Bayesian and Classical Machine Learning Methods: A Comparison for Tree Species Classification with LiDAR Waveform Signatures
Bayesian and Classical Machine Learning Methods: A Comparison for Tree Species Classification with LiDAR Waveform Signatures
复制标题
DOI：
10.3390/rs10010039
复制
发表时间：
2017-12
期刊：
影响因子：
--
通讯作者：
Tan Zhou;S. Popescu;A. M. Lawing;Marian Eriksson;Bogdan M. Strimbu;P. Bürkner
中科院分区：
文献类型：
--
作者：
Tan Zhou;S. Popescu;A. M. Lawing;Marian Eriksson;Bogdan M. Strimbu;P. Bürkner
全波形（FW）光探测和测距（LiDAR）数据中包含的大量信息为表征植被结构提供了前景。本研究旨在通过将波形指标与机器学习方法和贝叶斯推理相结合，研究FW激光雷达数据单独用于树种识别的能力。具体而言，我们首先基于基于波形的冠层高度模型（CHM），采用TreeVaW、流域算法以及TreeVaW与流域（TW）算法相结合的三种方法进行了树木自动分割。随后，使用随机森林（RF）和条件推理森林（CF）模型来识别来自三个不同来源的重要树级波形指标，如原始波形、复合波形、基于波形的点云和来自这三个来源的组合变量。此外，我们利用本研究确定的重要波形指标，通过RF、CF和贝叶斯多项逻辑回归（BMLR）对乔木（灰松、蓝橡木、室内活橡木）和灌木物种进行了区分。树木分割的结果表明，TW算法在划分单个树冠方面优于其他算法。CF模型克服了RF模型导致的波形指标选择偏差，有利于相关指标，提高了后续分类的准确性。我们还发现，在我们的研究区域，复合波形比原始波形和基于波形的点云更能提供信息。经典机器学习方法（RF和CF）和BMLR都产生了令人满意的平均总体精度（RF为74%，CF为77%，BMLR为81%），BMLR略优于其他两种方法。然而，这三种方法对蓝橡木的个体分类精度较低，由于蓝橡木与室内活橡木的特性相似，容易被误分类为室内活橡木。BMLR方法的不确定性估计弥补了这一缺点，它提供了概率意义上的分类结果，并使用户更有信心解释和应用分类结果到森林清册等现实世界的任务中。总的来说，本研究推荐CF方法进行特征选择，并表明BMLR可能是经典加工学习方法的更好选择。
A plethora of information contained in full-waveform (FW) Light Detection and Ranging (LiDAR) data offers prospects for characterizing vegetation structures. This study aims to investigate the capacity of FW LiDAR data alone for tree species identification through the integration of waveform metrics with machine learning methods and Bayesian inference. Specifically, we first conducted automatic tree segmentation based on the waveform-based canopy height model (CHM) using three approaches including TreeVaW, watershed algorithms and the combination of TreeVaW and watershed (TW) algorithms. Subsequently, the Random forests (RF) and Conditional inference forests (CF) models were employed to identify important tree-level waveform metrics derived from three distinct sources, such as raw waveforms, composite waveforms, the waveform-based point cloud and the combined variables from these three sources. Further, we discriminated tree (gray pine, blue oak, interior live oak) and shrub species through the RF, CF and Bayesian multinomial logistic regression (BMLR) using important waveform metrics identified in this study. Results of the tree segmentation demonstrated that the TW algorithms outperformed other algorithms for delineating individual tree crowns. The CF model overcomes waveform metrics selection bias caused by the RF model which favors correlated metrics and enhances the accuracy of subsequent classification. We also found that composite waveforms are more informative than raw waveforms and waveform-based point cloud for characterizing tree species in our study area. Both classical machine learning methods (the RF and CF) and the BMLR generated satisfactory average overall accuracy (74% for the RF, 77% for the CF and 81% for the BMLR) and the BMLR slightly outperformed the other two methods. However, these three methods suffered from low individual classification accuracy for the blue oak which is prone to being misclassified as the interior live oak due to the similar characteristics of blue oak and interior live oak. Uncertainty estimates from the BMLR method compensate for this downside by providing classification results in a probabilistic sense and rendering users with more confidence in interpreting and applying classification results to real-world tasks such as forest inventory. Overall, this study recommends the CF method for feature selection and suggests that BMLR could be a superior alternative to classical machining learning methods.