Estimation of six leaf traits of East Asian forest tree species by leaf spectroscopy and partial least square regression
Estimation of six leaf traits of East Asian forest tree species by leaf spectroscopy and partial least square regression
复制标题
DOI：
10.1016/j.rse.2019.111381
复制
发表时间：
2019-11
影响因子：
13.5
通讯作者：
T. Nakaji;H. Oguma;Masahiro Nakamura;Panida Kachina;L. Asanok;Dokrak Marod;Masahiro Aiba;H. Kurokawa;Y. Kosugi;A. Kassim;T. Hiura
中科院分区：
文献类型：
--
作者：
T. Nakaji;H. Oguma;Masahiro Nakamura;Panida Kachina;L. Asanok;Dokrak Marod;Masahiro Aiba;H. Kurokawa;Y. Kosugi;A. Kassim;T. Hiura
为阐明高光谱遥感技术在东亚森林叶功能性状评价中的应用潜力，对太平洋地区寒温带至热带气候带广泛分布的141种树种的叶片光谱和偏最小二乘回归(PLSR)模型进行了研究。在日本、泰国和马来西亚的14个森林中，从两种植物功能类型(落叶和常绿树种)采集了不同发育阶段(幼龄、成熟和衰老)的叶片样本。目标叶片性状是单位面积叶质量和氮(N)、碳(C)、总酚、纤维素和木质素的浓度。用光谱辐射计测量了叶片在3-10 nm间隔(351个波段)上的可见光-短波红外光谱反射率，范围为400 nm到2395 nm。首先，我们比较了PLSR模型在基于干质量(DM)的浓度和基于叶面积(LA)的浓度方面的性能。然后，我们考察了基于不同植物功能类型和叶片发育阶段的模型在训练数据集上的适用性。最后，我们通过根据可变投影重要性(VIP)和等间隔重采样改变所使用的波段来评估PLSR模型稳定运行所需的最小光谱波段数，对于基于LA的所有化学性状浓度的决定系数(R2)高于基于DM的决定系数(R2)，而归一化均方误差(NRMSE)趋于低于基于DM的决定系数(R2)，除了叶片N的情况外，训练数据中的植物功能类型对估计模型的适用性有很大的影响。当训练数据集中的功能类型与目标树种的功能类型不同时，估计误差增加。在利用常绿树种数据集校正的PLSR模型中，常绿树种6个性状的NRMSE为8.0~12.6%，落叶树种为10.4~21.1%。在针对落叶树种进行校正的模型中也观察到了类似的趋势。使用两种功能类型校准的模型对两种树种都显示出中等精度(NRMSE = 5.3-13.5%)。训练数据的发育阶段对模型性能也有影响，使用所有阶段校准的模型对幼叶和衰老叶片的估计精度高于单独使用成熟叶片数据校准的模型；等间隔重采样法的估计精度高于使用VIP值阈值的情况，但两种方法的模型性能都随着输入波段的减少而下降。与具有全波段的模型相比，以20 nm的间隔使用至少104个具有相等间隔移除的频段可以提供与PLSR模型类似的性能。这项研究首次描述了叶片光谱在东亚森林树种特征估计中的潜力，我们的发现表明，典型功能类型和不同发育阶段的训练数据集对于通过东亚广泛分布的几个生物群来估计叶片特征是重要的。
To elucidate the potential of hyperspectral remote sensing for estimating the functional leaf traits in East Asian forests, we investigated the utility of leaf spectroscopy and partial least square regression (PLSR) models for 141 tree species distributed widely across cool temperate to tropical climate zones in the Pacific region. In 14 forests in Japan, Thailand, and Malaysia, leaf samples at various developmental stages (young, mature, and senescent) were collected from two plant functional types (deciduous and evergreen species). The target leaf traits were leaf mass per unit area and concentrations of nitrogen (N), carbon (C), total phenol, cellulose, and lignin. The leaf reflectance at visible-short wave infrared spectral reflectance from 400 nm to 2395 nm was measured at 3–10-nm intervals (351 bands) using a spectral radiometer. First, we compared the performance of the PLSR models in terms of dry mass (DM) -based concentration and leaf area (LA) –based concentration. Then, we investigated the applicability of the models based on the different plant functional types and leaf developmental stages of the training dataset. Finally, we evaluated the minimum number of spectral bands needed for stable performance of the PLSR model by changing the used bands in accordance with the variable importance of projection (VIP) and equal interval resampling.The coefficient of determination (R2) was higher and the normalized root mean square error (NRMSE) tended to be lower for all chemical trait concentrations expressed based on LA than for those based on DM, except for the case of leaf N. Plant functional type in the training data affected the applicability of the estimation model strongly. The estimation error increased when the functional type in the training dataset differed from that of the target tree species. In the PLSR model calibrated using datasets of evergreen tree species, the NRMSEs of the six traits were 8.0–12.6% for evergreen tree species but 10.4–21.1% for deciduous tree species. A similar trend was observed in the model calibrated for deciduous tree species. The model calibrated using both functional types showed intermediate accuracy for both types of tree species (NRMSE = 5.3–13.5%). The developmental stage of training data also affected the model performance, and the model calibrated using all of the stages showed better estimation accuracies for young and senescent leaves than the model calibrated from the data of mature leaves alone.The equal interval resampling provided better estimation accuracy than the case using the threshold of the VIP value although the model performance tended to be diminished with the reduction of input waveband in both methods. Using at least 104 bands with equal interval removal at a 20-nm interval confers similar performance of the PLSR model compared to the model with full wavebands. This study is the first to describe the potential of leaf spectroscopy for trait estimation of East Asian forest tree species, and our findings suggest that training datasets of typical functional types and varied developmental stages are important for estimation of leaf traits through the several biomes distributed widely in East Asia.