严正声明:近期,《海洋技术学报》编辑部发现某些中介组织利用假冒网站、电子信箱、电话等方式,以可以快速刊发稿件为由,向作者提供假冒录用通知并收取费用。对此,我刊发布严正声明,详见动态信息公告。
VEAU-Net:一种低覆盖度碱蓬遥感识别的深度学习模型研究
VEAU-Net: A Deep Learning Model for Remote Sensing Identification of Low-Coverage Suaeda Salsa
  
DOI:10.3969/j.issn.1003-2029.2026.03.006
中文关键词:  遥感  低覆盖度碱蓬  黄河三角洲  注意力机制  U-Ne
英文关键词:remote sensing  low-coverage Suaeda salsa  Yellow River Delta  attention mechanism  U-Net
基金项目:山东省自然科学基金资助项目(ZR2025MS676);黄河三角洲智能生态监测技术与应用系统研究项目(2024TSGC0147)
作者单位
梁宽1,2,胡亚斌1,3,姬生月2,任广波1,3 (1.自然资源部第一海洋研究所,山东 青岛 266061
2.中国石油大学(华东),山东 青岛 266580
3.自然资源部海洋遥测技术创新中心,山东 青岛 266061) 
摘要点击次数: 77
全文下载次数: 23
中文摘要:
      碱蓬是滨海潮间带湿地的重要物种,对维持生态平衡和促进生态系统修复发挥着关键作用。针对滨海湿地碱蓬分布稀疏、光谱特征复杂且易与背景混淆导致遥感识别精度低的问题,提出了一种融合U-Net与视觉鹰注意力机制(Vision Eagle Attention,VEA)的深度学习模型VEAU-Net,该模型通过引入空间-通道注意力机制,强化了对稀疏斑块、边缘区域及混生背景下目标特征的聚焦能力。本文基于覆盖黄河口潮间带低覆盖度碱蓬区域的Seninel-2遥感数据和实地调查数据,开展了碱蓬识别实验。结果表明:提出的VEAU-Net算法能够实现低覆盖度碱蓬识别,准确率为95.33%,F1分数达0.9755,召回率达0.9988。相较传统模型,本文方法F1分数提升了0.0372~0.3154。该模型显著改善了低覆盖度碱蓬的漏检与错分问题,可获取高精度碱蓬分布结果,为滨海湿地生态监测与植被恢复评估提供方法支撑。
英文摘要:
      Suaeda salsa, a key species in coastal intertidal wetlands, plays a crucial role in maintaining ecological balance and promoting ecosystem restoration. To address the low accuracy of remote sensing identification caused by sparse distribution, complex spectral features, and background confusion of Suaeda salsa in coastal wetlands, this study proposes a deep learning model named VEAU-Net, which integrates U-Net with a Vision Eagle Attention (VEA) mechanism. By incorporating a spatial-channel attention strategy, the model enhances its ability to focus on target features in sparse patches, edge areas, and mixed background regions. Based on Sentinel-2 remote sensing imagery and field survey data covering the low-coverage Suaeda salsa areas of the Yellow River Estuary intertidal zone, classification experiments were conducted. The results show that the proposed VEAU-Net effectively identifies low-coverage Suaeda salsa, achieving an accuracy of 95.33%, an F1-score of 0.975 5, and a recall rate of 0.998 8. Compared with conventional models, the proposed method improves the F1-score by 0.037 2~0.315 4. This model substantially reduces omission and misclassification of low-coverage Suaeda salsa, providing high-precision distribution mapping and offering methodological support for coastal wetland ecological monitoring and vegetation restoration assessment.
查看全文  查看/发表评论  下载PDF阅读器