How to automate timely large-scale mangrove mapping with remote sensing
How to automate timely large-scale mangrove mapping with remote sensing
复制标题
DOI：
10.1016/j.rse.2021.112584
复制
发表时间：
2021-10
影响因子：
13.5
通讯作者：
Ying Lu;Le Wang
中科院分区：
文献类型：
--
作者：
Ying Lu;Le Wang
在过去四十年中，由于人为和自然干扰，红树林发生了重大变化。虽然有一些尝试的报道，但由于难以在大的地理区域收集足够的训练样本，仍然缺乏能够及时重复生成大尺度红树林地图的有效方法。在本研究中，我们通过以下方式实现了三个目标：(1)我们的目标是开发一种自动收集充足红树林训练样本的方法；相应地，我们开发了一种自动训练样本采集方法，从历史红树林地图中提取未改变的红树林样本；此外，我们采用了区域生长法，使训练样本更加多样化；(2)我们努力培育兼容的分类器，可以利用收集到的单类训练样本；为此，我们提出了两种具有代表性的单类分类器：支持向量数据描述（SVDD）和积极无标记学习算法（PUL）；(3)我们努力比较训练样本、分类器和输入图像的各种组合的有效性；因此，我们通过改变四个不同的变量开发了32个分类模型：训练样本（不变vs扩展）、输入数据（Landsat 8、Sentinel-1和Sentinel-2）、分类器（SVDD vs. PUL）和研究地点（美国佛罗里达州和中国广西）。我们发现我们开发的自动训练样本采集方法表现良好（用户准确率为97%）。年际NDVI结合几何限制保证了未改变训练样本的有效提取，而区域生长法由于增加了新出现的红树林而进一步减少了遗漏。此外，PUL优于SVDD。这是由于PUL不仅利用红树林样本，而且还利用SVDD中未标记的样本。最后，在所有比较模型中推荐Sentinel-1和Sentinel-2的组合。综上所述，我们开发了一种有效的红树林训练样本自动提取方法，在此基础上实现了大规模红树林制图的一类分类方法。我们设想我们的方法将有助于广泛的及时大规模红树林测绘任务。
Mangrove forests have witnessed significant changes resulted from both anthropogenic and natural disturbances in the last four decades. Although a few attempts have been reported, effective methods that can repeatedly generate large-scale mangrove maps on a timely basis are still lacking due to the difficulty in gathering sufficient training samples in large geographical areas. In this study, we have addressed three objectives in the following manner: (1) we aim to develop a method to automatically collect ample mangrove training samples; Correspondingly, we developed an automatic training sample collection method which extracted unchanged mangrove samples from a historical mangrove map; In addition, we employed a region growing method to include more diversified training samples; (2) we strive to foster compatible classifiers that can leverage the collected one-class training samples; To this end, we came up with two representative one-class classifiers: the Support Vector Data Description (SVDD), and the Positive and Unlabeled Learning algorithm (PUL); (3) we endeavor to compare the effectiveness of various combinations of training samples, classifiers, and input images; As a result, we developed 32 classification models by varying four different variables: training samples (unchanged vs. expanded), input data (Landsat 8, Sentinel-1, and Sentinel-2), classifiers (SVDD vs. PUL), and study sites (Florida, the United States and Guangxi, China). We found that our developed automatic training sample collection methods performed well (user's accuracy >97%). Inter-annual NDVI combined with geometric restrictions warranted the effective extraction of unchanged training samples while the region growing method further reduced the omission due to its addition of recently emerged mangroves. In addition, PUL performed better than SVDD. This is attributed to the fact that PUL draws upon not only mangrove samples, but also unlabeled ones unaccounted for in SVDD. Lastly, the combination of Sentinel-1 and Sentinel-2 is recommended among all the compared models. In summary, we developed an effective method to automatically extract mangrove training samples, based on which a one-class classification method for large-scale mangrove mapping is made possible. We envision our methods will contribute to a wide spectrum of timely large-scale mangrove mapping tasks.