| 引用本文： |
-
董迪,黄华梅,高晴,李小维,雷学铁,魏征,孙玉超,邹智垒,李亢.耦合SAM与地学知识的红树林-盐沼生态交错带无人机精细制图与碳储量估算[J].广西科学,2026,33(2):233-242. [点击复制]
- DONG Di,HUANG Huamei,GAO Qing,LI Xiaowei,LEI Xuetie,WEI Zheng,SUN Yuchao,ZOU Zhilei,LI Kang.Fine-scale Mapping of Unmanned Aerial Vehicle and Carbon Stock Estimation in Mangrove-Salt Marsh Ecotones by Coupling SAM and Geoscientific Knowledge[J].Guangxi Sciences,2026,33(2):233-242. [点击复制]
|
|
| |
|
|
| 本文已被：浏览 123次 下载 20次 |
码上扫一扫！ |
| 耦合SAM与地学知识的红树林-盐沼生态交错带无人机精细制图与碳储量估算 |
|
董迪1,2, 黄华梅2, 高晴2, 李小维3, 雷学铁3, 魏征2, 孙玉超2, 邹智垒2, 李亢2
|
|
|
| (1.自然资源部海洋环境探测技术与应用重点实验室, 广东广州 510300;2.自然资源部南海发展研究院(自然资源部南海遥感技术应用中心), 自然资源部南海遥感测绘协同应用技术创新中心, 广东广州 510300;3.自然资源部北海海洋中心, 广西北海 536000) |
|
| 摘要: |
| 针对传统分类方法在红树林-盐沼生态交错带精细制图中难以平衡“椒盐噪声”抑制与复杂边界提取的问题，本研究提出一种耦合分割一切大模型(Segment Anything Model，SAM)与地学知识的精细制图方法，并以广东省汕头市义丰溪湿地公园为研究区进行验证。采用多尺度层级约束策略，首先利用重采样后2 m空间分辨率的无人机数字正射影像(Digital Orthophoto Map,DOM)，通过面向对象随机森林分类法获取湿地植被的先验分布，构建宏观尺度地学背景约束；其次将先验分布范围作为SAM分割的区域约束，对重采样后20 cm空间分辨率的无人机DOM进行零样本分割，提取精细的地物图斑；最后耦合光谱和纹理特征以实现红树林与盐沼的精细识别。结果表明，本研究方法总体精度(Overall Accuracy,OA)为98.67%，Kappa系数为0.98，优于传统的面向对象随机森林分类法，在复杂边界提取方面表现较好。联合精细制图结果和InVEST模型，估算出研究区红树林生态系统和盐沼生态系统的总碳储量分别为14 953.53、923.58 Mg C。本研究证实耦合SAM与地学知识的无人机精细制图方法可为蓝碳生态系统的精细化管护提供可靠的数据支撑。 |
| 关键词: 红树林-盐沼生态交错带 SAM 蓝碳 精细制图 地学知识 |
| DOI：10.13656/j.cnki.gxkx.20260603.003 |
| 投稿时间：2026-01-14修订日期：2026-03-25 |
| 基金项目:自然资源部南海局科技发展基金项目(230206)和广东省林业局2025年度自然资源事务专项“广东省滨海湿地资源监测和生态价值评估”资助。 |
|
| Fine-scale Mapping of Unmanned Aerial Vehicle and Carbon Stock Estimation in Mangrove-Salt Marsh Ecotones by Coupling SAM and Geoscientific Knowledge |
|
DONG Di1,2, HUANG Huamei2, GAO Qing2, LI Xiaowei3, LEI Xuetie3, WEI Zheng2, SUN Yuchao2, ZOU Zhilei2, LI Kang2
|
| (1.Key Laboratory of Marine Environmental Survey Technology and Application, Ministry of Natural Resources, Guangzhou, Guangdong, 510300, China;2.Technology Innovation Center for South China Sea Remote Sensing, Surveying and Mapping Collaborative Application, South China Sea Development Research Institute (Remote Sensing Technology Application Center of South China Sea), Ministry of Natural Resources, Guangzhou, Guangdong, 510300, China;3.Beihai Marine Center, Ministry of Natural Resources, Beihai, Guangxi, 536000, China) |
| Abstract: |
| To address the challenges faced by conventional classification methods in balancing “salt and pepper noise” suppression and complex boundary extraction for fine-scale mapping of mangrove-salt marsh ecotones,this study proposes a mapping framework by coupling the Segment Anything Model (SAM) with geoscientific knowledge.The Yifengxi Wetland Park in Shantou,Guangdong was selected as the study area.A multi-scale hierarchical constraint strategy was utilized.First,the object-oriented random forest classification was adopted to the resampled unmanned aerial vehicle Digital Orthophoto Map (DOM,2 m spatial resolution) to obtain a coastal vegetation mask,serving as a macro-scale geospatial background constraint.Subsequently,this mask was applied as a regional constraint for the SAM model to perform zero-shot segmentation on the resampled unmanned aerial vehicle DOM (20 cm spatial resolution),and fine-scale object patches were extracted.Finally,spectral and textural features were coupled to achieve precise identification of mangroves and salt marshes.The results demonstrated that the proposed method achieved the Overall Accuracy (OA) of 98.67% and Kappa coefficient of 0.98,outperforming the conventional object-oriented random forest classification,and it showed a superior capability in delineating complex boundaries.Combination of the fine-scale mapping results and the InVEST model estimated the carbon stocks of mangroves ecosystem and salt marshes ecosystem in the study area as 14 953.53 and 923.58 Mg C,respectively.This study proves that the fine-scale mapping of unmanned aerial vehicle by coupling SAM and geoscientific knowledge provides reliable baseline data for the refined management of blue carbon ecosystems,offering significant application value. |
| Key words: mangrove-salt marsh ecotone SAM blue carbon fine-scale mapping geoscientific knowledge |
|
|
|
|
|
