Combining UAV-based hyperspectral and LiDAR data for mangrove species classification using the rotation forest algorithm
Combining UAV-based hyperspectral and LiDAR data for mangrove species classification using the rotation forest algorithm
复制标题
DOI：
10.1016/j.jag.2021.102414
复制
发表时间：
2021-10-01
影响因子：
7.5
通讯作者：
Peng, Liheng
中科院分区：
文献类型：
--
作者：
Cao, Jingjing;Liu, Kai;Peng, Liheng
对红树林物种信息进行准确及时的监测对于精确管理和实际保护至关重要。在红树林监测中使用的传统高光谱技术往往难以实现对红树林物种的精细分类，这是由于星载图像的空间分辨率低以及机载图像成本高。此外，仅使用光谱信息不足以对复杂生态系统中的红树林物种进行精细分类，因为红树林物种的光谱可区分性通常受到复杂冠层结构的限制。为了解决这些局限性，本研究提出了一种新的红树林物种分类方法，该方法综合使用基于无人机（UAV）的Nano - hyperspec高光谱图像、光探测和测距（LiDAR）数据以及旋转森林（RoF）集成学习算法。所提出的方法在中国最大的人工种植红树林——淇澳岛进行了测试。首先，我们从基于无人机的高光谱数据中提取光谱特征，并从LiDAR数据中提取结构信息；然后我们利用RoF算法根据光谱和结构特征对红树林物种进行分类，并与另外两种流行的集成学习算法，即随机森林（RF）和逻辑模型树（LMT）进行比较。结果表明，高光谱和LiDAR数据相结合对所有三个分类器都产生了令人满意的结果，总体精度（OA）高于95%，并且所提出的方法获得了最高的OA为97.22%，卡帕系数为0.9686。我们的研究证明，纳入冠层高度信息可以提高分类精度，OA和卡帕系数分别比单独使用原始光谱波段高2.43%和0.0274。还发现RoF算法在对红树林物种进行分类时比RF和LMT更准确和稳定。这些研究结果表明，所提出的方法可以实现对红树林的精细监测，并进一步促进红树林的恢复和管理。
Accurate and timely monitoring of mangrove species information is crucial for precise management and practical conservation. Conventional hyperspectral techniques employed in mangrove monitoring are often limited to achieve the fine classification of mangrove species, due to the low spatial resolution of space-borne images and the high cost of airborne images. Moreover, using the spectral information alone is not adequate for fine-scale classification of mangrove species in complex ecosystems, because the spectral discriminability of mangrove species is generally restricted by complex canopy structures. To address these limitations, this study proposes a novel mangrove species classification method that integratively uses unmanned aerial vehicle (UAV)-based Nano-hyperspec hyperspectral imagery, light detection and ranging (LiDAR) data, and the rotation forest (RoF) ensemble learning algorithm. The proposed method was tested in China's largest artificially planted mangroves, Qi'ao Island. First, we extracted spectral features from UAV-based hyperspectral data and structural information from LiDAR data; then we utilized the RoF algorithm to classify mangrove species based on the spectral and structural features and compared with two other popular ensemble learning algorithms, namely random forest (RF) and logistic model tree (LMT). Results showed that the combined hyperspectral and LiDAR data produced satisfactory results for all three classifiers with overall accuracy (OA) higher than 95%, and the proposed method achieved the highest OA of 97.22% and Kappa coefficient of 0.9686. Our study proved that incorporating the canopy height information can improve the classification accuracy, with the OA and Kappa coefficient being 2.43% and 0.0274 higher than using the original spectral bands alone, respectively. It is also found that the RoF algorithm is more accurate and stable in classifying mangrove species than those of RF and LMT. These findings indicated that the proposed approach could achieve fine-scale mangrove monitoring and further facilitate mangrove forest restoration and management.