。这些改进显著提高了分割的精度，尤其在小区域和边缘区域，为红树林保护和生物多样性研究提供了有效支持。 |
| 英文摘要: |
| Mangrove is an ecologically sensitive area that is of great significance to global ecological balance and biodiversity. Due to the complex terrain, changing lighting conditions and diverse vegetation, the remote sensing image segmentation of mangrove areas faces many challenges. Traditional image segmentation methods have poor accuracy for small areas and edge areas of mangrove remote sensing image segmentation. This paper proposes a mangrove remote sensing image segmentation algorithm based on deep learning: a semantic segmentation network of mangrove remote sensing image combining multi-feature fusion and attention mechanism. Firstly, the U-Net structure was improved by combining the structural characteristics of the Darknet-53 model, and the triple decoding features were utilized to build a triple structure U-Net network (T-Unet), which improved the segmentation accuracy of small areas. Secondly, on the basis of T-Unet, by introducing the SE Net and CBAM attention module, the SET Unet (Squeeze Exception Triple Unet) network structure was constructed, which further improves the segmentation effect of boundary areas. The experimental results show that the T-Unet model has significantly improved the effect compared with the standard U-Net model in the small area segmentation scene, and the recall rate of SET Unet model in the boundary region is significantly better than that of the comparison model. These improvements have significantly improved the accuracy of segmentation, especially in small areas and marginal areas, providing effective support for mangrove conservation and biodiversity research. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
