Abstract Surface roughness lengths, including the aerodynamic roughness length (z0m) and the thermodynamic roughness length (z0h, represented by excess resistance kB−1), are crucial parameters in the accurate simulation of surface turbulent fluxes. However, due to insufficient knowledge in the physical mechanisms of surface roughness lengths, there exist considerable uncertainties in physically-based models. In this study, we attempt to overcome this issue by establishing the data-driven surface roughness lengths models, which are based on global observations from the FLUXNET2015 dataset and Moderate Resolution Imaging Spectroradiometer (MODIS). Four machine learning algorithms, including random forest (RF), single hidden layer artificial neural network (ANN), multilayer perceptron (MLP), deep belief network (DBN) are explored. A large number of data from 45 flux tower sites (as many as 44,662 daily z0m and 583,484 half-hour kB−1 observations) are utilized to train and test the data-driven models. Our results show that the data-driven models surprisingly achieve significantly improved estimation of surface roughness lengths and turbulent fluxes than physical models, which indicated the model inadequacy of physical models. RF-driven models achieve the best results. MLP and DBN-driven models of higher complexity are slightly superior to ANN-driven models but exhibit unstable performance. RF and ANN accurately reproduce the unimodal function relationship between leaf area index and z0m, thus demonstrating that the machine learning methods can extract physical rules from vast numbers of observations. In contrast, MLP and DBN fail to capture this relationship, possibly because of too complicated architecture. It implies that a suitable complexity of machine learning algorithm is critical to excavate true physical mechanism. To the best of our knowledge, this is the first study to demonstrate that machine learning technique can contribute to highly accurate estimations of surface turbulent fluxes by building data-driven surface roughness lengths models.

中文翻译：

---

使用机器学习算法改进表面粗糙度长度估计

摘要 表面粗糙度长度，包括气动粗糙度长度（z0m）和热力学粗糙度长度（z0h，用超阻 kB−1 表示）是精确模拟表面湍流通量的关键参数。然而，由于对表面粗糙度长度的物理机制了解不足，基于物理的模型存在相当大的不确定性。在这项研究中，我们试图通过建立数据驱动的表面粗糙度长度模型来克服这个问题，该模型基于来自 FLUXNET2015 数据集和中分辨率成像光谱仪 (MODIS) 的全球观测。探索了四种机器学习算法，包括随机森林（RF）、单隐藏层人工神经网络（ANN）、多层感知器（MLP）、深度信念网络（DBN）。来自 45 个通量塔站点的大量数据（多达 44,662 个每日 z0m 和 583,484 个半小时 kB-1 观测）用于训练和测试数据驱动模型。我们的结果表明，数据驱动的模型令人惊讶地实现了比物理模型显着改善的表面粗糙度长度和湍流通量的估计，这表明模型物理模型的不足。射频驱动模型达到最佳效果。更高复杂度的 MLP 和 DBN 驱动模型略优于 ANN 驱动模型，但表现出不稳定的性能。RF 和 ANN 准确再现了叶面积指数与 z0m 之间的单峰函数关系，从而证明机器学习方法可以从大量观察中提取物理规则。相比之下，MLP 和 DBN 未能捕捉到这种关系，可能是因为架构太复杂。这意味着合适的机器学习算法复杂度对于挖掘真正的物理机制至关重要。据我们所知，这是第一项证明机器学习技术可以通过构建数据驱动的表面粗糙度长度模型来高精度估计表面湍流通量的研究。