{"title":"基于Sentinel-2图像的混合神经网络红树林测绘","authors":"Longjie Ye , Qihao Weng","doi":"10.1016/j.rse.2025.114917","DOIUrl":null,"url":null,"abstract":"<div><div>Mangroves, as cradles of biodiversity and blue carbon reservoirs, are facing survival challenges due to climate change and anthropogenic disturbance. Precise and rapid mapping of mangrove forests has thus become highly relevant, which can provide essential information to support the conservation practices of such blue carbon resources. Existing machine learning algorithms for mangrove mapping are incapable of delivering precise cartographic solutions under dynamic tidal conditions because of poor transferability. This study developed a generalized approach for large-area mangrove mapping using a hybrid neural network integrated with a vision transformer to effectively capture representative features. To adapt mangrove mapping to the variety of tidal conditions, a vision transformer architecture was developed by encoding the fusion of three Sentinel-2 bands: Green, NIR, and SWIR. The ground truth dataset for the year 2021 was created from the composited Sentinel-2 images after interpreting Google Earth and drone images, which comprised 88,645 training samples (256 × 256 pixels per sample) and 24,969 test samples. We selected 30 coastal counties as test dataset in China to evaluate the effectiveness of the proposed network and produced a 10 m mangrove map that reported a total mangrove area of 28,006.24 ha in China in 2021, yielding an overall accuracy (OA) of 95.91 %. Compared to existing data products, Global Mangrove Watch 3.0, ESA WorldCover V200 and HGMF, our method outperformed the second-best product HGMF in mixed tide regions, by a margin of 9.19 % in OA and by 9.36 % in mean F1 score. Despite fluctuations in tide levels captured by Sentinel-2 imagery, the proposed method consistently yielded robust mangrove mapping results, highlighting effective derivation of tidal information. In comparison with previous mapping methods, the superior efficacy of the proposed network is distinctly discernible in distinguishing and delineating mangrove ecosystems in mixed tide regions, presenting prospects for improved monitoring of mangroves at regional and global scales.</div></div>","PeriodicalId":417,"journal":{"name":"Remote Sensing of Environment","volume":"329 ","pages":"Article 114917"},"PeriodicalIF":11.1000,"publicationDate":"2025-07-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"A hybrid neural network for mangrove mapping considering tide states using Sentinel-2 imagery\",\"authors\":\"Longjie Ye , Qihao Weng\",\"doi\":\"10.1016/j.rse.2025.114917\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Mangroves, as cradles of biodiversity and blue carbon reservoirs, are facing survival challenges due to climate change and anthropogenic disturbance. Precise and rapid mapping of mangrove forests has thus become highly relevant, which can provide essential information to support the conservation practices of such blue carbon resources. Existing machine learning algorithms for mangrove mapping are incapable of delivering precise cartographic solutions under dynamic tidal conditions because of poor transferability. This study developed a generalized approach for large-area mangrove mapping using a hybrid neural network integrated with a vision transformer to effectively capture representative features. To adapt mangrove mapping to the variety of tidal conditions, a vision transformer architecture was developed by encoding the fusion of three Sentinel-2 bands: Green, NIR, and SWIR. The ground truth dataset for the year 2021 was created from the composited Sentinel-2 images after interpreting Google Earth and drone images, which comprised 88,645 training samples (256 × 256 pixels per sample) and 24,969 test samples. We selected 30 coastal counties as test dataset in China to evaluate the effectiveness of the proposed network and produced a 10 m mangrove map that reported a total mangrove area of 28,006.24 ha in China in 2021, yielding an overall accuracy (OA) of 95.91 %. Compared to existing data products, Global Mangrove Watch 3.0, ESA WorldCover V200 and HGMF, our method outperformed the second-best product HGMF in mixed tide regions, by a margin of 9.19 % in OA and by 9.36 % in mean F1 score. Despite fluctuations in tide levels captured by Sentinel-2 imagery, the proposed method consistently yielded robust mangrove mapping results, highlighting effective derivation of tidal information. In comparison with previous mapping methods, the superior efficacy of the proposed network is distinctly discernible in distinguishing and delineating mangrove ecosystems in mixed tide regions, presenting prospects for improved monitoring of mangroves at regional and global scales.</div></div>\",\"PeriodicalId\":417,\"journal\":{\"name\":\"Remote Sensing of Environment\",\"volume\":\"329 \",\"pages\":\"Article 114917\"},\"PeriodicalIF\":11.1000,\"publicationDate\":\"2025-07-19\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Remote Sensing of Environment\",\"FirstCategoryId\":\"5\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0034425725003219\",\"RegionNum\":1,\"RegionCategory\":\"地球科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"ENVIRONMENTAL SCIENCES\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Remote Sensing of Environment","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0034425725003219","RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENVIRONMENTAL SCIENCES","Score":null,"Total":0}
A hybrid neural network for mangrove mapping considering tide states using Sentinel-2 imagery
Mangroves, as cradles of biodiversity and blue carbon reservoirs, are facing survival challenges due to climate change and anthropogenic disturbance. Precise and rapid mapping of mangrove forests has thus become highly relevant, which can provide essential information to support the conservation practices of such blue carbon resources. Existing machine learning algorithms for mangrove mapping are incapable of delivering precise cartographic solutions under dynamic tidal conditions because of poor transferability. This study developed a generalized approach for large-area mangrove mapping using a hybrid neural network integrated with a vision transformer to effectively capture representative features. To adapt mangrove mapping to the variety of tidal conditions, a vision transformer architecture was developed by encoding the fusion of three Sentinel-2 bands: Green, NIR, and SWIR. The ground truth dataset for the year 2021 was created from the composited Sentinel-2 images after interpreting Google Earth and drone images, which comprised 88,645 training samples (256 × 256 pixels per sample) and 24,969 test samples. We selected 30 coastal counties as test dataset in China to evaluate the effectiveness of the proposed network and produced a 10 m mangrove map that reported a total mangrove area of 28,006.24 ha in China in 2021, yielding an overall accuracy (OA) of 95.91 %. Compared to existing data products, Global Mangrove Watch 3.0, ESA WorldCover V200 and HGMF, our method outperformed the second-best product HGMF in mixed tide regions, by a margin of 9.19 % in OA and by 9.36 % in mean F1 score. Despite fluctuations in tide levels captured by Sentinel-2 imagery, the proposed method consistently yielded robust mangrove mapping results, highlighting effective derivation of tidal information. In comparison with previous mapping methods, the superior efficacy of the proposed network is distinctly discernible in distinguishing and delineating mangrove ecosystems in mixed tide regions, presenting prospects for improved monitoring of mangroves at regional and global scales.
期刊介绍:
Remote Sensing of Environment (RSE) serves the Earth observation community by disseminating results on the theory, science, applications, and technology that contribute to advancing the field of remote sensing. With a thoroughly interdisciplinary approach, RSE encompasses terrestrial, oceanic, and atmospheric sensing.
The journal emphasizes biophysical and quantitative approaches to remote sensing at local to global scales, covering a diverse range of applications and techniques.
RSE serves as a vital platform for the exchange of knowledge and advancements in the dynamic field of remote sensing.